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

Fix formatting

Browse source
This commit is contained in:
Kunnis 2024-05-14 17:15:47 -05:00
parent bd80601ea8
commit d9ba30a204
1 changed files with 27 additions and 28 deletions
Download patch file Download diff file Expand all files Collapse all files

13
ggml.c
View file

@ -2435,7 +2435,6 @@ static void ggml_setup_op_has_task_pass(void) {
} }
} }
// //
// NUMA support // NUMA support
// //
@ -12085,8 +12084,9 @@ UseGgmlGemm2:;
int chunk_size = 16; int chunk_size = 16;
// We need to step up the size if it's small // We need to step up the size if it's small
if (nr0 == 1 || nr1 == 1) if (nr0 == 1 || nr1 == 1) {
chunk_size = 64; chunk_size = 64;
}
// distribute the work across the inner or outer loop based on which one is larger // distribute the work across the inner or outer loop based on which one is larger
// The number of chunks in the 0/1 dim. // The number of chunks in the 0/1 dim.
@ -12097,8 +12097,7 @@ UseGgmlGemm2:;
// If the chunking is poor for the number of threads on this setup, scrap the whole plan. Re-chunk it by thread. // If the chunking is poor for the number of threads on this setup, scrap the whole plan. Re-chunk it by thread.
// Also, chunking by thread was measured to have perform better on NUMA systems. See https://github.com/ggerganov/llama.cpp/pull/6915 // Also, chunking by thread was measured to have perform better on NUMA systems. See https://github.com/ggerganov/llama.cpp/pull/6915
// In theory, chunking should be just as useful on NUMA and non NUMA systems, but testing disagreed with that. // In theory, chunking should be just as useful on NUMA and non NUMA systems, but testing disagreed with that.
if (nchunk0 * nchunk1 < nth * 4 || ggml_is_numa()) if (nchunk0 * nchunk1 < nth * 4 || ggml_is_numa()) {
{
// distribute the thread work across the inner or outer loop based on which one is larger // distribute the thread work across the inner or outer loop based on which one is larger
nchunk0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows nchunk0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
nchunk1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows nchunk1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
@ -12114,8 +12113,7 @@ UseGgmlGemm2:;
//The first chunk comes from our thread_id, the rest will get auto-assigned. //The first chunk comes from our thread_id, the rest will get auto-assigned.
int current_chunk = ith; int current_chunk = ith;
while (current_chunk < nchunk0 * nchunk1) while (current_chunk < nchunk0 * nchunk1) {
{
const int64_t ith0 = current_chunk % nchunk0; const int64_t ith0 = current_chunk % nchunk0;
const int64_t ith1 = current_chunk / nchunk0; const int64_t ith1 = current_chunk / nchunk0;
@ -12131,8 +12129,9 @@ UseGgmlGemm2:;
chunks_executed++; chunks_executed++;
#endif #endif
if (nth >= nchunk0 * nchunk1) if (nth >= nchunk0 * nchunk1) {
break; break;
}
current_chunk = atomic_fetch_add(&state->shared->current_chunk, 1); current_chunk = atomic_fetch_add(&state->shared->current_chunk, 1);
} }