Skip to content

GitLab

Commit d3ad6274 authored by xuxzh1's avatar xuxzh1 🎱
Browse files

init

parent 97b02a89
Hide whitespace changes
Inline Side-by-side
Showing with 0 additions and 25671 deletions
llm/llama.cpp/ggml-cuda/sumrows.cu deleted 100644 → 0
View file @ 97b02a89
#include "sumrows.cuh"
static __global__ __launch_bounds__(1024) void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
const int row = blockIdx.x;
const int col = threadIdx.x;
float sum = 0.0f;
for (int i = col; i < ncols; i += blockDim.x) {
sum += x[row * ncols + i];
}
sum = warp_reduce_sum(sum);
if (col == 0) {
dst[row] = sum;
}
}
static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
const dim3 block_dims(WARP_SIZE, 1, 1);
const dim3 block_nums(nrows, 1, 1);
k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
}
void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(src0));
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream);
}
llm/llama.cpp/ggml-cuda/sumrows.cuh deleted 100644 → 0
View file @ 97b02a89
#include "common.cuh"
void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
llm/llama.cpp/ggml-cuda/tsembd.cu deleted 100644 → 0
View file @ 97b02a89
#include "tsembd.cuh"
static __global__ __launch_bounds__(1024) void timestep_embedding_f32(const float * timesteps, float * dst, const int nb1, const int dim, const int max_period) {
// blockIDx.y: idx of timesteps->ne[0]
// blockIDx.x: idx of ((dim + 1) / 2) / BLOCK_SIZE
int i = blockIdx.y;
int j = threadIdx.x + blockIdx.x * blockDim.x;
float * embed_data = (float *)((char *)dst + i*nb1);
if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
embed_data[dim] = 0.f;
}
int half = dim / 2;
if (j >= half) {
return;
}
float timestep = timesteps[i];
float freq = (float)expf(-logf(max_period) * j / half);
float arg = timestep * freq;
embed_data[j] = cosf(arg);
embed_data[j + half] = sinf(arg);
}
static void timestep_embedding_f32_cuda(const float * x, float * dst, const int ne00, const int nb1,
const int dim, const int max_period, cudaStream_t stream) {
int half_ceil = (dim + 1) / 2;
int num_blocks = (half_ceil + CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE;
dim3 gridDim(num_blocks, ne00, 1);
timestep_embedding_f32<<<gridDim, CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE, 0, stream>>>(x, dst, nb1, dim, max_period);
}
void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
const int dim = dst->op_params[0];
const int max_period = dst->op_params[1];
timestep_embedding_f32_cuda(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
}
llm/llama.cpp/ggml-cuda/tsembd.cuh deleted 100644 → 0
View file @ 97b02a89
#include "common.cuh"
#define CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE 256
void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
llm/llama.cpp/ggml-cuda/unary.cu deleted 100644 → 0
View file @ 97b02a89
#include "unary.cuh"
static __global__ __launch_bounds__(1024) void gelu_f32(const float * x, float * dst, const int k) {
const float GELU_COEF_A = 0.044715f;
const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
float xi = x[i];
dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi)));
}
static __global__ __launch_bounds__(1024) void gelu_quick_f32(const float * x, float * dst, int k) {
const float GELU_QUICK_COEF = -1.702f;
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i])));
}
static __global__ __launch_bounds__(1024) void silu_f32(const float * x, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = x[i] / (1.0f + expf(-x[i]));
}
static __global__ __launch_bounds__(1024) void tanh_f32(const float * x, float * dst, int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = tanhf(x[i]);
}
static __global__ __launch_bounds__(1024) void relu_f32(const float * x, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = fmaxf(x[i], 0);
}
static __global__ __launch_bounds__(1024) void sigmoid_f32(const float * x, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = 1.0f / (1.0f + expf(-x[i]));
}
static __global__ __launch_bounds__(1024) void hardsigmoid_f32(const float * x, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
}
static __global__ __launch_bounds__(1024) void hardswish_f32(const float * x, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
}
static __global__ __launch_bounds__(1024) void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope;
}
static __global__ __launch_bounds__(1024) void sqr_f32(const float * x, float * dst, const int k) {
const int i = blockDim.x*blockIdx.x + threadIdx.x;
if (i >= k) {
return;
}
dst[i] = x[i] * x[i];
}
static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
gelu_quick_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void sigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_SIGMOID_BLOCK_SIZE - 1) / CUDA_SIGMOID_BLOCK_SIZE;
sigmoid_f32<<<num_blocks, CUDA_SIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE;
hardsigmoid_f32<<<num_blocks, CUDA_HARDSIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_HARDSWISH_BLOCK_SIZE - 1) / CUDA_HARDSWISH_BLOCK_SIZE;
hardswish_f32<<<num_blocks, CUDA_HARDSWISH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
}
static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE;
sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
}
void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
gelu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
silu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
gelu_quick_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
tanh_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
sigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
hardsigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
hardswish_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
float negative_slope;
memcpy(&negative_slope, dst->op_params, sizeof(float));
leaky_relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), negative_slope, stream);
}
void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
}
llm/llama.cpp/ggml-cuda/unary.cuh deleted 100644 → 0
View file @ 97b02a89
#include "common.cuh"
#define CUDA_GELU_BLOCK_SIZE 256
#define CUDA_SILU_BLOCK_SIZE 256
#define CUDA_TANH_BLOCK_SIZE 256
#define CUDA_RELU_BLOCK_SIZE 256
#define CUDA_SIGMOID_BLOCK_SIZE 256
#define CUDA_HARDSIGMOID_BLOCK_SIZE 256
#define CUDA_HARDSWISH_BLOCK_SIZE 256
#define CUDA_SQR_BLOCK_SIZE 256
void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
llm/llama.cpp/ggml-cuda/upscale.cu deleted 100644 → 0
View file @ 97b02a89
#include "upscale.cuh"
static __global__ __launch_bounds__(1024) void upscale_f32(const float * x, float * dst,
const int nb00, const int nb01, const int nb02, const int nb03,
const int ne10, const int ne11, const int ne12, const int ne13,
const float sf0, const float sf1, const float sf2, const float sf3) {
int index = threadIdx.x + blockIdx.x * blockDim.x;
if (index >= ne10 * ne11 * ne12 * ne13) {
return;
}
int i10 = index % ne10;
int i11 = (index / ne10) % ne11;
int i12 = (index / (ne10 * ne11)) % ne12;
int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
int i00 = i10 / sf0;
int i01 = i11 / sf1;
int i02 = i12 / sf2;
int i03 = i13 / sf3;
dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
}
static void upscale_f32_cuda(const float * x, float * dst,
const int nb00, const int nb01, const int nb02, const int nb03,
const int ne10, const int ne11, const int ne12, const int ne13,
const float sf0, const float sf1, const float sf2, const float sf3,
cudaStream_t stream) {
int dst_size = ne10 * ne11 * ne12 * ne13;
int num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
upscale_f32<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3);
}
void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
const float * src0_d = (const float *)src0->data;
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
const float sf0 = (float)dst->ne[0]/src0->ne[0];
const float sf1 = (float)dst->ne[1]/src0->ne[1];
const float sf2 = (float)dst->ne[2]/src0->ne[2];
const float sf3 = (float)dst->ne[3]/src0->ne[3];
upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
}
llm/llama.cpp/ggml-cuda/upscale.cuh deleted 100644 → 0
View file @ 97b02a89
#include "common.cuh"
#define CUDA_UPSCALE_BLOCK_SIZE 256
void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
llm/llama.cpp/ggml-cuda/vecdotq.cuh deleted 100644 → 0
View file @ 97b02a89
#include "common.cuh"
static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) {
const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
int x32 = 0;
x32 |= x16[0] << 0;
x32 |= x16[1] << 16;
return x32;
}
static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) {
const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
int x32 = 0;
x32 |= x16[0] << 0;
x32 |= x16[1] << 16;
return x32;
}
static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) {
return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
}
static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) {
return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
}
// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
#define VDR_Q4_0_Q8_1_MMVQ 2
#define VDR_Q4_0_Q8_1_MMQ 4
template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl(
const int * v, const int * u, const float & d4, const half2 & ds8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
// SIMD dot product of quantized values
sumi = __dp4a(vi0, u[2*i+0], sumi);
sumi = __dp4a(vi1, u[2*i+1], sumi);
}
const float2 ds8f = __half22float2(ds8);
// second part effectively subtracts 8 from each quant value
return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y);
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q4_1_Q8_1_MMVQ 2
#define VDR_Q4_1_Q8_1_MMQ 4
template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl(
const int * v, const int * u, const half2 & dm4, const half2 & ds8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
// SIMD dot product of quantized values
sumi = __dp4a(vi0, u[2*i+0], sumi);
sumi = __dp4a(vi1, u[2*i+1], sumi);
}
#ifdef GGML_CUDA_F16
const float2 tmp = __half22float2(__hmul2(dm4, ds8));
const float d4d8 = tmp.x;
const float m4s8 = tmp.y;
#else
const float2 dm4f = __half22float2(dm4);
const float2 ds8f = __half22float2(ds8);
const float d4d8 = dm4f.x * ds8f.x;
const float m4s8 = dm4f.y * ds8f.y;
#endif // GGML_CUDA_F16
// scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q5_0_Q8_1_MMVQ 2
#define VDR_Q5_0_Q8_1_MMQ 4
template <int vdr> static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl(
const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4
vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12
vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20
vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28
sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4
vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12
vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20
vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28
sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
}
const float2 ds8f = __half22float2(ds8);
// second part effectively subtracts 16 from each quant value
return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y);
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q5_1_Q8_1_MMVQ 2
#define VDR_Q5_1_Q8_1_MMQ 4
template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl(
const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4
vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12
vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20
vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28
sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4
vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12
vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20
vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28
sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
}
#ifdef GGML_CUDA_F16
const float2 tmp = __half22float2(__hmul2(dm5, ds8));
const float d5d8 = tmp.x;
const float m5s8 = tmp.y;
#else
const float2 dm5f = __half22float2(dm5);
const float2 ds8f = __half22float2(ds8);
const float d5d8 = dm5f.x * ds8f.x;
const float m5s8 = dm5f.y * ds8f.y;
#endif // GGML_CUDA_F16
// scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q8_0_Q8_1_MMVQ 2
#define VDR_Q8_0_Q8_1_MMQ 8
template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_q8_1_impl(
const int * v, const int * u, const float & d8_0, const float & d8_1) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
// SIMD dot product of quantized values
sumi = __dp4a(v[i], u[i], sumi);
}
return d8_0*d8_1 * sumi;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl(
const int * v, const int * u, const half2 & dm8, const half2 & ds8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
#pragma unroll
for (int i = 0; i < vdr; ++i) {
// SIMD dot product of quantized values
sumi = __dp4a(v[i], u[i], sumi);
}
#ifdef GGML_CUDA_F16
const float2 tmp = __half22float2(__hmul2(dm8, ds8));
const float d8d8 = tmp.x;
const float m8s8 = tmp.y;
#else
const float2 dm8f = __half22float2(dm8);
const float2 ds8f = __half22float2(ds8);
const float d8d8 = dm8f.x * ds8f.x;
const float m8s8 = dm8f.y * ds8f.y;
#endif // GGML_CUDA_F16
// scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q2_K_Q8_1_MMVQ 1
#define VDR_Q2_K_Q8_1_MMQ 2
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq(
const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
const half2 & dm2, const float * __restrict__ d8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR2_K; ++i) {
const int sc = scales[2*i];
const int vi = (v >> (2*i)) & 0x03030303;
sumf_d += d8[i] * (__dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product
// fill int with 4x m
int m = sc >> 4;
m |= m << 8;
m |= m << 16;
sumf_m += d8[i] * __dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values
}
const float2 dm2f = __half22float2(dm2);
return dm2f.x*sumf_d - dm2f.y*sumf_m;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
const half2 & dm2, const float & d8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi_d = 0;
int sumi_m = 0;
#pragma unroll
for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) {
int sumi_d_sc = 0;
const int sc = scales[i0 / (QI8_1/2)];
// fill int with 4x m
int m = sc >> 4;
m |= m << 8;
m |= m << 16;
#pragma unroll
for (int i = i0; i < i0 + QI8_1/2; ++i) {
sumi_d_sc = __dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product
sumi_m = __dp4a(m, u[i], sumi_m); // multiply sum of q8_1 values with m
}
sumi_d += sumi_d_sc * (sc & 0xF);
}
const float2 dm2f = __half22float2(dm2);
return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m);
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q3_K_Q8_1_MMVQ 1
#define VDR_Q3_K_Q8_1_MMQ 2
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq(
const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales,
const int & scale_offset, const float & d3, const float * __restrict__ d8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
float sumf = 0.0f;
#pragma unroll
for (int i = 0; i < QR3_K; ++i) {
const int isc = scale_offset + 2*i;
const int isc_low = isc % (QK_K/32);
const int sc_shift_low = 4 * (isc / (QK_K/32));
const int sc_low = (scales[isc_low] >> sc_shift_low) & 0xF;
const int isc_high = isc % (QK_K/64);
const int sc_shift_high = 2 * (isc / (QK_K/64));
const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4;
const int sc = (sc_low | sc_high) - 32;
const int vil = (vl >> (2*i)) & 0x03030303;
const int vih = ((vh >> i) << 2) & 0x04040404;
const int vi = __vsubss4(vil, vih);
sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product
}
return d3 * sumf;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales,
const float & d3, const float & d8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
int sumi = 0;
#pragma unroll
for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) {
int sumi_sc = 0;
for (int i = i0; i < i0 + QI8_1/2; ++i) {
sumi_sc = __dp4a(v[i], u[i], sumi_sc); // SIMD dot product
}
sumi += sumi_sc * scales[i0 / (QI8_1/2)];
}
return d3*d8 * sumi;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q4_K_Q8_1_MMVQ 2
#define VDR_Q4_K_Q8_1_MMQ 8
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq(
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR4_K; ++i) {
const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F;
const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F;
const int dot1 = __dp4a(v1i, u[2*i+1], __dp4a(v0i, u[2*i+0], 0)); // SIMD dot product
const int dot2 = __dp4a(0x01010101, u[2*i+1], __dp4a(0x01010101, u[2*i+0], 0)); // sum of u
sumf_d += d8[i] * (dot1 * sc[i]);
sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values
}
const float2 dm4f = __half22float2(dm4);
return dm4f.x*sumf_d - dm4f.y*sumf_m;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) {
int sumi_d = 0;
#pragma unroll
for (int j = 0; j < QI8_1; ++j) {
sumi_d = __dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product
}
const float2 ds8f = __half22float2(ds8[i]);
sumf_d += ds8f.x * (sc[i] * sumi_d);
sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val
}
const float2 dm4f = __half22float2(dm4);
return dm4f.x*sumf_d - dm4f.y*sumf_m;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q5_K_Q8_1_MMVQ 2
#define VDR_Q5_K_Q8_1_MMQ 8
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq(
const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc,
const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR5_K; ++i) {
const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F;
const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F;
const int vh0i = ((vh[0] >> i) << 4) & 0x10101010;
const int vh1i = ((vh[1] >> i) << 4) & 0x10101010;
const int v0i = vl0i | vh0i;
const int v1i = vl1i | vh1i;
const int dot1 = __dp4a(v0i, u[2*i+0], __dp4a(v1i, u[2*i+1], 0)); // SIMD dot product
const int dot2 = __dp4a(0x01010101, u[2*i+0], __dp4a(0x01010101, u[2*i+1], 0)); // sum of u
sumf_d += d8[i] * (dot1 * sc[i]);
sumf_m += d8[i] * (dot2 * m[i]);
}
const float2 dm5f = __half22float2(dm5);
return dm5f.x*sumf_d - dm5f.y*sumf_m;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
float sumf_d = 0.0f;
float sumf_m = 0.0f;
#pragma unroll
for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) {
int sumi_d = 0;
#pragma unroll
for (int j = 0; j < QI8_1; ++j) {
sumi_d = __dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product
}
const float2 ds8f = __half22float2(ds8[i]);
sumf_d += ds8f.x * (sc[i] * sumi_d);
sumf_m += ds8f.y * m[i]; // sum of q8_1 block * q4_K min val
}
const float2 dm4f = __half22float2(dm4);
return dm4f.x*sumf_d - dm4f.y*sumf_m;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
#define VDR_Q6_K_Q8_1_MMVQ 1
#define VDR_Q6_K_Q8_1_MMQ 8
// contiguous v/x values
static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq(
const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales,
const float & d, const float * __restrict__ d8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
float sumf = 0.0f;
#pragma unroll
for (int i = 0; i < QR6_K; ++i) {
const int sc = scales[4*i];
const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
const int vih = ((vh >> (4*i)) << 4) & 0x30303030;
const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32
sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product
}
return d*sumf;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
// contiguous u/y values
static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc,
const float & d6, const float * __restrict__ d8) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
float sumf_d = 0.0f;
#pragma unroll
for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) {
int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale
#pragma unroll
for (int i = i0; i < i0 + 2; ++i) {
sumi_d.x = __dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product
sumi_d.x = __dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product
sumi_d.y = __dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product
sumi_d.y = __dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product
}
sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y);
}
return d6 * sumf_d;
#else
NO_DEVICE_CODE;
#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
}
static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
int v[VDR_Q4_0_Q8_1_MMVQ];
int u[2*VDR_Q4_0_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) {
v[i] = get_int_from_uint8(bq4_0->qs, iqs + i);
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0);
}
return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);
}
static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
int v[VDR_Q4_1_Q8_1_MMVQ];
int u[2*VDR_Q4_1_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) {
v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i);
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1);
}
return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMVQ>(v, u, bq4_1->dm, bq8_1->ds);
}
static __device__ __forceinline__ float vec_dot_q5_0_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;
int vl[VDR_Q5_0_Q8_1_MMVQ];
int vh[VDR_Q5_0_Q8_1_MMVQ];
int u[2*VDR_Q5_0_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) {
vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i);
vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i));
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0);
}
return vec_dot_q5_0_q8_1_impl<VDR_Q5_0_Q8_1_MMVQ>(vl, vh, u, bq5_0->d, bq8_1->ds);
}
static __device__ __forceinline__ float vec_dot_q5_1_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;
int vl[VDR_Q5_1_Q8_1_MMVQ];
int vh[VDR_Q5_1_Q8_1_MMVQ];
int u[2*VDR_Q5_1_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) {
vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i);
vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i));
u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1);
}
return vec_dot_q5_1_q8_1_impl<VDR_Q5_1_Q8_1_MMVQ>(vl, vh, u, bq5_1->dm, bq8_1->ds);
}
static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;
int v[VDR_Q8_0_Q8_1_MMVQ];
int u[VDR_Q8_0_Q8_1_MMVQ];
#pragma unroll
for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) {
v[i] = get_int_from_int8(bq8_0->qs, iqs + i);
u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
}
return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
}
static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q2_K * bq2_K = (const block_q2_K *) vbq;
const int bq8_offset = QR2_K * (iqs / QI8_1);
const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
const uint8_t * scales = bq2_K->scales + scale_offset;
const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs);
int u[QR2_K];
float d8[QR2_K];
#pragma unroll
for (int i = 0; i < QR2_K; ++ i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
}
return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
}
static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q3_K * bq3_K = (const block_q3_K *) vbq;
const int bq8_offset = QR3_K * (iqs / (QI3_K/2));
const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
const float d = bq3_K->d;
const int vl = get_int_from_uint8(bq3_K->qs, iqs);
// invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset;
int u[QR3_K];
float d8[QR3_K];
#pragma unroll
for (int i = 0; i < QR3_K; ++i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
}
return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
}
static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q4_K * bq4_K = (const block_q4_K *) vbq;
int v[2];
int u[2*QR4_K];
float d8[QR4_K];
// iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
// iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
// iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
// iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
// iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
v[0] = q4[0];
v[1] = q4[4];
const uint16_t * scales = (const uint16_t *)bq4_K->scales;
uint16_t aux[2];
const int j = bq8_offset/2;
if (j < 2) {
aux[0] = scales[j+0] & 0x3f3f;
aux[1] = scales[j+2] & 0x3f3f;
} else {
aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
}
const uint8_t * sc = (const uint8_t *)aux;
const uint8_t * m = sc + 2;
for (int i = 0; i < QR4_K; ++i) {
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
d8[i] = __low2float(bq8i->ds);
const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
u[2*i+0] = q8[0];
u[2*i+1] = q8[4];
}
return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
}
static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q5_K * bq5_K = (const block_q5_K *) vbq;
int vl[2];
int vh[2];
int u[2*QR5_K];
float d8[QR5_K];
const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2));
const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4));
vl[0] = ql[0];
vl[1] = ql[4];
vh[0] = qh[0] >> bq8_offset;
vh[1] = qh[4] >> bq8_offset;
const uint16_t * scales = (const uint16_t *)bq5_K->scales;
uint16_t aux[2];
const int j = bq8_offset/2;
if (j < 2) {
aux[0] = scales[j+0] & 0x3f3f;
aux[1] = scales[j+2] & 0x3f3f;
} else {
aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
}
const uint8_t * sc = (const uint8_t *)aux;
const uint8_t * m = sc + 2;
#pragma unroll
for (int i = 0; i < QR5_K; ++i) {
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
d8[i] = __low2float(bq8i->ds);
const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
u[2*i+0] = q8[0];
u[2*i+1] = q8[4];
}
return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
}
static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_q6_K * bq6_K = (const block_q6_K *) vbq;
const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4);
const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8);
const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4));
const int vl = get_int_from_uint8(bq6_K->ql, iqs);
const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift;
const int8_t * scales = bq6_K->scales + scale_offset;
int u[QR6_K];
float d8[QR6_K];
#pragma unroll
for (int i = 0; i < QR6_K; ++i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds);
}
return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
}
static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
#if QR2_XXS == 8
const int ib32 = iqs;
const uint16_t * q2 = bq2->qs + 4*ib32;
const uint8_t * aux8 = (const uint8_t *)q2;
const int8_t * q8 = bq8_1[ib32].qs;
uint32_t aux32 = q2[2] | (q2[3] << 16);
int sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
const uint8_t signs = ksigns_iq2xs[aux32 & 127];
for (int j = 0; j < 8; ++j) {
sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
}
q8 += 8;
aux32 >>= 7;
}
const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f;
return d * sumi;
#else
// iqs is 0...15
const int ib32 = iqs/2;
const int il = iqs%2;
const uint16_t * q2 = bq2->qs + 4*ib32;
const uint8_t * aux8 = (const uint8_t *)q2;
const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
const uint32_t aux32 = q2[2] | (q2[3] << 16);
const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f;
const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127];
const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127];
const int8_t * q8 = bq8_1[ib32].qs + 16*il;
int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < 8; ++j) {
sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1);
sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1);
}
return d * (sumi1 + sumi2);
#endif
}
static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
const int ib32 = iqs;
const uint16_t * q2 = bq2->qs + 4*ib32;
const int8_t * q8 = bq8_1[ib32].qs;
const uint8_t ls1 = bq2->scales[ib32] & 0xf;
const uint8_t ls2 = bq2->scales[ib32] >> 4;
int sumi1 = 0;
for (int l = 0; l < 2; ++l) {
const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]);
const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]);
sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1);
sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1);
q8 += 8;
}
int sumi2 = 0;
for (int l = 2; l < 4; ++l) {
const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
const int grid_l = __vsub4(grid[0] ^ signs[0], signs[0]);
const int grid_h = __vsub4(grid[1] ^ signs[1], signs[1]);
sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2);
sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2);
q8 += 8;
}
const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
#else
GGML_UNUSED(ksigns64);
NO_DEVICE_CODE;
#endif
}
// TODO
static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
const int ib32 = iqs;
const int8_t * q8 = bq8_1[ib32].qs;
const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32;
const uint8_t ls1 = bq2->scales[ib32] & 0xf;
const uint8_t ls2 = bq2->scales[ib32] >> 4;
int sumi1 = 0;
for (int l = 0; l < 2; ++l) {
const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201);
const int grid_l = __vsub4(grid[0] ^ signs0, signs0);
const int grid_h = __vsub4(grid[1] ^ signs1, signs1);
sumi1 = __dp4a(grid_l, *((const int *)q8 + 0), sumi1);
sumi1 = __dp4a(grid_h, *((const int *)q8 + 1), sumi1);
q8 += 8;
}
int sumi2 = 0;
for (int l = 2; l < 4; ++l) {
const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
const uint32_t signs0 = __vcmpeq4(((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
const uint32_t signs1 = __vcmpeq4(((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201);
const int grid_l = __vsub4(grid[0] ^ signs0, signs0);
const int grid_h = __vsub4(grid[1] ^ signs1, signs1);
sumi2 = __dp4a(grid_l, *((const int *)q8 + 0), sumi2);
sumi2 = __dp4a(grid_h, *((const int *)q8 + 1), sumi2);
q8 += 8;
}
const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f;
return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
#else
GGML_UNUSED(ksigns64);
NO_DEVICE_CODE;
#endif
}
static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
const int ib32 = iqs;
const uint8_t * q3 = bq2->qs + 8*ib32;
const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32;
const int8_t * q8 = bq8_1[ib32].qs;
uint32_t aux32 = gas[0] | (gas[1] << 16);
int sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0];
const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1];
const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127));
const int grid_l = __vsub4(grid1[0] ^ signs[0], signs[0]);
const int grid_h = __vsub4(grid2[0] ^ signs[1], signs[1]);
sumi = __dp4a(grid_l, *((int *)q8+0), sumi);
sumi = __dp4a(grid_h, *((int *)q8+1), sumi);
q8 += 8;
aux32 >>= 7;
}
const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.5f;
return d * sumi;
#else
NO_DEVICE_CODE;
#endif
}
// TODO: don't use lookup table for signs
static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
const int ib32 = iqs;
const uint8_t * qs = bq2->qs + 8*ib32;
const int8_t * q8 = bq8_1[ib32].qs;
int sumi = 0;
for (int l = 0; l < 4; ++l) {
const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256));
const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256));
uint32_t signs0 = __vcmpeq4(((bq2->signs[4*ib32+l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201);
uint32_t signs1 = __vcmpeq4(((bq2->signs[4*ib32+l] >> 4) * 0x01010101) & 0x08040201, 0x08040201);
const int grid_l = __vsub4(grid1[0] ^ signs0, signs0);
const int grid_h = __vsub4(grid2[0] ^ signs1, signs1);
sumi = __dp4a(grid_l, *((int *)q8+0), sumi);
sumi = __dp4a(grid_h, *((int *)q8+1), sumi);
q8 += 8;
}
const float d = (float)bq2->d * (1 + 2*((bq2->scales[ib32/2] >> 4*(ib32%2)) & 0xf)) * __low2float(bq8_1[ib32].ds);
return d * sumi;
#else
NO_DEVICE_CODE;
#endif
}
static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
const int ib32 = iqs;
int sumi = 0;
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const int * q8 = (const int *)bq8_1[ib32].qs;
for (int l = 0; l < 4; ++l) {
const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
int grid0 = grid[0] & 0x0f0f0f0f;
int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
sumi = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi));
}
#else
const int8_t * q8 = bq8_1[ib32].qs;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
for (int j = 0; j < 4; ++j) {
sumi += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4);
}
q8 += 8;
}
#endif
const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA;
const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1);
const float d = d1q * __low2float (bq8_1[ib32].ds);
const float m = d1q * __high2float(bq8_1[ib32].ds);
return d * sumi + m * delta;
}
static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
const int ib32 = iqs;
int sumi[2] = {0, 0};
float sumf[2] = {0.f, 0.f};
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const int * q8 = (const int *)bq8_1[ib32].qs;
for (int l = 0; l < 4; ++l) {
const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8)));
int grid0 = grid[0] & 0x0f0f0f0f;
int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
sumi[l/2] = __dp4a(q8[2*l+1], grid1, __dp4a(q8[2*l+0], grid0, sumi[l/2]));
const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
const int sumy = __dp4a(q8[2*l+1], 0x01010101, __dp4a(q8[2*l+0], 0x01010101, 0));
sumf[l/2] += delta*sumy;
}
#else
const int8_t * q8 = bq8_1[ib32].qs;
for (int l = 0; l < 4; ++l) {
const uint8_t * grid = (const uint8_t *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
int sumy = 0;
for (int j = 0; j < 4; ++j) {
sumi[l/2] += q8[j] * (grid[j] & 0xf) + q8[j+4] * (grid[j] >> 4);
sumy += q8[j] + q8[j+4];
}
const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
sumf[l/2] += delta*sumy;
q8 += 8;
}
#endif
iq1m_scale_t scale;
const uint16_t * sc = (const uint16_t *)bq1->scales;
scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds);
return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
}
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
static __device__ __forceinline__ void get_int_from_table_16(const uint32_t & q4, const uint8_t * values,
int & val1, int & val2) {
uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32;
aux32 = q4 & 0x0f0f0f0f;
uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8);
uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8);
val1 = v1 | (v2 << 16);
aux32 = (q4 >> 4) & 0x0f0f0f0f;
v1 = values[q8[0]] | (values[q8[1]] << 8);
v2 = values[q8[2]] | (values[q8[3]] << 8);
val2 = v1 | (v2 << 16);
}
#endif
static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
const block_iq4_nl * bq = (const block_iq4_nl *) vbq;
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs;
const int32_t * q8 = (const int32_t *)bq8_1->qs + iqs;
const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
int v1, v2;
int sumi1 = 0, sumi2 = 0;
for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16);
get_int_from_table_16(aux, values, v1, v2);
sumi1 = __dp4a(v1, q8[l+0], sumi1);
sumi2 = __dp4a(v2, q8[l+4], sumi2);
}
#else
const uint8_t * q4 = bq->qs + 4*iqs;
const int8_t * q8 = bq8_1->qs + 4*iqs;
int sumi1 = 0, sumi2 = 0;
for (int l = 0; l < 4*VDR_Q4_0_Q8_1_MMVQ; ++l) {
sumi1 += q8[l+ 0] * kvalues_iq4nl[q4[l] & 0xf];
sumi2 += q8[l+16] * kvalues_iq4nl[q4[l] >> 4];
}
#endif
const float d = (float)bq->d * __low2float(bq8_1->ds);
return d * (sumi1 + sumi2);
}
static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
// iqs is 0...7
const int ib32 = iqs;
const int32_t * q8 = (const int *)bq8_1[ib32].qs;
const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32;
const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
const float d = (float)bq4->d * (ls - 32) * __low2float(bq8_1[ib32].ds);
int v1, v2;
int sumi1 = 0, sumi2 = 0;
for (int j = 0; j < 4; ++j) {
get_int_from_table_16(q4[j], values, v1, v2);
sumi1 = __dp4a(v1, q8[j+0], sumi1);
sumi2 = __dp4a(v2, q8[j+4], sumi2);
}
return d * (sumi1 + sumi2);
#else
return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);
#endif
}
llm/llama.cpp/ggml-impl.h deleted 100644 → 0
View file @ 97b02a89
#pragma once
#include "ggml.h"
// GGML internal header
#include <assert.h>
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
#include <stddef.h>
#include <stdbool.h>
#include <string.h> // memcpy
#include <math.h> // fabsf
#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#if defined(_WIN32)
#define m512bh(p) p
#define m512i(p) p
#else
#define m512bh(p) (__m512bh)(p)
#define m512i(p) (__m512i)(p)
#endif
/**
* Converts brain16 to float32.
