nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 4 Packages Projects Releases Wiki Activity

ggml-cpu: Enable tiled matmul on AIX (#25199)

Browse Source 
The matmul_tiled path uses large local stack buffers for A_pack and B_pack. On AIX this can trigger a segmentation fault, so reduce the buffer footprint there to keep the tiled path usable.

 Performance Impact:
    ~ 2x gains in PP_Speed for FP32, Q4_0 and Q8_0 models tested with llama-bench, llama-batched-bench and llama-cli.
    Models used: Llama3.2 3b Instruct F32, qwen 2.5 3b Q4_0 and Q8_0
This commit is contained in:
shalinib-ibm
2026-07-06 15:48:17 +05:30
committed by GitHub
parent 86961efd56
commit 3b4fca11ac
1 changed files with 20 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
ggml/src/ggml-cpu/llamafile/sgemm.cpp
View File

@@ -2321,24 +2321,28 @@ class tinyBLAS_Q0_PPC {
}
void matmul(int64_t m, int64_t n) {
#if defined(_AIX) || defined(__BIG_ENDIAN__)
mnpack(0, m, 0, n);
#else
const int64_t mc = 64;
const int64_t kc = 64;
int64_t mc = 64;
int64_t nc = 64;
int64_t kc = 64;
int64_t n_chunk = 64;
#if defined(_AIX) || defined(__BIG_ENDIAN__)
mc = 32;
nc = 32;
kc = 32;
n_chunk = 32
#endif
int64_t n_aligned = 0;
if (n % 64 == 0) {
if (n % n_chunk == 0) {
n_aligned = n;
} else if (n == 4) {
n_aligned = 4;
} else if (n < 64) {
} else if (n < n_chunk) {
n_aligned = (n / 8) * 8;
} else {
n_aligned = (n / 64) * 64;
n_aligned = (n / n_chunk) * n_chunk;
}
if (n_aligned > 0) {
if (n_aligned % 64 == 0) nc = 64;
if (n_aligned % n_chunk == 0) nc = n_chunk;
else if (n_aligned == n) nc = n;
else if (n_aligned % 32 == 0) nc = 32;
else if (n_aligned % 24 == 0) nc = 24;
@@ -2354,7 +2358,6 @@ class tinyBLAS_Q0_PPC {
} else {
mnpack(0, m, 0, n);
}
#endif
}
private:
@@ -3195,16 +3198,19 @@ class tinyBLAS_PPC {
}
void matmul(int64_t m, int64_t n) {
int64_t mc = 256;
int64_t nc = 256;
int64_t kc = 256;
#if defined(_AIX) || defined(__BIG_ENDIAN__)
mnpack(0, m, 0, n);
#else
int64_t mc = 256; int64_t nc = 256; int64_t kc = 256;
mc = 128;
nc = 128;
kc = 128;
#endif
if (m % mc == 0 && n % nc == 0 && k % kc == 0) {
matmul_tiled(m, n, mc, nc, kc);
} else {
mnpack(0, m, 0, n);
}
#endif
}
private: