Skip to content

GitLab

Unverified Commit 0cf7794b authored by Daniel Hiltgen's avatar Daniel Hiltgen Committed by GitHub
Browse files

ggml update to b7108 (#12992) 

* Revert "vulkan: temporary cary of vulkan fixes (#12971)"

This reverts commit 3a9e8e9f.

* ggml update to b7087

* fix argsort on metal

* update to b7108

* fix bakllava regression

This model lacks the metadata for the projector type.

* update to b7209

* fix TopK perf

* only build arm code on arm
parent 854d40ed
Show whitespace changes
Inline Side-by-side
Showing with 342 additions and 73 deletions
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp
View file @ 0cf7794b
......@@ -7,6 +7,7 @@ layout (push_constant) uniform parameter
uint nb00; uint nb01; uint nb02; uint nb03;
uint ne10; uint ne11; uint ne12; uint ne13;
float sf0; float sf1; float sf2; float sf3;
float pixel_offset;
} p;
#include "types.glsl"
......@@ -19,7 +20,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
// from ggml.h: enum ggml_scale_mode, enum ggml_scale_flag
#define NEAREST 0
#define BILINEAR 1
#define ALIGN_CORNERS (1 << 8)
#define BICUBIC 2
layout (constant_id = 0) const uint scale_mode = 0;
......@@ -52,7 +53,7 @@ float fetch_bilinear(ivec2 c0, ivec2 c1, vec2 d, uint i12, uint i13) {
float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
const ivec2 ne0 = ivec2(p.ne00, p.ne01);
const vec2 c = (vec2(i10, i11) + 0.5) / vec2(p.sf0, p.sf1) - 0.5;
const vec2 c = (vec2(i10, i11) + p.pixel_offset) / vec2(p.sf0, p.sf1) - p.pixel_offset;
const vec2 c0f = floor(c);
const vec2 d = c - c0f;
const ivec2 c0 = max(ivec2(c0f), 0);
......@@ -61,14 +62,37 @@ float interpolate_bilinear(uint i10, uint i11, uint i12, uint i13) {
return fetch_bilinear(c0, c1, d, i12, i13);
}
float interpolate_bilinear_align_corners(uint i10, uint i11, uint i12, uint i13) {
const vec2 c = vec2(i10, i11) / vec2(p.sf0, p.sf1);
const vec2 c0f = floor(c);
const vec2 d = c - c0f;
const ivec2 c0 = ivec2(c0f);
const ivec2 c1 = c0 + 1;
// Bicubic interpolation with alpha = -0.75
// https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
const vec4 bcoeffs1 = vec4( 1.25, -2.25, 0.0, 1.0);
const vec4 bcoeffs2 = vec4(-0.75, 3.75, -6.0, 3.0);
vec4 powers(float x) { return vec4(x*x*x, x*x, x, 1); }
float bicubic(float p0, float p1, float p2, float p3, float x) {
return p0 * dot(bcoeffs2, powers(x + 1)) +
p1 * dot(bcoeffs1, powers(x )) +
p2 * dot(bcoeffs1, powers(1 - x)) +
p3 * dot(bcoeffs2, powers(2 - x));
}
return fetch_bilinear(c0, c1, d, i12, i13);
#define FETCH(a,b) data_a[base + clamp(i.x+(a), 0, res.x) * p.nb00 + clamp(i.y+(b), 0, res.y) * p.nb01]
float interpolate_bicubic(uint i10, uint i11, uint i12, uint i13) {
const ivec2 res = ivec2(p.ne00 - 1, p.ne01 - 1);
const vec2 coord = (vec2(i10, i11) + p.pixel_offset) / vec2(p.sf0, p.sf1) - p.pixel_offset;
const vec2 d = fract(coord);
const ivec2 i = ivec2(floor(coord));
const uint i02 = uint(i12 / p.sf2);
const uint i03 = uint(i13 / p.sf3);
const uint base = p.a_offset + i03 * p.nb03 + i02 * p.nb02;
return bicubic(
bicubic(FETCH(-1,-1), FETCH(0,-1), FETCH(1,-1), FETCH(2,-1), d.x),
bicubic(FETCH(-1, 0), FETCH(0, 0), FETCH(1, 0), FETCH(2, 0), d.x),
bicubic(FETCH(-1, 1), FETCH(0, 1), FETCH(1, 1), FETCH(2, 1), d.x),
bicubic(FETCH(-1, 2), FETCH(0, 2), FETCH(1, 2), FETCH(2, 2), d.x), d.y);
}
void main() {
......@@ -91,8 +115,8 @@ void main() {
case BILINEAR:
result = interpolate_bilinear(i10, i11, i12, i13);
break;
case BILINEAR | ALIGN_CORNERS:
result = interpolate_bilinear_align_corners(i10, i11, i12, i13);
case BICUBIC:
result = interpolate_bicubic(i10, i11, i12, i13);
break;
}
......
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View file @ 0cf7794b
......@@ -18,6 +18,7 @@
#include <algorithm>
#include <sys/stat.h>
#include <sys/types.h>
#include <filesystem>
#ifdef _WIN32
#define NOMINMAX
......@@ -75,7 +76,7 @@ enum MatMulIdType {
namespace {
void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {
void execute_command(std::vector<std::string>& command, std::string& stdout_str, std::string& stderr_str) {
#ifdef _WIN32
HANDLE stdout_read, stdout_write;
HANDLE stderr_read, stderr_write;
