diff options
| author | Georgi Gerganov <ggerganov@gmail.com> | 2024-06-14 17:14:09 +0300 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-06-14 17:14:09 +0300 |
| commit | 66ef1ceedf983773c8ceb4d925285d41d4e50e2a (patch) | |
| tree | 5730d00a2043129897aecdc1a9e1649a625f1349 | |
| parent | e65bbf606c61f49dc06c7ac060cd5ba7ae446025 (diff) | |
metal : utilize max shared memory for mul_mat_id (#7935)
| -rw-r--r-- | ggml-metal.m | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/ggml-metal.m b/ggml-metal.m index ec9e9530..f894274c 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -1862,9 +1862,10 @@ static enum ggml_status ggml_metal_graph_compute( // ne21 = n_rows const int dst_rows = ne20*ne21; const int dst_rows_min = n_as; + const int dst_rows_max = (ctx->device.maxThreadgroupMemoryLength - 32 - 8192)/4; // max size of the rowids array in the kernel shared buffer - GGML_ASSERT(dst_rows <= 2048); + GGML_ASSERT(dst_rows <= dst_rows_max); // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel |