Skip to content

GitLab

Commit 4c39a76a authored by Casper Hansen's avatar Casper Hansen
Browse files

NVCC: Use asm volatile

parent acdf7d9a
Hide whitespace changes
Inline Side-by-side
Showing with 36 additions and 45 deletions
awq_cuda/quantization/dequantize.cuh
View file @ 4c39a76a
......@@ -34,19 +34,19 @@ __device__ uint4 dequantize_s4_to_fp16x2(uint32_t const& source)
// immediately before required.
const uint32_t top_i4s = i4s >> 8;
// Extract elt_01 - (i4s & 0x000f000f) | 0x64006400
ASM("lop3.b32 %0, %1, %2, %3, %4;\n"
asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
: "=r"(h[0])
: "r"(i4s), "n"(BOTTOM_MASK), "n"(I4s_TO_F16s_MAGIC_NUM), "n"(immLut));
// Extract elt_23 (i4s & 0x00f000f0) | 0x64006400
ASM("lop3.b32 %0, %1, %2, %3, %4;\n"
asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
: "=r"(h[1])
: "r"(i4s), "n"(TOP_MASK), "n"(I4s_TO_F16s_MAGIC_NUM), "n"(immLut));
// Extract elt_45 (top_i4s & 0x000f000f) | 0x64006400
ASM("lop3.b32 %0, %1, %2, %3, %4;\n"
asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
: "=r"(h[2])
: "r"(top_i4s), "n"(BOTTOM_MASK), "n"(I4s_TO_F16s_MAGIC_NUM), "n"(immLut));
// Extract elt_67 (top_i4s & 0x00f000f0) | 0x64006400
ASM("lop3.b32 %0, %1, %2, %3, %4;\n"
asm volatile("lop3.b32 %0, %1, %2, %3, %4;\n"
: "=r"(h[3])
: "r"(top_i4s), "n"(TOP_MASK), "n"(I4s_TO_F16s_MAGIC_NUM), "n"(immLut));
......@@ -66,13 +66,13 @@ __device__ uint4 dequantize_s4_to_fp16x2(uint32_t const& source)
// Finally, we construct the output numbers.
// Convert elt_01
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(h[0]) : "r"(h[0]), "r"(FP16_TOP_MAGIC_NUM));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[0]) : "r"(h[0]), "r"(FP16_TOP_MAGIC_NUM));
// Convert elt_23
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(h[1]) : "r"(h[1]), "r"(ONE_SIXTEENTH), "r"(NEG_64));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(h[1]) : "r"(h[1]), "r"(ONE_SIXTEENTH), "r"(NEG_64));
// Convert elt_45
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(h[2]) : "r"(h[2]), "r"(FP16_TOP_MAGIC_NUM));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(h[2]) : "r"(h[2]), "r"(FP16_TOP_MAGIC_NUM));
// Convert elt_67
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(h[3]) : "r"(h[3]), "r"(ONE_SIXTEENTH), "r"(NEG_64));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(h[3]) : "r"(h[3]), "r"(ONE_SIXTEENTH), "r"(NEG_64));
return result;
}
......
awq_cuda/quantization/gemm_cuda_gen.cu
View file @ 4c39a76a
......@@ -9,12 +9,6 @@
*/
#ifdef _MSC_VER
#define ASM __asm
#else
#define ASM asm volatile
#endif
#include <torch/extension.h>
#include "gemm_cuda.h"
#include "dequantize.cuh"
......@@ -36,9 +30,6 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
float C_warp[32];
__shared__ half A_shared[16 * (32 + 8)];
__shared__ half B_shared[32 * (128 + 8)];
__shared__ half scaling_factors_shared[128];
__shared__ half zeros_shared[128];
int j_factors1 = ((OC + 128 - 1) / 128);
int blockIdx_x = 0;
......@@ -137,14 +128,14 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
// uint4 B_loaded_scale = *(uint4*)(scaling_factors_shared + (threadIdx.x % (cta_N / 8)) * 8);
// - zero and * scale
// TODO (Haotian): can save 4 assembly instructions if sormulate as deq = q * scale - zero * scale.
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
/*
if (ax0_ax1_fused_0 == 0 && blockIdx_z == 0 && blockIdx_y == 0 && k_0_0 == 0 && threadIdx.x == 17 && threadIdx.y == 0){
printf("[x] %X %X %X %X\n", B_loaded_fp16.x, B_loaded_fp16.y, B_loaded_fp16.z, B_loaded_fp16.w);
......@@ -159,14 +150,14 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
for (int k_0_1 = 0; k_0_1 < 2; ++k_0_1) {
{
unsigned int addr;
ASM(
asm volatile(
"{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
: "=r"(addr)
: "l"((void *)((&(A_shared[(k_0_1 * 16)])) + (((((int)threadIdx.x) & 15) * 40) + ((((int)threadIdx.x) >> 4) * 8))))
);
ASM(
asm volatile(
"ldmatrix.sync.aligned.m8n8.x4.shared.b16"
"{%0, %1, %2, %3}, [%4];\n"