......@@ -98,8 +99,10 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
si.hStdOutput = stdout_write;
si.hStdError = stderr_write;
std::vector<char> cmd(command.begin(), command.end());
cmd.push_back('\0');
std::string cmd;
for (const auto& part : command) {
cmd += part + " ";
}
if (!CreateProcessA(NULL, cmd.data(), NULL, NULL, TRUE, 0, NULL, NULL, &si, &pi)) {
throw std::runtime_error("Failed to create process");
......@@ -137,6 +140,12 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
throw std::runtime_error("Failed to fork process");
}
std::vector<char*> argv;
for (std::string& part : command) {
argv.push_back(part.data());
}
argv.push_back(nullptr);
if (pid == 0) {
close(stdout_pipe[0]);
close(stderr_pipe[0]);
......@@ -144,7 +153,7 @@ void execute_command(const std::string& command, std::string& stdout_str, std::s
dup2(stderr_pipe[1], STDERR_FILENO);
close(stdout_pipe[1]);
close(stderr_pipe[1]);
execl("/bin/sh", "sh", "-c", command.c_str(), (char*) nullptr);
execvp(argv[0], argv.data());
_exit(EXIT_FAILURE);
} else {
close(stdout_pipe[1]);
......@@ -315,20 +324,27 @@ compile_count_guard acquire_compile_slot() {
void string_to_spv_func(std::string name, std::string in_path, std::string out_path, std::map<std::string, std::string> defines, bool coopmat, bool dep_file, compile_count_guard slot) {
std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2";
// disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
// disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
std::string opt_level = (coopmat || name.find("bf16") != std::string::npos) ? "" : "-O";
#ifdef _WIN32
std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_path + "\""};
std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, "\"" + in_path + "\"", "-o", "\"" + out_path + "\""};
#else
std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, in_path, "-o", out_path};
std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, in_path, "-o", out_path};
#endif
// disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
// disable spirv-opt for bf16 shaders for https://github.com/ggml-org/llama.cpp/issues/15344
// disable spirv-opt for rope shaders for https://github.com/ggml-org/llama.cpp/issues/16860
if (!coopmat && name.find("bf16") == std::string::npos && name.find("rope") == std::string::npos) {
cmd.push_back("-O");
}
if (dep_file) {
cmd.push_back("-MD");
cmd.push_back("-MF");
#ifdef _WIN32
cmd.push_back("\"" + target_cpp + ".d\"");
#else
cmd.push_back(target_cpp + ".d");
#endif
}
#ifdef GGML_VULKAN_SHADER_DEBUG_INFO
......@@ -352,9 +368,13 @@ void string_to_spv_func(std::string name, std::string in_path, std::string out_p
// }
// std::cout << std::endl;
execute_command(command, stdout_str, stderr_str);
execute_command(cmd, stdout_str, stderr_str);
if (!stderr_str.empty()) {
std::cerr << "cannot compile " << name << "\n\n" << command << "\n\n" << stderr_str << std::endl;
std::cerr << "cannot compile " << name << "\n\n";
for (const auto& part : cmd) {
std::cerr << part << " ";
}
std::cerr << "\n\n" << stderr_str << std::endl;
return;
}
......@@ -428,7 +448,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
base_dict["ACC_TYPE" ] = f16acc ? "float16_t" : "float";
base_dict["ACC_TYPE_VEC2"] = f16acc ? "f16vec2" : "vec2";
if (f16acc) {
base_dict["ACC_TYPE_MAX"] = "\"float16_t(65504.0)\"";
base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
}
if (coopmat) {
......@@ -608,7 +628,7 @@ void process_shaders() {
fa_base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
fa_base_dict["ACC_TYPEV4"] = f16acc ? "f16vec4" : "vec4";
if (f16acc) {
fa_base_dict["ACC_TYPE_MAX"] = "\"float16_t(65504.0)\"";
fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
}
for (const auto& tname : type_names) {
......@@ -659,14 +679,20 @@ void process_shaders() {
string_to_spv("mul_mat_vec_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
string_to_spv("mul_mat_vec_" + tname + "_f16_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
string_to_spv("mul_mat_vec_id_" + tname + "_f32_f32_subgroup_no_shmem", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
// mul mat vec with integer dot product
