Merge pull request #7176 from stuartcarnie/sgc-metal

Metal: Add support for screen shots

gfx/common/metal/Context.h

@@ -34,6 +34,9 @@ typedef struct
 /*! @brief Specifies whether rendering is synchronized with the display */
 @property (nonatomic, readwrite) bool displaySyncEnabled;
 
+/*! @brief captureEnabled allows previous frames to be read */
+@property (nonatomic, readwrite) bool captureEnabled;
+
 /*! @brief Returns the command buffer used for pre-render work,
  * such as mip maps and shader effects
  * */
@@ -70,4 +73,6 @@ typedef struct
 /*! @brief end commits the command buffer */
 - (void)end;
 
+- (bool)readBackBuffer:(uint8_t *)buffer;
+
 @end

gfx/common/metal/Context.m

@@ -54,6 +54,9 @@
    id<MTLRenderPipelineState> _states[GFX_MAX_SHADERS][2];
    id<MTLRenderPipelineState> _clearState;
    Uniforms _uniforms;
+   
+   bool _captureEnabled;
+   id<MTLTexture> _backBuffer;
 }
 
 - (instancetype)initWithDevice:(id<MTLDevice>)d
@@ -470,11 +473,65 @@
    [_chain[_currentChain] discard];
 }
 
+- (void)setCaptureEnabled:(bool)captureEnabled
+{
+   if (_captureEnabled == captureEnabled)
+      return;
+   
+   _captureEnabled = captureEnabled;
+   //_layer.framebufferOnly = !captureEnabled;
+}
+
+- (bool)captureEnabled
+{
+   return _captureEnabled;
+}
+
+- (bool)readBackBuffer:(uint8_t *)buffer
+{
+   if (!_captureEnabled || _backBuffer == nil)
+      return NO;
+   
+   if (_backBuffer.pixelFormat != MTLPixelFormatBGRA8Unorm)
+   {
+      RARCH_WARN("[Metal]: unexpected pixel format %d\n", _backBuffer.pixelFormat);
+      return NO;
+   }
+   
+   uint8_t *tmp = malloc(_backBuffer.width * _backBuffer.height * 4);
+   
+   [_backBuffer getBytes:tmp
+             bytesPerRow:4 * _backBuffer.width
+              fromRegion:MTLRegionMake2D(0, 0, _backBuffer.width, _backBuffer.height)
+             mipmapLevel:0];
+   
+   NSUInteger srcStride = _backBuffer.width * 4;
+   uint8_t const *src = tmp + (_viewport.y * srcStride);
+   
+   NSUInteger dstStride = _viewport.width * 3;
+   uint8_t *dst = buffer + (_viewport.height - 1) * dstStride;
+   
+   for (int y = _viewport.y; y < _viewport.height; y++, src += srcStride, dst -= dstStride)
+   {
+      for (int x = _viewport.x; x < _viewport.width; x++)
+      {
+         dst[3 * x + 0] = src[4 * x + 0];
+         dst[3 * x + 1] = src[4 * x + 1];
+         dst[3 * x + 2] = src[4 * x + 2];
+      }
+   }
+   
+   free(tmp);
+   
+   return YES;
+}
+
 - (void)begin
 {
    assert(_commandBuffer == nil);
    dispatch_semaphore_wait(_inflightSemaphore, DISPATCH_TIME_FOREVER);
    _commandBuffer = [_commandQueue commandBuffer];
+   _backBuffer = nil;
 }
 
 - (id<MTLRenderCommandEncoder>)rce
@@ -486,6 +543,10 @@
       rpd.colorAttachments[0].clearColor = _clearColor;
       rpd.colorAttachments[0].loadAction = MTLLoadActionClear;
       rpd.colorAttachments[0].texture = self.nextDrawable.texture;
+      if (_captureEnabled)
+      {
+         _backBuffer = self.nextDrawable.texture;
+      }
       _rce = [_commandBuffer renderCommandEncoderWithDescriptor:rpd];
    }
    return _rce;
@@ -643,7 +704,7 @@ static const NSUInteger kConstantAlignment = 4;
 - (bool)allocRange:(BufferRange *)range length:(NSUInteger)length
 {
    bzero(range, sizeof(*range));
-   
+
 #if TARGET_OS_OSX
    MTLResourceOptions opts = MTLResourceStorageModeManaged;
 #else

gfx/common/metal/ShaderTypes.h

@@ -55,6 +55,12 @@ typedef struct
    vector_float2 texCoord METAL_ATTRIBUTE(VertexAttributeTexcoord);
 } Vertex;
 
+typedef struct
+{
+   vector_float4 position;
+   vector_float2 texCoord;
+} VertexSlang;
+
 typedef struct
 {
    vector_float4 position METAL_POSITION;

gfx/common/metal_common.h

@@ -40,6 +40,7 @@ extern MTLPixelFormat SelectOptimalPixelFormat(MTLPixelFormat fmt);
 - (void)setFilteringIndex:(int)index smooth:(bool)smooth;
 - (BOOL)setShaderFromPath:(NSString *)path;
 - (void)updateFrame:(void const *)src pitch:(NSUInteger)pitch;
+- (bool)readViewport:(uint8_t *)buffer isIdle:(bool)isIdle;
 
 @end
 

gfx/common/metal_common.m

@@ -546,6 +546,7 @@ typedef struct MTLALIGN(16)
    Context *_context;
    id<MTLTexture> _texture; // final render texture
    Vertex _v[4];
+   VertexSlang _vertex[4];
    CGSize _size; // size of view in pixels
    CGRect _frame;
    NSUInteger _bpp;
@@ -587,6 +588,15 @@ typedef struct MTLALIGN(16)
       self.size = d.size;
       self.frame = CGRectMake(0, 0, 1, 1);
       resize_render_targets = YES;
+   
+      // init slang vertex buffer
+      VertexSlang v[4] = {
+         {simd_make_float4(0, 1, 0, 1), simd_make_float2(0, 1)},
+         {simd_make_float4(1, 1, 0, 1), simd_make_float2(1, 1)},
+         {simd_make_float4(0, 0, 0, 1), simd_make_float2(0, 0)},
+         {simd_make_float4(1, 0, 0, 1), simd_make_float2(1, 0)},
+      };
+      memcpy(_vertex, v, sizeof(_vertex));
    }
    return self;
 }
@@ -750,6 +760,24 @@ typedef struct MTLALIGN(16)
    }
 }
 
+- (bool)readViewport:(uint8_t *)buffer isIdle:(bool)isIdle
+{
+   RARCH_LOG("[Metal]: readViewport is_idle = %s\n", isIdle ? "YES" : "NO");
+   
+   bool enabled = _context.captureEnabled;
+   if (!enabled)
+      _context.captureEnabled = YES;
+   
+   video_driver_cached_frame();
+   
+   bool res = [_context readBackBuffer:buffer];
+   
+   if (!enabled)
+      _context.captureEnabled = NO;
+   
+   return res;
+}
+
 - (void)updateFrame:(void const *)src pitch:(NSUInteger)pitch
 {
    if (_shader && (_engine.frame.output_size.x != _viewport->width ||
@@ -816,19 +844,6 @@ typedef struct MTLALIGN(16)
    init_history = NO;
 }
 
-typedef struct vertex
-{
-   simd_float4 pos;
-   simd_float2 tex;
-} vertex_t;
-
-static vertex_t vertex_bytes[] = {
-   {{0, 1, 0, 1}, {0, 1}},
-   {{1, 1, 0, 1}, {1, 1}},
-   {{0, 0, 0, 1}, {0, 0}},
-   {{1, 0, 0, 1}, {1, 0}},
-};
-
 - (void)drawWithEncoder:(id<MTLRenderCommandEncoder>)rce
 {
    if (_texture)
@@ -942,7 +957,7 @@ static vertex_t vertex_bytes[] = {
       
       [rce setFragmentTextures:textures withRange:NSMakeRange(0, SLANG_NUM_BINDINGS)];
       [rce setFragmentSamplerStates:samplers withRange:NSMakeRange(0, SLANG_NUM_BINDINGS)];
-      [rce setVertexBytes:vertex_bytes length:sizeof(vertex_bytes) atIndex:4];
+      [rce setVertexBytes:_vertex length:sizeof(_vertex) atIndex:4];
       [rce drawPrimitives:MTLPrimitiveTypeTriangleStrip vertexStart:0 vertexCount:4];
       
       if (!backBuffer)
@@ -1162,13 +1177,13 @@ static vertex_t vertex_bytes[] = {
          @try
          {
             MTLVertexDescriptor *vd = [MTLVertexDescriptor new];
-            vd.attributes[0].offset = offsetof(vertex_t, pos);
+            vd.attributes[0].offset = offsetof(VertexSlang, position);
             vd.attributes[0].format = MTLVertexFormatFloat4;
             vd.attributes[0].bufferIndex = 4;
-            vd.attributes[1].offset = offsetof(vertex_t, tex);
+            vd.attributes[1].offset = offsetof(VertexSlang, texCoord);
             vd.attributes[1].format = MTLVertexFormatFloat2;
             vd.attributes[1].bufferIndex = 4;
-            vd.layouts[4].stride = sizeof(vertex_t);
+            vd.layouts[4].stride = sizeof(VertexSlang);
             vd.layouts[4].stepFunction = MTLVertexStepFunctionPerVertex;
             
             MTLRenderPipelineDescriptor *psd = [MTLRenderPipelineDescriptor new];

gfx/drivers/metal.m

@@ -163,7 +163,8 @@ static void metal_viewport_info(void *data, struct video_viewport *vp)
 
 static bool metal_read_viewport(void *data, uint8_t *buffer, bool is_idle)
 {
-   return true;
+   MetalDriver *md = (__bridge MetalDriver *)data;
+   return [md.frameView readViewport:buffer isIdle:is_idle];
 }
 
 static uintptr_t metal_load_texture(void *video_data, void *data,