*
* The bfloat16 floating point format has the following structure:
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───┐
* 0b0000000000000000 brain16
*
* Since bf16 has the same number of exponent bits as a 32bit float,
* encoding and decoding numbers becomes relatively straightforward.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───────────────────┐
* 0b00000000000000000000000000000000 IEEE binary32
*
* For comparison, the standard fp16 format has fewer exponent bits.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌─┴─┐┌─┴──────┐
* 0b0000000000000000 IEEE binary16
*
* @see IEEE 754-2008
*/
static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
union {
float f;
uint32_t i;
} u;
u.i = (uint32_t)h.bits << 16;
return u.f;
}
/**
* Converts float32 to brain16.
*
* This function is binary identical to AMD Zen4 VCVTNEPS2BF16.
* Subnormals shall be flushed to zero, and NANs will be quiet.
* This code should vectorize nicely if using modern compilers.
*/
static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
ggml_bf16_t h;
union {
float f;
uint32_t i;
} u;
u.f = s;
if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
h.bits = (u.i >> 16) | 64; /* force to quiet */
return h;
}
if (!(u.i & 0x7f800000)) { /* subnormal */
h.bits = (u.i & 0x80000000) >> 16; /* flush to zero */
return h;
}
h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
return h;
}
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
#ifdef __cplusplus
extern "C" {
#endif
// static_assert should be a #define, but if it's not,
// fall back to the _Static_assert C11 keyword.
// if C99 - static_assert is noop
// ref: https://stackoverflow.com/a/53923785/4039976
#ifndef __cplusplus
#ifndef static_assert
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
#define static_assert(cond, msg) _Static_assert(cond, msg)
#else
#define static_assert(cond, msg) struct global_scope_noop_trick
#endif
#endif
#endif
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
#define __FMA__
#endif
#ifndef __F16C__
#define __F16C__
#endif
#endif
// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
#ifndef __SSE3__
#define __SSE3__
#endif
#ifndef __SSSE3__
#define __SSSE3__
#endif
#endif
#if defined(__ARM_FEATURE_SVE)
#include <arm_sve.h>
#endif
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
#if defined(__ARM_NEON)
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>
#ifdef _MSC_VER
typedef uint16_t ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
#else
typedef __fp16 ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
#endif // _MSC_VER
#if !defined(__aarch64__)
// 32-bit ARM compatibility
// vaddvq_s16
// vpaddq_s16
// vpaddq_s32
// vaddvq_s32
// vaddvq_f32
// vmaxvq_f32
// vcvtnq_s32_f32
// vzip1_u8
// vzip2_u8
inline static int32_t vaddvq_s16(int16x8_t v) {
return
(int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
(int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
(int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
(int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
return vcombine_s32(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
inline static float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
// vld1q_s16_x2
// vld1q_u8_x2
// vld1q_u8_x4
// vld1q_s8_x2
// vld1q_s8_x4
// TODO: double-check these work correctly
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vld1q_s16(ptr + 0);
res.val[1] = vld1q_s16(ptr + 8);
return res;
}
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
res.val[2] = vld1q_u8(ptr + 32);
res.val[3] = vld1q_u8(ptr + 48);
return res;
}
typedef struct ggml_int8x16x2_t {
int8x16_t val[2];
} ggml_int8x16x2_t;
inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
ggml_int8x16x2_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
res.val[2] = vld1q_s8(ptr + 32);
res.val[3] = vld1q_s8(ptr + 48);
return res;
}
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
// NOTE: not tested
inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
uint8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else
#define ggml_int16x8x2_t int16x8x2_t
#define ggml_uint8x16x2_t uint8x16x2_t
#define ggml_uint8x16x4_t uint8x16x4_t
#define ggml_int8x16x2_t int8x16x2_t
#define ggml_int8x16x4_t int8x16x4_t
#define ggml_vld1q_s16_x2 vld1q_s16_x2
#define ggml_vld1q_u8_x2 vld1q_u8_x2
#define ggml_vld1q_u8_x4 vld1q_u8_x4
#define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8
#define ggml_vqtbl1q_u8 vqtbl1q_u8
#endif // !defined(__aarch64__)
#if !defined(__ARM_FEATURE_DOTPROD)
inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}
#else
#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
#endif // !defined(__ARM_FEATURE_DOTPROD)
#endif // defined(__ARM_NEON)
#if defined(__ARM_NEON) && !defined(_MSC_VER)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
ggml_fp16_internal_t tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
ggml_fp16_internal_t tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
#else
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#undef bool
#define bool _Bool
#else
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#else
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
#if !defined(__riscv)
#include <immintrin.h>
#endif
#endif
#endif
#endif
#endif
#ifdef __riscv_v_intrinsic
#include <riscv_vector.h>
#endif
#if defined(__loongarch64)
#if defined(__loongarch_asx)
#include <lasxintrin.h>
#endif
#if defined(__loongarch_sx)
#include <lsxintrin.h>
#endif
#endif
#if defined(__loongarch_asx)
typedef union {
int32_t i;
float f;
} ft_union;
/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(float val) {
ft_union fi_tmpval = {.f = val};
return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
}
static __m256 __lasx_xvreplfr2vr_s(float val) {
ft_union fi_tmpval = {.f = val};
return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
}
#endif
#ifdef __F16C__
#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
register float f;
register double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
register double d;
register ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#else
// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
static inline float fp32_from_bits(uint32_t w) {
union {
uint32_t as_bits;
float as_value;
} fp32;
fp32.as_bits = w;
return fp32.as_value;
}
static inline uint32_t fp32_to_bits(float f) {
union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#endif // __F16C__
#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()
extern float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return ggml_table_f32_f16[s];
}
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#endif
#if !defined(GGML_FP32_TO_FP16)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif
#define GGML_HASHTABLE_FULL ((size_t)-1)
#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
struct ggml_hash_set ggml_hash_set_new(size_t size);
bool ggml_hash_contains (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
size_t ggml_hash_find (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
size_t ggml_hash_insert ( struct ggml_hash_set hash_set, struct ggml_tensor * key);
// return index, asserts if table is full
size_t ggml_hash_find_or_insert( struct ggml_hash_set hash_set, struct ggml_tensor * key);
#ifdef __cplusplus
}
#endif
llm/llama.cpp/ggml-kompute.cpp deleted 100644 → 0
View file @ 97b02a89
#include "ggml.h"
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-kompute.h"
// These are generated at build time by cmake custom command
#include "shaderop_scale.h"
#include "shaderop_scale_8.h"
#include "shaderop_add.h"
#include "shaderop_addrow.h"
#include "shaderop_mul.h"
#include "shaderop_silu.h"
#include "shaderop_relu.h"
#include "shaderop_gelu.h"
#include "shaderop_softmax.h"
#include "shaderop_norm.h"
#include "shaderop_rmsnorm.h"
#include "shaderop_diagmask.h"
#include "shaderop_mul_mat_f16.h"
#include "shaderop_mul_mat_q8_0.h"
#include "shaderop_mul_mat_q4_0.h"
#include "shaderop_mul_mat_q4_1.h"
#include "shaderop_mul_mat_q6_k.h"
#include "shaderop_mul_mat_mat_f32.h"
#include "shaderop_getrows_f16.h"
#include "shaderop_getrows_q4_0.h"
#include "shaderop_getrows_q4_1.h"
#include "shaderop_getrows_q6_k.h"
#include "shaderop_rope_f16.h"
#include "shaderop_rope_f32.h"
#include "shaderop_cpy_f16_f16.h"
#include "shaderop_cpy_f16_f32.h"
#include "shaderop_cpy_f32_f16.h"
#include "shaderop_cpy_f32_f32.h"
#include <algorithm>
#include <array>
#include <cassert>
#include <cstdint>
#include <cstdio>
#include <cstring>
#include <iostream>
#include <memory>
#include <stdexcept>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
#include <kompute/Kompute.hpp>
#include <vulkan/vulkan.hpp>
#ifdef __linux__
#include <cstdlib> // for setenv
#endif
#define QK4_0 32
#define QR4_0 2
#define QK4_1 32
#define QK_NL 16
typedef ggml_fp16_t half;
static std::string ggml_kompute_format_name(int device) {
return "Kompute" + std::to_string(device);
}
struct ggml_kompute_context {
int device;
std::string name;
std::shared_ptr<vk::DescriptorPool> pool;
ggml_kompute_context(int device)
: device(device), name(ggml_kompute_format_name(device)) {}
};
// FIXME: It would be good to consolidate the kompute manager and the kompute context into one object
// and consolidate the init functions and simplify object lifetime management. As it currently stands,
// we *have* to have the kompute manager no matter what for device discovery, but the kompute context
// is only created when a device is set and vulkan is explicitly turned on.
static ggml_kompute_context *s_kompute_context = nullptr;
class kompute_manager {
kp::Manager *s_mgr = nullptr;
public:
kp::Manager *operator()() {
if (s_mgr && !s_mgr->hasInstance()) {
destroy();
}
if (!s_mgr) {
s_mgr = new kp::Manager;
}
return s_mgr;
}
void destroy() {
delete s_mgr;
s_mgr = nullptr;
}
};
static kompute_manager komputeManager;
struct ggml_vk_memory {
void *data = nullptr;
size_t size = 0;
vk::DeviceMemory *primaryMemory = nullptr;
vk::Buffer *primaryBuffer = nullptr;
vk::DeviceMemory *stagingMemory = nullptr;
vk::Buffer *stagingBuffer = nullptr;
};
#ifdef __linux__
__attribute__((constructor))
static void enable_sam() {
setenv("RADV_PERFTEST", "sam", false);
}
#endif
static bool ggml_vk_checkPhysicalDeviceFeatures(vk::PhysicalDevice physical_device) {
vk::PhysicalDeviceFeatures availableFeatures;
physical_device.getFeatures(&availableFeatures);
if (!availableFeatures.shaderInt16)
return false;
vk::PhysicalDeviceVulkan11Features availableFeatures11;
vk::PhysicalDeviceVulkan12Features availableFeatures12;
availableFeatures11.pNext = &availableFeatures12;
availableFeatures12.pNext = nullptr;
vk::PhysicalDeviceFeatures2 features2;
features2.pNext = &availableFeatures11;
physical_device.getFeatures2(&features2);
if (!availableFeatures11.uniformAndStorageBuffer16BitAccess ||
!availableFeatures11.storageBuffer16BitAccess) {
return false;
}
if (!availableFeatures12.storageBuffer8BitAccess ||
!availableFeatures12.uniformAndStorageBuffer8BitAccess ||
!availableFeatures12.shaderFloat16 ||
!availableFeatures12.shaderInt8) {
return false;
}
return true;
}
static const char * ggml_vk_getVendorName(uint32_t vendorID) {
switch (vendorID) {
case 0x10DE:
return "nvidia";
case 0x1002:
return "amd";
case 0x8086:
return "intel";
default:
return "unknown";
}
}
static std::vector<ggml_vk_device> ggml_vk_available_devices_internal(size_t memoryRequired) {
std::vector<ggml_vk_device> results;
if (!komputeManager()->hasVulkan() || !komputeManager()->hasInstance())
return results;
std::vector<vk::PhysicalDevice> physical_devices;
try {
physical_devices = komputeManager()->listDevices();
} catch (vk::SystemError & err) {
std::cerr << __func__ << ": ignoring Vulkan exception: " << err.what() << "\n";
return results;
}
uint32_t deviceCount = physical_devices.size();
if (deviceCount == 0)
return results;
std::unordered_map<std::string, size_t> count_by_name;
for (uint32_t i = 0; i < deviceCount; i++) {
const auto & physical_device = physical_devices[i];
VkPhysicalDeviceProperties dev_props = physical_device.getProperties();
VkPhysicalDeviceMemoryProperties memoryProperties = physical_device.getMemoryProperties();
const uint32_t major = VK_VERSION_MAJOR(dev_props.apiVersion);
const uint32_t minor = VK_VERSION_MINOR(dev_props.apiVersion);
if (major < 1 || minor < 2)
continue;
if (!ggml_vk_checkPhysicalDeviceFeatures(physical_device))
continue;
size_t heapSize = 0;
for (uint32_t j = 0; j < memoryProperties.memoryHeapCount; ++j) {
VkMemoryHeap heap = memoryProperties.memoryHeaps[j];
if (heap.flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) {
heapSize = heap.size;
break;
}
}
if (heapSize < memoryRequired)
continue;
auto ext_props = physical_device.enumerateDeviceExtensionProperties();
bool has_maintenance4 = false;
// Check if maintenance4 is supported
for (const auto & properties : ext_props) {
if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
has_maintenance4 = true;
}
}
vk::PhysicalDeviceSubgroupProperties subgroup_props;
vk::PhysicalDeviceProperties2 dev_props2;
vk::PhysicalDeviceMaintenance3Properties dev_props3;
vk::PhysicalDeviceMaintenance4Properties dev_props4;
dev_props2.pNext = &dev_props3;
dev_props3.pNext = &subgroup_props;
if (has_maintenance4) {
subgroup_props.pNext = &dev_props4;
}
physical_device.getProperties2(&dev_props2);
if (subgroup_props.subgroupSize < 32)
continue;
ggml_vk_device d;
d.index = i;
d.type = dev_props.deviceType;
d.heapSize = heapSize;
d.vendor = strdup(ggml_vk_getVendorName(dev_props.vendorID));
d.subgroupSize = subgroup_props.subgroupSize;
d.bufferAlignment = dev_props.limits.minStorageBufferOffsetAlignment;
if (has_maintenance4) {
d.maxAlloc = std::min(dev_props3.maxMemoryAllocationSize, dev_props4.maxBufferSize);
} else {
d.maxAlloc = dev_props3.maxMemoryAllocationSize;
}
std::string name(dev_props.deviceName);
size_t n_idx = ++count_by_name[name];
if (n_idx > 1) {
name += " (" + std::to_string(n_idx) + ")";
}
d.name = strdup(name.c_str());
results.push_back(d);
}
std::stable_sort(results.begin(), results.end(),
[](const ggml_vk_device& lhs, const ggml_vk_device& rhs) -> bool {
if (lhs.type != rhs.type) {
if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return true;
if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return false;
if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return true;
if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return false;
}
return lhs.heapSize < rhs.heapSize;
}
);
return results;
}
// public API returns a C-style array
ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count) {
auto devices = ggml_vk_available_devices_internal(memoryRequired);
*count = devices.size();
if (devices.empty()) {
return nullptr;
}
size_t nbytes = sizeof (ggml_vk_device) * (devices.size());
auto * arr = static_cast<ggml_vk_device *>(malloc(nbytes));
memcpy(arr, devices.data(), nbytes);
return arr;
}
static void ggml_vk_filterByVendor(std::vector<ggml_vk_device>& devices, const std::string& targetVendor) {
devices.erase(
std::remove_if(devices.begin(), devices.end(),
[&targetVendor](const ggml_vk_device& device) {
return device.vendor != targetVendor;
}),
devices.end()
);
}
static void ggml_vk_filterByName(std::vector<ggml_vk_device>& devices, const std::string& targetName) {
devices.erase(
std::remove_if(devices.begin(), devices.end(),
[&targetName](const ggml_vk_device& device) {
return device.name != targetName;
}),
devices.end()
);
}
static bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const std::string & name) {
if (name.empty())
return false;
auto devices = ggml_vk_available_devices_internal(memoryRequired);
if (name == "amd" || name == "nvidia" || name == "intel") {
ggml_vk_filterByVendor(devices, name);
} else if (name != "gpu") {
ggml_vk_filterByName(devices, name);
}
if (devices.empty())
return false;
*device = devices.front();
return true;
}
bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const char * name) {
return ggml_vk_get_device(device, memoryRequired, std::string(name));
}
bool ggml_vk_has_vulkan() {
return komputeManager()->hasVulkan();
}
bool ggml_vk_has_device() {
return komputeManager()->hasDevice();
}
ggml_vk_device ggml_vk_current_device() {
if (!komputeManager()->hasDevice())
return ggml_vk_device();
auto devices = ggml_vk_available_devices_internal(0);
ggml_vk_filterByName(devices, komputeManager()->physicalDevice()->getProperties().deviceName.data());
GGML_ASSERT(!devices.empty());
return devices.front();
}
static
void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t size) {
std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
vk::DescriptorPoolSize(
vk::DescriptorType::eStorageBuffer,
3 * size // Descriptor count is number of possible tensors to pass into an algorithm
)
};
vk::DescriptorPoolCreateInfo descriptorPoolInfo(
vk::DescriptorPoolCreateFlags(),
size, // Max sets
static_cast<uint32_t>(descriptorPoolSizes.size()),
descriptorPoolSizes.data());
ctx->pool = std::make_shared<vk::DescriptorPool>();
vk::Result r = komputeManager()->device()->createDescriptorPool(
&descriptorPoolInfo, nullptr, ctx->pool.get());
if (r != vk::Result::eSuccess)
std::cerr << "Error allocating descriptor pool" << vk::to_string(r);
}
static
void ggml_vk_free_descriptor_pool(struct ggml_kompute_context * ctx) {
if (ctx->pool) {
komputeManager()->device()->destroy(
*ctx->pool,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
ctx->pool = nullptr;
}
}
static
vk::Buffer *ggml_vk_allocate_buffer(size_t size) {
vk::BufferCreateInfo bufferCreateInfo;
bufferCreateInfo.size = size;
bufferCreateInfo.usage = vk::BufferUsageFlagBits::eStorageBuffer |
vk::BufferUsageFlagBits::eTransferSrc |
vk::BufferUsageFlagBits::eTransferDst;
bufferCreateInfo.sharingMode = vk::SharingMode::eExclusive;
vk::Buffer *vkBuffer = new vk::Buffer;
vk::Result r = komputeManager()->device()->createBuffer(&bufferCreateInfo, nullptr, vkBuffer);
if (r != vk::Result::eSuccess)
std::cerr << "Error allocating buffer " << vk::to_string(r) << std::endl;
return vkBuffer;
}
static
vk::DeviceMemory *ggml_vk_allocate(size_t size, vk::MemoryPropertyFlags flags, vk::MemoryRequirements requirements, bool *isHostVisible) {
uint32_t memoryTypeIndex = -1;
bool memoryTypeIndexFound = false;
vk::PhysicalDeviceMemoryProperties memoryProperties = komputeManager()->physicalDevice()->getMemoryProperties();
for (uint32_t i = 0; i < memoryProperties.memoryTypeCount; i++) {
const vk::MemoryType &memoryType = memoryProperties.memoryTypes[i];
const vk::MemoryHeap &memoryHeap = memoryProperties.memoryHeaps[memoryType.heapIndex];
if (memoryHeap.size < size) {
continue;
}
if (requirements.memoryTypeBits & (1 << i)) {
if (((memoryProperties.memoryTypes[i]).propertyFlags &
flags) == flags) {
memoryTypeIndex = i;
memoryTypeIndexFound = true;
if (isHostVisible && (memoryProperties.memoryTypes[i].propertyFlags & vk::MemoryPropertyFlagBits::eHostVisible)) {
*isHostVisible = true;
}
break;
}
}
}
if (!memoryTypeIndexFound) {
throw std::runtime_error(
"Memory type index for buffer creation not found");
}
vk::MemoryAllocateInfo allocInfo;
allocInfo.allocationSize = size;
allocInfo.memoryTypeIndex = memoryTypeIndex;
vk::DeviceMemory *vkDeviceMemory = new vk::DeviceMemory;
vk::Result r = komputeManager()->device()->allocateMemory(&allocInfo, nullptr, vkDeviceMemory);
if (r != vk::Result::eSuccess) {
std::cerr << "Error allocating memory " << vk::to_string(r) << std::endl;
throw std::runtime_error("Error allocating vulkan memory.");
}
return vkDeviceMemory;
}
static size_t ggml_vk_aligned_offset(ggml_backend_buffer_t buffer, size_t offset) {
size_t minStorageBufferOffsetAlignment = ggml_backend_buffer_get_alignment(buffer);
// If offset is already aligned, return it directly
if (offset % minStorageBufferOffsetAlignment == 0) {
return offset;
}
// Otherwise, return the largest multiple of minStorageBufferOffsetAlignment less than offset
return (offset / minStorageBufferOffsetAlignment) * minStorageBufferOffsetAlignment;
}
static ggml_vk_memory ggml_vk_allocate(size_t size) {
ggml_vk_memory memory;
bool isHostVisible = false;
{
memory.primaryBuffer = ggml_vk_allocate_buffer(size);
vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.primaryBuffer);
vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eDeviceLocal;
memory.primaryMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
komputeManager()->device()->bindBufferMemory(*memory.primaryBuffer, *memory.primaryMemory, 0);
if (isHostVisible) {
vk::Result r = komputeManager()->device()->mapMemory(*memory.primaryMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
if (r != vk::Result::eSuccess)
std::cerr << "Error mapping memory" << vk::to_string(r);
}
}
if (!isHostVisible) {
memory.stagingBuffer = ggml_vk_allocate_buffer(size);
vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.stagingBuffer);
vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eHostVisible |
vk::MemoryPropertyFlagBits::eHostCoherent |
vk::MemoryPropertyFlagBits::eHostCached;
memory.stagingMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
komputeManager()->device()->bindBufferMemory(*memory.stagingBuffer, *memory.stagingMemory, 0);
vk::Result r = komputeManager()->device()->mapMemory(*memory.stagingMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
if (r != vk::Result::eSuccess)
std::cerr << "Error mapping memory" << vk::to_string(r);
}
memory.size = size;
return memory;
}
static void ggml_vk_free_memory(ggml_vk_memory &memory)
{
komputeManager()->device()->destroy(
*memory.primaryBuffer,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
if (memory.stagingBuffer) {
komputeManager()->device()->destroy(
*memory.stagingBuffer,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
}
komputeManager()->device()->freeMemory(
*memory.primaryMemory,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
if (memory.stagingMemory) {
komputeManager()->device()->freeMemory(
*memory.stagingMemory,
(vk::Optional<const vk::AllocationCallbacks>)nullptr);
}
}
static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft);
static
ggml_vk_memory * ggml_vk_find_tensor(const struct ggml_tensor * t, uint64_t & offset) {
ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
// compatibility with ggml-backend
GGML_ASSERT(buffer && buffer->buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name);
ggml_vk_memory * buf_ctx = static_cast<ggml_vk_memory *>(buffer->context);
const intptr_t ioffs = intptr_t(t->data) - intptr_t(buf_ctx->data);
GGML_ASSERT(ioffs >= 0 && ioffs + int64_t(ggml_nbytes(t)) <= int64_t(buffer->size));
offset = uint64_t(ioffs);
return buf_ctx;
}
static
const std::shared_ptr<kp::Tensor> ggml_vk_get_tensor(const struct ggml_tensor * t, uint32_t * alignedOffset = nullptr) {
uint64_t originalOffset = 0;
auto * res = ggml_vk_find_tensor(t, originalOffset);
if (!res) {
static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
return nullTensor;
}
// Create a tensor whose memory will be composed of our buffers at the correct offset
const size_t nelements = ggml_nelements(t);
size_t nbytes = ggml_nbytes(t);
size_t vulkanOffset = ggml_vk_aligned_offset(t->buffer, originalOffset);
if (alignedOffset) {
*alignedOffset = originalOffset - vulkanOffset;
nbytes += *alignedOffset;
}
return komputeManager()->tensor(
t->data,
nelements,
nbytes, kp::Tensor::TensorDataTypes::eFloat,
res->primaryMemory, res->primaryBuffer,
res->stagingMemory, res->stagingBuffer,
vulkanOffset);
}
static std::vector<uint32_t> getSpirvShader(const unsigned char* rawData, size_t size) {
if (size % sizeof(uint32_t) != 0) {
throw std::runtime_error("Invalid size: must be divisible by sizeof(uint32_t)");
}
const uint32_t* data_ptr = reinterpret_cast<const uint32_t*>(rawData);
size_t count = size / sizeof(uint32_t);
return std::vector<uint32_t>(data_ptr, data_ptr + count);
}
inline static
uint32_t safe_divide(uint32_t a, uint32_t b) {
if (b <= 1) {
return a;
}
if ((a % b) != 0) {
fprintf(stderr, "((%u %% %u) == %u) != 0\n", a, b, a % b);
GGML_ASSERT(!"safe_divide result would've had remainder");
}
return a / b;
}
static void ggml_vk_add(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
int32_t ne0,
int32_t nb0, int32_t nb1, int32_t nb2, int32_t nb3
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_add_comp_spv,
kp::shader_data::op_add_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00;
int32_t nb00, nb01, nb02, nb03;
int32_t ne10, ne11, ne12, ne13;
int32_t nb10, nb11, nb12, nb13;
int32_t ne0;
int32_t nb0, nb1, nb2, nb3;
} const pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
ne0,
nb0, nb1, nb2, nb3
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_addrow(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
uint32_t size, uint32_t row = 0) {
const static auto spirv = getSpirvShader(kp::shader_data::op_addrow_comp_spv,
kp::shader_data::op_addrow_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
uint32_t row;
} const pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
row
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__))
s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({size});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
int32_t ne0,
int32_t nb0, int32_t nb1, int32_t nb2, int32_t nb3
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_comp_spv,
kp::shader_data::op_mul_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00;
int32_t nb00, nb01, nb02, nb03;
int32_t ne10, ne11, ne12, ne13;
int32_t nb10, nb11, nb12, nb13;
int32_t ne0;
int32_t nb0, nb1, nb2, nb3;
} const pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
ne0,
nb0, nb1, nb2, nb3
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_scale(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
uint32_t size, float scale) {
const static auto spirv_1 = getSpirvShader(
kp::shader_data::op_scale_comp_spv, kp::shader_data::op_scale_comp_spv_len
);
const static auto spirv_8 = getSpirvShader(
kp::shader_data::op_scale_8_comp_spv, kp::shader_data::op_scale_8_comp_spv_len
);
struct PushConstants {
uint32_t inOff, outOff;
float scale;
} const pushConsts {
safe_divide(inOff, 4), safe_divide(outOff, 4),
scale
};
const auto * spirv = &spirv_1;
std::string name(__func__);
if (size % 8 == 0) {
size /= 8;
name += "_8";
spirv = &spirv_8;
}
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, *spirv, {size}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({size});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_xxlu(
const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
uint32_t size
) {
struct PushConstants {
uint32_t inOff, outOff;
} const pushConsts {
safe_divide(inOff, 4), safe_divide(outOff, 4),
};
auto name = std::string(__func__) + "_" + suffix;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {size}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({size});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_silu(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_silu_comp_spv,
kp::shader_data::op_silu_comp_spv_len);
ggml_vk_xxlu(spirv, "silu", std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_relu(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_relu_comp_spv,
kp::shader_data::op_relu_comp_spv_len);
ggml_vk_xxlu(spirv, "relu", std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_gelu(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_gelu_comp_spv,
kp::shader_data::op_gelu_comp_spv_len);
ggml_vk_xxlu(spirv, "gelu", std::forward<Args>(args)...);
}
static void ggml_vk_soft_max(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, uint32_t ne03,
float scale
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_softmax_comp_spv,
kp::shader_data::op_softmax_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne01, ne02;
float scale;
int32_t mask;
} pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne01, ne02,
scale,
bool(inB)
};
auto & inB_ = inB ? inB : inA;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
// FIXME: The softmax kernel needs to be fixed to use the subgroupsize which can vary by device
const uint32_t local_x = 32;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {local_x}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB_, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_norm_(
const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
int32_t ne00, int32_t nb01,
int32_t nrows, float epsilon
) {
GGML_ASSERT(nb01%sizeof(float) == 0);
GGML_ASSERT(ne00%sizeof(float) == 0);
struct PushConstants {
uint32_t inOff, outOff;
uint32_t ne00, nb01;
float eps;
} pushConsts {
safe_divide(inOff, 4), safe_divide(outOff, 4),
(uint32_t)ne00, (uint32_t)nb01, epsilon
};
auto name = std::string(__func__) + "_" + suffix;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {(uint32_t)nrows}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({(uint32_t)nrows});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_norm(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_norm_comp_spv,
kp::shader_data::op_norm_comp_spv_len);