: "=r"(((unsigned *)(A_shared_warp + 0))[0]), "=r"(((unsigned *)(A_shared_warp + 0))[1]), "=r"(((unsigned *)(A_shared_warp + 0))[2]), "=r"(((unsigned *)(A_shared_warp + 0))[3])
......@@ -177,12 +168,12 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
for (int ax1_0 = 0; ax1_0 < 4; ++ax1_0) {
{
unsigned int addr;
ASM(
asm volatile(
"{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
: "=r"(addr)
: "l"((void *)((&(B_shared[(((k_0_1 * 2176) + (((int)threadIdx.y) * 64)) + (ax1_0 * 16))])) + (((((int)threadIdx.x) & 15) * 136) + ((((int)threadIdx.x) >> 4) * 8))))
);
ASM(
asm volatile(
"ldmatrix.sync.aligned.m8n8.x4.trans.shared.b16"
"{%0, %1, %2, %3}, [%4];\n"
: "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[0]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[1]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[2]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[3])
......@@ -192,7 +183,7 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
}
for (int j_0_4 = 0; j_0_4 < 4; ++j_0_4) {
{
ASM(
asm volatile(
"mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
"{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
: "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
......@@ -200,7 +191,7 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
}
{
ASM(
asm volatile(
"mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
"{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
: "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
......@@ -331,14 +322,14 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, in
// uint4 B_loaded_scale = *(uint4*)(scaling_factors_shared + (threadIdx.x % (cta_N / 8)) * 8);
// - zero and * scale
// TODO (Haotian): can save 4 assembly instructions if sormulate as deq = q * scale - zero * scale.
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
ASM("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
ASM("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
/*
if (ax0_ax1_fused_0 == 0 && blockIdx_z == 0 && blockIdx_y == 0 && k_0_0 == 0 && threadIdx.x == 17 && threadIdx.y == 0){
printf("[x] %X %X %X %X\n", B_loaded_fp16.x, B_loaded_fp16.y, B_loaded_fp16.z, B_loaded_fp16.w);
......@@ -354,12 +345,12 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, in
{
{
unsigned int addr;
ASM(
asm volatile(
"{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
: "=r"(addr)
: "l"((void *)((&(A_shared[(k_0_1 * 16)])) + (((((int)threadIdx.x) & 15) * 40) + ((((int)threadIdx.x) >> 4) * 8))))
);
ASM(
asm volatile(
"ldmatrix.sync.aligned.m8n8.x4.shared.b16"
"{%0, %1, %2, %3}, [%4];\n"
: "=r"(((unsigned *)(A_shared_warp + 0))[0]), "=r"(((unsigned *)(A_shared_warp + 0))[1]), "=r"(((unsigned *)(A_shared_warp + 0))[2]), "=r"(((unsigned *)(A_shared_warp + 0))[3])
......@@ -372,12 +363,12 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, in
{
{
unsigned int addr;
ASM(
asm volatile(
"{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
: "=r"(addr)
: "l"((void *)((&(B_shared[(((k_0_1 * 1152) + (((int)threadIdx.y) * 32)) + (ax1_0 * 16))])) + (((((int)threadIdx.x) & 15) * 72) + ((((int)threadIdx.x) >> 4) * 8))))
);
ASM(
asm volatile(
"ldmatrix.sync.aligned.m8n8.x4.trans.shared.b16"
"{%0, %1, %2, %3}, [%4];\n"
: "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[0]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[1]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[2]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[3])
......@@ -390,7 +381,7 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, in
{
{
ASM(
asm volatile(
"mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
"{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
: "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
......@@ -398,7 +389,7 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, in
}
{
ASM(
asm volatile(
"mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
"{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
: "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
......
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