diff options
| author | Kawrakow <48489457+ikawrakow@users.noreply.github.com> | 2023-07-20 18:19:45 +0300 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2023-07-20 18:19:45 +0300 |
| commit | e782c9e735f93ab4767ffc37462c523b73a17ddc (patch) | |
| tree | b5c87fd34707ec9aa7a7a9716b2ef96157c033c5 /ggml-metal.m | |
| parent | 785829dfe8baf0213f2ff66963d28c62f92d7930 (diff) | |
Faster Q5_K and Q6_K on Metal (#2294)
* Faster Q6_K on Metal
* Faster Q5_K on Metal
* Another Q5_K speedup
---------
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'ggml-metal.m')
| -rw-r--r-- | ggml-metal.m | 19 |
1 files changed, 11 insertions, 8 deletions
diff --git a/ggml-metal.m b/ggml-metal.m index 5e2a211..44d0468 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -703,8 +703,8 @@ void ggml_metal_graph_compute( GGML_ASSERT(ne02 == 1); GGML_ASSERT(ne12 == 1); - nth0 = 4; - nth1 = 16; + nth0 = 2; + nth1 = 32; [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32]; } break; case GGML_TYPE_Q6_K: @@ -712,8 +712,8 @@ void ggml_metal_graph_compute( GGML_ASSERT(ne02 == 1); GGML_ASSERT(ne12 == 1); - nth0 = 4; - nth1 = 16; + nth0 = 2; + nth1 = 32; [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32]; } break; default: @@ -743,11 +743,14 @@ void ggml_metal_graph_compute( src0t == GGML_TYPE_Q4_K) { [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } + else if (src0t == GGML_TYPE_Q5_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } + else if (src0t == GGML_TYPE_Q6_K) { + [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + } else if (src0t == GGML_TYPE_Q2_K || - src0t == GGML_TYPE_Q3_K || - src0t == GGML_TYPE_Q4_K || - src0t == GGML_TYPE_Q5_K || - src0t == GGML_TYPE_Q6_K) { + src0t == GGML_TYPE_Q3_K) { [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; } else { |