mirror of
https://github.com/ggerganov/whisper.cpp.git
synced 2024-12-19 04:37:51 +00:00
Add OpenBLAS support
Supported via CMake - just add: cmake .. -DWHISPER_SUPPORT_OPENBLAS=ON On Ubuntu, you have to install the library like this: apt install libopenblas-dev Unfortunately, I don't observe any benefit compared to the original AVX2 + FP16 implementation. Maybe I'm missing something
This commit is contained in:
parent
ebb01b9e33
commit
fbd513b813
@ -41,8 +41,13 @@ option(WHISPER_BUILD_EXAMPLES "whisper: build examples" ${WHISPER_STAND
|
|||||||
|
|
||||||
option(WHISPER_SUPPORT_SDL2 "whisper: support for libSDL2" OFF)
|
option(WHISPER_SUPPORT_SDL2 "whisper: support for libSDL2" OFF)
|
||||||
|
|
||||||
|
if (APPLE)
|
||||||
|
option(WHISPER_NO_ACCELERATE "whisper: disable Accelerate framework" OFF)
|
||||||
|
else()
|
||||||
|
option(WHISPER_SUPPORT_OPENBLAS "whisper: support for OpenBLAS" OFF)
|
||||||
|
endif()
|
||||||
|
|
||||||
option(WHISPER_PERF "whisper: enable perf timings" OFF)
|
option(WHISPER_PERF "whisper: enable perf timings" OFF)
|
||||||
option(WHISPER_NO_ACCELERATE "whisper: disable Accelerate framework" OFF)
|
|
||||||
|
|
||||||
# sanitizers
|
# sanitizers
|
||||||
|
|
||||||
@ -86,6 +91,18 @@ if (APPLE AND NOT WHISPER_NO_ACCELERATE)
|
|||||||
endif()
|
endif()
|
||||||
endif()
|
endif()
|
||||||
|
|
||||||
|
if (WHISPER_SUPPORT_OPENBLAS)
|
||||||
|
find_library(OPENBLAS_LIB openblas)
|
||||||
|
if (OPENBLAS_LIB)
|
||||||
|
message(STATUS "OpenBLAS found")
|
||||||
|
|
||||||
|
set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${OPENBLAS_LIB})
|
||||||
|
set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
|
||||||
|
else()
|
||||||
|
message(WARNING "OpenBLAS not found")
|
||||||
|
endif()
|
||||||
|
endif()
|
||||||
|
|
||||||
# compiler flags
|
# compiler flags
|
||||||
|
|
||||||
if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
|
if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
|
||||||
|
86
ggml.c
86
ggml.c
@ -76,6 +76,8 @@ typedef void* thread_ret_t;
|
|||||||
|
|
||||||
#ifdef GGML_USE_ACCELERATE
|
#ifdef GGML_USE_ACCELERATE
|
||||||
#include <Accelerate/Accelerate.h>
|
#include <Accelerate/Accelerate.h>
|
||||||
|
#elif GGML_USE_OPENBLAS
|
||||||
|
#include <cblas.h>
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
// floating point type used to accumulate sums
|
// floating point type used to accumulate sums
|
||||||
@ -4055,46 +4057,44 @@ void ggml_compute_forward_mul_mat_f32(
|
|||||||
// nb00 < nb01 - src0 is transposed
|
// nb00 < nb01 - src0 is transposed
|
||||||
// compute by src0 columns
|
// compute by src0 columns
|
||||||
|
|
||||||
//#ifdef GGML_USE_ACCELERATE
|
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
|
||||||
// if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
|
if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
|
||||||
// GGML_ASSERT(ggml_is_contiguous(src0));
|
GGML_ASSERT(ggml_is_contiguous(src0));
|
||||||
// GGML_ASSERT(nb10 == sizeof(float));
|
GGML_ASSERT(nb10 == sizeof(float));
|
||||||
//
|
|
||||||
// if (params->ith != 0) return;
|
if (params->ith != 0) return;
|
||||||
//
|
|
||||||
// if (params->type == GGML_TASK_INIT) {
|
if (params->type == GGML_TASK_INIT) {
|
||||||
// return;
|
return;
|
||||||
// }