ggml_vk_norm_(spirv, "norm", std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_rms_norm(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_rmsnorm_comp_spv,
kp::shader_data::op_rmsnorm_comp_spv_len);
ggml_vk_norm_(spirv, "rms", std::forward<Args>(args)...);
}
static void ggml_vk_diag_mask_inf(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
uint32_t n_past,
int32_t ne00, int32_t ne01, int32_t ne02) {
const static auto spirv = getSpirvShader(kp::shader_data::op_diagmask_comp_spv,
kp::shader_data::op_diagmask_comp_spv_len);
struct PushConstants {
uint32_t inOff, outOff;
uint32_t n_past;
int32_t ne00, ne01;
} pushConsts {
safe_divide(inOff, 4), safe_divide(outOff, 4),
n_past,
ne00, ne01
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__))
s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne00), unsigned(ne01), unsigned(ne02)}, {}, {pushConsts});
else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({unsigned(ne00), unsigned(ne01), unsigned(ne02)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul_mat_f16(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02,
uint32_t nb00, uint32_t nb01, uint32_t nb02,
int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
uint32_t nb10, uint32_t nb11, uint32_t nb12,
int32_t ne0, int32_t ne1,
uint32_t r2, uint32_t r3
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_f16_comp_spv,
kp::shader_data::op_mul_mat_f16_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne01, ne02;
uint32_t nb00, nb01, nb02;
int32_t ne10, ne11, ne12;
uint32_t nb10, nb11, nb12;
int32_t ne0, ne1;
uint32_t r2, r3;
} pushConsts {
safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne01, ne02,
nb00, nb01, nb02,
ne10, ne11, ne12,
nb10, nb11, nb12,
ne0, ne1,
r2, r3
};
const unsigned ny = unsigned((ne11 + 4 - 1)/4);
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), ny, unsigned(ne12*ne13)}, {local_x}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), ny, unsigned(ne12*ne13)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02,
uint32_t nb01, uint32_t nb02,
int32_t ne11, int32_t ne12,
uint32_t nb11, uint32_t nb12,
uint32_t nb1, uint32_t nb2) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_mat_f32_comp_spv,
kp::shader_data::op_mul_mat_mat_f32_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne01, ne02, ne11, ne12;
uint32_t nb01, nb02;
uint32_t nb11, nb12;
uint32_t nb1, nb2;
} pushConsts {
safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne01, ne02, ne11, ne12,
nb01, nb02, nb11, nb12,
nb1, nb2
};
const uint32_t local_x = ggml_vk_current_device().subgroupSize;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(),
{inA, inB, out}, spirv,
{unsigned(ne01),
unsigned(ne11),
unsigned(std::max(ne12, ne02))
},
{local_x},
{pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01),
unsigned(ne11),
unsigned(std::max(ne12, ne02)),
});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_mul_mat_impl(
const std::vector<uint32_t>& spirv, const char * suffix, uint32_t block_size, kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02,
int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
int32_t ne0, int32_t ne1,
uint32_t r2, uint32_t r3
) {
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne01, ne02;
int32_t ne10, ne12;
int32_t ne0, ne1;
uint32_t r2, r3;
} pushConsts {
safe_divide(inAOff, block_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne01, ne02,
ne10, ne12,
ne0, ne1,
r2, r3
};
auto name = std::string(__func__) + "_" + suffix;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_mul_mat_q4_0(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_0_comp_spv,
kp::shader_data::op_mul_mat_q4_0_comp_spv_len);
ggml_vk_mul_mat_impl(spirv, "q4_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_mul_mat_q4_1(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_1_comp_spv,
kp::shader_data::op_mul_mat_q4_1_comp_spv_len);
ggml_vk_mul_mat_impl(spirv, "q4_1", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_mul_mat_q8_0(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q8_0_comp_spv,
kp::shader_data::op_mul_mat_q8_0_comp_spv_len);
ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
}
static void ggml_vk_mul_mat_q6_k(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne10, int32_t ne0, int32_t ne1,
int32_t ne01, int32_t ne11, int32_t ne12, int32_t ne02
) {
const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q6_k_comp_spv,
kp::shader_data::op_mul_mat_q6_k_comp_spv_len);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, ne10, ne0, ne1, ne01, gqa;
} pushConsts {
inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, ne10, ne0, ne1, ne01, ne12/ne02
};
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) {
const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_get_rows(
const std::vector<uint32_t>& spirv,
const char * suffix,
unsigned element_size, unsigned qk,
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t nb01, int32_t nb1,
uint32_t size
) {
GGML_ASSERT(nb01%element_size == 0);
GGML_ASSERT(nb1%sizeof(float) == 0);
if (qk) GGML_ASSERT(ne00%qk == 0);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t ne00, nb01, nb1;
} pushConsts {
safe_divide(inAOff, element_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, nb01, nb1
};
auto name = std::string(__func__) + "_" + suffix;
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({size});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_get_rows_f16(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f16_comp_spv,
kp::shader_data::op_getrows_f16_comp_spv_len);
ggml_vk_get_rows(spirv, "f16", sizeof(half), 0, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_get_rows_q4_0(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_0_comp_spv,
kp::shader_data::op_getrows_q4_0_comp_spv_len);
ggml_vk_get_rows(spirv, "q4_0", 1/*We access blocks unaligned*/, QK4_0, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_get_rows_q4_1(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_1_comp_spv,
kp::shader_data::op_getrows_q4_1_comp_spv_len);
ggml_vk_get_rows(spirv, "q4_1", 1/*We access blocks unaligned*/, QK4_1, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_get_rows_q6_k(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q6_k_comp_spv,
kp::shader_data::op_getrows_q6_k_comp_spv_len);
ggml_vk_get_rows(spirv, "q6_k", 1/*We access blocks unaligned*/, QK_NL, std::forward<Args>(args)...);
}
static void ggml_vk_rope(
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inA,
const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
ggml_type src0t, int32_t n_dims, int32_t mode, int32_t n_orig_ctx,
float freq_base, float freq_scale, float ext_factor, float attn_factor, float beta_fast, float beta_slow,
int32_t ne01, int32_t ne02, int32_t ne03,
uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
int32_t ne0,
uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
) {
GGML_ASSERT(src0t == GGML_TYPE_F16 || src0t == GGML_TYPE_F32);
static const auto spirv_f16 = getSpirvShader(
kp::shader_data::op_rope_f16_comp_spv, kp::shader_data::op_rope_f16_comp_spv_len
);
static const auto spirv_f32 = getSpirvShader(
kp::shader_data::op_rope_f32_comp_spv, kp::shader_data::op_rope_f32_comp_spv_len
);
int type_size = src0t == GGML_TYPE_F16 ? 2 : 4;
GGML_ASSERT(nb03 % type_size == 0);
GGML_ASSERT(nb02 % type_size == 0);
GGML_ASSERT(nb01 % type_size == 0);
GGML_ASSERT(nb00 % type_size == 0);
GGML_ASSERT(nb3 % type_size == 0);
GGML_ASSERT(nb2 % type_size == 0);
GGML_ASSERT(nb1 % type_size == 0);
GGML_ASSERT(nb0 % type_size == 0);
struct PushConstants {
uint32_t inAOff, inBOff, outOff;
int32_t n_dims, mode, n_orig_ctx;
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
uint32_t nb00, nb01, nb02, nb03;
int32_t ne0;
uint32_t nb0, nb1, nb2, nb3;
} pushConsts {
safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(outOff, type_size),
n_dims, mode, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
nb00, nb01, nb02, nb03,
ne0,
nb0, nb1, nb2, nb3
};
auto name = std::string(__func__) + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32");
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name)) {
s_algo = komputeManager()->algorithm<float, PushConstants>(
name, s_kompute_context->pool.get(), {inA, inB, out},
src0t == GGML_TYPE_F16 ? spirv_f16 : spirv_f32,
{unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}
);
} else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
static void ggml_vk_cpy(
const std::vector<uint32_t>& spirv,
uint32_t in_element_size, uint32_t out_element_size,
kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& in,
const std::shared_ptr<kp::Tensor>& out,
uint32_t inOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
int32_t ne0, int32_t ne1, int32_t ne2,
uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
) {
struct PushConstants {
uint32_t inOff, outOff;
int32_t ne00, ne01, ne02;
uint32_t nb00, nb01, nb02, nb03;
int32_t ne0, ne1, ne2;
uint32_t nb0, nb1, nb2, nb3;
} pushConsts {
safe_divide(inOff, in_element_size), safe_divide(outOff, out_element_size),
ne00, ne01, ne02,
nb00, nb01, nb02, nb03,
ne0, ne1, ne2,
nb0, nb1, nb2, nb3
};
std::string name = std::string(__func__)
+ "_i_" + std::to_string(in_element_size)
+ "_o_" + std::to_string(out_element_size);
std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(name))
s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
else {
s_algo = komputeManager()->getAlgorithm(name);
s_algo->setTensors({in, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get());
}
seq.record<kp::OpAlgoDispatch>(s_algo);
}
template <typename... Args>
static void ggml_vk_cpy_f32_f16(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f16_comp_spv,
kp::shader_data::op_cpy_f32_f16_comp_spv_len);
ggml_vk_cpy(spirv, 4, 2, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_cpy_f32_f32(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f32_comp_spv,
kp::shader_data::op_cpy_f32_f32_comp_spv_len);
ggml_vk_cpy(spirv, 4, 4, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_cpy_f16_f16(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f16_comp_spv,
kp::shader_data::op_cpy_f16_f16_comp_spv_len);
ggml_vk_cpy(spirv, 2, 2, std::forward<Args>(args)...);
}
template <typename... Args>
static void ggml_vk_cpy_f16_f32(Args&&... args) {
const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f32_comp_spv,
kp::shader_data::op_cpy_f16_f32_comp_spv_len);
ggml_vk_cpy(spirv, 2, 4, std::forward<Args>(args)...);
}
static bool ggml_vk_supports_op(const struct ggml_tensor * op) {
switch (op->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
break;
default:
return false;
}
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_SILU:
return true;
default:
;
}
break;
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
case GGML_OP_ADD:
case GGML_OP_MUL:
case GGML_OP_SCALE:
case GGML_OP_SOFT_MAX:
case GGML_OP_RMS_NORM:
case GGML_OP_NORM:
case GGML_OP_ROPE:
return true;
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
switch (op->src[0]->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
break;
default:
return false;
}
switch (op->type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
break;
default:
return false;
}
return true;
case GGML_OP_DIAG_MASK_INF:
return op->ne[3] == 1;
case GGML_OP_GET_ROWS:
switch (op->src[0]->type) {
case GGML_TYPE_F16:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q6_K:
return op->ne[2] == 1 && op->ne[3] == 1;
default:
;
}
return false;
case GGML_OP_MUL_MAT:
if (op->src[1]->type != GGML_TYPE_F32 || ggml_is_transposed(op->src[0]) || ggml_is_transposed(op->src[1]))
return false;
switch (op->src[0]->type) {
case GGML_TYPE_F32:
case GGML_TYPE_Q6_K:
return op->ne[3] == 1;
case GGML_TYPE_F16:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
return true;
default:
;
}
default:
;
}
return false;
}
static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) {
const int n_seq = 8;
// FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting
// it to the size of the graph, but I think it can be made smaller?
ggml_vk_allocate_descriptor_pool(ctx, gf->n_nodes);
std::vector<std::shared_ptr<kp::Sequence>> sequences(n_seq);
for (auto& sequence : sequences) {
sequence = komputeManager()->sequence();
}
for (int seq_idx = 0; seq_idx < n_seq; ++seq_idx) {
const int n_nodes_per_seq = (gf->n_nodes + n_seq - 1) / n_seq;
auto& seq = *sequences[seq_idx];
const int node_start = (seq_idx + 0) * n_nodes_per_seq;
const int node_end = std::min((seq_idx == n_seq - 1) ? gf->n_nodes : (seq_idx + 1) * n_nodes_per_seq, gf->n_nodes);
bool any_commands_recorded = false;
for (int i = node_start; i < node_end; ++i) {
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
struct ggml_tensor * src2 = gf->nodes[i]->src[2]; GGML_UNUSED(src2);
struct ggml_tensor * dst = gf->nodes[i];
GGML_ASSERT(dst->data != nullptr);
if (ggml_is_empty(dst)) {
continue;
}
switch (dst->op) {
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
continue; // noop -> next node
default:
break;
}
any_commands_recorded = true;
if (!ggml_vk_supports_op(dst)) {
fprintf(stderr, "%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
GGML_ASSERT(!"unsupported op");
}
const int32_t ne00 = src0 ? src0->ne[0] : 0;
const int32_t ne01 = src0 ? src0->ne[1] : 0;
const int32_t ne02 = src0 ? src0->ne[2] : 0;
const int32_t ne03 = src0 ? src0->ne[3] : 0;
const uint32_t nb00 = src0 ? src0->nb[0] : 0;
const uint32_t nb01 = src0 ? src0->nb[1] : 0;
const uint32_t nb02 = src0 ? src0->nb[2] : 0;
const uint32_t nb03 = src0 ? src0->nb[3] : 0;
const int32_t ne10 = src1 ? src1->ne[0] : 0;
const int32_t ne11 = src1 ? src1->ne[1] : 0;
const int32_t ne12 = src1 ? src1->ne[2] : 0;
const int32_t ne13 = src1 ? src1->ne[3] : 0;
const uint32_t nb10 = src1 ? src1->nb[0] : 0;
const uint32_t nb11 = src1 ? src1->nb[1] : 0;
const uint32_t nb12 = src1 ? src1->nb[2] : 0;
const uint32_t nb13 = src1 ? src1->nb[3] : 0;
const int32_t ne0 = dst ? dst->ne[0] : 0;
const int32_t ne1 = dst ? dst->ne[1] : 0;
const int32_t ne2 = dst ? dst->ne[2] : 0;
// const int32_t ne3 = dst ? dst->ne[3] : 0;
const uint32_t nb0 = dst ? dst->nb[0] : 0;
const uint32_t nb1 = dst ? dst->nb[1] : 0;
const uint32_t nb2 = dst ? dst->nb[2] : 0;
const uint32_t nb3 = dst ? dst->nb[3] : 0;
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
const static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
uint32_t off_src0 = 0;
uint32_t off_src1 = 0;
uint32_t off_dst = 0;
const std::shared_ptr<kp::Tensor>& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor;
const std::shared_ptr<kp::Tensor>& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor;
const std::shared_ptr<kp::Tensor>& id_dst = dst ? ggml_vk_get_tensor(dst, &off_dst) : nullTensor;
switch (dst->op) {
case GGML_OP_ADD:
{
if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
// src1 is a row
ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
} else {
ggml_vk_add(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne03,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
ne0,
nb0, nb1, nb2, nb3
);
}
} break;
case GGML_OP_MUL:
{
ggml_vk_mul(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne03,
nb00, nb01, nb02, nb03,
ne10, ne11, ne12, ne13,
nb10, nb11, nb12, nb13,
ne0,
nb0, nb1, nb2, nb3
);
} break;
case GGML_OP_SCALE:
{
float scale; memcpy(&scale, dst->op_params, sizeof(float));
ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale);
} break;
case GGML_OP_UNARY:
{
int64_t n = ggml_nelements(dst);
GGML_ASSERT(n % 4 == 0);
switch (ggml_get_unary_op(gf->nodes[i])) {
case GGML_UNARY_OP_SILU:
{
ggml_vk_silu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
} break;
case GGML_UNARY_OP_RELU:
{
ggml_vk_relu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
} break;
case GGML_UNARY_OP_GELU:
{
GGML_ASSERT(n % 8 == 0);
ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, n/8);
} break;
default:
{
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
}
} break;
case GGML_OP_SOFT_MAX:
{
float scale;
float max_bias;
memcpy(&scale, (float *)dst->op_params + 0, sizeof(float));
memcpy(&max_bias, (float *)dst->op_params + 1, sizeof(float));
#pragma message("TODO: add ggml_vk_soft_max() F16 src1 support")
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/5021")
GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
#pragma message("TODO: add ALiBi support")
#pragma message("ref: https://github.com/ggerganov/llama.cpp/pull/7192")
GGML_ASSERT(max_bias == 0.0f);
ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale);
} break;
case GGML_OP_DIAG_MASK_INF:
{
const int n_past = ((int32_t *)(dst->op_params))[0];
ggml_vk_diag_mask_inf(seq, id_src0, id_dst, off_src0, off_dst, n_past, ne00, ne01, ne02);
} break;
case GGML_OP_NORM:
{
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
ggml_vk_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
} break;
case GGML_OP_RMS_NORM:
{
GGML_ASSERT(ne00 % 4 == 0);
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
ggml_vk_rms_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
} break;
case GGML_OP_MUL_MAT:
{
GGML_ASSERT(ne00 == ne10);
GGML_ASSERT(ne12 % ne02 == 0);
GGML_ASSERT(ne13 % ne03 == 0);
const uint32_t r2 = ne12/ne02;
const uint32_t r3 = ne13/ne03;
if (src1t != GGML_TYPE_F32) {
fprintf(stderr, "%s: %s: Unsupported src1 type: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
goto not_implemented;
}
if (ggml_is_transposed(src0) ||
ggml_is_transposed(src1)) {
fprintf(stderr, "%s: %s: matmul on tranposed tensor not supported: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
goto not_implemented;
}
switch (src0t) {
case GGML_TYPE_F32:
ggml_vk_mul_mat_mat_f32(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, nb01, nb02, ne11, ne12, nb11, nb12, nb1, nb2
);
break;
case GGML_TYPE_F16:
ggml_vk_mul_mat_f16(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, ne13, nb10, nb11, nb12,
ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q8_0:
ggml_vk_mul_mat_q8_0(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q4_0:
ggml_vk_mul_mat_q4_0(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q4_1:
ggml_vk_mul_mat_q4_1(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
);
break;
case GGML_TYPE_Q6_K:
ggml_vk_mul_mat_q6_k(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02
);
break;
default: {
fprintf(stderr, "%s: %s: Unsupported quantization: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
goto not_implemented;
}
}
} break;
case GGML_OP_GET_ROWS:
{
if (src0t == GGML_TYPE_F16) {
ggml_vk_get_rows_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
} else if (src0t == GGML_TYPE_Q4_0) {
ggml_vk_get_rows_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
} else if (src0t == GGML_TYPE_Q4_1) {
ggml_vk_get_rows_q4_1(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
} else if (src0t == GGML_TYPE_Q6_K) {
ggml_vk_get_rows_q6_k(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
} else {
fprintf(stderr, "%s: %s: Unsupported quantization: %u\n", __func__, ggml_op_name(dst->op), src0t);
goto not_implemented;
}
} break;
case GGML_OP_ROPE:
{
#pragma message("TODO: implement phi3 frequency factors support")
#pragma message(" https://github.com/ggerganov/llama.cpp/pull/7225")
GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
GGML_ASSERT(ne10 == ne02);
GGML_ASSERT(src0t == dstt);
// const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
// skip 3, n_ctx used in GLM RoPE, unimplemented in Vulkan
const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
ggml_vk_rope(
seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, src0t, n_dims, mode, n_orig_ctx,
freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3
);
} break;
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
{
switch (src0t) {
case GGML_TYPE_F32:
{
switch (dstt) {
case GGML_TYPE_F16: ggml_vk_cpy_f32_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
case GGML_TYPE_F32: ggml_vk_cpy_f32_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
default: goto not_implemented;
}
} break;
case GGML_TYPE_F16:
{
switch (dstt) {
case GGML_TYPE_F16: ggml_vk_cpy_f16_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
case GGML_TYPE_F32: ggml_vk_cpy_f16_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
default: goto not_implemented;
} break;
default: goto not_implemented;
}
}
} break;
default: goto not_implemented;
}
continue;
not_implemented: {}
fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
//GGML_ASSERT(false);
}
// Evaluate sequence
if (any_commands_recorded) {
seq.evalAsync();
}
}
// Wait for all sequences to finish
for (auto& sequence : sequences) {
if (sequence->isRunning())
sequence->evalAwait();
}
ggml_vk_free_descriptor_pool(ctx);
}
template<>
kp::Tensor::TensorDataTypes
kp::TensorT<half>::dataType()
{
return TensorDataTypes::eFloat;
}
template<>
kp::Tensor::TensorDataTypes
kp::TensorT<uint8_t>::dataType()
{
return TensorDataTypes::eUnsignedInt;
}
////////////////////////////////////////////////////////////////////////////////
// backend interface
struct ggml_backend_kompute_buffer_type_context {
int device;
int device_ref = 0;
uint64_t buffer_alignment;
uint64_t max_alloc;
std::string name;
ggml_backend_kompute_buffer_type_context(int device, uint64_t buffer_alignment, uint64_t max_alloc)
: device(device), buffer_alignment(buffer_alignment), max_alloc(max_alloc), name(ggml_kompute_format_name(device)) {}
};
static void ggml_backend_kompute_device_ref(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
if (!ctx->device_ref) {
komputeManager()->initializeDevice(
ctx->device, {}, {
"VK_KHR_shader_float16_int8", "VK_KHR_8bit_storage",
"VK_KHR_16bit_storage", "VK_KHR_shader_non_semantic_info"
}
);
}
assert(ggml_vk_has_device());
ctx->device_ref++;
}
static void ggml_backend_kompute_device_unref(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
assert(ctx->device_ref > 0);
ctx->device_ref--;
if (!ctx->device_ref) {
komputeManager.destroy();
}
}
static const char * ggml_backend_kompute_buffer_get_name(ggml_backend_buffer_t buffer) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buffer->buft->context);
return ctx->name.c_str();
}
static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer) {
auto * memory = (ggml_vk_memory *)buffer->context;
if (ggml_vk_has_device()) {
ggml_vk_free_memory(*memory);
}
delete memory;
}
static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
return ((ggml_vk_memory *)buffer->context)->data;
}
static void ggml_backend_kompute_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
GGML_UNUSED(buffer);
const auto res = ggml_vk_get_tensor(tensor);
GGML_ASSERT(res);
memcpy((char *)tensor->data + offset, data, size);
komputeManager()->sequence()->eval<kp::OpTensorSyncDevice>({res});
}
static void ggml_backend_kompute_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
GGML_UNUSED(buffer);
const auto res = ggml_vk_get_tensor(tensor);
GGML_ASSERT(res);
komputeManager()->sequence()->eval<kp::OpTensorSyncLocal>({res});
memcpy(data, (const char *)tensor->data + offset, size);
}
static void ggml_backend_kompute_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
auto * memory = (ggml_vk_memory *)buffer->context;
memset(memory->data, value, buffer->size);
if (memory->stagingBuffer)
komputeManager()->sequence()->eval<kp::OpBufferSyncDevice>(memory->primaryBuffer, memory->stagingBuffer, memory->size);
}
static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = {
/* .get_name = */ ggml_backend_kompute_buffer_get_name,
/* .free_buffer = */ ggml_backend_kompute_buffer_free_buffer,
/* .get_base = */ ggml_backend_kompute_buffer_get_base,
/* .init_tensor = */ NULL,
/* .set_tensor = */ ggml_backend_kompute_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_kompute_buffer_get_tensor,
/* .cpy_tensor = */ NULL,
/* .clear = */ ggml_backend_kompute_buffer_clear,
/* .reset = */ NULL,
};
// default buffer type
static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
return ctx->name.c_str();
}
static ggml_backend_buffer_t ggml_backend_kompute_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
ggml_backend_kompute_device_ref(buft);
auto * ctx = new ggml_vk_memory(ggml_vk_allocate(size));
return ggml_backend_buffer_init(buft, ggml_backend_kompute_buffer_i, ctx, size);
}
static size_t ggml_backend_kompute_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
return ctx->buffer_alignment;
}
static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
return ctx->max_alloc;
}
static bool ggml_backend_kompute_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
GGML_UNUSED(buft);
return ggml_backend_is_kompute(backend);
}
static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = {
/* .get_name = */ ggml_backend_kompute_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_kompute_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_kompute_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_vk_buffer_type_get_max_size,
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .supports_backend = */ ggml_backend_kompute_buffer_type_supports_backend,
/* .is_host = */ NULL,
};
ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) {
static std::vector<ggml_backend_buffer_type> bufts = []() {
std::vector<ggml_backend_buffer_type> vec;
auto devices = ggml_vk_available_devices_internal(0);
vec.reserve(devices.size());
for (const auto & dev : devices) {
vec.push_back({
/* .iface = */ ggml_backend_kompute_buffer_type_interface,
/* .context = */ new ggml_backend_kompute_buffer_type_context(dev.index, dev.bufferAlignment, dev.maxAlloc)
});
}
return vec;
}();
auto it = std::find_if(bufts.begin(), bufts.end(), [device](const ggml_backend_buffer_type & t) {
return device == static_cast<ggml_backend_kompute_buffer_type_context *>(t.context)->device;
});
return it < bufts.end() ? &*it : nullptr;
}
// backend
static const char * ggml_backend_kompute_name(ggml_backend_t backend) {
auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
return ctx->name.c_str();
}
static void ggml_backend_kompute_free(ggml_backend_t backend) {
auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
assert(ctx == s_kompute_context);
s_kompute_context = nullptr;
if (ctx != nullptr) {
delete ctx;
}
delete backend;
}
static ggml_backend_buffer_type_t ggml_backend_kompute_get_default_buffer_type(ggml_backend_t backend) {
auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
return ggml_backend_kompute_buffer_type(ctx->device);
}
static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
ggml_vk_graph_compute(ctx, cgraph);
return GGML_STATUS_SUCCESS;
}
static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
GGML_UNUSED(backend);
return ggml_vk_supports_op(op);
}
static struct ggml_backend_i kompute_backend_i = {
/* .get_name = */ ggml_backend_kompute_name,
/* .free = */ ggml_backend_kompute_free,
/* .get_default_buffer_type = */ ggml_backend_kompute_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_kompute_graph_compute,
/* .supports_op = */ ggml_backend_kompute_supports_op,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .event_synchronize = */ NULL,
};
static ggml_guid_t ggml_backend_kompute_guid() {
static ggml_guid guid = { 0x7b, 0x57, 0xdc, 0xaf, 0xde, 0x12, 0x1d, 0x49, 0xfb, 0x35, 0xfa, 0x9b, 0x18, 0x31, 0x1d, 0xca };
return &guid;
}
ggml_backend_t ggml_backend_kompute_init(int device) {
GGML_ASSERT(s_kompute_context == nullptr);
s_kompute_context = new ggml_kompute_context(device);
ggml_backend_t kompute_backend = new ggml_backend {
/* .guid = */ ggml_backend_kompute_guid(),
/* .interface = */ kompute_backend_i,
/* .context = */ s_kompute_context,
};
return kompute_backend;
}
bool ggml_backend_is_kompute(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
}
static ggml_backend_t ggml_backend_reg_kompute_init(const char * params, void * user_data) {
GGML_UNUSED(params);
return ggml_backend_kompute_init(intptr_t(user_data));
}
extern "C" int ggml_backend_kompute_reg_devices();
int ggml_backend_kompute_reg_devices() {
auto devices = ggml_vk_available_devices_internal(0);
for (const auto & device : devices) {
ggml_backend_register(
ggml_kompute_format_name(device.index).c_str(),
ggml_backend_reg_kompute_init,
ggml_backend_kompute_buffer_type(device.index),
reinterpret_cast<void *>(intptr_t(device.index))
);
}
return devices.size();
}
llm/llama.cpp/ggml-kompute.h deleted 100644 → 0
View file @ 97b02a89
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
struct ggml_vk_device {
int index;
int type; // same as VkPhysicalDeviceType
size_t heapSize;
const char * name;
const char * vendor;
int subgroupSize;
uint64_t bufferAlignment;
uint64_t maxAlloc;
};
struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
bool ggml_vk_has_vulkan(void);
bool ggml_vk_has_device(void);
struct ggml_vk_device ggml_vk_current_device(void);
//
// backend API
//
// forward declaration
typedef struct ggml_backend * ggml_backend_t;
GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
#ifdef __cplusplus
}
#endif
llm/llama.cpp/ggml-metal.h deleted 100644 → 0
View file @ 97b02a89
// An interface allowing to compute ggml_cgraph with Metal
//
// This is a fully functional interface that extends ggml with GPU support for Apple devices.
// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
//
// How it works?
//
// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
//
// You only need to make sure that all memory buffers that you used during the graph creation
// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
// used during the graph evaluation to determine the arguments of the compute kernels.
//
// Synchronization between device and host memory (for example for input and output tensors)
// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
//
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#include <stddef.h>
#include <stdbool.h>
// max memory buffers that can be mapped to the device
#define GGML_METAL_MAX_BUFFERS 64
struct ggml_tensor;
struct ggml_cgraph;
#ifdef __cplusplus
extern "C" {
#endif
//
// backend API
// user-code should use only these functions
//
GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
GGML_API ggml_backend_t ggml_backend_metal_init(void);
GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
// helper to check if the device supports a specific family
// ideally, the user code should be doing these checks
// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
#ifdef __cplusplus
}
#endif
llm/llama.cpp/ggml-metal.m deleted 100644 → 0
View file @ 97b02a89
#import "ggml-metal.h"
#import "ggml-backend-impl.h"
#import "ggml.h"
#import <Foundation/Foundation.h>
#import <Metal/Metal.h>
#undef MIN
#undef MAX
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#ifdef GGML_METAL_NDEBUG
#define GGML_METAL_LOG_INFO(...)
#define GGML_METAL_LOG_WARN(...)
#define GGML_METAL_LOG_ERROR(...)
#else
#define GGML_METAL_LOG_INFO(...) ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
#define GGML_METAL_LOG_WARN(...) ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
#define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
#endif
#define UNUSED(x) (void)(x)
struct ggml_metal_kernel {
id<MTLComputePipelineState> pipeline;
};
enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_ADD,
GGML_METAL_KERNEL_TYPE_ADD_ROW,
GGML_METAL_KERNEL_TYPE_MUL,
GGML_METAL_KERNEL_TYPE_MUL_ROW,
GGML_METAL_KERNEL_TYPE_DIV,
GGML_METAL_KERNEL_TYPE_DIV_ROW,
GGML_METAL_KERNEL_TYPE_REPEAT_F32,
GGML_METAL_KERNEL_TYPE_REPEAT_F16,
GGML_METAL_KERNEL_TYPE_REPEAT_I32,
GGML_METAL_KERNEL_TYPE_REPEAT_I16,
GGML_METAL_KERNEL_TYPE_SCALE,
GGML_METAL_KERNEL_TYPE_SCALE_4,
GGML_METAL_KERNEL_TYPE_CLAMP,
GGML_METAL_KERNEL_TYPE_TANH,
GGML_METAL_KERNEL_TYPE_RELU,
GGML_METAL_KERNEL_TYPE_SIGMOID,
GGML_METAL_KERNEL_TYPE_GELU,
GGML_METAL_KERNEL_TYPE_GELU_4,
GGML_METAL_KERNEL_TYPE_GELU_QUICK,
GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
GGML_METAL_KERNEL_TYPE_SILU,
GGML_METAL_KERNEL_TYPE_SILU_4,
GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4,
GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,
GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,
GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,
GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
GGML_METAL_KERNEL_TYPE_RMS_NORM,
GGML_METAL_KERNEL_TYPE_GROUP_NORM,
GGML_METAL_KERNEL_TYPE_NORM,
GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
GGML_METAL_KERNEL_TYPE_ROPE_F32,
GGML_METAL_KERNEL_TYPE_ROPE_F16,
GGML_METAL_KERNEL_TYPE_IM2COL_F16,
GGML_METAL_KERNEL_TYPE_IM2COL_F32,
GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
GGML_METAL_KERNEL_TYPE_PAD_F32,
GGML_METAL_KERNEL_TYPE_ARANGE_F32,
GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,
GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
//GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
//GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,
GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,
GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,
GGML_METAL_KERNEL_TYPE_CPY_F16_F16,
GGML_METAL_KERNEL_TYPE_CPY_F16_F32,
GGML_METAL_KERNEL_TYPE_CONCAT,
GGML_METAL_KERNEL_TYPE_SQR,
GGML_METAL_KERNEL_TYPE_SUM_ROWS,
GGML_METAL_KERNEL_TYPE_COUNT
};
struct ggml_metal_context {
int n_cb;
id<MTLDevice> device;
id<MTLCommandQueue> queue;
dispatch_queue_t d_queue;
struct ggml_metal_kernel kernels[GGML_METAL_KERNEL_TYPE_COUNT];
bool support_simdgroup_reduction;
bool support_simdgroup_mm;
bool should_capture_next_compute;
};
// MSL code
// TODO: move the contents here when ready
// for now it is easier to work in a separate file
// static NSString * const msl_library_source = @"see metal.metal";
// Here to assist with NSBundle Path Hack
@interface GGMLMetalClass : NSObject
@end
@implementation GGMLMetalClass
@end
static void ggml_metal_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
fprintf(stderr, "%s", msg);
UNUSED(level);
UNUSED(user_data);
}
ggml_log_callback ggml_metal_log_callback = ggml_metal_default_log_callback;
void * ggml_metal_log_user_data = NULL;
GGML_ATTRIBUTE_FORMAT(2, 3)
static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
if (ggml_metal_log_callback != NULL) {
va_list args;
va_start(args, format);
char buffer[128];
int len = vsnprintf(buffer, 128, format, args);
if (len < 128) {
ggml_metal_log_callback(level, buffer, ggml_metal_log_user_data);
} else {
char* buffer2 = malloc(len+1);
va_end(args);
va_start(args, format);
vsnprintf(buffer2, len+1, format, args);
buffer2[len] = 0;
ggml_metal_log_callback(level, buffer2, ggml_metal_log_user_data);
free(buffer2);
}
va_end(args);
}
}
static void * ggml_metal_host_malloc(size_t n) {
void * data = NULL;
#if TARGET_OS_OSX
kern_return_t err = vm_allocate((vm_map_t) mach_task_self(), (void *) &data, n, VM_FLAGS_ANYWHERE);
if (err != KERN_SUCCESS) {
GGML_METAL_LOG_ERROR("%s: error: vm_allocate failed\n", __func__);
return NULL;
}
#else
const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
if (result != 0) {
GGML_METAL_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
return NULL;
}
#endif
return data;
}
static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
#if TARGET_OS_OSX && !GGML_METAL_NDEBUG
// Show all the Metal device instances in the system
NSArray * devices = MTLCopyAllDevices();
for (id<MTLDevice> device in devices) {
GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
}
[devices release]; // since it was created by a *Copy* C method
#endif
// Pick and show default Metal device
id<MTLDevice> device = MTLCreateSystemDefaultDevice();
GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
// Configure context
struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
ctx->device = device;
ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
ctx->queue = [ctx->device newCommandQueue];
ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
id<MTLLibrary> metal_library;
// load library
//
// - first check if the library is embedded
// - then check if the library is in the bundle
// - if not found, load the source and compile it
// - if that fails, return NULL
{
NSBundle * bundle = nil;
#ifdef SWIFT_PACKAGE
bundle = SWIFTPM_MODULE_BUNDLE;
#else
bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
#endif
NSError * error = nil;
#if GGML_METAL_EMBED_LIBRARY
const bool try_metallib = false;
#else
const bool try_metallib = true;
#endif
NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
if (try_metallib && path_lib != nil) {
// pre-compiled library found
NSURL * libURL = [NSURL fileURLWithPath:path_lib];
GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
metal_library = [ctx->device newLibraryWithURL:libURL error:&error];
if (error) {
GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
} else {
#if GGML_METAL_EMBED_LIBRARY
GGML_METAL_LOG_INFO("%s: using embedded metal library\n", __func__);
extern const char ggml_metallib_start[];
extern const char ggml_metallib_end[];
NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
#else
GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
NSString * path_source;
NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
if (path_resource) {
path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
} else {
path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
}
if (path_source == nil) {
GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
path_source = @"ggml-metal.metal";
}
GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
if (error) {
GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
#endif // GGML_METAL_EMBED_LIBRARY
@autoreleasepool {
// dictionary of preprocessor macros
NSMutableDictionary * prep = [NSMutableDictionary dictionary];
MTLCompileOptions* options = [MTLCompileOptions new];
options.preprocessorMacros = prep;
//[options setFastMathEnabled:false];
metal_library = [ctx->device newLibraryWithSource:src options:options error:&error];
if (error) {
GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
}
}
}
// print MTL GPU family:
GGML_METAL_LOG_INFO("%s: GPU name: %s\n", __func__, [[ctx->device name] UTF8String]);
const NSInteger MTLGPUFamilyMetal3 = 5001;
// determine max supported GPU family
// https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
// https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
{
for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
if ([ctx->device supportsFamily:i]) {
GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
break;
}
}
for (int i = MTLGPUFamilyCommon1 + 5; i >= MTLGPUFamilyCommon1; --i) {
if ([ctx->device supportsFamily:i]) {
GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyCommon%d (%d)\n", __func__, i - (int) MTLGPUFamilyCommon1 + 1, i);
break;
}
}
for (int i = MTLGPUFamilyMetal3 + 5; i >= MTLGPUFamilyMetal3; --i) {
if ([ctx->device supportsFamily:i]) {
GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyMetal%d (%d)\n", __func__, i - (int) MTLGPUFamilyMetal3 + 3, i);
break;
}
}
}
ctx->support_simdgroup_reduction = [ctx->device supportsFamily:MTLGPUFamilyApple7];
ctx->support_simdgroup_reduction |= [ctx->device supportsFamily:MTLGPUFamilyMetal3];
ctx->support_simdgroup_mm = [ctx->device supportsFamily:MTLGPUFamilyApple7];
GGML_METAL_LOG_INFO("%s: simdgroup reduction support = %s\n", __func__, ctx->support_simdgroup_reduction ? "true" : "false");
GGML_METAL_LOG_INFO("%s: simdgroup matrix mul. support = %s\n", __func__, ctx->support_simdgroup_mm ? "true" : "false");
GGML_METAL_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
ctx->should_capture_next_compute = false;
#if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
if (@available(macOS 10.12, iOS 16.0, *)) {
GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1e6);
}
#elif TARGET_OS_OSX
if (ctx->device.maxTransferRate != 0) {
GGML_METAL_LOG_INFO("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1e6);
} else {
GGML_METAL_LOG_INFO("%s: maxTransferRate = built-in GPU\n", __func__);
}
#endif
// load kernels
{
NSError * error = nil;
for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
ctx->kernels[i].pipeline = nil;
}
/*
GGML_METAL_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
(int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
(int) kernel->pipeline.threadExecutionWidth); \
*/
#define GGML_METAL_ADD_KERNEL(e, name, supported) \
if (supported) { \
struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \
id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \
[metal_function release]; \
if (error) { \
GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
[metal_library release]; \
return NULL; \
} \
} else { \
GGML_METAL_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \
}
// simd_sum and simd_max requires MTLGPUFamilyApple7
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32, repeat_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16, repeat_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32, repeat_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16, repeat_i16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID, sigmoid, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M, get_rows_iq1_m, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32, rope_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16, rope_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, ctx->support_simdgroup_mm);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, ctx->support_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, ctx->support_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, ctx->support_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true);
}
[metal_library release];
return ctx;
}
static void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
[ctx->kernels[i].pipeline release];
}
[ctx->queue release];
[ctx->device release];
dispatch_release(ctx->d_queue);
free(ctx);
}
// temporarily defined here for compatibility between ggml-backend and the old API
struct ggml_backend_metal_buffer {
void * data;
size_t size;
id<MTLBuffer> metal;
};
struct ggml_backend_metal_buffer_context {
void * all_data;
size_t all_size;
bool owned;
// multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
int n_buffers;
struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
};
// finds the Metal buffer that contains the tensor data on the GPU device
// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
// Metal buffer based on the host memory pointer
//
static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs) {
//GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
const int64_t tsize = ggml_nbytes(t);
ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
// find the view that contains the tensor fully
for (int i = 0; i < buf_ctx->n_buffers; ++i) {
const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
//GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
*offs = (size_t) ioffs;
//GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
return buf_ctx->buffers[i].metal;
}
}
GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
return nil;
}
static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
switch (op->op) {
case GGML_OP_UNARY:
switch (ggml_get_unary_op(op)) {
case GGML_UNARY_OP_TANH:
case GGML_UNARY_OP_RELU:
case GGML_UNARY_OP_SIGMOID:
case GGML_UNARY_OP_GELU:
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_SILU:
return true;
default:
return false;
}
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
case GGML_OP_CONCAT:
case GGML_OP_ADD:
case GGML_OP_ACC:
case GGML_OP_MUL:
case GGML_OP_DIV:
case GGML_OP_REPEAT:
case GGML_OP_SCALE:
case GGML_OP_CLAMP:
case GGML_OP_SQR:
case GGML_OP_SUM_ROWS:
return true;
case GGML_OP_SOFT_MAX:
case GGML_OP_RMS_NORM:
case GGML_OP_GROUP_NORM:
return ctx->support_simdgroup_reduction;
case GGML_OP_NORM:
case GGML_OP_ROPE:
case GGML_OP_IM2COL:
return true;
case GGML_OP_POOL_1D:
case GGML_OP_POOL_2D:
return false;
case GGML_OP_UPSCALE:
case GGML_OP_PAD:
case GGML_OP_ARANGE:
case GGML_OP_TIMESTEP_EMBEDDING:
case GGML_OP_ARGSORT:
case GGML_OP_LEAKY_RELU:
return true;
case GGML_OP_FLASH_ATTN_EXT:
if (op->src[0]->ne[0] == 256) {
return false;
}
return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
return ctx->support_simdgroup_reduction &&
(op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32);
case GGML_OP_CPY:
case GGML_OP_DUP:
case GGML_OP_CONT:
{
switch (op->src[0]->type) {
case GGML_TYPE_F32:
switch (op->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_IQ4_NL:
return true;
default:
return false;
}
case GGML_TYPE_F16:
switch (op->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
return true;
default:
return false;
}
default:
return false;
};
}
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_GET_ROWS:
{
return op->src[0]->type != GGML_TYPE_BF16 && op->ne[3] == 1;
}
default:
return false;
}
}
static enum ggml_status ggml_metal_graph_compute(
struct ggml_metal_context * ctx,
struct ggml_cgraph * gf) {
@autoreleasepool {
MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
edesc.dispatchType = MTLDispatchTypeSerial;
// create multiple command buffers and enqueue them
// then, we encode the graph into the command buffers in parallel
const int n_nodes = gf->n_nodes;
const int n_cb = ctx->n_cb;
const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
const bool should_capture = ctx->should_capture_next_compute;
if (should_capture) {
ctx->should_capture_next_compute = false;
MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new];
descriptor.captureObject = ctx->queue;
NSError * error = nil;
if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) {
GGML_METAL_LOG_ERROR("%s: error: unable to start capture '%s'\n", __func__, [[error localizedDescription] UTF8String]);
GGML_ASSERT(!"capture failed");
}
}
id<MTLCommandBuffer> command_buffer_builder[n_cb];
for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
command_buffer_builder[cb_idx] = command_buffer;
// enqueue the command buffers in order to specify their execution order
[command_buffer enqueue];
}
const id<MTLCommandBuffer> *command_buffers = command_buffer_builder;
dispatch_apply(n_cb, ctx->d_queue, ^(size_t iter) {
const int cb_idx = iter;
size_t offs_src0 = 0;
size_t offs_src1 = 0;
size_t offs_src2 = 0;
size_t offs_dst = 0;
id<MTLCommandBuffer> command_buffer = command_buffers[cb_idx];
id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoderWithDescriptor: edesc];
const int node_start = (cb_idx + 0) * n_nodes_per_cb;
const int node_end = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes);
for (int i = node_start; i < node_end; ++i) {
if (i == -1) {
[encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
continue;
}
//GGML_METAL_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
struct ggml_tensor * src2 = gf->nodes[i]->src[2];
struct ggml_tensor * dst = gf->nodes[i];
if (ggml_is_empty(dst)) {
continue;
}
switch (dst->op) {
case GGML_OP_NONE:
case GGML_OP_RESHAPE:
case GGML_OP_VIEW:
case GGML_OP_TRANSPOSE:
case GGML_OP_PERMUTE:
{
// noop -> next node
} continue;
default:
{
} break;
}
if (!ggml_metal_supports_op(ctx, dst)) {
GGML_METAL_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
GGML_ASSERT(!"unsupported op");
}
if (should_capture) {
[encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(dst) encoding:NSUTF8StringEncoding]];
}
const int64_t ne00 = src0 ? src0->ne[0] : 0;
const int64_t ne01 = src0 ? src0->ne[1] : 0;
const int64_t ne02 = src0 ? src0->ne[2] : 0;
const int64_t ne03 = src0 ? src0->ne[3] : 0;
const uint64_t nb00 = src0 ? src0->nb[0] : 0;
const uint64_t nb01 = src0 ? src0->nb[1] : 0;
const uint64_t nb02 = src0 ? src0->nb[2] : 0;
const uint64_t nb03 = src0 ? src0->nb[3] : 0;
const int64_t ne10 = src1 ? src1->ne[0] : 0;
const int64_t ne11 = src1 ? src1->ne[1] : 0;
const int64_t ne12 = src1 ? src1->ne[2] : 0;
const int64_t ne13 = src1 ? src1->ne[3] : 0;
const uint64_t nb10 = src1 ? src1->nb[0] : 0;
const uint64_t nb11 = src1 ? src1->nb[1] : 0;
const uint64_t nb12 = src1 ? src1->nb[2] : 0;
const uint64_t nb13 = src1 ? src1->nb[3] : 0;
const int64_t ne20 = src2 ? src2->ne[0] : 0;
const int64_t ne21 = src2 ? src2->ne[1] : 0;
const int64_t ne22 = src2 ? src2->ne[2] : 0; GGML_UNUSED(ne22);
const int64_t ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
const uint64_t nb21 = src2 ? src2->nb[1] : 0;
const uint64_t nb22 = src2 ? src2->nb[2] : 0;
const uint64_t nb23 = src2 ? src2->nb[3] : 0;
const int64_t ne0 = dst ? dst->ne[0] : 0;
const int64_t ne1 = dst ? dst->ne[1] : 0;
const int64_t ne2 = dst ? dst->ne[2] : 0;
const int64_t ne3 = dst ? dst->ne[3] : 0;
const uint64_t nb0 = dst ? dst->nb[0] : 0;
const uint64_t nb1 = dst ? dst->nb[1] : 0;
const uint64_t nb2 = dst ? dst->nb[2] : 0;
const uint64_t nb3 = dst ? dst->nb[3] : 0;
const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(dst, &offs_dst) : nil;
//GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
//if (src0) {
// GGML_METAL_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
// ggml_is_contiguous(src0), src0->name);
//}
//if (src1) {
// GGML_METAL_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
// ggml_is_contiguous(src1), src1->name);
//}
//if (dst) {
// GGML_METAL_LOG_INFO("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
// dst->name);
//}
switch (dst->op) {
case GGML_OP_CONCAT:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline;
const int32_t dim = ((int32_t *) dst->op_params)[0];
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26];
[encoder setBytes:&dim length:sizeof(dim) atIndex:27];
const int nth = MIN(1024, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ADD:
case GGML_OP_MUL:
case GGML_OP_DIV:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
GGML_ASSERT(src1t == GGML_TYPE_F32);
const size_t offs = 0;
bool bcast_row = false;
int64_t nb = ne00; // used by the "row" kernels
id<MTLComputePipelineState> pipeline = nil;
if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
GGML_ASSERT(ggml_is_contiguous(src0));
// src1 is a row
GGML_ASSERT(ne11 == 1);
nb = ne00 / 4;
switch (dst->op) {
case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break;
case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break;
default: GGML_ASSERT(false);
}
bcast_row = true;
} else {
switch (dst->op) {
case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break;
case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break;
case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break;
default: GGML_ASSERT(false);
}
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:24];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:25];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:26];
[encoder setBytes:&offs length:sizeof(offs) atIndex:27];
[encoder setBytes:&nb length:sizeof(nb) atIndex:28];
if (bcast_row) {
const int64_t n = ggml_nelements(dst)/4;
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} else {
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
}
} break;
case GGML_OP_REPEAT:
{
id<MTLComputePipelineState> pipeline;
switch (src0t) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break;
case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break;
default: GGML_ASSERT(false);
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ACC:
{
GGML_ASSERT(src0t == GGML_TYPE_F32);
GGML_ASSERT(src1t == GGML_TYPE_F32);
GGML_ASSERT(dstt == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(src0));
GGML_ASSERT(ggml_is_contiguous(src1));
const size_t pnb1 = ((int32_t *) dst->op_params)[0];
const size_t pnb2 = ((int32_t *) dst->op_params)[1];
const size_t pnb3 = ((int32_t *) dst->op_params)[2];
const size_t offs = ((int32_t *) dst->op_params)[3];
const bool inplace = (bool) ((int32_t *) dst->op_params)[4];
if (!inplace) {
// run a separete kernel to cpy src->dst
// not sure how to avoid this
// TODO: make a simpler cpy_bytes kernel
const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
}
const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
[encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8];
[encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9];
[encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:19];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:20];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:21];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:22];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:23];
[encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24];
[encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25];
[encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
[encoder setBytes:&offs length:sizeof(offs) atIndex:27];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
[encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_SCALE:
{
GGML_ASSERT(ggml_is_contiguous(src0));
float scale;
memcpy(&scale, dst->op_params, sizeof(scale));
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
n /= 4;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE_4].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&scale length:sizeof(scale) atIndex:2];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_CLAMP:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
float min;
float max;
memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&min length:sizeof(min) atIndex:2];
[encoder setBytes:&max length:sizeof(max) atIndex:3];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_UNARY:
switch (ggml_get_unary_op(gf->nodes[i])) {
// we are not taking into account the strides, so for now require contiguous tensors
GGML_ASSERT(ggml_is_contiguous(src0));
case GGML_UNARY_OP_TANH:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TANH].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_RELU:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RELU].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_SIGMOID:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIGMOID].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_GELU:
{
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_GELU_QUICK:
{
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_SILU:
{
int64_t n = ggml_nelements(dst);
id<MTLComputePipelineState> pipeline = nil;
if (n % 4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU_4].pipeline;
n /= 4;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
default:
{
GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
} break;
case GGML_OP_SQR:
{
GGML_ASSERT(ggml_is_contiguous(src0));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQR].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_SUM_ROWS:
{
GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:18];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:19];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:20];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:21];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:22];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:23];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:24];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:25];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_SOFT_MAX:
{
GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
int nth = 32; // SIMD width
id<MTLComputePipelineState> pipeline = nil;
const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
if (ne00%4 == 0) {
while (nth < ne00/4 && nth*ne01*ne02*ne03 < 256) {
nth *= 2;
}
if (use_f16) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline;
}
} else {
while (nth < ne00 && nth*ne01*ne02*ne03 < 256) {
nth *= 2;
}
if (use_f16) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32].pipeline;
}
}
float scale;
float max_bias;
memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale));
memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
const int64_t nrows_x = ggml_nrows(src0);
const int64_t nrows_y = src0->ne[1];
const uint32_t n_head = nrows_x/nrows_y;
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
if (id_src1) {
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&scale length:sizeof(scale) atIndex:6];
[encoder setBytes:&max_bias length:sizeof(max_bias) atIndex:7];
[encoder setBytes:&m0 length:sizeof(m0) atIndex:8];
[encoder setBytes:&m1 length:sizeof(m1) atIndex:9];
[encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10];
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_DIAG_MASK_INF:
{
const int n_past = ((int32_t *)(dst->op_params))[0];
id<MTLComputePipelineState> pipeline = nil;
if (ne00%8 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8].pipeline;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline;
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&n_past length:sizeof(int) atIndex:4];
if (ne00%8 == 0) {
[encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
}
else {
[encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
}
} break;
case GGML_OP_MUL_MAT:
{
GGML_ASSERT(ne00 == ne10);
GGML_ASSERT(ne12 % ne02 == 0);
GGML_ASSERT(ne13 % ne03 == 0);
const uint r2 = ne12/ne02;
const uint r3 = ne13/ne03;
// find the break-even point where the matrix-matrix kernel becomes more efficient compared
// to the matrix-vector kernel
int ne11_mm_min = 1;
#if 0
// the numbers below are measured on M2 Ultra for 7B and 13B models
// these numbers do not translate to other devices or model sizes
// TODO: need to find a better approach
if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
switch (src0t) {
case GGML_TYPE_F16: ne11_mm_min = 2; break;
case GGML_TYPE_Q8_0: ne11_mm_min = 7; break;
case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
case GGML_TYPE_Q3_K: ne11_mm_min = 7; break;
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
case GGML_TYPE_Q5_0: // not tested yet
case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
case GGML_TYPE_Q5_K: ne11_mm_min = 7; break;
case GGML_TYPE_Q6_K: ne11_mm_min = 7; break;
default: ne11_mm_min = 1; break;
}
}
#endif
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
!ggml_is_transposed(src0) &&
!ggml_is_transposed(src1) &&
src1t == GGML_TYPE_F32 &&
ne00 % 32 == 0 && ne00 >= 64 &&
(ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
//printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
// some Metal matrix data types require aligned pointers
// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
switch (src0->type) {
case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break;
default: break;
}
id<MTLComputePipelineState> pipeline = nil;
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32 ].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32 ].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32 ].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32 ].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32 ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32 ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32 ].pipeline; break;
case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32 ].pipeline; break;
case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32 ].pipeline; break;
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32 ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32 ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break;
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:5];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:6];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:7];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:8];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:9];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:10];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:11];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:12];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:13];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:14];
[encoder setThreadgroupMemoryLength:8192 atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
int nth0 = 32;
int nth1 = 1;
int nrows = 1;
//printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
id<MTLComputePipelineState> pipeline = nil;
// use custom matrix x vector kernel
switch (src0t) {
case GGML_TYPE_F32:
{
GGML_ASSERT(src1t == GGML_TYPE_F32);
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
nrows = 4;
} break;
case GGML_TYPE_F16:
{
nth0 = 32;
nth1 = 1;
if (src1t == GGML_TYPE_F32) {
if (ne11 * ne12 < 4) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
} else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
nrows = ne11;
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
nrows = 4;
}
} else {
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
nrows = 4;
}
} break;
case GGML_TYPE_Q4_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
} break;
case GGML_TYPE_Q4_1:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
} break;
case GGML_TYPE_Q5_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
} break;
case GGML_TYPE_Q5_1:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
} break;
case GGML_TYPE_Q8_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
} break;
case GGML_TYPE_Q2_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
} break;
case GGML_TYPE_Q3_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
} break;
case GGML_TYPE_Q4_K:
{
nth0 = 4; //1;
nth1 = 8; //32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
} break;
case GGML_TYPE_Q5_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
} break;
case GGML_TYPE_Q6_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
} break;
case GGML_TYPE_IQ2_XXS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ2_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
} break;
case GGML_TYPE_IQ3_XXS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ3_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
} break;
case GGML_TYPE_IQ2_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
} break;
case GGML_TYPE_IQ1_M:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
} break;
case GGML_TYPE_IQ4_NL:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
} break;
case GGML_TYPE_IQ4_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
GGML_ASSERT(false && "not implemented");
}
};
if (ggml_is_quantized(src0t)) {
GGML_ASSERT(ne00 >= nth0*nth1);
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:15];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:16];
[encoder setBytes:&r2 length:sizeof(r2) atIndex:17];
[encoder setBytes:&r3 length:sizeof(r3) atIndex:18];
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 ||
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
const int mem_size = 32*sizeof(float);
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
const int64_t ny = (ne11 + nrows - 1)/nrows;
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
}
} break;
case GGML_OP_MUL_MAT_ID:
{
const int n_as = src0->ne[2];
// src2 = ids
const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
GGML_ASSERT(src2t == GGML_TYPE_I32);
GGML_ASSERT(!ggml_is_transposed(src0));
GGML_ASSERT(!ggml_is_transposed(src1));
GGML_ASSERT(src1t == GGML_TYPE_F32);
// find the break-even point where the matrix-matrix kernel becomes more efficient compared
// to the matrix-vector kernel
// ne20 = n_used_experts
// ne21 = n_rows
const int dst_rows = ne20*ne21;
const int dst_rows_min = n_as;
// max size of the rowids array in the kernel shared buffer
GGML_ASSERT(dst_rows <= 2048);
// for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
// AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
// !!!
