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

metal : optimize pad + cpy (#23354)

Browse Source 
* metal : optimize pad

* metal : optinmize cpy

* cont : better row packing in threadgroup
This commit is contained in:
Georgi Gerganov
2026-05-20 09:42:00 +03:00
committed by GitHub
parent 871b0b70f8
commit 57ebaf4edd
4 changed files with 94 additions and 67 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-device.cpp
+7 -1
View File
@@ -1897,7 +1897,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad(ggml_metal_l
char base[256];
char name[256];
snprintf(base, 256, "kernel_pad_%s", ggml_type_name(op->src[0]->type));
// note: this is slower
//const bool is_c4 = op->src[0]->ne[0] % 4 == 0 && op->ne[0] % 4 == 0;
const bool is_c4 = false;
snprintf(base, 256, "kernel_pad_%s%s", ggml_type_name(op->src[0]->type), is_c4 ? "_4" : "");
snprintf(name, 256, "%s", base);
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
@@ -1907,6 +1911,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad(ggml_metal_l
res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
res.c4 = is_c4;
return res;
}