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

adding the looping structure based on the chunk configuration.

Browse source
This commit is contained in:
Kunnis 2024-05-09 23:29:49 -05:00
parent 4762d79d3d
commit fc7dc515f1
1 changed files with 6 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml.c
View file

@ -12091,8 +12091,10 @@ UseGgmlGemm2:;
const int64_t nchunk0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
const int64_t nchunk1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
//The first chunk comes from our thread_id, the rest will get auto-assigned.
int current_chunk = ith;
while (current_chunk < nchunk0 * nchunk1)
{
const int64_t ith0 = current_chunk % nchunk0;
const int64_t ith1 = current_chunk / nchunk0;
@ -12112,6 +12114,10 @@ UseGgmlGemm2:;
chunks_executed++;
#endif
if (nth >= nchunk0 * nchunk1)
break;
current_chunk = atomic_fetch_add(&state->shared->current_chunk, 1);
}
#ifdef GGML_PERF