Add profiler/metal to profile gpu on macos

Author: junjihashimoto

examples/hello_world/Makefile

@@ -9,7 +9,7 @@ ifeq ($(shell $(CXX) -std=c++17 -x c++ -E -include array - < /dev/null > /dev/nu
 else
     STDLIB := -stdlib=libc++
 endif
-FLAGS=-std=c++17 $(STDLIB) -I$(GPUCPP) -I$(GPUCPP)/third_party/headers -L$(GPUCPP)/third_party/lib run.cpp -ldl -ldawn
+FLAGS=-std=c++17 $(STDLIB) -I$(GPUCPP) -I$(GPUCPP)/third_party/headers -L$(GPUCPP)/third_party/lib -ldl -ldawn
 
 run: ./build/$(TARGET) dawnlib
 	$(LIBSPEC) && ./build/$(TARGET)
@@ -20,7 +20,7 @@ run_setup: check-python
 	cd $(GPUCPP) && python3 setup.py
 
 build/$(TARGET): run.cpp
-	mkdir -p build && $(CXX) $(FLAGS) -DNO_LOG -o ./build/$(TARGET)
+	mkdir -p build && $(CXX) $(FLAGS) -DNO_LOG -o $@ $<
 
 debug: run.cpp
 	mkdir -p build && $(CXX) $(FLAGS) -g -o ./build/$(TARGET)

examples/matmul/Makefile

@@ -15,11 +15,17 @@ FLAGS=-std=c++17 $(STDLIB) -I$(GPUCPP) -I$(GPUCPP)/third_party/headers -L$(GPUCP
 run: ./build/$(TARGET)
 	$(LIBSPEC) && ./build/$(TARGET)
 
+run_with_profile: ./build/$(TARGET)_with_profile
+	$(LIBSPEC) && export METAL_CAPTURE_ENABLED=1 && ./build/$(TARGET)_with_profile
+
 # Use clang -v to see the include paths
 # Note in this example optimization is turned on
 build/$(TARGET): run.cpp
 	mkdir -p build && $(CXX) $(FLAGS) -o ./build/$(TARGET)
 
+build/$(TARGET)_with_profile: run.cpp
+	mkdir -p build && $(CXX) $(FLAGS) -o ./build/$(TARGET)_with_profile  $(GPUCPP)/experimental/profiler/metal.mm -framework metal -framework Foundation -DMETAL_PROFILER
+
 watch: 
 	@command -v entr >/dev/null 2>&1 || { echo >&2 "Please install entr with 'brew install entr' or 'sudo apt-get install entr'"; exit 1; }
 	mkdir -p build && $(CODEPATH) | entr -s "$(LIBSPEC) && rm -f ./build/$(TARGET) && make -j$(NUM_JOBS) ./build/$(TARGET) && ./build/$(TARGET)"

examples/matmul/run.cpp

@@ -13,6 +13,10 @@
 #include "experimental/wgsl.h"      // loopUnrolling
 #include "numeric_types/half.hpp"
 
+#ifdef METAL_PROFILER
+#include "experimental/profiler/metal.hpp"
+#endif
+
 using namespace gpu;
 
 const std::string versionToStr(int version);
@@ -815,8 +819,10 @@ void runTest(int version, size_t M, size_t K, size_t N,
     kernels[i] = selectMatmul(ctx, version, {input, weights, outputs[i]}, M, K, N, numtype);
   }
 
+#ifndef METAL_PROFILER
   printf("[ Press enter to start tests ... ]\n");
   getchar();
+#endif
   LOG(kDefLog, kInfo, "Dispatching Kernel version %d: %s, %d iterations ...",
       version, versionToStr(version).c_str(), nIter);
 
@@ -930,11 +936,17 @@ int main() {
     N = 2 * 4096;
   }
 
+#ifdef METAL_PROFILER
+  startCapture();
+#endif
   if (enableF16) {
     runTestWithCheck<half>(version, M, K, N, transposedInput, kTestSize, numtype);
   } else {
     runTestWithCheck<float>(version, M, K, N, transposedInput, kTestSize, numtype);
   }
+#ifdef METAL_PROFILER
+  stopCapture();
+#endif
 
   LOG(kDefLog, kInfo, "Done.");
   return 0;

experimental/profiler/metal.hpp

@@ -0,0 +1,6 @@
+#ifdef __APPLE__
+extern "C" {
+  void startCapture();
+  void stopCapture();
+}
+#endif

experimental/profiler/metal.mm

@@ -0,0 +1,46 @@
+#import <Foundation/Foundation.h>
+#import <Metal/Metal.h>
+#import <QuartzCore/CAMetalLayer.h>
+
+
+extern "C" {
+  void startCapture() {
+    if (![[NSProcessInfo processInfo].environment[@"METAL_CAPTURE_ENABLED"] boolValue]) {
+      NSLog(@"METAL_CAPTURE_ENABLED is not set. Please set it to 1 to enable Metal capture.");
+      return;
+    }
+    
+    MTLCaptureDescriptor *descriptor = [[MTLCaptureDescriptor alloc] init];
+    descriptor.destination = MTLCaptureDestinationGPUTraceDocument;
+    descriptor.outputURL = [NSURL fileURLWithPath:@"gpu.cpp.gputrace"];
+
+    NSFileManager *fileManager = [NSFileManager defaultManager];
+    if ([fileManager fileExistsAtPath:@"gpu.cpp.gputrace"]) {
+      NSError *error = nil;
+      [fileManager removeItemAtPath:@"gpu.cpp.gputrace" error:&error];
+      if (error) {
+        NSLog(@"Error deleting existing gpu.cpp.gputrace directory: %@", error);
+        return;
+      } else {
+        NSLog(@"Deleted existing gpu.cpp.gputrace directory.");
+      }
+    }
+
+    NSError *error = nil;
+    id<MTLDevice> device = MTLCreateSystemDefaultDevice();
+    if (!device) {
+      NSLog(@"MTLCreateSystemDefaultDevice returned nil. Metal may not be supported on this system.");
+      return;
+    }
+    descriptor.captureObject = device;
+    
+    BOOL success = [MTLCaptureManager.sharedCaptureManager startCaptureWithDescriptor:descriptor error:&error];
+    if (!success) {
+        NSLog(@" error capturing mtl => %@ ", [error localizedDescription] );
+    }
+  }
+
+  void stopCapture() {
+    [MTLCaptureManager.sharedCaptureManager stopCapture];
+  }
+}