#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
if (is_legacy_quant(tname)) {
if (is_legacy_quant(tname) || tname == "mxfp4" || is_k_quant(tname)) {
string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
string_to_spv("mul_mat_vec_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}}));
string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}));
string_to_spv("mul_mat_vec_id_" + tname + "_q8_1_f32_subgroup_no_shmem", "mul_mat_vecq.comp", merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}, {"ACC_TYPE", "float"}, {"USE_SUBGROUP_ADD_NO_SHMEM", "1"}}));
}
#endif
......@@ -694,6 +720,8 @@ void process_shaders() {
string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("rms_norm_partials_f32", "rms_norm_partials.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("rms_norm_mul_rope_f32_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float"}, {"RMS_NORM_ROPE_FUSION", "1"}}));
string_to_spv("rms_norm_mul_rope_f32_f16_rte", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RMS_NORM_ROPE_FUSION", "1"}, {"RTE16", "1"}}));
string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
......@@ -712,6 +740,9 @@ void process_shaders() {
string_to_spv("cpy_f32_i32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "int"}});
string_to_spv("cpy_i32_f32", "copy.comp", {{"A_TYPE", "int"}, {"D_TYPE", "float"}});
string_to_spv("cpy_transpose_16", "copy_transpose.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
string_to_spv("cpy_transpose_32", "copy_transpose.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});
for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("cpy_f32_" + t + "_rte", "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
......@@ -794,6 +825,9 @@ void process_shaders() {
std::string suffix = rte ? "_rte" : "";
string_to_spv("exp_f16" + suffix, "exp.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
string_to_spv("exp_f32" + suffix, "exp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"} , {"RTE16", rte ? "1" : "0"}});
string_to_spv("log_f16" + suffix, "log.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", rte ? "1" : "0"}});
string_to_spv("log_f32" + suffix, "log.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"RTE16", rte ? "1" : "0"}});
}
string_to_spv("gelu_f16", "gelu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
......@@ -805,6 +839,8 @@ void process_shaders() {
string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("relu_f16", "relu.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("neg_f16", "neg.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("neg_f32", "neg.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("tanh_f16", "tanh.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("sigmoid_f16", "sigmoid.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
......@@ -813,6 +849,30 @@ void process_shaders() {
string_to_spv("hardsigmoid_f32","hardsigmoid.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("hardswish_f16", "hardswish.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("hardswish_f32", "hardswish.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("abs_f16", "abs.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("abs_f32", "abs.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("tri_f16", "tri.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("tri_f32", "tri.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("softplus_f16", "softplus.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("softplus_f32", "softplus.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("add1_f16_f16", "add1.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
string_to_spv("add1_f16_f32", "add1.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
string_to_spv("add1_f32_f32", "add1.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("arange_f32", "arange.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("fill_f32", "fill.comp", {{"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
