ggml : more perfo with llamafile tinyblas on x86_64 (#10714)

* more perfo with llamafile tinyblas on x86_64.

- add bf16 suport
- change dispache strategie (thanks:
https://github.com/ikawrakow/ik_llama.cpp/pull/71 )
- reduce memory bandwidth

simple tinyblas dispache and more cache freindly

* tinyblas dynamic dispaching

* sgemm: add M blocs.

* - git 2.47 use short id of len 9.
- show-progress is not part of GNU Wget2

* remove not stable test
This commit is contained in:
Djip007 2024-12-24 18:54:49 +01:00 committed by GitHub
parent 09fe2e7613
commit 2cd43f4900
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 287 additions and 278 deletions

View file

@ -7419,14 +7419,14 @@ static void ggml_compute_forward_mul_mat(
if (src1_cont) {
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(params,
ne01, ne11, ne00/ggml_blck_size(src0->type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
(const char *)src1->data + i12*nb12 + i13*nb13,
nb11/ggml_type_size(src1->type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
src1->type,
dst->type))
@ -7471,14 +7471,14 @@ UseGgmlGemm1:;
for (int64_t i13 = 0; i13 < ne13; i13++)
for (int64_t i12 = 0; i12 < ne12; i12++)
if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
if (!llamafile_sgemm(params,
ne01, ne11, ne00/ggml_blck_size(src0->type),
(const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
nb01/ggml_type_size(src0->type),
(const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
row_size/ggml_type_size(vec_dot_type),
(char *)dst->data + i12*nb2 + i13*nb3,
nb1/ggml_type_size(dst->type),
ith, nth,
src0->type,
vec_dot_type,
dst->type))