Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

metal: optimise GGML_OP_SUM (#16559)

Browse Source 
* optimise GGML_OP_SUM

* add non-contiguous tests by permuting the input

* change tests to require full contiguity of OP_SUM

* cuda : add check GGML_OP_SUM

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Sam/Samuel
2025-10-15 17:05:56 +03:00
committed by GitHub
co-authored by Georgi Gerganov
parent 17304cbcc1
commit f4ce81c45e
5 changed files with 71 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+14 -1
View File
@@ -866,12 +866,25 @@ int ggml_metal_op_sum(ggml_metal_op_t ctx, int idx) {
ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_sum(lib, op);
int nth = 32; // SIMD width
while (nth < (int) n && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
nth *= 2;
}
nth = std::min(nth, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
nth = std::min(nth, (int) n);
const int nsg = (nth + 31) / 32;
ggml_metal_encoder_set_pipeline(enc, pipeline);
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 2);
ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, 1, 1, 1);
ggml_metal_encoder_set_threadgroup_memory_size(enc, nsg * sizeof(float), 0);
ggml_metal_encoder_dispatch_threadgroups(enc, 1, 1, 1, nth, 1, 1);
return 1;
}