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

Making it much more likely to rechunk.

Browse source
This commit is contained in:
Kunnis 2024-05-10 15:17:33 -05:00
parent 807c8252ce
commit 974e43be25
1 changed files with 6 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

7
ggml.c
View file

@ -12097,8 +12097,10 @@ UseGgmlGemm2:;
//printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
//If the chunking is poor for the number of threads on this setup, scrap the whole plan. Re-chunk it by thread.
if (nchunk0 * nchunk1 < nth * 4)
if (nchunk0 * nchunk1 < nth * 400)
{
//if (ith == 0)
// printf("rechunked");
// 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
nchunk1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
@ -12108,6 +12110,9 @@ UseGgmlGemm2:;
const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;
//if (ith == 0)
// printf("MUL_MAT = [%d, %d, %d, %d] x [%d, %d, %d, %d] = %d x %d = %d. Fp/Ch %d\n", ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nchunk0, nchunk1, nchunk0 * nchunk1, ne00 * nr0 * nr1 / nchunk0 / nchunk1);
//The first chunk comes from our thread_id, the rest will get auto-assigned.
int current_chunk = ith;