extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2024-12-26 16:48:50 +01:00
Code Issues Packages Projects Releases Wiki Activity

metal : add F32 support + update bench output

Browse Source
This commit is contained in:
Georgi Gerganov 2023-09-15 13:56:08 +03:00
parent 93935980f8
commit b8432f28f4
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
7 changed files with 111 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
Makefile
View File

@ -186,6 +186,7 @@ ifndef WHISPER_NO_METAL
ifeq ($(UNAME_S),Darwin)
WHISPER_METAL := 1
CFLAGS += -DGGML_USE_METAL
CXXFLAGS += -DGGML_USE_METAL
LDFLAGS += -framework Foundation -framework Metal -framework MetalKit
endif

10
extra/bench-all.sh
View File

@ -44,8 +44,8 @@ if [ "$encoder_only" -eq 0 ]; then
printf "\n"
fi
printf "| %6s | %6s | %12s | %9s | %3s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "Enc." "Dec." "PP" "Commit"
printf "| %6s | %6s | %12s | %9s | %3s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---"
printf "| %6s | %6s | %16s | %11s | %3s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "Enc." "Dec." "PP" "Commit"
printf "| %6s | %6s | %16s | %11s | %3s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---"
for model in "${models[@]}"; do
# actual run
@ -83,9 +83,13 @@ for model in "${models[@]}"; do
config="$config COREML"
fi
if [[ $system_info == *"METAL = 1"* ]]; then
config="$config METAL"
fi
commit=$(git rev-parse --short HEAD)
if [ $ret -eq 0 ]; then
printf "| <todo> | <todo> | %12s | %9s | %3s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$encode_time" "$decode_time" "$prompt_time" "$commit"
printf "| <todo> | <todo> | %16s | %11s | %3s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$encode_time" "$decode_time" "$prompt_time" "$commit"
fi
done

12
ggml-metal.m
View File

@ -78,6 +78,7 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(get_rows_q6_K);
GGML_METAL_DECL_KERNEL(rms_norm);
GGML_METAL_DECL_KERNEL(norm);
GGML_METAL_DECL_KERNEL(mul_mat_f32_f32);
GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_1row);
GGML_METAL_DECL_KERNEL(mul_mat_f16_f32_l4);
@ -89,6 +90,7 @@ struct ggml_metal_context {
GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
@ -237,6 +239,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(get_rows_q6_K);
GGML_METAL_ADD_KERNEL(rms_norm);
GGML_METAL_ADD_KERNEL(norm);
GGML_METAL_ADD_KERNEL(mul_mat_f32_f32);
GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_1row);
GGML_METAL_ADD_KERNEL(mul_mat_f16_f32_l4);
@ -248,6 +251,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
@ -309,6 +313,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(get_rows_q6_K);
GGML_METAL_DEL_KERNEL(rms_norm);
GGML_METAL_DEL_KERNEL(norm);
GGML_METAL_DEL_KERNEL(mul_mat_f32_f32);
GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_1row);
GGML_METAL_DEL_KERNEL(mul_mat_f16_f32_l4);
@ -320,6 +325,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32);
GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32);
GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
@ -885,6 +891,7 @@ void ggml_metal_graph_compute(
ne00%32 == 0 &&
ne11 > 1) {
switch (src0->type) {
case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f32_f32]; break;
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32]; break;
case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
@ -919,6 +926,11 @@ void ggml_metal_graph_compute(
// use custom matrix x vector kernel
switch (src0t) {
case GGML_TYPE_F32:
{
[encoder setComputePipelineState:ctx->pipeline_mul_mat_f32_f32];
nrows = 4;
} break;
case GGML_TYPE_F16:
{
nth0 = 32;

88
ggml-metal.metal
View File

@ -523,6 +523,79 @@ kernel void kernel_mul_mat_q8_0_f32(
}
}
#define N_F32_F32 4
kernel void kernel_mul_mat_f32_f32(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
const int64_t r0 = tgpig.x;
const int64_t rb = tgpig.y*N_F32_F32;
const int64_t im = tgpig.z;
device const float * x = (device const float *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
if (ne00 < 128) {
for (int row = 0; row < N_F32_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00; i += 32) {
sumf += (float) x[i] * (float) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
} else {
device const float4 * x4 = (device const float4 *)x;
for (int row = 0; row < N_F32_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
device const float4 * y4 = (device const float4 *) y;
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
}
}
kernel void kernel_mul_mat_f16_f32_1row(
device const char * src0,
device const char * src1,
@ -1399,13 +1472,13 @@ kernel void kernel_mul_mat_q4_K_f32(
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
constant int64_t & ne01 [[buffer(4)]],
constant int64_t & ne02 [[buffer(5)]],
constant int64_t & ne10 [[buffer(9)]],
constant int64_t & ne12 [[buffer(11)]],
constant int64_t & ne0 [[buffer(15)]],
constant int64_t & ne1 [[buffer(16)]],
constant uint & gqa [[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
@ -2268,6 +2341,7 @@ typedef void (mat_mm_t)(
constant uint & gqa,
threadgroup uchar *, uint3, uint, uint);
template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm<float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;

8
ggml.c
View File

@ -20753,6 +20753,14 @@ int ggml_cpu_has_arm_fma(void) {
#endif
}
int ggml_cpu_has_metal(void) {
#if defined(GGML_USE_METAL)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_f16c(void) {
#if defined(__F16C__)
return 1;

1
ggml.h
View File

@ -1961,6 +1961,7 @@ extern "C" {
GGML_API int ggml_cpu_has_fma (void);
GGML_API int ggml_cpu_has_neon (void);
GGML_API int ggml_cpu_has_arm_fma (void);
GGML_API int ggml_cpu_has_metal (void);
GGML_API int ggml_cpu_has_f16c (void);
GGML_API int ggml_cpu_has_fp16_va (void);
GGML_API int ggml_cpu_has_wasm_simd (void);

1
whisper.cpp
View File

@ -3669,6 +3669,7 @@ const char * whisper_print_system_info(void) {
s += "FMA = " + std::to_string(ggml_cpu_has_fma()) + " | ";
s += "NEON = " + std::to_string(ggml_cpu_has_neon()) + " | ";
s += "ARM_FMA = " + std::to_string(ggml_cpu_has_arm_fma()) + " | ";
s += "METAL = " + std::to_string(ggml_cpu_has_metal()) + " | ";
s += "F16C = " + std::to_string(ggml_cpu_has_f16c()) + " | ";
s += "FP16_VA = " + std::to_string(ggml_cpu_has_fp16_va()) + " | ";
s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";