// TODO: for now, always use mat-vec kernels until we figure out how to improve the
// indirect matrix multiplication
// !!!
if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
ne00 % 32 == 0 && ne00 >= 64 &&
dst_rows > dst_rows_min) {
// some Metal matrix data types require aligned pointers
// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
switch (src0->type) {
case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8 == 0); break;
default: break;
}
id<MTLComputePipelineState> pipeline = nil;
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32 ].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32 ].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32 ].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32 ].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32 ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32 ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32 ].pipeline; break;
case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32 ].pipeline; break;
case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32 ].pipeline; break;
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32 ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32 ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break;
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32 ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
[encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
[encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:8];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:9];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:10];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:17];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:18];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:19];
[encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else {
int nth0 = 32;
int nth1 = 1;
int nrows = 1;
//printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
id<MTLComputePipelineState> pipeline = nil;
// use custom matrix x vector kernel
switch (src0t) {
case GGML_TYPE_F32:
{
GGML_ASSERT(src1t == GGML_TYPE_F32);
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32].pipeline;
} break;
case GGML_TYPE_F16:
{
GGML_ASSERT(src1t == GGML_TYPE_F32);
nth0 = 32;
nth1 = 1;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32].pipeline;
} break;
case GGML_TYPE_Q4_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32].pipeline;
} break;
case GGML_TYPE_Q4_1:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32].pipeline;
} break;
case GGML_TYPE_Q5_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32].pipeline;
} break;
case GGML_TYPE_Q5_1:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32].pipeline;
} break;
case GGML_TYPE_Q8_0:
{
nth0 = 8;
nth1 = 8;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline;
} break;
case GGML_TYPE_Q2_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32].pipeline;
} break;
case GGML_TYPE_Q3_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32].pipeline;
} break;
case GGML_TYPE_Q4_K:
{
nth0 = 4; //1;
nth1 = 8; //32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32].pipeline;
} break;
case GGML_TYPE_Q5_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32].pipeline;
} break;
case GGML_TYPE_Q6_K:
{
nth0 = 2;
nth1 = 32;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32].pipeline;
} break;
case GGML_TYPE_IQ2_XXS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ2_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32].pipeline;
} break;
case GGML_TYPE_IQ3_XXS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
} break;
case GGML_TYPE_IQ3_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline;
} break;
case GGML_TYPE_IQ2_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline;
} break;
case GGML_TYPE_IQ1_S:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
} break;
case GGML_TYPE_IQ1_M:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline;
} break;
case GGML_TYPE_IQ4_NL:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
} break;
case GGML_TYPE_IQ4_XS:
{
nth0 = 4;
nth1 = 16;
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
} break;
default:
{
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
GGML_ASSERT(false && "not implemented");
}
};
if (ggml_is_quantized(src0t)) {
GGML_ASSERT(ne00 >= nth0*nth1);
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
[encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
[encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12];
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13];
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14];
[encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15];
[encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16];
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17];
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18];
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:20];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:21];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:22];
const int64_t _ne1 = 1;
const int tgz = dst_rows;
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 ||
src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K ||
src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
const int mem_size = 32*sizeof(float);
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q4_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q3_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q5_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
else if (src0t == GGML_TYPE_Q6_K) {
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
} else {
const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
}
}
} break;
case GGML_OP_GET_ROWS:
{
id<MTLComputePipelineState> pipeline = nil;
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F32 ].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F16 ].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0 ].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1 ].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0 ].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1 ].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0 ].pipeline; break;
case GGML_TYPE_Q2_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K ].pipeline; break;
case GGML_TYPE_Q3_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K ].pipeline; break;
case GGML_TYPE_Q4_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K ].pipeline; break;
case GGML_TYPE_Q5_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K ].pipeline; break;
case GGML_TYPE_Q6_K: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K ].pipeline; break;
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S ].pipeline; break;
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S ].pipeline; break;
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break;
case GGML_TYPE_IQ1_M: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M ].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5];
[encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7];
[encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:10];
[encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
} break;
case GGML_OP_RMS_NORM:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(ggml_is_contiguous_1(src0));
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
int nth = 32; // SIMD width
while (nth < ne00/4 && nth < 1024) {
nth *= 2;
}
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
[encoder setBytes:&eps length:sizeof( float) atIndex:4];
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
const int64_t nrows = ggml_nrows(src0);
[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_GROUP_NORM:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(ggml_is_contiguous(src0));
//float eps;
//memcpy(&eps, dst->op_params, sizeof(float));
const float eps = 1e-6f; // TODO: temporarily hardcoded
const int32_t n_groups = ((int32_t *) dst->op_params)[0];
int nth = 32; // SIMD width
//while (nth < ne00/4 && nth < 1024) {
// nth *= 2;
//}
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:5];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8];
[encoder setBytes:&eps length:sizeof( float) atIndex:9];
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_NORM:
{
GGML_ASSERT(ggml_is_contiguous_1(src0));
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
const int nth = MIN(256, ne00);
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NORM].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
[encoder setBytes:&eps length:sizeof( float) atIndex:4];
[encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0];
const int64_t nrows = ggml_nrows(src0);
[encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ROPE:
{
GGML_ASSERT(ne10 == ne02);
const int nth = MIN(1024, ne00);
const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
// skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
float freq_base;
float freq_scale;
float ext_factor;
float attn_factor;
float beta_fast;
float beta_slow;
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_glm = mode & 4;
GGML_ASSERT(!is_glm && "GLM RoPE not implemented in Metal");
if (!is_neox) {
GGML_ASSERT(id_src2 == nil && "TODO: freq_factors not implemented for !is_neox");
}
id<MTLComputePipelineState> pipeline = nil;
switch (src0->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_F16].pipeline; break;
default: GGML_ASSERT(false);
};
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
if (id_src2 != nil) {
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:2];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:3];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:10];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:11];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:14];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:15];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:18];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:19];
[encoder setBytes:&n_past length:sizeof( int) atIndex:20];
[encoder setBytes:&n_dims length:sizeof( int) atIndex:21];
[encoder setBytes:&mode length:sizeof( int) atIndex:22];
[encoder setBytes:&n_orig_ctx length:sizeof( int) atIndex:23];
[encoder setBytes:&freq_base length:sizeof( float) atIndex:24];
[encoder setBytes:&freq_scale length:sizeof( float) atIndex:25];
[encoder setBytes:&ext_factor length:sizeof( float) atIndex:26];
[encoder setBytes:&attn_factor length:sizeof( float) atIndex:27];
[encoder setBytes:&beta_fast length:sizeof( float) atIndex:28];
[encoder setBytes:&beta_slow length:sizeof( float) atIndex:29];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_IM2COL:
{
GGML_ASSERT(src0->type == GGML_TYPE_F16);
GGML_ASSERT(src1->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
const int32_t N = src1->ne[is_2D ? 3 : 2];
const int32_t IC = src1->ne[is_2D ? 2 : 1];
const int32_t IH = is_2D ? src1->ne[1] : 1;
const int32_t IW = src1->ne[0];
const int32_t KH = is_2D ? src0->ne[1] : 1;
const int32_t KW = src0->ne[0];
const int32_t OH = is_2D ? dst->ne[2] : 1;
const int32_t OW = dst->ne[1];
const int32_t CHW = IC * KH * KW;
const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4;
const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4;
id<MTLComputePipelineState> pipeline = nil;
switch (dst->type) {
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break;
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break;
default: GGML_ASSERT(false);
};
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ofs0 length:sizeof( int32_t) atIndex:2];
[encoder setBytes:&ofs1 length:sizeof( int32_t) atIndex:3];
[encoder setBytes:&IW length:sizeof( int32_t) atIndex:4];
[encoder setBytes:&IH length:sizeof( int32_t) atIndex:5];
[encoder setBytes:&CHW length:sizeof( int32_t) atIndex:6];
[encoder setBytes:&s0 length:sizeof( int32_t) atIndex:7];
[encoder setBytes:&s1 length:sizeof( int32_t) atIndex:8];
[encoder setBytes:&p0 length:sizeof( int32_t) atIndex:9];
[encoder setBytes:&p1 length:sizeof( int32_t) atIndex:10];
[encoder setBytes:&d0 length:sizeof( int32_t) atIndex:11];
[encoder setBytes:&d1 length:sizeof( int32_t) atIndex:12];
[encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
} break;
case GGML_OP_UPSCALE:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
const float sf0 = (float)ne0/src0->ne[0];
const float sf1 = (float)ne1/src0->ne[1];
const float sf2 = (float)ne2/src0->ne[2];
const float sf3 = (float)ne3/src0->ne[3];
const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17];
[encoder setBytes:&sf0 length:sizeof(sf0) atIndex:18];
[encoder setBytes:&sf1 length:sizeof(sf1) atIndex:19];
[encoder setBytes:&sf2 length:sizeof(sf2) atIndex:20];
[encoder setBytes:&sf3 length:sizeof(sf3) atIndex:21];
const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_PAD:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
[encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10];
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11];
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12];
[encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17];
const int nth = MIN(1024, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ARANGE:
{
GGML_ASSERT(dst->type == GGML_TYPE_F32);
float start;
float step;
memcpy(&start, ((int32_t *) dst->op_params) + 0, sizeof(float));
memcpy(&step, ((int32_t *) dst->op_params) + 2, sizeof(float));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_dst offset:offs_dst atIndex:0];
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:1];
[encoder setBytes:&start length:sizeof(start) atIndex:2];
[encoder setBytes:&step length:sizeof(step) atIndex:3];
const int nth = MIN(1024, ne0);
[encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_TIMESTEP_EMBEDDING:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
const int dim = dst->op_params[0];
const int max_period = dst->op_params[1];
const int half = dim / 2;
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:2];
[encoder setBytes:&dim length:sizeof(dim) atIndex:3];
[encoder setBytes:&max_period length:sizeof(max_period) atIndex:4];
const int nth = MIN(1024, half);
[encoder dispatchThreadgroups:MTLSizeMake(ne00, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
case GGML_OP_ARGSORT:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
const int nrows = ggml_nrows(src0);
enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
// bitonic sort requires the number of elements to be power of 2
int64_t ne00_padded = 1;
while (ne00_padded < ne00) {
ne00_padded *= 2;
}
// Metal kernels require the buffer size to be multiple of 16 bytes
// https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
const int mem_size = GGML_PAD(ne00_padded*sizeof(int32_t), 16);
id<MTLComputePipelineState> pipeline = nil;
switch (order) {
case GGML_SORT_ORDER_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break;
case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
default: GGML_ASSERT(false);
};
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne00_padded length:sizeof( int64_t) atIndex:3];
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00_padded, 1, 1)];
} break;
case GGML_OP_LEAKY_RELU:
{
GGML_ASSERT(src0->type == GGML_TYPE_F32);
float slope;
memcpy(&slope, dst->op_params, sizeof(float));
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&slope length:sizeof(slope) atIndex:2];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_OP_FLASH_ATTN_EXT:
{
GGML_ASSERT(ne00 % 4 == 0);
GGML_ASSERT(ne11 % 32 == 0);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_are_same_shape (src1, src2));
struct ggml_tensor * src3 = gf->nodes[i]->src[3];
size_t offs_src3 = 0;
id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);
GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
"the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
const int64_t ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
//const int64_t ne31 = src3 ? src3->ne[1] : 0;
const int64_t ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);
const int64_t ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33);
const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30);
const uint64_t nb31 = src3 ? src3->nb[1] : 0;
const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32);
const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33);
const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
float scale;
float max_bias;
memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(scale));
memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
const uint32_t n_head = src0->ne[2];
const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
const float m0 = powf(2.0f, -(max_bias ) / n_head_log2);
const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
id<MTLComputePipelineState> pipeline = nil;
bool use_vec_kernel = false;
if (ne01 >= 4 || (ne00%128 != 0)) {
switch (ne00) {
case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
//case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
GGML_METAL_LOG_ERROR("add template specialization for this size\n");
GGML_ASSERT(false && "add template specialization for this size");
}
}
} else {
use_vec_kernel = true;
switch (ne00) {
case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break;
//case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
default:
{
GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
GGML_METAL_LOG_ERROR("add template specialization for this size\n");
GGML_ASSERT(false && "add template specialization for this size");
}
}
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
if (id_src3) {
[encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
} else {
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:3];
}
[encoder setBuffer:id_dst offset:offs_dst atIndex:4];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:6];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:7];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10];
[encoder setBytes:&ne11 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne12 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne13 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb11 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb12 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb13 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb21 length:sizeof(uint64_t) atIndex:17];
[encoder setBytes:&nb22 length:sizeof(uint64_t) atIndex:18];
[encoder setBytes:&nb23 length:sizeof(uint64_t) atIndex:19];
[encoder setBytes:&nb31 length:sizeof(uint64_t) atIndex:20];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:21];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:22];
[encoder setBytes:&scale length:sizeof( float) atIndex:23];
[encoder setBytes:&max_bias length:sizeof( float) atIndex:24];
[encoder setBytes:&m0 length:sizeof(m0) atIndex:25];
[encoder setBytes:&m1 length:sizeof(m1) atIndex:26];
[encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:27];
if (!use_vec_kernel) {
// half8x8 kernel
const int64_t nqptg = 8; // queries per threadgroup !! sync with kernel template arguments !!
const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
GGML_ASSERT(nqptg <= 32);
GGML_ASSERT(nqptg % 8 == 0);
GGML_ASSERT(ncpsg % 32 == 0);
int64_t nsgmax = 2;
while (true) {
const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2);
if (smem > ctx->device.maxThreadgroupMemoryLength) {
break;
}
nsgmax *= 2;
}
nsgmax /= 2;
// simdgroups per threadgroup (a.k.a. warps)
const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2);
//printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
[encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
} else {
// half1x4 kernel
const int64_t nqptg = 1; // queries per threadgroup !! sync with kernel template arguments !!
const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
GGML_ASSERT(nqptg <= 32);
GGML_ASSERT(nqptg % 1 == 0);
GGML_ASSERT(ncpsg % 32 == 0);
// simdgroups per threadgroup (a.k.a. warps)
const int64_t nsgt = MAX(2, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32));
int64_t nsg = 1;
while (nsg <= nsgt) {
nsg *= 2;
}
nsg /= 2;
const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2);
//printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
[encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
}
} break;
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
{
GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
int nth = MIN(1024, ne00/ggml_blck_size(src0->type));
id<MTLComputePipelineState> pipeline = nil;
switch (src0t) {
case GGML_TYPE_F32:
{
GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0);
switch (dstt) {
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline; break;
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; break;
case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break;
case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break;
case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break;
case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break;
case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break;
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
};
} break;
case GGML_TYPE_F16:
{
switch (dstt) {
case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline; break;
case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F32].pipeline; break;
default: GGML_ASSERT(false && "not implemented");
};
} break;
default: GGML_ASSERT(false && "not implemented");
}
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
[encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
[encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
[encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
[encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
[encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
[encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
[encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
[encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
[encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:10];
[encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:11];
[encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:12];
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:13];
[encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:14];
[encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:15];
[encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:16];
[encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:17];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
} break;
default:
{
GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
}
if (should_capture) {
[encoder popDebugGroup];
}
}
[encoder endEncoding];
[command_buffer commit];
});
// Wait for completion and check status of each command buffer
// needed to detect if the device ran out-of-memory for example (#1881)
for (int i = 0; i < n_cb; ++i) {
id<MTLCommandBuffer> command_buffer = command_buffers[i];
[command_buffer waitUntilCompleted];
MTLCommandBufferStatus status = [command_buffer status];
if (status != MTLCommandBufferStatusCompleted) {
GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
if (status == MTLCommandBufferStatusError) {
NSString * error_code = [command_buffer error].localizedDescription;
GGML_METAL_LOG_INFO("error: %s\n", [error_code UTF8String]);
}
return GGML_STATUS_FAILED;
}
}
if (should_capture) {
[[MTLCaptureManager sharedCaptureManager] stopCapture];
}
}
return GGML_STATUS_SUCCESS;
}
////////////////////////////////////////////////////////////////////////////////
// backend interface
// default buffer
static id<MTLDevice> g_backend_device = nil;
static int g_backend_device_ref_count = 0;
static id<MTLDevice> ggml_backend_metal_get_device(void) {
if (g_backend_device == nil) {
g_backend_device = MTLCreateSystemDefaultDevice();
}
g_backend_device_ref_count++;
return g_backend_device;
}
static void ggml_backend_metal_free_device(void) {
assert(g_backend_device_ref_count > 0);
g_backend_device_ref_count--;
if (g_backend_device_ref_count == 0) {
[g_backend_device release];
g_backend_device = nil;
}
}
GGML_CALL static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
return "Metal";
UNUSED(buffer);
}
GGML_CALL static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
for (int i = 0; i < ctx->n_buffers; i++) {
[ctx->buffers[i].metal release];
}
ggml_backend_metal_free_device();
if (ctx->owned) {
#if TARGET_OS_OSX
vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)ctx->all_data, ctx->all_size);
#else
free(ctx->all_data);
#endif
}
free(ctx);
}
GGML_CALL static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
return ctx->all_data;
}
GGML_CALL static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
memcpy((char *)tensor->data + offset, data, size);
UNUSED(buffer);
}
GGML_CALL static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
memcpy(data, (const char *)tensor->data + offset, size);
UNUSED(buffer);
}
GGML_CALL static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
if (ggml_backend_buffer_is_host(src->buffer)) {
memcpy(dst->data, src->data, ggml_nbytes(src));
return true;
}
return false;
UNUSED(buffer);
}
GGML_CALL static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
memset(ctx->all_data, value, ctx->all_size);
}
static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
/* .get_name = */ ggml_backend_metal_buffer_get_name,
/* .free_buffer = */ ggml_backend_metal_buffer_free_buffer,
/* .get_base = */ ggml_backend_metal_buffer_get_base,
/* .init_tensor = */ NULL,
/* .set_tensor = */ ggml_backend_metal_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_metal_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_metal_buffer_cpy_tensor,
/* .clear = */ ggml_backend_metal_buffer_clear,
/* .reset = */ NULL,
};
// default buffer type
GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
return "Metal";
UNUSED(buft);
}
static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) {
#ifndef GGML_METAL_NDEBUG
#if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
if (@available(macOS 10.12, iOS 16.0, *)) {
GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)",
__func__,
size_aligned / 1024.0 / 1024.0,
device.currentAllocatedSize / 1024.0 / 1024.0,
device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
} else {
GGML_METAL_LOG_INFO("\n");
}
} else {
GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n",
__func__,
size_aligned / 1024.0 / 1024.0,
device.currentAllocatedSize / 1024.0 / 1024.0);
}
#endif
#endif
UNUSED(device);
UNUSED(size_aligned);
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
const size_t size_page = sysconf(_SC_PAGESIZE);
size_t size_aligned = size;
if ((size_aligned % size_page) != 0) {
size_aligned += (size_page - (size_aligned % size_page));
}
id<MTLDevice> device = ggml_backend_metal_get_device();
ctx->all_data = ggml_metal_host_malloc(size_aligned);
ctx->all_size = size_aligned;
ctx->owned = true;
ctx->n_buffers = 1;
if (ctx->all_data != NULL) {
ctx->buffers[0].data = ctx->all_data;
ctx->buffers[0].size = size;
ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
length:size_aligned
options:MTLResourceStorageModeShared
deallocator:nil];
}
if (ctx->all_data == NULL || ctx->buffers[0].metal == nil) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
free(ctx);
ggml_backend_metal_free_device();
return NULL;
}
//ggml_backend_metal_log_allocated_size(device, size_aligned);
return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
}
GGML_CALL static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
return 32;
UNUSED(buft);
}
GGML_CALL static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
id<MTLDevice> device = ggml_backend_metal_get_device();
size_t max_size = device.maxBufferLength;
ggml_backend_metal_free_device();
return max_size;
UNUSED(buft);
}
GGML_CALL static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
UNUSED(buft);
}
GGML_CALL static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
return true;
UNUSED(buft);
}
GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
/* .iface = */ {
/* .get_name = */ ggml_backend_metal_buffer_type_get_name,
/* .alloc_buffer = */ ggml_backend_metal_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_metal_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_metal_buffer_type_get_max_size,
/* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
/* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
/* .is_host = */ ggml_backend_metal_buffer_type_is_host,
},
/* .context = */ NULL,
};
return &ggml_backend_buffer_type_metal;
}
// buffer from ptr
GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
ctx->all_data = data;
ctx->all_size = size;
ctx->owned = false;
ctx->n_buffers = 0;
const size_t size_page = sysconf(_SC_PAGESIZE);
// page-align the data ptr
{
const uintptr_t offs = (uintptr_t) data % size_page;
data = (void *) ((char *) data - offs);
size += offs;
}
size_t size_aligned = size;
if ((size_aligned % size_page) != 0) {
size_aligned += (size_page - (size_aligned % size_page));
}
id<MTLDevice> device = ggml_backend_metal_get_device();
// the buffer fits into the max buffer size allowed by the device
if (size_aligned <= device.maxBufferLength) {
ctx->buffers[ctx->n_buffers].data = data;
ctx->buffers[ctx->n_buffers].size = size;
ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
return false;
}
ggml_backend_metal_log_allocated_size(device, size_aligned);
++ctx->n_buffers;
} else {
// this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
// one of the views
const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
const size_t size_step = device.maxBufferLength - size_ovlp;
const size_t size_view = device.maxBufferLength;
for (size_t i = 0; i < size; i += size_step) {
const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
ctx->buffers[ctx->n_buffers].size = size_step_aligned;
ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
return false;
}
ggml_backend_metal_log_allocated_size(device, size_step_aligned);
if (i + size_step < size) {
GGML_METAL_LOG_INFO("\n");
}
++ctx->n_buffers;
}
}
return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
}
// backend
GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
return "Metal";
UNUSED(backend);
}
GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
ggml_metal_free(ctx);
free(backend);
}
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_metal_buffer_type();
UNUSED(backend);
}
GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
return ggml_metal_graph_compute(metal_ctx, cgraph);
}
GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
return ggml_metal_supports_op(metal_ctx, op);
}
static struct ggml_backend_i ggml_backend_metal_i = {
/* .get_name = */ ggml_backend_metal_name,
/* .free = */ ggml_backend_metal_free,
/* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_metal_graph_compute,
