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

reuse main thread

Browse source
This commit is contained in:
slaren 2024-06-07 04:35:11 +02:00
parent 2dd049ed45
commit 63e06d0c28
1 changed files with 10 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

10
ggml-blas.cpp
View file

@ -98,7 +98,7 @@ static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct gg
to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
}
#else
for (int i = 0; i < ctx->n_threads; i++) {
for (int i = 0; i < ctx->n_threads - 1; i++) {
ctx->tasks.push_back(std::async(std::launch::async, [=]() {
const int64_t start = i*ne01/ctx->n_threads;
const int64_t end = (i + 1)*ne01/ctx->n_threads;
@ -107,6 +107,14 @@ static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct gg
}
}));
}
{
// reuse the current thread for the last task
const int64_t start = (ctx->n_threads - 1)*ne01/ctx->n_threads;
const int64_t end = ne01;
for (int64_t i01 = start; i01 < end; i01++) {
to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
}
}
#endif
}
}