feat: support macOS arm64

5 files changed, 193 insertions, 45 deletions

diff --git a/premake5.lua b/premake5.lua
index 98e5757..236347c 100644
--- a/premake5.lua
+++ b/premake5.lua
@@ -8,7 +8,7 @@ workspace "Donut"
 		"Release",
 		"Dist"
 	}
-	
+
 	flags
 	{
 		"MultiProcessorCompile"
@@ -20,6 +20,10 @@ workspace "Donut"
 		defines "DONUT_LINUX"
 	filter "system:macosx"
 		defines "DONUT_MACOS"
+		-- Build natively for Apple Silicon (was x86_64 under Rosetta). Native
+		-- arm64 is required for clean Metal work and better performance.
+		architecture "ARM64"
+	filter {}
 
 outputdir = "%{cfg.buildcfg}-%{cfg.system}/%{prj.name}"
 
@@ -55,7 +59,7 @@ project "GLAD"
     {
         "ext/glad/include"
     }
-    
+
     filter "system:windows"
         systemversion "latest"
 
@@ -101,7 +105,7 @@ project "GLFW"
 	filter "system:linux"
 		pic "On"
 		systemversion "latest"
-		
+
 		files
 		{
 			"ext/glfw/src/x11_init.c",
@@ -110,22 +114,22 @@ project "GLFW"
 			"ext/glfw/src/x11_window.c",
 			"ext/glfw/src/xkb_unicode.c",
 			"ext/glfw/src/xkb_unicode.h",
-			
+
 			"ext/glfw/src/wl_init.c",
 			"ext/glfw/src/wl_monitor.c",
 			"ext/glfw/src/wl_platform.h",
 			"ext/glfw/src/wl_window.c",
-			
+
 			"ext/glfw/src/posix_module.c",
 			"ext/glfw/src/posix_time.c",
 			"ext/glfw/src/posix_time.h",
 			"ext/glfw/src/posix_thread.c",
 			"ext/glfw/src/posix_thread.h",
-			
+
 			"ext/glfw/src/glx_context.c",
 			"ext/glfw/src/egl_context.c",
 			"ext/glfw/src/osmesa_context.c",
-			
+
 			"ext/glfw/src/linux_joystick.c",
 			"ext/glfw/src/linux_joystick.h"
 		}
@@ -148,11 +152,11 @@ project "GLFW"
 			"ext/glfw/src/cocoa_time.c",
 			"ext/glfw/src/cocoa_time.h",
 			"ext/glfw/src/cocoa_window.m",
-			
+
 			"ext/glfw/src/nsgl_context.m",
 			"ext/glfw/src/egl_context.c",
 			"ext/glfw/src/osmesa_context.c",
-			
+
 			"ext/glfw/src/posix_module.c",
 			"ext/glfw/src/posix_thread.c",
 			"ext/glfw/src/posix_thread.h"
@@ -184,8 +188,8 @@ project "GLFW"
 			"ext/glfw/src/osmesa_context.c"
 		}
 
-		defines 
-		{ 
+		defines
+		{
 			"_GLFW_WIN32",
 			"_CRT_SECURE_NO_WARNINGS",
 		}
@@ -194,7 +198,7 @@ project "GLFW"
 		runtime "Debug"
 		symbols "on"
 