/* .supports_op = */ ggml_backend_metal_supports_op,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .event_synchronize = */ NULL,
};
void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
ggml_metal_log_callback = log_callback;
ggml_metal_log_user_data = user_data;
}
static ggml_guid_t ggml_backend_metal_guid(void) {
static ggml_guid guid = { 0x81, 0xa1, 0x8b, 0x1e, 0x71, 0xec, 0x79, 0xed, 0x2b, 0x85, 0xdc, 0x8a, 0x61, 0x98, 0x30, 0xe6 };
return &guid;
}
ggml_backend_t ggml_backend_metal_init(void) {
struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
if (ctx == NULL) {
return NULL;
}
ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
*metal_backend = (struct ggml_backend) {
/* .guid = */ ggml_backend_metal_guid(),
/* .interface = */ ggml_backend_metal_i,
/* .context = */ ctx,
};
return metal_backend;
}
bool ggml_backend_is_metal(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_metal_guid());
}
void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
}
bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
}
void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
ctx->should_capture_next_compute = true;
}
GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
return ggml_backend_metal_init();
GGML_UNUSED(params);
GGML_UNUSED(user_data);
}
llm/llama.cpp/ggml-opencl.cpp deleted 100644 → 0
View file @ 97b02a89
#include "ggml.h"
#include "ggml-opencl.h"
#include "ggml-backend-impl.h"
#include <array>
#include <atomic>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <limits>
#include <sstream>
#include <vector>
#define CL_TARGET_OPENCL_VERSION 120
#include <clblast.h>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#define CL_DMMV_LOCAL_SIZE 32
#ifndef K_QUANTS_PER_ITERATION
#define K_QUANTS_PER_ITERATION 1
#else
static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
#endif
#define MULTILINE_QUOTE(...) #__VA_ARGS__
static std::string program_source = MULTILINE_QUOTE(
typedef char int8_t;
typedef uchar uint8_t;
typedef short int16_t;
typedef ushort uint16_t;
typedef int int32_t;
typedef uint uint32_t;
struct __attribute__ ((packed)) block_q4_0
{
half d;
uint8_t qs[QK4_0 / 2];
};
struct __attribute__ ((packed)) block_q4_1
{
half d;
half m;
uint8_t qs[QK4_1 / 2];
};
struct __attribute__ ((packed)) block_q5_0
{
half d;
uint32_t qh;
uint8_t qs[QK5_0 / 2];
};
struct __attribute__ ((packed)) block_q5_1
{
half d;
half m;
uint32_t qh;
uint8_t qs[QK5_1 / 2];
};
struct __attribute__ ((packed)) block_q8_0
{
half d;
int8_t qs[QK8_0];
};
struct __attribute__((packed)) block_q2_K
{
uint8_t scales[16];
uint8_t qs[64];
half d;
half dmin;
};
struct __attribute__((packed)) block_q3_K
{
uint8_t hmask[32];
uint8_t qs[64];
uint8_t scales[12];
half d;
};
struct __attribute__((packed)) block_q4_K
{
half d;
half dmin;
uint8_t scales[12];
uint8_t qs[128];
};
struct __attribute__((packed)) block_q5_K
{
half d;
half dmin;
uint8_t scales[12];
uint8_t qh[32];
uint8_t qs[128];
};
struct __attribute__((packed)) block_q6_K
{
uint8_t ql[128];
uint8_t qh[64];
int8_t scales[16];
half d;
};
__kernel void convert_fp16_to_fp32(__global half* x, __global float* y) {
const uint i = get_global_id(0);
y[i] = vload_half(0, &x[i]);
}
void dequantize_q4_0(__global const struct block_q4_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
*v0 = (vi0 - 8)*d;
*v1 = (vi1 - 8)*d;
}
void dequantize_q4_1(__global const struct block_q4_1* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const float m = vload_half(0, &x[ib].m);
const uint8_t vui = x[ib].qs[iqs];
const int8_t vi0 = vui & 0xF;
const int8_t vi1 = vui >> 4;
*v0 = vi0*d + m;
*v1 = vi1*d + m;
}
void dequantize_q5_0(__global const struct block_q5_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
uint32_t qh = x[ib].qh;
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1) - 16;
*v0 = x0*d;
*v1 = x1*d;
}
void dequantize_q5_1(__global const struct block_q5_1* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const float m = vload_half(0, &x[ib].m);
uint32_t qh = x[ib].qh;
const uint8_t xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10;
const uint8_t xh_1 = ((qh >> (iqs + 12)) ) & 0x10;
const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
const int32_t x1 = ((x[ib].qs[iqs] >> 4) | xh_1);
*v0 = x0*d + m;
*v1 = x1*d + m;
}
void dequantize_q8_0(__global const struct block_q8_0* x, const int ib, const int iqs, float* v0, float* v1) {
const float d = vload_half(0, &x[ib].d);
const int8_t vi0 = x[ib].qs[iqs + 0];
const int8_t vi1 = x[ib].qs[iqs + 1];
*v0 = vi0*d;
*v1 = vi1*d;
}
void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float* v1){
*v0 = vload_half(0, &x[ib + 0]);
*v1 = vload_half(0, &x[ib + 1]);
}
);
static std::string k_quants_source = MULTILINE_QUOTE(
inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8_t *m)
{
if (j < 4)
{
*d = q[j] & 63;
*m = q[j + 4] & 63;
}
else
{
*d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
*m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
}
}
__kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __global float *yy)
{
const int i = get_group_id(0) + get_global_offset(0);
const int tid = get_local_id(0);
const int n = tid / 32;
const int l = tid - 32 * n;
const int is = 8 * n + l / 16;
const uint8_t q = x[i].qs[32 * n + l];
__global float *y = yy + get_group_id(0) * QK_K + 128 * n;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
y[l + 0] = dall * (x[i].scales[is + 0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is + 0] >> 4);
y[l + 32] = dall * (x[i].scales[is + 2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is + 2] >> 4);
y[l + 64] = dall * (x[i].scales[is + 4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is + 4] >> 4);
y[l + 96] = dall * (x[i].scales[is + 6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is + 6] >> 4);
}
__kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __global float *yy)
{
int r = get_local_id(0) / 4;
int i = get_group_id(0) + get_global_offset(0);
int tid = r / 2;
int is0 = r % 2;
int l0 = 16 * is0 + 4 * (get_local_id(0) % 4);
int n = tid / 4;
int j = tid - 4 * n;
uint8_t m = 1 << (4 * n + j);
int is = 8 * n + 2 * j + is0;
int shift = 2 * j;
int8_t us = is < 4 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 8] >> 0) & 3) << 4)
: is < 8 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 4] >> 2) & 3) << 4)
: is < 12 ? (x[i].scales[is - 8] >> 4) | (((x[i].scales[is + 0] >> 4) & 3) << 4)
: (x[i].scales[is - 8] >> 4) | (((x[i].scales[is - 4] >> 6) & 3) << 4);
float d_all = vload_half(0, &x[i].d);
float dl = d_all * (us - 32);
__global float *y = yy + get_group_id(0) * QK_K + 128 * n + 32 * j;
const __global uint8_t *q = x[i].qs + 32 * n;
const __global uint8_t *hm = x[i].hmask;
for (int l = l0; l < l0 + 4; ++l)
y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
}
__kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __global float *yy)
{
const int i = get_group_id(0) + get_global_offset(0);
const int tid = get_local_id(0);
const int il = tid / 8;
const int ir = tid % 8;
const int is = 2 * il;
const int n = 4;
__global float *y = yy + get_group_id(0) * QK_K + 64 * il + n * ir;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint8_t *q = x[i].qs + 32 * il + n * ir;
uint8_t sc, m;
get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
float d1 = dall * sc;
float m1 = dmin * m;
get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
float d2 = dall * sc;
float m2 = dmin * m;
for (int l = 0; l < n; ++l)
{
y[l + 0] = d1 * (q[l] & 0xF) - m1;
y[l + 32] = d2 * (q[l] >> 4) - m2;
}
}
__kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __global float *yy)
{
const int i = get_group_id(0) + get_global_offset(0);
const int tid = get_local_id(0);
const int il = tid / 16;
const int ir = tid % 16;
const int is = 2 * il;
__global float *y = yy + get_group_id(0) * QK_K + 64 * il + 2 * ir;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint8_t *ql = x[i].qs + 32 * il + 2 * ir;
__global const uint8_t *qh = x[i].qh + 2 * ir;
uint8_t sc, m;
get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
const float d1 = dall * sc;
const float m1 = dmin * m;
get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
const float d2 = dall * sc;
const float m2 = dmin * m;
uint8_t hm = 1 << (2 * il);
y[0] = d1 * ((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0)) - m1;
y[1] = d1 * ((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0)) - m1;
hm <<= 1;
y[32] = d2 * ((ql[0] >> 4) + (qh[0] & hm ? 16 : 0)) - m2;
y[33] = d2 * ((ql[1] >> 4) + (qh[1] & hm ? 16 : 0)) - m2;
}
__kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __global float *yy)
{
const int i = get_group_id(0) + get_global_offset(0);
const int tid = get_local_id(0);
const int ip = tid / 32;
const int il = tid - 32 * ip;
const int is = 8 * ip + il / 16;
__global float *y = yy + get_group_id(0) * QK_K + 128 * ip + il;
const float d = vload_half(0, &x[i].d);
__global const uint8_t *ql = x[i].ql + 64 * ip + il;
const uint8_t qh = x[i].qh[32 * ip + il];
__global const int8_t *sc = x[i].scales + is;
y[0] = d * sc[0] * ((int8_t)((ql[0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
y[64] = d * sc[4] * ((int8_t)((ql[0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
}
__kernel void dequantize_mul_mat_vec_q2_K(__global const struct block_q2_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
__global const struct block_q2_K * x = xx + ib0;
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...15
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0,1
const int step = 16/K_QUANTS_PER_ITERATION;
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0...15 or 0...7
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2
const int q_offset = 32*im + l0;
const int s_offset = 8*im;
const int y_offset = 128*im + l0;
tmp[16 * ix + tid] = 0;
uint32_t aux[4];
const uint8_t * d = (const uint8_t *)aux;
const uint8_t * m = (const uint8_t *)(aux + 2);
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
__global const float * y = yy + i * QK_K + y_offset;
__global const uint8_t * q = x[i].qs + q_offset;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint32_t * a = (__global const uint32_t *)(x[i].scales + s_offset);
aux[0] = a[0] & 0x0f0f0f0f;
aux[1] = a[1] & 0x0f0f0f0f;
aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
float sum1 = 0, sum2 = 0;
for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
+ y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
+ y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
+ y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
+ y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
+ y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
+ y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
+y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
+ y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
}
tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
__kernel void dequantize_mul_mat_vec_q3_K(__global const struct block_q3_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
const uint16_t kmask1 = 0x0303;
const uint16_t kmask2 = 0x0f0f;
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
__global const struct block_q3_K * x = xx + ib0;
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0,1
const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop
const int step = 16/K_QUANTS_PER_ITERATION;
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0....15 or 0...7
const uint8_t m = 1 << (4*im);
const int l0 = n*in; // 0...15 or 0...14 in steps of 2
const int q_offset = 32*im + l0;
const int y_offset = 128*im + l0;
uint16_t utmp[4];
const int8_t * s = (const int8_t *)utmp;
const uint16_t s_shift = 4*im;
tmp[16 * ix + tid] = 0;
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
__global const float * y = yy + i * QK_K + y_offset;
__global const uint8_t * q = x[i].qs + q_offset;
__global const uint8_t * h = x[i].hmask + l0;
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
const float d = vload_half(0, &x[i].d);
float sum = 0;
for (int l = 0; l < n; ++l) {
sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
+ y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
+ y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
+ y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
+ y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
+ y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
+ y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
}
tmp[16 * ix + tid] += d * sum;
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
__kernel void dequantize_mul_mat_vec_q4_K(__global const struct block_q4_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
//to rename it later, just to test now
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...15
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION;
const int step = 8/K_QUANTS_PER_ITERATION;
const int il = tid/step; // 0...3
const int ir = tid - step*il;// 0...3
const int n = 2*K_QUANTS_PER_ITERATION;
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int in = il%2;
const int l0 = n*(2*ir + in);
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
uint16_t aux[4];
const uint8_t * sc = (const uint8_t *)aux;
__global const struct block_q4_K * x = xx + ib0;
tmp[16 * ix + tid] = 0;
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
__global const uint8_t * q1 = x[i].qs + q_offset;
__global const uint8_t * q2 = q1 + 64;
__global const float * y1 = yy + i*QK_K + y_offset;
__global const float * y2 = y1 + 128;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
aux[0] = a[im+0] & kmask1;
aux[1] = a[im+2] & kmask1;
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
float4 s = (float4)(0.f);
float smin = 0;
for (int l = 0; l < n; ++l) {
s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
}
tmp[16 * ix + tid] += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
__kernel void dequantize_mul_mat_vec_q5_K(__global const struct block_q5_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
const uint16_t kmask1 = 0x3f3f;
const uint16_t kmask2 = 0x0f0f;
const uint16_t kmask3 = 0xc0c0;
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
const int tid = get_local_id(0)/2; // 0...15
const int ix = get_local_id(0)%2;
const int il = tid/4; // 0...3
const int ir = tid - 4*il;// 0...3
const int n = 2;
const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
const int in = il%2;
const int l0 = n*(2*ir + in);
const int q_offset = 32*im + l0;
const int y_offset = 64*im + l0;
const uint8_t hm1 = 1 << (2*im);
const uint8_t hm2 = hm1 << 4;
uint16_t aux[4];
const uint8_t * sc = (const uint8_t *)aux;
__global const struct block_q5_K * x = xx + ib0;
tmp[16 * ix + tid] = 0;
for (int i = ix; i < num_blocks_per_row; i += 2) {
__global const uint8_t * ql1 = x[i].qs + q_offset;
__global const uint8_t * ql2 = ql1 + 64;
__global const uint8_t * qh = x[i].qh + l0;
__global const float * y1 = yy + i*QK_K + y_offset;
__global const float * y2 = y1 + 128;
const float dall = vload_half(0, &x[i].d);
const float dmin = vload_half(0, &x[i].dmin);
__global const uint16_t * a = (__global const uint16_t *)x[i].scales;
aux[0] = a[im+0] & kmask1;
aux[1] = a[im+2] & kmask1;
aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
float4 sum = (float4)(0.f);
float smin = 0;
for (int l = 0; l < n; ++l) {
sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+ y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
sum.y += y1[l+32] * ((ql1[l+ 0] >> 4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+ y1[l+48] * ((ql1[l+16] >> 4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+ y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
sum.w += y2[l+32] * ((ql2[l+ 0] >> 4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+ y2[l+48] * ((ql2[l+16] >> 4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
+ (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
}
tmp[16 * ix + tid] += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
__kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx, __local float* tmp, __global const float * yy, __global float * dst, const int ncols) {
const int row = get_group_id(0);
const int num_blocks_per_row = ncols / QK_K;
const int ib0 = row*num_blocks_per_row + get_global_offset(0);
__global const struct block_q6_K * x = xx + ib0;
const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION; // 0...31 or 0...16
const int ix = get_local_id(0)%K_QUANTS_PER_ITERATION; // 0 or 0, 1
const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8
const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128...
const int in = tid - step*im; // 0...15 or 0...7
\n#if K_QUANTS_PER_ITERATION == 1\n
const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15
const int is = 0;
\n#else\n
const int l0 = 4 * in; // 0, 4, 8, ..., 28
const int is = in / 4;
\n#endif\n
const int ql_offset = 64*im + l0;
const int qh_offset = 32*im + l0;
const int s_offset = 8*im + is;
const int y_offset = 128*im + l0;
tmp[16 * ix + tid] = 0; // partial sum for thread in warp
for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
__global const float * y = yy + i * QK_K + y_offset;
__global const uint8_t * ql = x[i].ql + ql_offset;
__global const uint8_t * qh = x[i].qh + qh_offset;
__global const int8_t * s = x[i].scales + s_offset;
const float d = vload_half(0, &x[i].d);
\n#if K_QUANTS_PER_ITERATION == 1\n
float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+ y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+ y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+ y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+ y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+ y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+ y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
tmp[16 * ix + tid] += sum;
\n#else\n
float sum = 0;
for (int l = 0; l < 4; ++l) {
sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+ y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+ y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+ y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
}
tmp[16 * ix + tid] += sum;
\n#endif\n
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=16; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
);
std::string dequant_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0)*2;
if (i >= get_global_size(0)) {
return;
}
const uint qk = QUANT_K;
const uint qr = QUANT_R;
const int ib = i/qk + get_global_offset(0); // block index
const int iqs = (i%qk)/qr; // quant index
const int iybs = i - i%qk; // y block start index
const int y_offset = qr == 1 ? 1 : qk/2;
// dequantize
float v0, v1;
DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
y[iybs + iqs + 0] = v0;
y[iybs + iqs + y_offset] = v1;
}
);
std::string dequant_mul_mat_vec_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
const int local_size = get_local_size(0);
const int row = get_group_id(0);
const int tid = get_local_id(0);
const uint qk = QUANT_K;
const uint qr = QUANT_R;
const int col_step = local_size * 2;
const int y_offset = qr == 1 ? 1 : qk/2;
x += get_global_offset(0);
tmp[tid] = 0;
for (int col = tid*2; col < ncols; col += col_step) {
const int ib = (row*ncols + col)/qk; // block index
const int iqs = (col%qk)/qr; // quant index
const int iybs = col - col%qk; // y block start index
// dequantize
float v0, v1;
DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
// matrix multiplication
tmp[tid] += v0 * y[iybs + iqs + 0];
tmp[tid] += v1 * y[iybs + iqs + y_offset];
}
// sum up partial sums and write back result
barrier(CLK_LOCAL_MEM_FENCE);
for (int s=local_size/2; s>0; s>>=1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (tid == 0) {
dst[row] = tmp[0];
}
}
);
std::string mul_template = MULTILINE_QUOTE(
__kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) {
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
if (i >= get_global_size(0)) {
return;
}
dst[dst_offset + i] = x[x_offset + i] * y[y_offset + i%ky];
}
);
std::string add_template = MULTILINE_QUOTE(
__kernel void add_f32(__global float * x, const int x_offset, __global float * y, const int y_offset, __global float * dst, const int dst_offset, const int ky) {
const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
if (i >= get_global_size(0)) {
return;
}
dst[dst_offset + i] = x[x_offset + i] + y[y_offset + i%ky];
}
);
#define CL_CHECK(err) \
do { \
cl_int err_ = (err); \
if (err_ != CL_SUCCESS) { \
fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n", \
#err, err_, __FILE__, __LINE__); \
exit(1); \
} \
} while (0)
#define CLBLAST_CHECK(err) \
do { \
CLBlastStatusCode err_ = (err); \
if (err_ != CLBlastSuccess) { \
fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n", \
#err, err_, __FILE__, __LINE__); \
exit(1); \
} \
} while (0)
std::array<std::string, 5> dequant_str_keys = {
"KERNEL_NAME", "X_TYPE", "QUANT_K", "QUANT_R", "DEQUANT_FUNC"
};
std::array<std::string, 30> dequant_str_values = {
"dequantize_row_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
"dequantize_row_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
"dequantize_row_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
"dequantize_row_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
"dequantize_row_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
"convert_row_f16", "half", "1", "1", "convert_f16"
};
std::array<std::string, 30> dequant_mul_mat_vec_str_values = {
"dequantize_mul_mat_vec_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
"dequantize_mul_mat_vec_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
"dequantize_mul_mat_vec_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
"dequantize_mul_mat_vec_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
"dequantize_mul_mat_vec_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
"convert_mul_mat_vec_f16", "half", "1", "1", "convert_f16"
};
std::array<std::string, 2> mul_str_keys = {
"KERNEL_NAME", "TYPE"
};
std::array<std::string, 2> mul_str_values = {
"mul_f32", "float"
};
static std::string& replace(std::string& s, const std::string& from, const std::string& to) {
size_t pos = 0;
while ((pos = s.find(from, pos)) != std::string::npos) {
s.replace(pos, from.length(), to);
pos += to.length();
}
return s;
}
static std::string generate_kernels() {
std::stringstream src;
src << program_source << '\n';
src << k_quants_source << '\n';
for (size_t i = 0; i < dequant_str_values.size(); i += dequant_str_keys.size()) {
std::string dequant_kernel = dequant_template;
std::string dmmv_kernel = dequant_mul_mat_vec_template;
for (size_t j = 0; j < dequant_str_keys.size(); j++) {
replace(dequant_kernel, dequant_str_keys[j], dequant_str_values[i + j]);
replace(dmmv_kernel, dequant_str_keys[j], dequant_mul_mat_vec_str_values[i + j]);
}
src << dequant_kernel << '\n';
src << dmmv_kernel << '\n';
}
for (size_t i = 0; i < mul_str_values.size(); i += mul_str_keys.size()) {
std::string mul_kernel = mul_template;
for (size_t j = 0; j < mul_str_keys.size(); j++) {
replace(mul_kernel, mul_str_keys[j], mul_str_values[i + j]);
}
src << mul_kernel << '\n';
}
src << add_template << '\n';
return src.str();
}
static cl_platform_id platform;
static cl_device_id device;
static cl_context context;
static cl_command_queue queue;
static cl_program program;
static cl_kernel convert_row_f16_cl;
static cl_kernel dequantize_row_q4_0_cl, dequantize_row_q4_1_cl, dequantize_row_q5_0_cl, dequantize_row_q5_1_cl, dequantize_row_q8_0_cl;
static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, dequantize_mul_mat_vec_q5_0_cl, dequantize_mul_mat_vec_q5_1_cl, dequantize_mul_mat_vec_q8_0_cl, convert_mul_mat_vec_f16_cl;
static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
static cl_kernel mul_f32_cl;
static cl_kernel add_f32_cl;
static bool fp16_support;
static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
cl_program p;
char *program_log;
size_t program_size;
size_t log_size;
int err;
program_size = strlen(program_buffer);
p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
if(err < 0) {
fprintf(stderr, "OpenCL error creating program");
exit(1);
}
std::string compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math "
"-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1 "
"-DQK_K=256 -DK_QUANTS_PER_ITERATION=" + std::to_string(K_QUANTS_PER_ITERATION);
err = clBuildProgram(p, 0, NULL, compile_opts.c_str(), NULL, NULL);
if(err < 0) {
clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
program_log = (char*) malloc(log_size + 1);
program_log[log_size] = '\0';
clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
fprintf(stderr, "ggml_opencl: kernel compile error:\n\n%s\n", program_log);
free(program_log);
exit(1);
}
return p;
}
void ggml_cl_init(void) {
static bool initialized = false;
if (initialized) {
return;
}
initialized = true;
cl_int err;
struct cl_device;
struct cl_platform {
cl_platform_id id;
unsigned number;
char name[128];
char vendor[128];
struct cl_device * devices;
unsigned n_devices;
struct cl_device * default_device;
};
struct cl_device {
struct cl_platform * platform;
cl_device_id id;
unsigned number;
cl_device_type type;
char name[128];
};
enum { NPLAT = 16, NDEV = 16 };
struct cl_platform platforms[NPLAT];
unsigned n_platforms = 0;
struct cl_device devices[NDEV];
unsigned n_devices = 0;
struct cl_device * default_device = NULL;
platform = NULL;
device = NULL;
cl_platform_id platform_ids[NPLAT];
CL_CHECK(clGetPlatformIDs(NPLAT, platform_ids, &n_platforms));
for (unsigned i = 0; i < n_platforms; i++) {
struct cl_platform * p = &platforms[i];
p->number = i;
p->id = platform_ids[i];
CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_NAME, sizeof(p->name), &p->name, NULL));
CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_VENDOR, sizeof(p->vendor), &p->vendor, NULL));
cl_device_id device_ids[NDEV];
cl_int clGetDeviceIDsError = clGetDeviceIDs(p->id, CL_DEVICE_TYPE_ALL, NDEV, device_ids, &p->n_devices);
if (clGetDeviceIDsError == CL_DEVICE_NOT_FOUND) {
p->n_devices = 0;
} else {
CL_CHECK(clGetDeviceIDsError);
}
p->devices = p->n_devices > 0 ? &devices[n_devices] : NULL;
p->default_device = NULL;
for (unsigned j = 0; j < p->n_devices; j++) {
struct cl_device * d = &devices[n_devices];
d->number = n_devices++;
d->id = device_ids[j];
d->platform = p;
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL));
CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL));
if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) {
p->default_device = d;
}
}
if (default_device == NULL && p->default_device != NULL) {
default_device = p->default_device;
}
}
if (n_devices == 0) {
fprintf(stderr, "ggml_opencl: could find any OpenCL devices.\n");
exit(1);
}
char * user_platform_string = getenv("GGML_OPENCL_PLATFORM");
char * user_device_string = getenv("GGML_OPENCL_DEVICE");
int user_platform_number = -1;
int user_device_number = -1;
unsigned n;
if (user_platform_string != NULL && sscanf(user_platform_string, " %u", &n) == 1 && n < n_platforms) {
user_platform_number = (int)n;
}
if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1 && n < n_devices) {
user_device_number = (int)n;
}
if (user_platform_number != -1 && user_device_number != -1) {
cl_platform* platform = &platforms[user_platform_number];
if ((unsigned)user_device_number >= platform->n_devices) {
fprintf(stderr, "ggml_opencl: invalid device number %d\n", user_device_number);
exit(1);
}
default_device = &platform->devices[user_device_number];
} else {
struct cl_device * selected_devices = devices;
unsigned n_selected_devices = n_devices;
if (user_platform_number == -1 && user_platform_string != NULL && user_platform_string[0] != 0) {
for (unsigned i = 0; i < n_platforms; i++) {
struct cl_platform * p = &platforms[i];
if (strstr(p->name, user_platform_string) != NULL ||
strstr(p->vendor, user_platform_string) != NULL) {
user_platform_number = (int)i;
break;
}
}
if (user_platform_number == -1) {
fprintf(stderr, "ggml_opencl: no platform matching '%s' was found.\n", user_platform_string);
exit(1);
}
}
if (user_platform_number != -1) {
struct cl_platform * p = &platforms[user_platform_number];
selected_devices = p->devices;
n_selected_devices = p->n_devices;
default_device = p->default_device;
if (n_selected_devices == 0) {
fprintf(stderr, "ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
exit(1);
}
}
if (user_device_number == -1 && user_device_string != NULL && user_device_string[0] != 0) {
for (unsigned i = 0; i < n_selected_devices; i++) {
struct cl_device * d = &selected_devices[i];
if (strstr(d->name, user_device_string) != NULL) {
user_device_number = d->number;
break;
}
}
if (user_device_number == -1) {
fprintf(stderr, "ggml_opencl: no device matching '%s' was found.\n", user_device_string);
exit(1);
}
}
if (user_device_number != -1) {
selected_devices = &devices[user_device_number];
n_selected_devices = 1;
default_device = &selected_devices[0];
}
GGML_ASSERT(n_selected_devices > 0);
if (default_device == NULL) {
default_device = &selected_devices[0];
}
}
fprintf(stderr, "ggml_opencl: selecting platform: '%s'\n", default_device->platform->name);
fprintf(stderr, "ggml_opencl: selecting device: '%s'\n", default_device->name);
if (default_device->type != CL_DEVICE_TYPE_GPU) {
fprintf(stderr, "ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name);
}
platform = default_device->platform->id;
device = default_device->id;
size_t ext_str_size;
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
char *ext_buffer = (char *)alloca(ext_str_size + 1);
clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
// Disabled due to faulty outputs
// Check if ext_buffer contains cl_khr_fp16
fp16_support = false; // strstr(ext_buffer, "cl_khr_fp16") != NULL;
// fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
cl_context_properties properties[] = {
(intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
};
CL_CHECK((context = clCreateContext(properties, 1, &device, NULL, NULL, &err), err));
CL_CHECK((queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err),
(err != CL_INVALID_QUEUE_PROPERTIES && err != CL_INVALID_VALUE ? err :
(queue = clCreateCommandQueue(context, device, 0, &err), err)
)));
const std::string kernel_src = generate_kernels();
program = build_program_from_source(context, device, kernel_src.c_str());
// FP16 to FP32 kernel
CL_CHECK((convert_row_f16_cl = clCreateKernel(program, "convert_row_f16", &err), err));
// Dequantize kernels
CL_CHECK((dequantize_row_q4_0_cl = clCreateKernel(program, "dequantize_row_q4_0", &err), err));
CL_CHECK((dequantize_row_q4_1_cl = clCreateKernel(program, "dequantize_row_q4_1", &err), err));
CL_CHECK((dequantize_row_q5_0_cl = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
CL_CHECK((dequantize_row_q5_1_cl = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
CL_CHECK((dequantize_block_q2_k_cl = clCreateKernel(program, "dequantize_block_q2_K", &err), err));
CL_CHECK((dequantize_block_q3_k_cl = clCreateKernel(program, "dequantize_block_q3_K", &err), err));
CL_CHECK((dequantize_block_q4_k_cl = clCreateKernel(program, "dequantize_block_q4_K", &err), err));
CL_CHECK((dequantize_block_q5_k_cl = clCreateKernel(program, "dequantize_block_q5_K", &err), err));
CL_CHECK((dequantize_block_q6_k_cl = clCreateKernel(program, "dequantize_block_q6_K", &err), err));
// dequant mul mat kernel
CL_CHECK((dequantize_mul_mat_vec_q4_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_0", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q4_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_1", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q5_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_0", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q5_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_1", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q8_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q8_0", &err), err));
CL_CHECK((convert_mul_mat_vec_f16_cl = clCreateKernel(program, "convert_mul_mat_vec_f16", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q2_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q2_K", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q3_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q3_K", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q4_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_K", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q5_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_K", &err), err));
CL_CHECK((dequantize_mul_mat_vec_q6_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q6_K", &err), err));
// mul kernel
CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
CL_CHECK((add_f32_cl = clCreateKernel(program, "add_f32", &err), err));
}
static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return &dequantize_row_q4_0_cl;
case GGML_TYPE_Q4_1:
return &dequantize_row_q4_1_cl;
case GGML_TYPE_Q5_0:
return &dequantize_row_q5_0_cl;
case GGML_TYPE_Q5_1:
return &dequantize_row_q5_1_cl;
case GGML_TYPE_Q8_0:
return &dequantize_row_q8_0_cl;
case GGML_TYPE_Q2_K:
return &dequantize_block_q2_k_cl;
case GGML_TYPE_Q3_K:
return &dequantize_block_q3_k_cl;
case GGML_TYPE_Q4_K:
return &dequantize_block_q4_k_cl;
case GGML_TYPE_Q5_K:
return &dequantize_block_q5_k_cl;
case GGML_TYPE_Q6_K:
return &dequantize_block_q6_k_cl;
case GGML_TYPE_F16:
return &convert_row_f16_cl;
default:
return nullptr;
}
}
static size_t ggml_cl_global_denom(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
return 1;
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
return 4;
case GGML_TYPE_Q4_K:
return 8;
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
return 4;
case GGML_TYPE_F16:
default:
return 1;
}
}
static size_t ggml_cl_local_size(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q5_0:
case GGML_TYPE_Q5_1:
case GGML_TYPE_Q8_0:
return 0;
case GGML_TYPE_Q2_K:
case GGML_TYPE_Q3_K:
return 64;
case GGML_TYPE_Q4_K:
return 32;
case GGML_TYPE_Q5_K:
case GGML_TYPE_Q6_K:
return 64;
case GGML_TYPE_F16:
default:
return 0;
}
}
static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
switch (type) {
case GGML_TYPE_Q4_0:
return &dequantize_mul_mat_vec_q4_0_cl;
case GGML_TYPE_Q4_1:
return &dequantize_mul_mat_vec_q4_1_cl;
case GGML_TYPE_Q5_0:
return &dequantize_mul_mat_vec_q5_0_cl;
case GGML_TYPE_Q5_1:
return &dequantize_mul_mat_vec_q5_1_cl;
case GGML_TYPE_Q8_0:
return &dequantize_mul_mat_vec_q8_0_cl;
case GGML_TYPE_F16:
return &convert_mul_mat_vec_f16_cl;
case GGML_TYPE_Q2_K:
return &dequantize_mul_mat_vec_q2_K_cl;
case GGML_TYPE_Q3_K:
return &dequantize_mul_mat_vec_q3_K_cl;
case GGML_TYPE_Q4_K:
return &dequantize_mul_mat_vec_q4_K_cl;
case GGML_TYPE_Q5_K:
return &dequantize_mul_mat_vec_q5_K_cl;
case GGML_TYPE_Q6_K:
return &dequantize_mul_mat_vec_q6_K_cl;
default:
return nullptr;
}
}
// buffer pool for cl
#define MAX_CL_BUFFERS 256
struct scoped_spin_lock {
std::atomic_flag& lock;
scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
while (lock.test_and_set(std::memory_order_acquire)) {
; // spin
}
}
~scoped_spin_lock() {
lock.clear(std::memory_order_release);
}
scoped_spin_lock(const scoped_spin_lock&) = delete;
scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
};
struct cl_buffer {
cl_mem mem;
size_t size = 0;
};
static cl_buffer g_cl_buffer_pool[MAX_CL_BUFFERS];
static std::atomic_flag g_cl_pool_lock = ATOMIC_FLAG_INIT;
static cl_mem ggml_cl_pool_malloc(size_t size, size_t * actual_size) {
scoped_spin_lock lock(g_cl_pool_lock);
cl_int err;
int best_i = -1;
size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
int worst_i = -1;
size_t worst_size = 0; //largest unused buffer seen so far
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
cl_buffer &b = g_cl_buffer_pool[i];
if (b.size > 0 && b.size >= size && b.size < best_size)
{
best_i = i;
best_size = b.size;
}
if (b.size > 0 && b.size > worst_size)
{
worst_i = i;
worst_size = b.size;
}
}
if(best_i!=-1) //found the smallest buffer that fits our needs
{
cl_buffer& b = g_cl_buffer_pool[best_i];
cl_mem mem = b.mem;
*actual_size = b.size;
b.size = 0;
return mem;
}
if(worst_i!=-1) //no buffer that fits our needs, resize largest one to save memory
{
cl_buffer& b = g_cl_buffer_pool[worst_i];
cl_mem mem = b.mem;
b.size = 0;
clReleaseMemObject(mem);
}
cl_mem mem;
CL_CHECK((mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err), err));
*actual_size = size;
return mem;
}
static void ggml_cl_pool_free(cl_mem mem, size_t size) {
scoped_spin_lock lock(g_cl_pool_lock);
for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
cl_buffer& b = g_cl_buffer_pool[i];
if (b.size == 0) {
b.mem = mem;
b.size = size;
return;
}
}
fprintf(stderr, "WARNING: cl buffer pool full, increase MAX_CL_BUFFERS\n");
clReleaseMemObject(mem);
}
void ggml_cl_free_data(const struct ggml_tensor* tensor) {
if (tensor->backend != GGML_BACKEND_TYPE_GPU) {
return;
}
cl_mem mem = (cl_mem)tensor->extra;
clReleaseMemObject(mem);
}
static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t offset, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cl_event* ev) {
cl_int err;
const uint64_t ne0 = src->ne[0];
const uint64_t ne1 = src->ne[1];
const uint64_t nb0 = src->nb[0];
const uint64_t nb1 = src->nb[1];
const uint64_t nb2 = src->nb[2];
const uint64_t nb3 = src->nb[3];
const enum ggml_type type = src->type;
const size_t ts = ggml_type_size(type);
const size_t bs = ggml_blck_size(type);
const uint64_t row_size = ts*ne0/bs;
const char * x = (const char *) src->data + i2*nb2 + i3*nb3;
if (nb0 == ts && nb1 == row_size) {
return clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*row_size, x, 0, NULL, ev);
}
if (nb0 == ts) {
const size_t buffer_origin[3] = { offset, 0, 0 };
const size_t host_origin[3] = { 0, 0, 0 };
const size_t region[3] = { row_size, ne1, 1 };
return clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, row_size, 0, nb1, 0, x, 0, NULL, ev);
}
std::vector<cl_event> events;
if (ev && ne1>1) events.reserve(ne1-1);
for (uint64_t i1 = 0; i1 < ne1; i1++) {
// pretend the row is a matrix with cols=1
const size_t buffer_origin[3] = { offset + i1*row_size, 0, 0 };
const size_t host_origin[3] = { 0, 0, 0 };
const size_t region[3] = { ts, ne0/bs, 1 };
// if an event is requested, make the last write wait for all previous writes to complete
if (ev && i1) {
events.push_back(*ev);
}
cl_uint nevents = i1 == ne1-1 ? events.size() : 0U;
err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts, 0, nb0, 0, x + i1*nb1, nevents, nevents ? events.data() : nullptr, ev);
if (err != CL_SUCCESS) {
for (auto event : events) {
clReleaseEvent(event);
}
return err;
}
}
for (auto event : events) {
CL_CHECK(clReleaseEvent(event));
}
return CL_SUCCESS;
}
static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
size_t x_size;
size_t d_size;
cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
cl_event ev;
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int i1 = i13*ne12*ne11 + i12*ne11;
cl_int x_offset = 0;
cl_int y_offset = i1*ne10;
cl_int d_offset = 0;
size_t global = ne00 * ne01;
cl_int ky = ne10 * ne11;
CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
CL_CHECK(clReleaseEvent(ev));
CL_CHECK(clFinish(queue));
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
}
}
ggml_cl_pool_free(d_X, x_size);
ggml_cl_pool_free(d_D, d_size);
}
void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
ggml_cl_mul_f32(src0, src1, dst);
}
static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
size_t x_size;
size_t d_size;
cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
cl_event ev;
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
const int64_t i13 = i03%ne13;
const int64_t i12 = i02%ne12;
const int i1 = i13*ne12*ne11 + i12*ne11;
cl_int x_offset = 0;
cl_int y_offset = i1*ne10;
cl_int d_offset = 0;
size_t global = ne00 * ne01;
cl_int ky = ne10 * ne11;
CL_CHECK(clSetKernelArg(add_f32_cl, 0, sizeof(cl_mem), &d_X));
CL_CHECK(clSetKernelArg(add_f32_cl, 1, sizeof(cl_int), &x_offset));
CL_CHECK(clSetKernelArg(add_f32_cl, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(add_f32_cl, 3, sizeof(cl_int), &y_offset));
CL_CHECK(clSetKernelArg(add_f32_cl, 4, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(add_f32_cl, 5, sizeof(cl_int), &d_offset));
CL_CHECK(clSetKernelArg(add_f32_cl, 6, sizeof(cl_int), &ky));
CL_CHECK(clEnqueueNDRangeKernel(queue, add_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
CL_CHECK(clReleaseEvent(ev));
CL_CHECK(clFinish(queue));
// copy dst to host
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
}
}
ggml_cl_pool_free(d_X, x_size);
ggml_cl_pool_free(d_D, d_size);
}
void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
ggml_cl_add_f32(src0, src1, dst);
}
static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const int64_t r2 = ne12 / ne02;
const int64_t r3 = ne13 / ne03;
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
size_t x_size;
size_t y_size;
size_t d_size;
cl_mem d_X;
if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
d_X = (cl_mem) src0->extra;
} else {
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
}
cl_mem d_Y = src1->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = dst->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
size_t x_offset = 0;
for (int64_t i03 = 0; i03 < ne03; i03++) {
// TODO: copy src0 here when r3>1
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
if (src0->backend == GGML_BACKEND_TYPE_GPU) {
x_offset = (i03 * ne02 + i02) * x_ne;
} else {
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
}
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
// copy src1 to device
if (src1->backend == GGML_BACKEND_TYPE_CPU) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
}
CL_CHECK(clFinish(queue));
// compute
cl_event ev_sgemm;
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, x_offset, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
// copy dst to host
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
}
}
}
}
}
if (src0->backend != GGML_BACKEND_TYPE_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
if (src1->backend != GGML_BACKEND_TYPE_GPU) {
ggml_cl_pool_free(d_Y, y_size);
}
if (dst->backend != GGML_BACKEND_TYPE_GPU) {
ggml_cl_pool_free(d_D, d_size);
}
}
static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
GGML_ASSERT(fp16_support);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb10 = src1->nb[0];
const int nb11 = src1->nb[1];
const int nb12 = src1->nb[2];
const int nb13 = src1->nb[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const int64_t r2 = ne12 / ne02;
const int64_t r3 = ne13 / ne03;
const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * y_ne);
GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * d_ne);
ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata;
size_t x_size;
size_t y_size;
size_t d_size;
cl_mem d_X;
if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
d_X = (cl_mem) src0->extra;
} else {
d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * d_ne, &d_size);
bool src1_cont_rows = nb10 == sizeof(float);
bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
size_t x_offset = 0;
for (int64_t i03 = 0; i03 < ne03; i03++) {
// TODO: copy src0 here when r3>1
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
if (src0->backend == GGML_BACKEND_TYPE_GPU) {
x_offset = (i03 * ne02 + i02) * x_ne;
} else {
// copy src0 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
}
// FIXME: convert on device
for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
// convert src1 to fp16
// TODO: use multiple threads
char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
if (src1_cont_rows) {
if (src1_cont_cols) {
ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
}
else {
for (int64_t i11 = 0; i11 < ne11; i11++) {
ggml_fp32_to_fp16_row((float *) (src1i + i11*nb11), tmp + i11*ne10, ne10);
}
}
}
else {
for (int64_t i11 = 0; i11 < ne11; i11++) {
for (int64_t i10 = 0; i10 < ne10; i10++) {
// very slow due to no inlining
tmp[i11*ne10 + i10] = ggml_fp32_to_fp16(*(float *) (src1i + i11*nb11 + i10*nb10));
}
}
}
// copy src1 to device
CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
// compute
cl_event ev_sgemm;
clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, x_offset, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, &ev_sgemm);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
// copy dst to host, then convert to float
if (dst->backend == GGML_BACKEND_TYPE_CPU) {
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
ggml_fp16_to_fp32_row(tmp, d, d_ne);
} else {
// FIXME: convert dst to fp32 on device
}
}
}
}
}
if (src0->backend != GGML_BACKEND_TYPE_GPU) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
}
static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int64_t ne02 = src0->ne[2];
const int64_t ne03 = src0->ne[3];
const int64_t ne10 = src1->ne[0];
const int64_t ne11 = src1->ne[1];
const int64_t ne12 = src1->ne[2];
const int64_t ne13 = src1->ne[3];
const int nb2 = dst->nb[2];
const int nb3 = dst->nb[3];
const ggml_type type = src0->type;
const bool mul_mat_vec = ne11 == 1 && ne00%2 == 0;
const int64_t r2 = ne12 / ne02;
const int64_t r3 = ne13 / ne03;
const float alpha = 1.0f;
const float beta = 0.0f;
const int x_ne = ne01 * ne00;
const int y_ne = ne11 * ne10;
const int d_ne = ne11 * ne01;
const int x_bps = x_ne / ggml_blck_size(type); // blocks per 2D slice
const size_t q_sz = ggml_type_size(type) * x_bps;
size_t x_size;
size_t y_size;
size_t d_size;
size_t q_size;
cl_mem d_X;
if (!mul_mat_vec) {
d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
}
cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
cl_mem d_Q;
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
}
cl_kernel* to_fp32_cl = ggml_get_to_fp32_cl(type);
cl_kernel* dmmv = ggml_get_dequantize_mul_mat_vec_cl(type);
GGML_ASSERT(to_fp32_cl != nullptr);
const size_t global_denom = ggml_cl_global_denom(type);
const size_t local = mul_mat_vec ? CL_DMMV_LOCAL_SIZE : ggml_cl_local_size(type);
size_t ev_idx = 0;
std::vector<cl_event> events;
for (int64_t i03 = 0; i03 < ne03; i03++) {
// TODO: copy and dequantize src0 here when r3>1
for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
// copy src0 to device if necessary
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
} else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
d_Q = (cl_mem) src0->extra;
} else {
GGML_ASSERT(false);
}
if (!mul_mat_vec) {
// convert src0 to fp32 on device
const size_t global = x_ne / global_denom;
const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
}
int64_t i12 = i02 * r2;
int64_t e12 = i12 + r2;
events.reserve(e12 - i12);
for (; i12 < e12; i12++) {
if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
// copy src1 to device
events.emplace_back();
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, events.data() + ev_idx++));
// compute
const size_t global = ne01 * local;
const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
const cl_int ncols = ne00;
events.emplace_back();
CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, &offset, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
} else { // CLBlast matrix matrix multiplication
// copy src1 to device
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
// wait for conversion
CL_CHECK(clFinish(queue));
// compute
events.emplace_back();
clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
clblast::Transpose::kYes, clblast::Transpose::kNo,
ne01, ne11, ne10,
alpha,
d_X, 0, ne00,
d_Y, 0, ne10,
beta,
d_D, 0, ne01,
&queue, events.data() + ev_idx++);
if (status != clblast::StatusCode::kSuccess) {
GGML_ASSERT(false);
}
}
// copy dst to host
float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
for (auto *event : events) {
clReleaseEvent(event);
}
ev_idx = 0;
events.clear();
}
}
}
}
if (!mul_mat_vec) {
ggml_cl_pool_free(d_X, x_size);
}
ggml_cl_pool_free(d_Y, y_size);
ggml_cl_pool_free(d_D, d_size);
if (src0->backend == GGML_BACKEND_TYPE_CPU) {
ggml_cl_pool_free(d_Q, q_size);
}
}
bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst) {
const int64_t ne10 = src1->ne[0];
const int64_t ne0 = dst->ne[0];
const int64_t ne1 = dst->ne[1];
// TODO: find the optimal values for these
if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
src1->type == GGML_TYPE_F32 &&
dst->type == GGML_TYPE_F32 &&
((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU)) {
return true;
}
return false;
}
static bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
// If device doesn't support FP16
if (!fp16_support) {
return false;
}
size_t src0_sz = ggml_nbytes(src0);
size_t src1_sz = ggml_nbytes(src1);
// mul_mat_q: src0 is converted to fp32 on device
size_t mul_mat_q_transfer = src0_sz + src1_sz;
// mul_mat_f16: src1 is converted to fp16 on cpu
size_t mul_mat_f16_transfer = src0_sz + sizeof(ggml_fp16_t) * ggml_nelements(src1);
// choose the smaller one to transfer to the device
// TODO: this is not always the best choice due to the overhead of converting to fp16
return mul_mat_f16_transfer < mul_mat_q_transfer;
}
void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) {
GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst));
if (src0->type == GGML_TYPE_F32) {
ggml_cl_mul_mat_f32(src0, src1, dst);
}
else if (src0->type == GGML_TYPE_F16) {
if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
ggml_cl_mul_mat_f16(src0, src1, dst, wdata, wsize);
}
else {
ggml_cl_mul_mat_q_f32(src0, src1, dst);
}
}
else if (ggml_is_quantized(src0->type)) {
ggml_cl_mul_mat_q_f32(src0, src1, dst);
}
else {
GGML_ASSERT(false);
}
}
size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
if (src0->type == GGML_TYPE_F16 && ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
return sizeof(ggml_fp16_t) * std::max(src1->ne[0] * src1->ne[1], dst->ne[0] * dst->ne[1]);
}
return 0;
}
void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
const int64_t ne0 = tensor->ne[0];
const int64_t ne1 = tensor->ne[1];
const int64_t ne2 = tensor->ne[2];
const int64_t ne3 = tensor->ne[3];
const ggml_type type = tensor->type;
const size_t s_sz = ggml_type_size(type) * (size_t) (ne0 * ne1 / ggml_blck_size(type));
const size_t q_sz = s_sz * (size_t) (ne2 * ne3);
size_t q_size;
cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
tensor->data = data;
// copy tensor to device
size_t offset = 0;
for (int64_t i3 = 0; i3 < ne3; i3++) {
for (int64_t i2 = 0; i2 < ne2; i2++) {
CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, offset, tensor, i3, i2, NULL));
offset += s_sz;
}
}
CL_CHECK(clFinish(queue));
tensor->extra = dst;
GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
}
// ggml-backend
// buffer
struct ggml_backend_opencl_buffer_context {
~ggml_backend_opencl_buffer_context() {
if (buffer) {
clReleaseMemObject(buffer);
}
for (auto * sub_buffer : sub_buffers) {
clReleaseMemObject(sub_buffer);
}
}
cl_mem buffer;
std::vector<cl_mem> sub_buffers;
};
static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000;
static const char * ggml_backend_opencl_buffer_get_name(ggml_backend_buffer_t buffer) {
return "OpenCL";
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
delete ctx;
}
static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
return cl_ptr_base;
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
if (tensor->view_src != NULL && tensor->view_offs == 0) {
tensor->extra = tensor->view_src->extra;
} else {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
cl_buffer_region region = {(size_t)((char *)tensor->data - (char *)cl_ptr_base), ggml_nbytes(tensor)};
cl_int err;
cl_mem sub_buffer = clCreateSubBuffer(ctx->buffer, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
CL_CHECK(err);
ctx->sub_buffers.push_back(sub_buffer);
tensor->extra = sub_buffer;
}
tensor->backend = GGML_BACKEND_TYPE_GPU;
}
static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
cl_mem tensor_buffer = (cl_mem) tensor->extra;
CL_CHECK(clEnqueueWriteBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
cl_mem tensor_buffer = (cl_mem) tensor->extra;
CL_CHECK(clEnqueueReadBuffer(queue, tensor_buffer, true, offset, size, data, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
CL_CHECK(clEnqueueFillBuffer(queue, ctx->buffer, &value, sizeof(value), 0, buffer->size, 0, NULL, NULL));
CL_CHECK(clFinish(queue));
}
static void ggml_backend_opencl_buffer_reset(ggml_backend_buffer_t buffer) {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
for (auto * sub_buffer : ctx->sub_buffers) {
clReleaseMemObject(sub_buffer);
}
ctx->sub_buffers.clear();
}
static ggml_backend_buffer_i ggml_backend_opencl_buffer_interface = {
/* .get_name = */ ggml_backend_opencl_buffer_get_name,
/* .free_buffer = */ ggml_backend_opencl_buffer_free_buffer,
/* .get_base = */ ggml_backend_opencl_buffer_get_base,
/* .init_tensor = */ ggml_backend_opencl_buffer_init_tensor,
/* .set_tensor = */ ggml_backend_opencl_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_opencl_buffer_get_tensor,
/* .cpy_tensor = */ NULL,
/* .clear = */ ggml_backend_opencl_buffer_clear,
/* .reset = */ ggml_backend_opencl_buffer_reset,
};
// buffer type
static const char * ggml_backend_opencl_buffer_type_name(ggml_backend_buffer_type_t buffer_type) {
return "OpenCL";
GGML_UNUSED(buffer_type);
}
static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buffer_type, size_t size) {
ggml_cl_init();
cl_int err;
cl_mem mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err);
if (err != CL_SUCCESS) {
fprintf(stderr, "%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);
return nullptr;
}
ggml_backend_opencl_buffer_context * ctx = new ggml_backend_opencl_buffer_context{mem, {}};
return ggml_backend_buffer_init(buffer_type, ggml_backend_opencl_buffer_interface, ctx, size);
}
static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_type_t buffer_type) {
// FIXME: not thread safe, device may not be initialized yet
static cl_uint alignment = -1;
if (alignment == (cl_uint)-1) {
ggml_cl_init();
clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &alignment, NULL);
alignment /= 8; // bits to bytes
}
return alignment;
GGML_UNUSED(buffer_type);
}
static size_t ggml_backend_opencl_buffer_type_get_max_size(ggml_backend_buffer_type_t buffer_type) {
static size_t max_size = -1;
if (max_size == (size_t)-1) {
ggml_cl_init();
clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &max_size, NULL);
}
return max_size;
}
static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buffer_type, ggml_backend_t backend) {
//return ggml_backend_is_opencl(backend); // opencl must be used through the cpu backend
return ggml_backend_is_cpu(backend);
GGML_UNUSED(buffer_type);
}
static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
/* .get_name = */ ggml_backend_opencl_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_opencl_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_opencl_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_opencl_buffer_type_get_max_size,
/* .get_alloc_size = */ NULL,
/* .supports_backend = */ ggml_backend_opencl_buffer_type_supports_backend,
/* .is_host = */ NULL,
};
ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type() {
static ggml_backend_buffer_type buffer_type = {
/* .iface = */ ggml_backend_opencl_buffer_type_interface,
/* .context = */ nullptr,
};
return &buffer_type;
}
#if 0
// host buffer type
static const char * ggml_backend_opencl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
return "CL_Host";
GGML_UNUSED(buft);
}
static const char * ggml_backend_opencl_host_buffer_name(ggml_backend_buffer_t buffer) {
return "CL_Host";
GGML_UNUSED(buffer);
}
static void ggml_backend_opencl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_cl_host_free(buffer->context);
}
static ggml_backend_buffer_t ggml_backend_opencl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * ptr = ggml_cl_host_malloc(size);
if (ptr == nullptr) {
// fallback to cpu buffer
return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
buffer->buft = buft;
buffer->iface.get_name = ggml_backend_opencl_host_buffer_name;
buffer->iface.free_buffer = ggml_backend_opencl_host_buffer_free_buffer;
return buffer;
}
ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type() {
static struct ggml_backend_buffer_type ggml_backend_opencl_buffer_type_host = {
/* .iface = */ {
/* .get_name = */ ggml_backend_opencl_host_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_opencl_host_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
/* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
},
/* .context = */ nullptr,
};
return &ggml_backend_opencl_buffer_type_host;
}
// backend
static const char * ggml_backend_opencl_name(ggml_backend_t backend) {
return "OpenCL";
GGML_UNUSED(backend);
}
static void ggml_backend_opencl_free(ggml_backend_t backend) {
GGML_UNUSED(backend);
}
static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(ggml_backend_t backend) {
return ggml_backend_opencl_buffer_type();
GGML_UNUSED(backend);
}
static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
for (int i = 0; i < graph->n_nodes; ++i) {
ggml_tensor * node = graph->nodes[i];
if (ggml_is_empty(node)) {
continue;
}
switch (node->op) {
case GGML_OP_MUL_MAT:
ggml_cl_mul_mat(node->src[0], node->src[1], node, nullptr, 0);
break;
case GGML_OP_MUL:
ggml_cl_mul(node->src[0], node->src[1], node);
break;
default:
GGML_ASSERT(false);
}
}
return GGML_STATUS_SUCCESS;
GGML_UNUSED(backend);
}
static bool ggml_backend_opencl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
switch (op->op) {
case GGML_OP_MUL_MAT:
return ggml_cl_can_mul_mat(op->src[0], op->src[1], op);
case GGML_OP_MUL:
// return ggml_can_repeat_rows(op->src[1], op->src[0]);
return true;
default:
return false;
}
GGML_UNUSED(backend);
}
static ggml_backend_i opencl_backend_i = {
/* .get_name = */ ggml_backend_opencl_name,
/* .free = */ ggml_backend_opencl_free,
/* .get_default_buffer_type = */ ggml_backend_opencl_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_from_async = */ NULL,
/* .cpy_tensor_to_async = */ NULL,
/* .synchronize = */ NULL,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_opencl_graph_compute,
/* .supports_op = */ ggml_backend_opencl_supports_op,
};
ggml_backend_t ggml_backend_opencl_init() {
ggml_backend_t backend = new ggml_backend {
/* .interface = */ opencl_backend_i,
/* .context = */ nullptr
};
return backend;
}
bool ggml_backend_is_opencl(ggml_backend_t backend) {
return backend && backend->iface.get_name == ggml_backend_opencl_name;
}
#endif
llm/llama.cpp/ggml-opencl.h deleted 100644 → 0
View file @ 97b02a89
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
#endif
GGML_API void ggml_cl_init(void);
GGML_API void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst);
GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
GGML_API void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
// GGML_API void * ggml_cl_host_malloc(size_t size);
// GGML_API void ggml_cl_host_free(void * ptr);
GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
// backend API
// GGML_API ggml_backend_t ggml_backend_opencl_init(void);
// GGML_API bool ggml_backend_is_opencl(ggml_backend_t backend);
GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
// GGML_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
#ifdef __cplusplus
}
#endif
llm/llama.cpp/ggml-quants.c deleted 100644 → 0
View file @ 97b02a89
This source diff could not be displayed because it is too large. You can view the blob instead.
llm/llama.cpp/ggml-quants.h deleted 100644 → 0
View file @ 97b02a89
#pragma once
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml.h"
// GGML internal header
#ifdef __cplusplus
extern "C" {
#endif
// Quantization
void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_s_reference (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_s_reference (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq3_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
void quantize_row_iq2_s (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
// Dequantization
void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq2_s (const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq1_m (const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq4_nl (const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq4_xs (const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
void dequantize_row_iq3_s (const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
// Dot product
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq2_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq1_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq1_m (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_iq3_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
void iq2xs_init_impl(enum ggml_type type);
void iq2xs_free_impl(enum ggml_type type);
void iq3xs_init_impl(int grid_size);
void iq3xs_free_impl(int grid_size);
#ifdef __cplusplus
}
#endif
llm/llama.cpp/ggml-rpc.cpp deleted 100644 → 0
View file @ 97b02a89
#include "ggml-rpc.h"
#include "ggml.h"
#include "ggml-backend-impl.h"
#include <cinttypes>
#include <string>
#include <vector>
#include <memory>
#include <mutex>
#include <unordered_map>
#include <unordered_set>
#ifdef _WIN32
# define WIN32_LEAN_AND_MEAN
# ifndef NOMINMAX
# define NOMINMAX
# endif
# include <windows.h>
# include <winsock2.h>
#else
# include <arpa/inet.h>
# include <sys/socket.h>
# include <sys/types.h>
# include <netinet/in.h>
# include <netinet/tcp.h>
# include <netdb.h>
# include <unistd.h>
#endif
#include <string.h>
#define UNUSED GGML_UNUSED
#define GGML_DEBUG 0
#if (GGML_DEBUG >= 1)
#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
#else
#define GGML_PRINT_DEBUG(...)
