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

metal : optimize Metal Tensor API usage for GGML_OP_MUL_MAT (#20962)

Browse Source 
* Optimize Metal Tensor API usage for matmul2d

Separates the Metal Tensor API (matmul2d) path in kernel_mul_mm into its own standalone kernel, gated by GGML_METAL_HAS_TENSOR.

The legacy simdgroup_matrix kernel is preserved under #else.

Previously both paths were interleaved via #ifdef blocks within a single kernel, forcing the tensor path to share the legacy kernel's data layout and threadgroup memory scheme. Splitting the kernel enabled memory and dispatch optimizations that weren't possible when the two paths shared code structure.

* cont : cleanup

* cont : cleanup

* cont : cleanup

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Developer-Ecosystem-Engineering
2026-04-25 15:14:28 +03:00
committed by GitHub
co-authored by Georgi Gerganov
parent 9d34231bb8
commit d1649047a3
6 changed files with 189 additions and 109 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-device.m
+11 -6
View File
@@ -95,8 +95,8 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
struct ggml_metal_library {
id<MTLLibrary> obj;
id<MTLDevice> device;
ggml_metal_device_t dev;
ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
NSLock * lock;
@@ -251,7 +251,7 @@ ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
res->obj = library;
res->device = device;
res->dev = dev;
res->pipelines = ggml_metal_pipelines_init();
res->lock = [NSLock new];
@@ -318,7 +318,7 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
}
res->obj = library;
res->device = device;
res->dev = dev;
res->pipelines = ggml_metal_pipelines_init();
res->lock = [NSLock new];
@@ -341,6 +341,10 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
free(lib);
}
ggml_metal_device_t ggml_metal_library_get_device(ggml_metal_library_t lib) {
return lib->dev;
}
struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_metal_library_t lib, const char * name) {
[lib->lock lock];
@@ -405,7 +409,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_
return res;
}
id<MTLComputePipelineState> obj = [lib->device newComputePipelineStateWithFunction:mtl_function error:&error];
id<MTLDevice> device = ggml_metal_device_get_obj(lib->dev);
id<MTLComputePipelineState> obj = [device newComputePipelineStateWithFunction:mtl_function error:&error];
[mtl_function release];
@@ -699,7 +704,7 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
" auto sB = tB.slice(0, 0); \n"
" mm.run(sB, sA, cT); \n"
" \n"
" auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(4, 4)); \n"
" auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(16, 16)); \n"
" \n"
" cT.store(tC); \n"
"}";
@@ -749,7 +754,7 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
" auto sB = tB.slice(0, 0); \n"
" mm.run(sB, sA, cT); \n"
" \n"
" auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(4, 4)); \n"
" auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(16, 16)); \n"
" \n"
" cT.store(tC); \n"
"}";