-	filter { "system:windows", "configurations:Debug-AS" }	
+	filter { "system:windows", "configurations:Debug-AS" }
 		runtime "Debug"
 		symbols "on"
 		sanitize { "Address" }
@@ -213,10 +217,10 @@ project "ImGui"
 	kind "StaticLib"
 	language "C++"
 	staticruntime "on"
-	
+
 	targetdir ("bin/" .. outputdir)
 	objdir ("bin-int/" .. outputdir)
-	
+
 	files
 	{
 		"ext/imgui/imgui.cpp",
@@ -225,24 +229,24 @@ project "ImGui"
 		"ext/imgui/imgui_widgets.cpp",
 		"ext/imgui/imgui_demo.cpp"
 	}
-	
+
 	filter "system:windows"
 		systemversion "latest"
 		cppdialect "C++17"
-	
+
 	filter "system:linux"
 		pic "On"
 		systemversion "latest"
 		cppdialect "C++17"
-	
+
 	filter "configurations:Debug"
 		runtime "Debug"
 		symbols "on"
-	
+
 	filter "configurations:Release"
 		runtime "Release"
 		optimize "on"
-	
+
 	filter "configurations:Dist"
 		runtime "Release"
 		optimize "on"
@@ -252,39 +256,39 @@ project "ImGuizmo"
 	kind "StaticLib"
 	language "C++"
 	staticruntime "on"
-	
+
 	targetdir ("bin/" .. outputdir)
 	objdir ("bin-int/" .. outputdir)
-	
+
 	files
 	{
 		"ext/ImGuizmo/ImGuizmo.cpp",
 		"ext/ImGuizmo/ImGuizmo.h"
 	}
-	
+
 	includedirs
 	{
 		"%{IncludeDir.imgui}",
 		"%{IncludeDir.imguizmo}"
 	}
-	
+
 	filter "system:windows"
 		systemversion "latest"
 		cppdialect "C++17"
-	
+
 	filter "system:linux"
 		pic "On"
 		systemversion "latest"
 		cppdialect "C++17"
-	
+
 	filter "configurations:Debug"
 		runtime "Debug"
 		symbols "on"
-	
+
 	filter "configurations:Release"
 		runtime "Release"
 		optimize "on"
-	
+
 	filter "configurations:Dist"
 		runtime "Release"
 		optimize "on"
@@ -318,10 +322,10 @@ project "Donut"
 		"src/rendering/framebuffer.cpp",
 		"src/platform/opengl/opengl_framebuffer.h",
 		"src/platform/opengl/opengl_framebuffer.cpp",
-		
+
 		"ext/imgui/backends/imgui_impl_glfw.cpp",
 		"ext/imgui/backends/imgui_impl_opengl3.cpp",
-		
+
 		"assets/fonts/inter/static/Inter_18pt-Regular.ttf",
 		"assets/fonts/inter/static/Inter_18pt-Bold.ttf",
 		"assets/fonts/inter/static/Inter_18pt-Light.ttf"
@@ -330,7 +334,7 @@ project "Donut"
 	includedirs
 	{
 		"src",
-		
+
 		"%{IncludeDir.glm}",
 		"%{IncludeDir.glfw}",
 		"%{IncludeDir.glad}",
@@ -359,12 +363,20 @@ project "Donut"
 		}
 
 	filter "system:macosx"
+		-- Objective-C++ (Metal) sources compile only on macOS.
+		files
+		{
+			"src/**.mm"
+		}
+
 		links
 		{
 			"Cocoa.framework",
 			"IOKit.framework",
 			"CoreFoundation.framework",
-			"QuartzCore.framework"
+			"QuartzCore.framework",
+			"Metal.framework",
+			"Foundation.framework"
 		}
 
 	filter "configurations:Debug"
diff --git a/src/core/application.cpp b/src/core/application.cpp
index 37d7fa0..fa5fd6e 100644
--- a/src/core/application.cpp
+++ b/src/core/application.cpp
@@ -10,6 +10,10 @@
 #include <GLFW/glfw3.h>
 #include <ImGuizmo.h>
 
+#ifdef __APPLE__
+#include "platform/metal/metal_context.h"
+#endif
+
 namespace Donut
 {
     Application* Application::s_instance = nullptr;
@@ -103,9 +107,14 @@ namespace Donut
         m_state_manager->on_event(event);
     }
 
-    void Application::on_init()     
-    { 
+    void Application::on_init()
+    {
         Logger::init();
+
+#ifdef __APPLE__
+        metal_probe();
+        metal_compute_self_test();
+#endif
         
         SettingsManager::initialize();
         
diff --git a/src/engine/engine.h b/src/engine/engine.h
index 0a50612..1c25466 100644
--- a/src/engine/engine.h
+++ b/src/engine/engine.h
@@ -35,7 +35,7 @@ namespace Donut
         double m_radius;
         double m_rs;
 