string_to_spv("step_f16", "step.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("step_f32", "step.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("round_f16", "round.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("round_f32", "round.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("ceil_f16", "ceil.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("ceil_f32", "ceil.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("floor_f16", "floor.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("floor_f32", "floor.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("trunc_f16", "trunc.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("trunc_f32", "trunc.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
for (auto rte : {false, true}) {
std::string suffix = rte ? "_rte" : "";
......@@ -839,31 +899,36 @@ void process_shaders() {
string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
string_to_spv("soft_max_back_f32", "soft_max_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_norm_f32_f16", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
string_to_spv("rope_norm_f32_f16_rte", "rope_norm.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_neox_f32_f16", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}});
string_to_spv("rope_neox_f32_f16_rte", "rope_neox.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"ROPE_D_TYPE", "float"}});
string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}});
string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"ROPE_D_TYPE", "float16_t"}, {"RTE16", "1"}});
string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
string_to_spv("argsort_large_f32", "argsort_large.comp", {{"A_TYPE", "float"}});
string_to_spv("topk_argsort_f32", "topk_argsort.comp", {{"A_TYPE", "float"}});
string_to_spv("topk_nary_search_f32", "topk_nary_search.comp", {{"A_TYPE", "float"}});
string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}}));
string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}}));
string_to_spv("cumsum_f32", "cumsum.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
for (std::string dim_str : {"", "_3d"}) {
for (bool bda : {false, true}) {
......@@ -888,6 +953,8 @@ void process_shaders() {
string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
string_to_spv("solve_tri_f32", "solve_tri.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
for (auto transpose : {false, true}) {
for (auto unroll : {false, true}) {
for (auto a_f16 : {false, true}) {
......@@ -1039,7 +1106,7 @@ void write_output_files() {
for (const std::string& btype : btypes) {
for (const auto& tname : type_names) {
if (btype == "q8_1" && !is_legacy_quant(tname)) {
if (btype == "q8_1" && !is_legacy_quant(tname) && tname != "mxfp4" && !is_k_quant(tname)) {
continue;
}
hdr << "extern const void * arr_dmmv_" << tname << "_" << btype << "_f32_data[3];\n";
......@@ -1048,6 +1115,16 @@ void write_output_files() {
src << "const void * arr_dmmv_" << tname << "_" << btype << "_f32_data[3] = {mul_mat_vec_" << tname << "_" << btype << "_f32_data, mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_data, mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_no_shmem_data};\n";
src << "const uint64_t arr_dmmv_" << tname << "_" << btype << "_f32_len[3] = {mul_mat_vec_" << tname << "_" << btype << "_f32_len, mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_len, mul_mat_vec_" << tname << "_" << btype << "_f32_subgroup_no_shmem_len};\n";
}
if (btype == "f16") {
continue;
}
hdr << "extern const void * arr_dmmv_id_" << tname << "_" << btype << "_f32_data[3];\n";
hdr << "extern const uint64_t arr_dmmv_id_" << tname << "_" << btype << "_f32_len[3];\n";
if (basename(input_filepath) == "mul_mat_vec.comp") {
src << "const void * arr_dmmv_id_" << tname << "_" << btype << "_f32_data[3] = {mul_mat_vec_id_" << tname << "_" << btype << "_f32_data, mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_data, mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_no_shmem_data};\n";
src << "const uint64_t arr_dmmv_id_" << tname << "_" << btype << "_f32_len[3] = {mul_mat_vec_id_" << tname << "_" << btype << "_f32_len, mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_len, mul_mat_vec_id_" << tname << "_" << btype << "_f32_subgroup_no_shmem_len};\n";
}
}
}
......
ml/backend/ggml/ggml/src/ggml.c
View file @ 0cf7794b
