vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Remove ggml_compute_forward_out_prod_use_blas(), fix compiling errors on cmake/zig, remove trailing whitespace

Browse source
This commit is contained in:
gwjr 2023-11-15 23:00:44 +00:00
parent 2f0c5dcaf5
commit e5c1f02645
1 changed files with 20 additions and 43 deletions
Download patch file Download diff file Expand all files Collapse all files

49
ggml.c
View file

@ -9598,35 +9598,6 @@ static void ggml_compute_forward_mul_mat(
// ggml_compute_forward_out_prod
#if defined(GGML_USE_ACCELERATE)
// helper function to determine if it is better to use BLAS or not
// based on ggml_compute_forward_mul_mat_use_blas()
// However, testing suggested that BLAS was never slower than the existing code
static bool ggml_compute_forward_out_prod_use_blas(
const struct ggml_tensor * src0,
const struct ggml_tensor * src1,
struct ggml_tensor * dst) {
UNUSED(dst);
// const int64_t ne10 = src1->ne[0];
//
// const int64_t ne0 = dst->ne[0];
// const int64_t ne1 = dst->ne[1];
if (ggml_is_matrix(src0) &&
ggml_is_matrix(src1) &&
ggml_is_contiguous(src0) &&
(ggml_is_contiguous(src1) || ggml_is_transposed(src1))){ //&&
//(ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
return true;
}
// if (ne0 >= 32 && ne1 >= 32 && ne10 >= 32) {
// printf("Cannot use BLAS for large matrix at %s; ne0: %lld, ne1: %lld, ne10:, %lld", dst->name, ne0, ne1, ne10);
// }
return false;
}
#endif
static void ggml_compute_forward_out_prod_f32(
const struct ggml_compute_params * params,
const struct ggml_tensor * src0,
@ -9660,16 +9631,22 @@ static void ggml_compute_forward_out_prod_f32(
// compute by src0 rows
// TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
// TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
// TODO: #if defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
#if defined(GGML_USE_ACCELERATE)
bool use_blas = ggml_is_matrix(src0) &&
ggml_is_matrix(src1) &&
ggml_is_contiguous(src0) &&
(ggml_is_contiguous(src1) || ggml_is_transposed(src1));
#endif
if (params->type == GGML_TASK_INIT) {
#if !defined(GGML_USE_ACCELERATE) // gemm beta will do this
if (!ggml_compute_forward_out_prod_use_blas(src0, src1, dst)) {
#endif
ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
#if !defined(GGML_USE_ACCELERATE)
#if defined(GGML_USE_ACCELERATE) // gemm beta will zero dst
if (use_blas) {
return;
}
#endif
ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
return;
}
@ -9678,7 +9655,7 @@ static void ggml_compute_forward_out_prod_f32(
}
#if defined(GGML_USE_ACCELERATE)
if (ggml_compute_forward_out_prod_use_blas(src0, src1, dst)) {
if (use_blas) {
if (params->ith != 0) { // All threads other than the first do no work.
return;
}