-        BlackHole(glm::vec3 pos, float mass) 
+        BlackHole(glm::vec3 pos, float mass)
             : m_position(pos), m_mass(mass)
         {
             m_rs = 2.0 * G * m_mass / (c * c);
@@ -99,33 +99,33 @@ namespace Donut
         int   get_compute_height()     const { return m_compute_height;                        }
         int   get_compute_width()      const { return (m_width * m_compute_height) / m_height; }
         void  update_compute_dimensions();
-        
+
         int   get_max_steps_moving()    const { return m_max_steps_moving; }
         int   get_max_steps_static()    const { return m_max_steps_static; }
         float get_early_exit_distance() const { return m_early_exit_distance; }
         void  set_max_steps_moving(int steps) { m_max_steps_moving = steps; }
         void  set_max_steps_static(int steps) { m_max_steps_static = steps; }
         void  set_early_exit_distance(float distance) { m_early_exit_distance = distance; }
-        
+
         float get_disk_thickness()         const  { return m_disk_thickness;      }
         void  set_disk_thickness(float thickness) { m_disk_thickness = thickness; }
-        
+
         float get_disk_density()            const { return m_disk_density;    }
         void  set_disk_density(float density)     { m_disk_density = density; }
-        
+
         float get_rotation_speed()          const { return m_rotation_speed;  }
         void  set_rotation_speed(float speed)     { m_rotation_speed = speed; }
-        
+
         float get_blur_strength()           const { return m_blur_strength;     }
         void  set_blur_strength(float strength)   { m_blur_strength = strength; }
-        
+
         float get_glow_intensity()          const { return m_glow_intensity;      }
         void  set_glow_intensity(float intensity) { m_glow_intensity = intensity; }
-        
+
         void load_objects_from_scene(const std::vector<Donut::Object>& objects);
         void export_high_res_frame(const std::string& filename, int width = 4096, int height = 3072);
         void print_object_info() const;
-        
+
         void set_hdri_environment(Ref<CubemapTexture> hdri) { m_hdri_environment = hdri; }
         Ref<CubemapTexture> get_hdri_environment()    const { return m_hdri_environment; }
     private:
@@ -161,12 +161,12 @@ namespace Donut
         BlackHole               m_sag_a;
         Camera                  m_camera;
         bool                    m_gravity = false;
-        
+
         int   m_max_steps_moving    = 60000;
         int   m_max_steps_static    = 30000;
         float m_early_exit_distance = 5.0e11f;
-        
-        float m_disk_thickness = 0.1f;    
+
+        float m_disk_thickness = 0.1f;
         float m_disk_density   = 0.1f;
         float m_rotation_speed = 1.0f;
         float m_blur_strength  = 2.0f;
diff --git a/src/platform/metal/metal_context.h b/src/platform/metal/metal_context.h
new file mode 100644
index 0000000..66e8ae2
--- /dev/null
+++ b/src/platform/metal/metal_context.h
@@ -0,0 +1,8 @@
+#pragma once
+
+namespace Donut
+{
+    bool metal_probe();
+
+    bool metal_compute_self_test();
+}
diff --git a/src/platform/metal/metal_context.mm b/src/platform/metal/metal_context.mm
new file mode 100644
index 0000000..a616865
--- /dev/null
+++ b/src/platform/metal/metal_context.mm
@@ -0,0 +1,119 @@
+#import <Metal/Metal.h>
+#import <Foundation/Foundation.h>
+
+#include "metal_context.h"
+#include "core/log.h"
+
+#include <vector>
+
+namespace Donut
+{
+    bool metal_probe()
+    {
+        @autoreleasepool
+        {
+            id<MTLDevice> device = MTLCreateSystemDefaultDevice();
+            if (device == nil)
+            {
+                DONUT_ERROR("Metal: no default device available");
+                return false;
+            }
+
+            id<MTLCommandQueue> queue = [device newCommandQueue];
+            const char* name = [[device name] UTF8String];
+
+            DONUT_INFO("Metal device: {}", name ? name : "(unknown)");
