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

metal : avoid divisions in bin kernel (#20426)

Browse Source 
* metal : avoid modulus in bin kernel when not broadcasting

* metal : fix capture_started flag
This commit is contained in:
Georgi Gerganov
2026-03-12 09:42:40 +02:00
committed by GitHub
parent 4cc6eb158c
commit e4cff0956b
4 changed files with 14 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-device.cpp
+3 -1
View File
@@ -1470,10 +1470,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(ggml_metal_l
const bool is_c4 = (op->src[0]->ne[0] % 4 == 0) && (op->src[1]->ne[0] % 4 == 0);
const bool is_cb = op->src[0]->ne[0] != op->src[1]->ne[0];
const bool is_rb = ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]) && (ggml_nrows(op->src[1]) == 1) && ggml_nelements(op) < 65536;
snprintf(base, 256, "kernel_bin_fuse_%s_%s_%s%s", t0_str, t1_str, t_str, is_c4 ? "_4" : "");
snprintf(name, 256, "%s_op=%d_nf=%d_rb=%d", base, op_num, n_fuse, is_rb);
snprintf(name, 256, "%s_op=%d_nf=%d_rb=%d_cb=%d", base, op_num, n_fuse, is_rb, is_cb);
ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
if (!res.pipeline) {
@@ -1482,6 +1483,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin(ggml_metal_l
ggml_metal_cv_set_int16(cv, op_num, FC_BIN + 0);
ggml_metal_cv_set_int16(cv, n_fuse, FC_BIN + 1);
ggml_metal_cv_set_bool (cv, is_rb, FC_BIN + 2);
ggml_metal_cv_set_bool (cv, is_cb, FC_BIN + 3);
res = ggml_metal_library_compile_pipeline(lib, base, name, cv);