metal : reduce command encoding overhead

ggml-ci

Author: ggerganov

examples/llava/clip.cpp

@@ -2446,7 +2446,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
 
 #ifdef GGML_USE_METAL
     if (ggml_backend_is_metal(ctx->backend)) {
-        ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
+        ggml_backend_metal_set_n_cb(ctx->backend, 1);
     }
 #endif
 

examples/perf-metal/perf-metal.cpp

@@ -104,16 +104,18 @@ int main(int argc, char ** argv) {
         const int64_t t_end = ggml_time_us();
 
         // actual trace
-        if (n_thread == 4) {
-            ggml_backend_metal_capture_next_compute(backend);
-            ggml_backend_graph_compute(backend, gf);
-            ggml_backend_metal_capture_next_compute(backend);
-            ggml_backend_graph_compute(backend, gf);
-            ggml_backend_metal_capture_next_compute(backend);
-            ggml_backend_graph_compute(backend, gf);
-
-            printf("%s: trace dumped\n", __func__);
-        }
+        //if (n_thread == 4) {
+        //    ggml_backend_metal_capture_next_compute(backend);
+        //    ggml_backend_graph_compute(backend, gf);
+        //    //std::this_thread::sleep_for(std::chrono::milliseconds(1000)); // NOTE: these intervals do not appear in the XCode trace!
+        //    ggml_backend_metal_capture_next_compute(backend);
+        //    ggml_backend_graph_compute(backend, gf);
+        //    //std::this_thread::sleep_for(std::chrono::milliseconds(1000)); // NOTE: these intervals do not appear in the XCode trace!
+        //    ggml_backend_metal_capture_next_compute(backend);
+        //    ggml_backend_graph_compute(backend, gf);
+
+        //    printf("%s: trace dumped\n", __func__);
+        //}
 
         printf("%s: n_thread = %d, time = %f ms\n", __func__, n_thread, (t_end - t_start) / 1000.0 / n_iter);
     }

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;