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

metal : extend l2_norm support for non-cont src0 (#19502)

Browse Source
This commit is contained in:
Georgi Gerganov
2026-02-11 14:53:19 +02:00
committed by GitHub
parent ada90bf2ba
commit 9ab072ebbe
5 changed files with 75 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+33 -13
View File
@@ -2979,39 +2979,59 @@ int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
float eps;
memcpy(&eps, op->op_params, sizeof(float));
int nth = 32; // SIMD width
ggml_metal_kargs_l2_norm args = {
/*.ne00 =*/ ne00,
/*.ne00_4 =*/ ne00/4,
/*.nb01 =*/ nb01,
/*.eps =*/ eps,
/*.ne00 =*/ ne00,
/*.ne01 =*/ ne01,
/*.ne02 =*/ ne02,
/*.ne03 =*/ ne03,
/*.nb00 =*/ nb00,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.ne0 =*/ ne0,
/*.ne1 =*/ ne1,
/*.ne2 =*/ ne2,
/*.ne3 =*/ ne3,
/*.nb0 =*/ nb0,
/*.nb1 =*/ nb1,
/*.nb2 =*/ nb2,
/*.nb3 =*/ nb3,
/*.eps =*/ eps,
};
auto pipeline = ggml_metal_library_get_pipeline_l2_norm(lib, op);
while (nth < ne00/4 && nth < ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
if (pipeline.c4) {
args.ne00 = ne00/4;
args.ne0 = ne0/4;
}
int nth = 32; // SIMD width
while (nth < ne00 && 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, ne00/4);
const size_t smem = pipeline.smem;
const int64_t nrows = ggml_nrows(op->src[0]);
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_set_buffer (enc, bid_src0, 1);
ggml_metal_encoder_set_buffer (enc, bid_dst, 2);
ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, nth, 1, 1);
ggml_metal_encoder_dispatch_threadgroups(enc, ne01, ne02, ne03, nth, 1, 1);
return 1;
}