|
}
|
||||||
//
|
|
||||||
// if (params->type == GGML_TASK_FINALIZE) {
|
if (params->type == GGML_TASK_FINALIZE) {
|
||||||
// return;
|
return;
|
||||||
// }
|
}
|
||||||
//
|
|
||||||
// float * const wdata = params->wdata;
|
for (int i03 = 0; i03 < ne03; i03++) {
|
||||||
//
|
for (int i02 = 0; i02 < ne02; i02++) {
|
||||||
// for (int i03 = 0; i03 < ne03; i03++) {
|
const float * x = (float *) (src0->data);
|
||||||
// for (int i02 = 0; i02 < ne02; i02++) {
|
const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
|
||||||
// const float * x = (float *) (src0->data);
|
|
||||||
// const float * y = (float *) ((char *) src1->data + i02*nb12 + i03*nb13);
|
float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
|
||||||
//
|
|
||||||
// float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
|
// zT = y * xT
|
||||||
//
|
{
|
||||||
// // zT = y * xT
|
cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
|
||||||
// {
|
ne11, ne01, ne10,
|
||||||
// cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
|
1.0f, y, ne10,
|
||||||
// ne11, ne01, ne10,
|
x, ne10,
|
||||||
// 1.0f, y, ne10,
|
0.0f, d, ne01);
|
||||||
// x, ne10,
|
}
|
||||||
// 0.0f, d, ne01);
|
}
|
||||||
// }
|
}
|
||||||
// }
|
|
||||||
// }
|
//printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
|
||||||
//
|
|
||||||
// //printf("CBLAS F32 = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
|
return;
|
||||||
//
|
}
|
||||||
// return;
|
#endif
|
||||||
// }
|
|
||||||
//#endif
|
|
||||||
|
|
||||||
if (params->type == GGML_TASK_INIT) {
|
if (params->type == GGML_TASK_INIT) {
|
||||||
if (nb01 >= nb00) {
|
if (nb01 >= nb00) {
|
||||||
@ -4301,7 +4301,7 @@ void ggml_compute_forward_mul_mat_f16_f32(
|
|||||||
// nb00 < nb01 - src0 is transposed
|
// nb00 < nb01 - src0 is transposed
|
||||||
// compute by src0 columns
|
// compute by src0 columns
|
||||||
|
|
||||||
#ifdef GGML_USE_ACCELERATE
|
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
|
||||||
if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
|
if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
|
||||||
GGML_ASSERT(nb10 == sizeof(float));
|
GGML_ASSERT(nb10 == sizeof(float));
|
||||||
|
|
||||||
@ -6857,7 +6857,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
|
|||||||
} else {
|
} else {
|
||||||
if (node->src0->type == GGML_TYPE_F16 &&
|
if (node->src0->type == GGML_TYPE_F16 &&
|
||||||
node->src1->type == GGML_TYPE_F32) {
|
node->src1->type == GGML_TYPE_F32) {
|
||||||
#ifdef GGML_USE_ACCELERATE
|
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
|
||||||
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
|
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
|
||||||
cur = sizeof(float)*(node->src0->ne[0]*node->src0->ne[1]);
|
cur = sizeof(float)*(node->src0->ne[0]*node->src0->ne[1]);
|
||||||
} else {
|
} else {
|
||||||
@ -8074,7 +8074,7 @@ int ggml_cpu_has_wasm_simd(void) {
|
|||||||
}
|
}
|
||||||
|
|
||||||
int ggml_cpu_has_blas(void) {
|
int ggml_cpu_has_blas(void) {
|
||||||
#if defined(GGML_USE_BLAS) || defined(GGML_USE_ACCELERATE)
|
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
|
||||||
return 1;
|
return 1;
|
||||||
#else
|
#else
|
||||||
return 0;
|
return 0;
|
||||||
|
Loading…
Reference in New Issue
Block a user