+            DONUT_INFO("Metal: unified memory = {}, max threads/threadgroup = {}",
+                       device.hasUnifiedMemory ? "yes" : "no",
+                       (unsigned long)device.maxThreadsPerThreadgroup.width);
+
+            if (queue == nil)
+            {
+                DONUT_WARN("Metal: failed to create command queue");
+                return false;
+            }
+
+            return true;
+        }
+    }
+
+    bool metal_compute_self_test()
+    {
+        @autoreleasepool
+        {
+            id<MTLDevice> device = MTLCreateSystemDefaultDevice();
+            id<MTLCommandQueue> queue = [device newCommandQueue];
+            if (device == nil || queue == nil)
+                return false;
+
+            NSString* src =
+                @"#include <metal_stdlib>\n"
+                 "using namespace metal;\n"
+                 "kernel void selfTest(texture2d<float, access::write> outTex [[texture(0)]],\n"
+                 "                     uint2 gid [[thread_position_in_grid]])\n"
+                 "{\n"
+                 "    uint w = outTex.get_width();\n"
+                 "    uint h = outTex.get_height();\n"
+                 "    if (gid.x >= w || gid.y >= h) return;\n"
+                 "    outTex.write(float4(float(gid.x) / float(w - 1),\n"
+                 "                        float(gid.y) / float(h - 1), 0.5, 1.0), gid);\n"
+                 "}\n";
+
+            NSError* err = nil;
+            id<MTLLibrary> lib = [device newLibraryWithSource:src options:nil error:&err];
+            if (lib == nil)
+            {
+                DONUT_ERROR("Metal self-test: kernel compile failed: {}",
+                            err ? [[err localizedDescription] UTF8String] : "unknown");
+                return false;
+            }
+
+            id<MTLFunction> fn = [lib newFunctionWithName:@"selfTest"];
+            id<MTLComputePipelineState> pipeline =
+                [device newComputePipelineStateWithFunction:fn error:&err];
+            if (pipeline == nil)
+            {
+                DONUT_ERROR("Metal self-test: pipeline creation failed");
+                return false;
+            }
+
+            const uint32_t W = 64, H = 64;
+            MTLTextureDescriptor* desc =
+                [MTLTextureDescriptor texture2DDescriptorWithPixelFormat:MTLPixelFormatRGBA8Unorm
+                                                                   width:W
+                                                                  height:H
+                                                               mipmapped:NO];
+            desc.usage = MTLTextureUsageShaderWrite | MTLTextureUsageShaderRead;
+            desc.storageMode = MTLStorageModeShared;
+            id<MTLTexture> tex = [device newTextureWithDescriptor:desc];
+
+            id<MTLCommandBuffer> cb = [queue commandBuffer];
+            id<MTLComputeCommandEncoder> enc = [cb computeCommandEncoder];
+            [enc setComputePipelineState:pipeline];
+            [enc setTexture:tex atIndex:0];
+
+            MTLSize tg = MTLSizeMake(16, 16, 1);
+            MTLSize grid = MTLSizeMake(W, H, 1);
+            [enc dispatchThreads:grid threadsPerThreadgroup:tg];
+            [enc endEncoding];
+            [cb commit];
+            [cb waitUntilCompleted];
+
+            std::vector<uint8_t> px(W * H * 4);
+            [tex getBytes:px.data()
+              bytesPerRow:W * 4
+               fromRegion:MTLRegionMake2D(0, 0, W, H)
+              mipmapLevel:0];
+
+            size_t corner = ((size_t)(H - 1) * W + (W - 1)) * 4;
+            DONUT_INFO("Metal compute self-test: corner pixel RGBA = ({}, {}, {}, {})",
+                       (int)px[corner + 0], (int)px[corner + 1],
+                       (int)px[corner + 2], (int)px[corner + 3]);
+
+            bool ok = px[corner + 0] > 250 && px[corner + 1] > 250 &&
+                      px[corner + 3] == 255;
+            DONUT_INFO("Metal compute self-test: {}", ok ? "PASS" : "FAIL");
+            return ok;
+        }
+    }
+}