......@@ -940,6 +940,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"COS",
"SUM",
"SUM_ROWS",
"CUMSUM",
"MEAN",
"ARGMAX",
"COUNT_EQUAL",
......@@ -994,7 +995,10 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"ARANGE",
"TIMESTEP_EMBEDDING",
"ARGSORT",
"TOP_K",
"LEAKY_RELU",
"TRI",
"FILL",
"FLASH_ATTN_EXT",
"FLASH_ATTN_BACK",
......@@ -1007,6 +1011,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"RWKV_WKV6",
"GATED_LINEAR_ATTN",
"RWKV_WKV7",
"SOLVE_TRI",
"UNARY",
......@@ -1024,7 +1029,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"GLU",
};
static_assert(GGML_OP_COUNT == 90, "GGML_OP_COUNT != 90");
static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
......@@ -1044,6 +1049,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"cos(x)",
"Σx",
"Σx_k",
"cumsum(x)",
"Σx/n",
"argmax(x)",
"count_equal(x)",
......@@ -1098,7 +1104,10 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"arange(start, stop, step)",
"timestep_embedding(timesteps, dim, max_period)",
"argsort(x)",
"top_k(x)",
"leaky_relu(x)",
"tri(x)",
"fill(x, c)",
"flash_attn_ext(x)",
"flash_attn_back(x)",
......@@ -1111,6 +1120,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"rwkv_wkv6(k, v, r, tf, td, s)",
"gated_linear_attn(k, v, q, gate, s)",
"rwkv_wkv7(r, w, k, v, a, b, s)",
"A X = B, A triangular, solve X",
"unary(x)",
......@@ -1128,7 +1138,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"glu(x)",
};
static_assert(GGML_OP_COUNT == 90, "GGML_OP_COUNT != 90");
static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
......@@ -1147,6 +1157,8 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
"HARDSWISH",
"HARDSIGMOID",
"EXP",
"EXPM1",
"SOFTPLUS",
"GELU_ERF",
"XIELU",
"FLOOR",
......@@ -1155,7 +1167,7 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
"TRUNC",
};
static_assert(GGML_UNARY_OP_COUNT == 20, "GGML_UNARY_OP_COUNT != 20");
static_assert(GGML_UNARY_OP_COUNT == 22, "GGML_UNARY_OP_COUNT != 22");
static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
"REGLU",
......@@ -2263,6 +2275,30 @@ struct ggml_tensor * ggml_log_inplace(
return ggml_log_impl(ctx, a, true);
}
struct ggml_tensor * ggml_expm1(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary(ctx, a, GGML_UNARY_OP_EXPM1);
}
struct ggml_tensor * ggml_expm1_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_EXPM1);
}
struct ggml_tensor * ggml_softplus(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary(ctx, a, GGML_UNARY_OP_SOFTPLUS);
}
struct ggml_tensor * ggml_softplus_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SOFTPLUS);
}
// ggml_sin
static struct ggml_tensor * ggml_sin_impl(
......@@ -2346,6 +2382,21 @@ struct ggml_tensor * ggml_sum_rows(
return result;
}
// ggml_cumsum
struct ggml_tensor * ggml_cumsum(
struct ggml_context * ctx,
struct ggml_tensor * a) {
GGML_ASSERT(a->type == GGML_TYPE_F32);
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
result->op = GGML_OP_CUMSUM;
result->src[0] = a;
return result;
}
// ggml_mean
struct ggml_tensor * ggml_mean(
......@@ -2673,8 +2724,8 @@ struct ggml_tensor * ggml_xielu(
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
ggml_set_op_params_i32(result, 0, (int32_t) GGML_UNARY_OP_XIELU);
ggml_set_op_params_f32(result, 1, beta + ggml_softplus(alpha_n));
ggml_set_op_params_f32(result, 2, ggml_softplus(alpha_p));
ggml_set_op_params_f32(result, 1, beta + ggml_compute_softplus_f32(alpha_n));
ggml_set_op_params_f32(result, 2, ggml_compute_softplus_f32(alpha_p));
ggml_set_op_params_f32(result, 3, beta);
ggml_set_op_params_f32(result, 4, eps);
......@@ -4992,28 +5043,6 @@ struct ggml_tensor * ggml_roll(
return result;
}
// ggml_arange
struct ggml_tensor * ggml_arange(
struct ggml_context * ctx,
float start,
float stop,
float step) {
GGML_ASSERT(stop > start);
const int64_t steps = (int64_t) ceilf((stop - start) / step);
struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
ggml_set_op_params_f32(result, 0, start);
ggml_set_op_params_f32(result, 1, stop);
ggml_set_op_params_f32(result, 2, step);
result->op = GGML_OP_ARANGE;
return result;
}
// ggml_timestep_embedding
struct ggml_tensor * ggml_timestep_embedding(