#endif
#ifdef _WIN32
typedef SOCKET sockfd_t;
using ssize_t = __int64;
#else
typedef int sockfd_t;
#endif
// cross-platform socket
struct socket_t {
sockfd_t fd;
socket_t(sockfd_t fd) : fd(fd) {}
~socket_t() {
GGML_PRINT_DEBUG("[%s] closing socket %d\n", __func__, this->fd);
#ifdef _WIN32
closesocket(this->fd);
#else
close(this->fd);
#endif
}
};
// ggml_tensor is serialized into rpc_tensor
#pragma pack(push, 1)
struct rpc_tensor {
uint64_t id;
uint32_t type;
uint64_t buffer;
uint32_t ne[GGML_MAX_DIMS];
uint32_t nb[GGML_MAX_DIMS];
uint32_t op;
int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
int32_t flags;
uint64_t src[GGML_MAX_SRC];
uint64_t view_src;
uint64_t view_offs;
uint64_t data;
char name[GGML_MAX_NAME];
};
#pragma pack(pop)
// RPC commands
enum rpc_cmd {
ALLOC_BUFFER = 0,
GET_ALIGNMENT,
GET_MAX_SIZE,
BUFFER_GET_BASE,
FREE_BUFFER,
BUFFER_CLEAR,
SET_TENSOR,
GET_TENSOR,
COPY_TENSOR,
GRAPH_COMPUTE,
GET_DEVICE_MEMORY,
};
// RPC data structures
static ggml_guid_t ggml_backend_rpc_guid() {
static ggml_guid guid = {0x99, 0x68, 0x5b, 0x6c, 0xd2, 0x83, 0x3d, 0x24, 0x25, 0x36, 0x72, 0xe1, 0x5b, 0x0e, 0x14, 0x03};
return &guid;
}
struct ggml_backend_rpc_buffer_type_context {
std::string endpoint;
std::string name;
size_t alignment;
size_t max_size;
};
struct ggml_backend_rpc_context {
std::string endpoint;
std::string name;
};
struct ggml_backend_rpc_buffer_context {
std::shared_ptr<socket_t> sock;
std::unordered_map<ggml_backend_buffer_t, void *> base_cache;
uint64_t remote_ptr;
std::string name;
};
// RPC helper functions
static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
#ifdef _WIN32
if (fd == INVALID_SOCKET) {
return nullptr;
}
#else
if (fd < 0) {
return nullptr;
}
#endif
return std::make_shared<socket_t>(fd);
}
static bool set_no_delay(sockfd_t sockfd) {
int flag = 1;
// set TCP_NODELAY to disable Nagle's algorithm
int ret = setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(int));
return ret == 0;
}
static bool set_reuse_addr(sockfd_t sockfd) {
int flag = 1;
int ret = setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, (char *)&flag, sizeof(int));
return ret == 0;
}
static std::shared_ptr<socket_t> socket_connect(const char * host, int port) {
struct sockaddr_in addr;
auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
auto sock_ptr = make_socket(sockfd);
if (sock_ptr == nullptr) {
return nullptr;
}
if (!set_no_delay(sockfd)) {
fprintf(stderr, "Failed to set TCP_NODELAY\n");
return nullptr;
}
addr.sin_family = AF_INET;
addr.sin_port = htons(port);
struct hostent * server = gethostbyname(host);
if (server == NULL) {
fprintf(stderr, "Cannot resolve host '%s'\n", host);
return nullptr;
}
memcpy(&addr.sin_addr.s_addr, server->h_addr, server->h_length);
if (connect(sock_ptr->fd, (struct sockaddr *)&addr, sizeof(addr)) < 0) {
return nullptr;
}
return sock_ptr;
}
static std::shared_ptr<socket_t> socket_accept(sockfd_t srv_sockfd) {
auto client_socket_fd = accept(srv_sockfd, NULL, NULL);
auto client_socket = make_socket(client_socket_fd);
if (client_socket == nullptr) {
return nullptr;
}
if (!set_no_delay(client_socket_fd)) {
fprintf(stderr, "Failed to set TCP_NODELAY\n");
return nullptr;
}
return client_socket;
}
static std::shared_ptr<socket_t> create_server_socket(const char * host, int port) {
auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
auto sock = make_socket(sockfd);
if (sock == nullptr) {
return nullptr;
}
if (!set_reuse_addr(sockfd)) {
fprintf(stderr, "Failed to set SO_REUSEADDR\n");
return nullptr;
}
struct sockaddr_in serv_addr;
serv_addr.sin_family = AF_INET;
serv_addr.sin_addr.s_addr = inet_addr(host);
serv_addr.sin_port = htons(port);
if (bind(sockfd, (struct sockaddr *) &serv_addr, sizeof(serv_addr)) < 0) {
return nullptr;
}
if (listen(sockfd, 1) < 0) {
return nullptr;
}
return sock;
}
static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
size_t bytes_sent = 0;
while (bytes_sent < size) {
ssize_t n = send(sockfd, (const char *)data + bytes_sent, size - bytes_sent, 0);
if (n < 0) {
return false;
}
bytes_sent += n;
}
return true;
}
static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
size_t bytes_recv = 0;
while (bytes_recv < size) {
ssize_t n = recv(sockfd, (char *)data + bytes_recv, size - bytes_recv, 0);
if (n <= 0) {
return false;
}
bytes_recv += n;
}
return true;
}
static bool parse_endpoint(const std::string & endpoint, std::string & host, int & port) {
size_t pos = endpoint.find(':');
if (pos == std::string::npos) {
return false;
}
host = endpoint.substr(0, pos);
port = std::stoi(endpoint.substr(pos + 1));
return true;
}
// RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
// RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
uint8_t cmd_byte = cmd;
if (!send_data(sock->fd, &cmd_byte, sizeof(cmd_byte))) {
return false;
}
uint64_t input_size = input.size();
if (!send_data(sock->fd, &input_size, sizeof(input_size))) {
return false;
}
if (!send_data(sock->fd, input.data(), input.size())) {
return false;
}
uint64_t output_size;
if (!recv_data(sock->fd, &output_size, sizeof(output_size))) {
return false;
}
if (output_size == 0) {
output.clear();
return true;
}
output.resize(output_size);
if (!recv_data(sock->fd, output.data(), output_size)) {
return false;
}
return true;
}
// RPC client-side implementation
static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
static std::unordered_map<std::string, std::weak_ptr<socket_t>> sockets;
static bool initialized = false;
auto it = sockets.find(endpoint);
if (it != sockets.end()) {
if (auto sock = it->second.lock()) {
return sock;
}
}
std::string host;
int port;
if (!parse_endpoint(endpoint, host, port)) {
return nullptr;
}
#ifdef _WIN32
if (!initialized) {
WSADATA wsaData;
int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
if (res != 0) {
return nullptr;
}
initialized = true;
}
#else
UNUSED(initialized);
#endif
auto sock = socket_connect(host.c_str(), port);
if (sock == nullptr) {
return nullptr;
}
GGML_PRINT_DEBUG("[%s] connected to %s, sockfd=%d\n", __func__, endpoint.c_str(), sock->fd);
sockets[endpoint] = sock;
return sock;
}
GGML_CALL static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
return ctx->name.c_str();
}
GGML_CALL static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// input serialization format: | remote_ptr (8 bytes) |
std::vector<uint8_t> input(sizeof(uint64_t), 0);
uint64_t remote_ptr = ctx->remote_ptr;
memcpy(input.data(), &remote_ptr, sizeof(remote_ptr));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, FREE_BUFFER, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.empty());
delete ctx;
}
GGML_CALL static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
if (ctx->base_cache.find(buffer) != ctx->base_cache.end()) {
return ctx->base_cache[buffer];
}
// input serialization format: | remote_ptr (8 bytes) |
std::vector<uint8_t> input(sizeof(uint64_t), 0);
uint64_t remote_ptr = ctx->remote_ptr;
memcpy(input.data(), &remote_ptr, sizeof(remote_ptr));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, BUFFER_GET_BASE, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == sizeof(uint64_t));
// output serialization format: | base_ptr (8 bytes) |
uint64_t base_ptr;
memcpy(&base_ptr, output.data(), sizeof(base_ptr));
void * base = reinterpret_cast<void *>(base_ptr);
ctx->base_cache[buffer] = base;
return base;
}
static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
rpc_tensor result;
result.id = reinterpret_cast<uint64_t>(tensor);
result.type = tensor->type;
if (tensor->buffer) {
ggml_backend_buffer_t buffer = tensor->buffer;
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
result.buffer = ctx->remote_ptr;
} else {
result.buffer = 0;
}
for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
result.ne[i] = tensor->ne[i];
result.nb[i] = tensor->nb[i];
}
result.op = tensor->op;
for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
result.op_params[i] = tensor->op_params[i];
}
result.flags = tensor->flags;
for (uint32_t i = 0; i < GGML_MAX_SRC; i++) {
result.src[i] = reinterpret_cast<uint64_t>(tensor->src[i]);
}
result.view_src = reinterpret_cast<uint64_t>(tensor->view_src);
result.view_offs = tensor->view_offs;
result.data = reinterpret_cast<uint64_t>(tensor->data);
snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
return result;
}
GGML_CALL static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
UNUSED(buffer);
if (ggml_is_quantized(tensor->type)) {
// TODO: this check is due to MATRIX_ROW_PADDING in CUDA and should be generalized
GGML_ASSERT(tensor->ne[0] % 512 == 0 && "unsupported quantized tensor");
}
}
GGML_CALL static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
std::vector<uint8_t> input(input_size, 0);
rpc_tensor rpc_tensor = serialize_tensor(tensor);
memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, SET_TENSOR, input, output);
GGML_ASSERT(status);
}
GGML_CALL static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// input serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) |
int input_size = sizeof(rpc_tensor) + 2*sizeof(uint64_t);
std::vector<uint8_t> input(input_size, 0);
rpc_tensor rpc_tensor = serialize_tensor(tensor);
memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &size, sizeof(size));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, GET_TENSOR, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == size);
// output serialization format: | data (size bytes) |
memcpy(data, output.data(), size);
}
GGML_CALL static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
// check if src and dst are on the same server
ggml_backend_buffer_t src_buffer = src->buffer;
ggml_backend_rpc_buffer_context * src_ctx = (ggml_backend_rpc_buffer_context *)src_buffer->context;
ggml_backend_buffer_t dst_buffer = dst->buffer;
ggml_backend_rpc_buffer_context * dst_ctx = (ggml_backend_rpc_buffer_context *)dst_buffer->context;
if (src_ctx->sock != dst_ctx->sock) {
return false;
}
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// input serialization format: | rpc_tensor src | rpc_tensor dst |
int input_size = 2*sizeof(rpc_tensor);
std::vector<uint8_t> input(input_size, 0);
rpc_tensor rpc_src = serialize_tensor(src);
rpc_tensor rpc_dst = serialize_tensor(dst);
memcpy(input.data(), &rpc_src, sizeof(rpc_src));
memcpy(input.data() + sizeof(rpc_src), &rpc_dst, sizeof(rpc_dst));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, COPY_TENSOR, input, output);
GGML_ASSERT(status);
// output serialization format: | result (1 byte) |
GGML_ASSERT(output.size() == 1);
return output[0];
}
GGML_CALL static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
// serialization format: | bufptr (8 bytes) | value (1 byte) |
int input_size = sizeof(uint64_t) + sizeof(uint8_t);
std::vector<uint8_t> input(input_size, 0);
memcpy(input.data(), &ctx->remote_ptr, sizeof(ctx->remote_ptr));
memcpy(input.data() + sizeof(ctx->remote_ptr), &value, sizeof(value));
std::vector<uint8_t> output;
bool status = send_rpc_cmd(ctx->sock, BUFFER_CLEAR, input, output);
GGML_ASSERT(status);
}
static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
/* .get_name = */ ggml_backend_rpc_buffer_get_name,
/* .free_buffer = */ ggml_backend_rpc_buffer_free_buffer,
/* .get_base = */ ggml_backend_rpc_buffer_get_base,
/* .init_tensor = */ ggml_backend_rpc_buffer_init_tensor,
/* .set_tensor = */ ggml_backend_rpc_buffer_set_tensor,
/* .get_tensor = */ ggml_backend_rpc_buffer_get_tensor,
/* .cpy_tensor = */ ggml_backend_rpc_buffer_cpy_tensor,
/* .clear = */ ggml_backend_rpc_buffer_clear,
/* .reset = */ NULL,
};
GGML_CALL static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t buft) {
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
return buft_ctx->name.c_str();
}
GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
// input serialization format: | size (8 bytes) |
int input_size = sizeof(uint64_t);
std::vector<uint8_t> input(input_size, 0);
memcpy(input.data(), &size, sizeof(size));
std::vector<uint8_t> output;
auto sock = get_socket(buft_ctx->endpoint);
bool status = send_rpc_cmd(sock, ALLOC_BUFFER, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == 2*sizeof(uint64_t));
// output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) |
uint64_t remote_ptr;
memcpy(&remote_ptr, output.data(), sizeof(remote_ptr));
size_t remote_size;
memcpy(&remote_size, output.data() + sizeof(uint64_t), sizeof(remote_size));
if (remote_ptr != 0) {
ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
ggml_backend_rpc_buffer_interface,
new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC"},
remote_size);
return buffer;
} else {
return nullptr;
}
}
static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
// input serialization format: | 0 bytes |
std::vector<uint8_t> input;
std::vector<uint8_t> output;
bool status = send_rpc_cmd(sock, GET_ALIGNMENT, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == sizeof(uint64_t));
// output serialization format: | alignment (8 bytes) |
uint64_t alignment;
memcpy(&alignment, output.data(), sizeof(alignment));
return alignment;
}
GGML_CALL static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
return buft_ctx->alignment;
}
static size_t get_max_size(const std::shared_ptr<socket_t> & sock) {
// input serialization format: | 0 bytes |
std::vector<uint8_t> input;
std::vector<uint8_t> output;
bool status = send_rpc_cmd(sock, GET_MAX_SIZE, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == sizeof(uint64_t));
// output serialization format: | max_size (8 bytes) |
uint64_t max_size;
memcpy(&max_size, output.data(), sizeof(max_size));
return max_size;
}
GGML_CALL static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) {
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
return buft_ctx->max_size;
}
GGML_CALL static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
UNUSED(buft);
return ggml_nbytes(tensor);
}
GGML_CALL static bool ggml_backend_rpc_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
if (!ggml_backend_is_rpc(backend)) {
return false;
}
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
return buft_ctx->endpoint == rpc_ctx->endpoint;
}
static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = {
/* .get_name = */ ggml_backend_rpc_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_rpc_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_rpc_buffer_type_get_alignment,
/* .get_max_size = */ ggml_backend_rpc_get_max_size,
/* .get_alloc_size = */ ggml_backend_rpc_buffer_type_get_alloc_size,
/* .supports_backend = */ ggml_backend_rpc_buffer_type_supports_backend,
/* .is_host = */ NULL,
};
GGML_CALL static const char * ggml_backend_rpc_name(ggml_backend_t backend) {
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
return rpc_ctx->name.c_str();
}
GGML_CALL static void ggml_backend_rpc_free(ggml_backend_t backend) {
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
delete rpc_ctx;
delete backend;
}
GGML_CALL static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) {
ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;
return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str());
}
GGML_CALL static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
UNUSED(backend);
// this is no-op because we don't have any async operations
}
static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
if (tensor == nullptr) {
return;
}
if (visited.find(tensor) != visited.end()) {
return;
}
visited.insert(tensor);
for (int i = 0; i < GGML_MAX_SRC; i++) {
add_tensor(tensor->src[i], tensors, visited);
}
add_tensor(tensor->view_src, tensors, visited);
tensors.push_back(serialize_tensor(tensor));
}
static void serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
uint32_t n_nodes = cgraph->n_nodes;
std::vector<rpc_tensor> tensors;
std::unordered_set<ggml_tensor*> visited;
for (uint32_t i = 0; i < n_nodes; i++) {
add_tensor(cgraph->nodes[i], tensors, visited);
}
// serialization format:
// | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
uint32_t n_tensors = tensors.size();
int output_size = sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(rpc_tensor);
output.resize(output_size, 0);
memcpy(output.data(), &n_nodes, sizeof(n_nodes));
uint64_t * out_nodes = (uint64_t *)(output.data() + sizeof(n_nodes));
for (uint32_t i = 0; i < n_nodes; i++) {
out_nodes[i] = reinterpret_cast<uint64_t>(cgraph->nodes[i]);
}
uint32_t * out_ntensors = (uint32_t *)(output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t));
*out_ntensors = n_tensors;
rpc_tensor * out_tensors = (rpc_tensor *)(output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t));
memcpy(out_tensors, tensors.data(), n_tensors * sizeof(rpc_tensor));
}
GGML_CALL static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
std::vector<uint8_t> input;
serialize_graph(cgraph, input);
std::vector<uint8_t> output;
auto sock = get_socket(rpc_ctx->endpoint);
bool status = send_rpc_cmd(sock, GRAPH_COMPUTE, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == 1);
return (enum ggml_status)output[0];
}
GGML_CALL static bool ggml_backend_rpc_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
UNUSED(backend);
UNUSED(op);
GGML_ASSERT(false && "not implemented");
return false;
}
static ggml_backend_i ggml_backend_rpc_interface = {
/* .get_name = */ ggml_backend_rpc_name,
/* .free = */ ggml_backend_rpc_free,
/* .get_default_buffer_type = */ ggml_backend_rpc_get_default_buffer_type,
/* .set_tensor_async = */ NULL,
/* .get_tensor_async = */ NULL,
/* .cpy_tensor_async = */ NULL,
/* .synchronize = */ ggml_backend_rpc_synchronize,
/* .graph_plan_create = */ NULL,
/* .graph_plan_free = */ NULL,
/* .graph_plan_compute = */ NULL,
/* .graph_compute = */ ggml_backend_rpc_graph_compute,
/* .supports_op = */ ggml_backend_rpc_supports_op,
/* .offload_op = */ NULL,
/* .event_new = */ NULL,
/* .event_free = */ NULL,
/* .event_record = */ NULL,
/* .event_wait = */ NULL,
/* .event_synchronize = */ NULL,
};
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
// NOTE: buffer types are allocated and never freed; this is by design
static std::unordered_map<std::string, ggml_backend_buffer_type_t> buft_map;
auto it = buft_map.find(endpoint);
if (it != buft_map.end()) {
return it->second;
}
auto sock = get_socket(endpoint);
if (sock == nullptr) {
return nullptr;
}
size_t alignment = get_alignment(sock);
size_t max_size = get_max_size(sock);
ggml_backend_rpc_buffer_type_context * buft_ctx = new ggml_backend_rpc_buffer_type_context {
/* .endpoint = */ endpoint,
/* .name = */ "RPC[" + std::string(endpoint) + "]",
/* .alignment = */ alignment,
/* .max_size = */ max_size
};
ggml_backend_buffer_type_t buft = new ggml_backend_buffer_type {
/* .iface = */ ggml_backend_rpc_buffer_type_interface,
/* .context = */ buft_ctx
};
buft_map[endpoint] = buft;
return buft;
}
GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
/* .endpoint = */ endpoint,
/* .name = */ "RPC",
};
ggml_backend_t backend = new ggml_backend {
/* .guid = */ ggml_backend_rpc_guid(),
/* .interface = */ ggml_backend_rpc_interface,
/* .context = */ ctx
};
return backend;
}
GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend) {
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid());
}
static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * free, size_t * total) {
// input serialization format: | 0 bytes |
std::vector<uint8_t> input;
std::vector<uint8_t> output;
bool status = send_rpc_cmd(sock, GET_DEVICE_MEMORY, input, output);
GGML_ASSERT(status);
GGML_ASSERT(output.size() == 2*sizeof(uint64_t));
// output serialization format: | free (8 bytes) | total (8 bytes) |
uint64_t free_mem;
memcpy(&free_mem, output.data(), sizeof(free_mem));
uint64_t total_mem;
memcpy(&total_mem, output.data() + sizeof(uint64_t), sizeof(total_mem));
*free = free_mem;
*total = total_mem;
}
GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
auto sock = get_socket(endpoint);
if (sock == nullptr) {
*free = 0;
*total = 0;
return;
}
get_device_memory(sock, free, total);
}
// RPC server-side implementation
class rpc_server {
public:
rpc_server(ggml_backend_t backend) : backend(backend) {}
~rpc_server();
bool alloc_buffer(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
void get_alignment(std::vector<uint8_t> & output);
void get_max_size(std::vector<uint8_t> & output);
bool buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool free_buffer(const std::vector<uint8_t> & input);
bool buffer_clear(const std::vector<uint8_t> & input);
bool set_tensor(const std::vector<uint8_t> & input);
bool get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
bool graph_compute(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
private:
ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor);
ggml_tensor * create_node(uint64_t id,
struct ggml_context * ctx,
const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map);
ggml_backend_t backend;
std::unordered_set<ggml_backend_buffer_t> buffers;
};
bool rpc_server::alloc_buffer(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// input serialization format: | size (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t size;
memcpy(&size, input.data(), sizeof(size));
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
uint64_t remote_ptr = 0;
uint64_t remote_size = 0;
if (buffer != nullptr) {
remote_ptr = reinterpret_cast<uint64_t>(buffer);
remote_size = buffer->size;
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> remote_ptr: %" PRIx64 ", remote_size: %" PRIu64 "\n", __func__, size, remote_ptr, remote_size);
buffers.insert(buffer);
} else {
GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> failed\n", __func__, size);
}
// output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) |
output.resize(2*sizeof(uint64_t), 0);
memcpy(output.data(), &remote_ptr, sizeof(remote_ptr));
memcpy(output.data() + sizeof(uint64_t), &remote_size, sizeof(remote_size));
return true;
}
void rpc_server::get_alignment(std::vector<uint8_t> & output) {
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
size_t alignment = ggml_backend_buft_get_alignment(buft);
GGML_PRINT_DEBUG("[%s] alignment: %lu\n", __func__, alignment);
// output serialization format: | alignment (8 bytes) |
output.resize(sizeof(uint64_t), 0);
memcpy(output.data(), &alignment, sizeof(alignment));
}
void rpc_server::get_max_size(std::vector<uint8_t> & output) {
ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
size_t max_size = ggml_backend_buft_get_max_size(buft);
GGML_PRINT_DEBUG("[%s] max_size: %lu\n", __func__, max_size);
// output serialization format: | max_size (8 bytes) |
output.resize(sizeof(uint64_t), 0);
memcpy(output.data(), &max_size, sizeof(max_size));
}
bool rpc_server::buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// input serialization format: | remote_ptr (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, remote_ptr);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
}
void * base = ggml_backend_buffer_get_base(buffer);
// output serialization format: | base_ptr (8 bytes) |
uint64_t base_ptr = reinterpret_cast<uint64_t>(base);
output.resize(sizeof(uint64_t), 0);
memcpy(output.data(), &base_ptr, sizeof(base_ptr));
return true;
}
bool rpc_server::free_buffer(const std::vector<uint8_t> & input) {
// input serialization format: | remote_ptr (8 bytes) |
if (input.size() != sizeof(uint64_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, remote_ptr);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
}
ggml_backend_buffer_free(buffer);
buffers.erase(buffer);
return true;
}
bool rpc_server::buffer_clear(const std::vector<uint8_t> & input) {
// input serialization format: | remote_ptr (8 bytes) | value (1 byte) |
if (input.size() != sizeof(uint64_t) + sizeof(uint8_t)) {
return false;
}
uint64_t remote_ptr;
memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
uint8_t value;
memcpy(&value, input.data() + sizeof(uint64_t), sizeof(value));
GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 ", value: %u\n", __func__, remote_ptr, value);
ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
if (buffers.find(buffer) == buffers.end()) {
GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
return false;
}
ggml_backend_buffer_clear(buffer, value);
return true;
}
ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor) {
ggml_tensor * result = ggml_new_tensor_4d(ctx, (ggml_type) tensor->type,
tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
result->nb[i] = tensor->nb[i];
}
result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
return nullptr;
}
result->op = (ggml_op) tensor->op;
for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
result->op_params[i] = tensor->op_params[i];
}
result->flags = tensor->flags;
result->data = reinterpret_cast<void *>(tensor->data);
ggml_set_name(result, tensor->name);
return result;
}
bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
// serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
if (input.size() < sizeof(rpc_tensor) + sizeof(uint64_t)) {
return false;
}
const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
uint64_t offset;
memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead(),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
if (tensor == nullptr) {
GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
ggml_free(ctx);
return false;
}
GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
ggml_backend_tensor_set(tensor, data, offset, size);
ggml_free(ctx);
return true;
}
bool rpc_server::get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) |
if (input.size() != sizeof(rpc_tensor) + 2*sizeof(uint64_t)) {
return false;
}
const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
uint64_t offset;
memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
uint64_t size;
memcpy(&size, input.data() + sizeof(rpc_tensor) + sizeof(offset), sizeof(size));
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead(),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
if (tensor == nullptr) {
GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
ggml_free(ctx);
return false;
}
GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
// output serialization format: | data (size bytes) |
output.resize(size, 0);
ggml_backend_tensor_get(tensor, output.data(), offset, size);
ggml_free(ctx);
return true;
}
bool rpc_server::copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// serialization format: | rpc_tensor src | rpc_tensor dst |
if (input.size() != 2*sizeof(rpc_tensor)) {
return false;
}
const rpc_tensor * rpc_src = (const rpc_tensor *)input.data();
const rpc_tensor * rpc_dst = (const rpc_tensor *)(input.data() + sizeof(rpc_src));
struct ggml_init_params params {
/*.mem_size =*/ 2*ggml_tensor_overhead(),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
ggml_tensor * src = deserialize_tensor(ctx, rpc_src);
ggml_tensor * dst = deserialize_tensor(ctx, rpc_dst);
if (src == nullptr || dst == nullptr) {
GGML_PRINT_DEBUG("[%s] error deserializing tensors\n", __func__);
ggml_free(ctx);
return false;
}
GGML_PRINT_DEBUG("[%s] src->buffer: %p, dst->buffer: %p\n", __func__, (void*)src->buffer, (void*)dst->buffer);
bool result = ggml_backend_buffer_copy_tensor(src, dst);
// output serialization format: | result (1 byte) |
output.resize(1, 0);
output[0] = result;
ggml_free(ctx);
return true;
}
ggml_tensor * rpc_server::create_node(uint64_t id,
struct ggml_context * ctx,
const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map) {
if (id == 0) {
return nullptr;
}
if (tensor_map.find(id) != tensor_map.end()) {
return tensor_map[id];
}
const rpc_tensor * tensor = tensor_ptrs.at(id);
struct ggml_tensor * result = deserialize_tensor(ctx, tensor);
if (result == nullptr) {
return nullptr;
}
tensor_map[id] = result;
for (int i = 0; i < GGML_MAX_SRC; i++) {
result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
}
result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
result->view_offs = tensor->view_offs;
return result;
}
bool rpc_server::graph_compute(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
// serialization format:
// | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
if (input.size() < sizeof(uint32_t)) {
return false;
}
uint32_t n_nodes;
memcpy(&n_nodes, input.data(), sizeof(n_nodes));
if (input.size() < sizeof(uint32_t) + n_nodes*sizeof(uint64_t) + sizeof(uint32_t)) {
return false;
}
const uint64_t * nodes = (const uint64_t *)(input.data() + sizeof(n_nodes));
uint32_t n_tensors;
memcpy(&n_tensors, input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t), sizeof(n_tensors));
if (input.size() < sizeof(uint32_t) + n_nodes*sizeof(uint64_t) + sizeof(uint32_t) + n_tensors*sizeof(rpc_tensor)) {
return false;
}
const rpc_tensor * tensors = (const rpc_tensor *)(input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t) + sizeof(n_tensors));
GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
static size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
struct ggml_init_params params = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
struct ggml_context * ctx = ggml_init(params);
struct ggml_cgraph * graph = ggml_new_graph_custom(ctx, n_nodes, false);
graph->n_nodes = n_nodes;
std::unordered_map<uint64_t, const rpc_tensor*> tensor_ptrs;
for (uint32_t i = 0; i < n_tensors; i++) {
tensor_ptrs[tensors[i].id] = &tensors[i];
}
std::unordered_map<uint64_t, ggml_tensor*> tensor_map;
for (uint32_t i = 0; i < n_nodes; i++) {
graph->nodes[i] = create_node(nodes[i], ctx, tensor_ptrs, tensor_map);
}
ggml_status status = ggml_backend_graph_compute(backend, graph);
// output serialization format: | status (1 byte) |
output.resize(1, 0);
output[0] = status;
ggml_free(ctx);
return true;
}
rpc_server::~rpc_server() {
for (auto buffer : buffers) {
ggml_backend_buffer_free(buffer);
}
}
static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) {
rpc_server server(backend);
while (true) {
uint8_t cmd;
if (!recv_data(sockfd, &cmd, 1)) {
break;
}
std::vector<uint8_t> input;
std::vector<uint8_t> output;
uint64_t input_size;
if (!recv_data(sockfd, &input_size, sizeof(input_size))) {
break;
}
input.resize(input_size);
if (!recv_data(sockfd, input.data(), input_size)) {
break;
}
bool ok = true;
switch (cmd) {
case ALLOC_BUFFER: {
ok = server.alloc_buffer(input, output);
break;
}
case GET_ALIGNMENT: {
server.get_alignment(output);
break;
}
case GET_MAX_SIZE: {
server.get_max_size(output);
break;
}
case BUFFER_GET_BASE: {
ok = server.buffer_get_base(input, output);
break;
}
case FREE_BUFFER: {
ok = server.free_buffer(input);
break;
}
case BUFFER_CLEAR: {
ok = server.buffer_clear(input);
break;
}
case SET_TENSOR: {
ok = server.set_tensor(input);
break;
}
case GET_TENSOR: {
ok = server.get_tensor(input, output);
break;
}
case COPY_TENSOR: {
ok = server.copy_tensor(input, output);
break;
}
case GRAPH_COMPUTE: {
ok = server.graph_compute(input, output);
break;
}
case GET_DEVICE_MEMORY: {
// output serialization format: | free (8 bytes) | total (8 bytes) |
output.resize(2*sizeof(uint64_t), 0);
memcpy(output.data(), &free_mem, sizeof(free_mem));
memcpy(output.data() + sizeof(uint64_t), &total_mem, sizeof(total_mem));
break;
}
default: {
fprintf(stderr, "Unknown command: %d\n", cmd);
ok = false;
}
}
if (!ok) {
break;
}
uint64_t output_size = output.size();
if (!send_data(sockfd, &output_size, sizeof(output_size))) {
break;
}
if (!send_data(sockfd, output.data(), output_size)) {
break;
}
}
}
void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem) {
std::string host;
int port;
if (!parse_endpoint(endpoint, host, port)) {
return;
}
#ifdef _WIN32
{
WSADATA wsaData;
int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
if (res != 0) {
fprintf(stderr, "WSAStartup failed: %d\n", res);
return;
}
}
#endif
auto server_socket = create_server_socket(host.c_str(), port);
if (server_socket == nullptr) {
fprintf(stderr, "Failed to create server socket\n");
return;
}
while (true) {
auto client_socket = socket_accept(server_socket->fd);
if (client_socket == nullptr) {
fprintf(stderr, "Failed to accept client connection\n");
return;
}
printf("Accepted client connection, free_mem=%zu, total_mem=%zu\n", free_mem, total_mem);
rpc_serve_client(backend, client_socket->fd, free_mem, total_mem);
printf("Client connection closed\n");
}
#ifdef _WIN32
WSACleanup();
#endif
}
llm/llama.cpp/ggml-rpc.h deleted 100644 → 0
View file @ 97b02a89
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_RPC_MAX_SERVERS 16
// backend API
GGML_API GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
GGML_API GGML_CALL bool ggml_backend_is_rpc(ggml_backend_t backend);
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
GGML_API GGML_CALL void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
GGML_API GGML_CALL void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
#ifdef __cplusplus
}
#endif
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment