metal : reduce command encoding overhead

ggml-ci
ggerganov · Oct 1, 2024 · 43b9d69 · 43b9d69
1 parent 1927378
commit 43b9d69
Show file tree

Hide file tree

Showing 5 changed files with 1,890 additions and 1,805 deletions.
diff --git a/examples/cvector-generator/pca.hpp b/examples/cvector-generator/pca.hpp
@@ -204,13 +204,6 @@ static ggml_status compute_piter(
         ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
     }
 
-// TODO: enable GPU support when support for GGML_OP_SQRT is added
-//#ifdef GGML_USE_METAL
-//    if (ggml_backend_is_metal(model.backend)) {
-//        ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
-//    }
-//#endif
-
     ggml_status res = ggml_backend_graph_compute(model.backend, gf);
     if (res == GGML_STATUS_SUCCESS) {
         auto extract_i = [](std::string prefix, std::string str) -> int {

diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
@@ -2444,12 +2444,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
     }
 
-#ifdef GGML_USE_METAL
-    if (ggml_backend_is_metal(ctx->backend)) {
-        ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
-    }
-#endif
-
     ggml_backend_graph_compute(ctx->backend, gf);
 
     // the last node is the embedding tensor

diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h
@@ -25,9 +25,6 @@
 #include <stddef.h>
 #include <stdbool.h>
 
-// max memory buffers that can be mapped to the device
-#define GGML_METAL_MAX_BUFFERS 64
-
 struct ggml_tensor;
 struct ggml_cgraph;
 
@@ -48,8 +45,6 @@ GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
 
-GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
 GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
 GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);