......@@ -5033,6 +5062,61 @@ struct ggml_tensor * ggml_timestep_embedding(
return result;
}
// ggml_tri
struct ggml_tensor * ggml_tri(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_tri_type type) {
GGML_ASSERT(a->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(a));
GGML_ASSERT(a->ne[0] == a->ne[1]);
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
ggml_set_op_params_i32(result, 0, type);
result->op = GGML_OP_TRI;
result->src[0] = a;
return result;
}
// ggml_fill
static struct ggml_tensor * ggml_fill_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
float c,
bool inplace) {
GGML_ASSERT(a->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(a));
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params_f32(result, 0, c);
result->op = GGML_OP_FILL;
result->src[0] = a;
return result;
}
struct ggml_tensor * ggml_fill(
struct ggml_context * ctx,
struct ggml_tensor * a,
float c) {
return ggml_fill_impl(ctx, a, c, false);
}
struct ggml_tensor * ggml_fill_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
float c) {
return ggml_fill_impl(ctx, a, c, true);
}
// ggml_argsort
struct ggml_tensor * ggml_argsort(
......@@ -5040,6 +5124,7 @@ struct ggml_tensor * ggml_argsort(
struct ggml_tensor * a,
enum ggml_sort_order order) {
GGML_ASSERT(a->ne[0] <= INT32_MAX);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
ggml_set_op_params_i32(result, 0, (int32_t) order);
......@@ -5050,9 +5135,9 @@ struct ggml_tensor * ggml_argsort(
return result;
}
// ggml_top_k
// ggml_argsort_top_k
struct ggml_tensor * ggml_top_k(
struct ggml_tensor * ggml_argsort_top_k(
struct ggml_context * ctx,
struct ggml_tensor * a,
int k) {
......@@ -5068,6 +5153,44 @@ struct ggml_tensor * ggml_top_k(
return result;
}
// ggml_top_k
struct ggml_tensor * ggml_top_k(
struct ggml_context * ctx,
struct ggml_tensor * a,
int k) {
GGML_ASSERT(a->ne[0] >= k);
struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_I32, k, a->ne[1], a->ne[2], a->ne[3]);
result->op = GGML_OP_TOP_K;
result->src[0] = a;
return result;
}
// ggml_arange
struct ggml_tensor * ggml_arange(
struct ggml_context * ctx,
float start,
float stop,
float step) {
GGML_ASSERT(stop > start);
const int64_t steps = (int64_t) ceilf((stop - start) / step);
struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
ggml_set_op_params_f32(result, 0, start);
ggml_set_op_params_f32(result, 1, stop);
ggml_set_op_params_f32(result, 2, step);
result->op = GGML_OP_ARANGE;
return result;
}
// ggml_flash_attn_ext
struct ggml_tensor * ggml_flash_attn_ext(
......@@ -5887,6 +6010,41 @@ struct ggml_tensor * ggml_opt_step_sgd(
return result;
}
// solve_tri
struct ggml_tensor * ggml_solve_tri(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
bool left,
bool lower,
bool uni) {
GGML_ASSERT(a->type == GGML_TYPE_F32);
GGML_ASSERT(b->type == GGML_TYPE_F32);
// A must be square and lower diagonal
GGML_ASSERT(a->ne[0] == a->ne[1]);
// B must have same outer dimension as A
GGML_ASSERT(a->ne[1] == b->ne[1]);
// batch dimensions must be equal
GGML_ASSERT(a->ne[2] == b->ne[2]);
GGML_ASSERT(a->ne[3] == b->ne[3]);
GGML_ASSERT(ggml_is_contiguous(a));
GGML_ASSERT(ggml_is_contiguous(b));
GGML_ASSERT(lower && left && !uni); // TODO: support other variants
struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, b->ne[0], b->ne[1], b->ne[2], b->ne[3]);
result->op = GGML_OP_SOLVE_TRI;
result->src[0] = a;
result->src[1] = b;
return result;
}
////////////////////////////////////////////////////////////////////////////////
struct ggml_hash_set ggml_hash_set_new(size_t size) {
......@@ -6459,6 +6617,16 @@ static void ggml_compute_backward(
ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, tensor, grad));
}
} break;
case GGML_UNARY_OP_EXPM1: {
if (src0_needs_grads) {
ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_exp(ctx, src0)));
}
} break;
case GGML_UNARY_OP_SOFTPLUS: {
if (src0_needs_grads) {
ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_sigmoid(ctx, src0)));
}
} break;
default: {
fprintf(stderr, "%s: unsupported unary op for backward pass: %s\n",
__func__, ggml_unary_op_name(ggml_get_unary_op(tensor)));
......
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