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

metal : restore im2col implementation for large kernels (#23901)

Browse Source
This commit is contained in:
Georgi Gerganov
2026-05-30 15:26:13 +03:00
committed by GitHub
parent e674b1279b
commit 2d9b7c8e98
4 changed files with 79 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+17 -7
View File
@@ -3635,16 +3635,26 @@ int ggml_metal_op_im2col(ggml_metal_op_t ctx, int idx) {
auto pipeline = ggml_metal_library_get_pipeline_im2col(lib, op);
GGML_ASSERT(KH*KW <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
if (KH*KW <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
const uint64_t ntptg0 = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)/(KH*KW), N);
const uint64_t ntptg0 = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)/(KH*KW), N);
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[1]), 1);
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 2);
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[1]), 1);
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 2);
ggml_metal_encoder_dispatch_threadgroups(enc, IC, OH, OW, ntptg0, KH, KW);
} else {
const uint64_t n_threads = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), N);
const int64_t quotient = N / n_threads + (N % n_threads > 0 ? 1 : 0);
ggml_metal_encoder_dispatch_threadgroups(enc, IC, OH, OW, ntptg0, KH, KW);
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[1]), 1);
ggml_metal_encoder_set_buffer (enc, ggml_metal_get_buffer_id(op), 2);
ggml_metal_encoder_dispatch_threadgroups(enc, quotient * CHW, OH, OW, n_threads, 1, 1);
}
return 1;
}