[mtl canvas] cull tiles queues outside viewport

examples/tiger/main.c

@@ -65,7 +65,7 @@ int main()
 
 	//NOTE: create surface
 	mg_surface surface = mg_surface_create_for_window(window, MG_BACKEND_DEFAULT);
-	mg_surface_swap_interval(surface, 1);
+	mg_surface_swap_interval(surface, 0);
 
 	//TODO: create canvas
 	mg_canvas canvas = mg_canvas_create(surface);

src/mtl_canvas.m

@@ -244,6 +244,7 @@ void mg_mtl_canvas_draw_batch(mg_canvas_backend* interface, mg_image_data* image
 		[shapeEncoder setBuffer: backend->tilesBuffer offset:0 atIndex: 2];
 		[shapeEncoder setBuffer: backend->tilesOffset offset:0 atIndex: 3];
 		[shapeEncoder setBytes: &scale length: sizeof(float) atIndex: 4];
+		[shapeEncoder setBytes: &viewportSize length: sizeof(vector_uint2) atIndex: 5];
 
 		MTLSize shapeGroupSize = MTLSizeMake(backend->shapePipeline.maxTotalThreadsPerThreadgroup, 1, 1);
 		MTLSize shapeGridSize = MTLSizeMake(shapeCount, 1, 1);

src/mtl_shader.h

@@ -65,12 +65,13 @@ typedef struct mg_triangle_data
 using namespace metal;
 #endif
 
-#define MG_TILE_CMD_MASK (1<<31)
+#define MG_TILE_CMD_MASK (3<<30)
 
 typedef enum mg_tile_cmd_kind
 {
 	mg_cmd_triangle = 0,
-	mg_cmd_color = 1<<31,
+	mg_cmd_color    = 1<<30,
+	mg_cmd_flip     = 2<<30
 } mg_tile_cmd_kind;
 
 typedef int mg_tile_cmd;

src/mtl_shader.metal

@@ -61,9 +61,10 @@ kernel void ShapeSetup(constant mg_shape* shapeBuffer [[buffer(0)]],
                        device mg_tile* tilesBuffer [[buffer(2)]],
                        device volatile atomic_uint* tilesOffset [[buffer(3)]],
                        constant float* scaling [[buffer(4)]],
+                       constant int2* viewport [[buffer(5)]],
                        uint gid [[thread_position_in_grid]])
 {
-
+	int2 tilesMatrixDim = (*viewport - 1) / RENDERER_TILE_SIZE + 1;
 	float4 box = shapeBuffer[gid].clip * scaling[0];
 
 	int2 firstTile = int2(box.xy)/RENDERER_TILE_SIZE;
@@ -72,8 +73,11 @@ kernel void ShapeSetup(constant mg_shape* shapeBuffer [[buffer(0)]],
 	//      any tile queue, the tileQueues pointer for that shape would alias the tileQueues pointer of another
 	//      shape, and we would have to detect that in the tiling and drawing kernels. Instead, just accept some
 	//      waste and keep the other kernels more uniforms for now...
-	int nTilesX = int(box.z)/RENDERER_TILE_SIZE - firstTile.x + 1;
+	//TODO limit to screen
-	int nTilesY = int(box.w)/RENDERER_TILE_SIZE - firstTile.y + 1;
+	int2 lastTile = max(firstTile, min(int2(box.zw)/RENDERER_TILE_SIZE, tilesMatrixDim));
+
+	int nTilesX = lastTile.x - firstTile.x + 1;
+	int nTilesY = lastTile.y - firstTile.y + 1;
 
 	int tileCount = nTilesX * nTilesY;
 
@@ -172,8 +176,8 @@ kernel void TriangleKernel(constant mg_vertex* vertexBuffer [[buffer(0)]],
 
 	int xMin = max(0, coarseBox.x - shapeQueue->area.x);
 	int yMin = max(0, coarseBox.y - shapeQueue->area.y);
-	int xMax = min(coarseBox.z - shapeQueue->area.x, shapeQueue->area.z);
+	int xMax = min(coarseBox.z - shapeQueue->area.x, shapeQueue->area.z-1);
-	int yMax = min(coarseBox.w - shapeQueue->area.y, shapeQueue->area.w);
+	int yMax = min(coarseBox.w - shapeQueue->area.y, shapeQueue->area.w-1);
 
 	//NOTE(martin): it's important to do the computation with signed int, so that we can have negative xMax/yMax
 	//              otherwise all triangles on the left or below the x/y axis are attributed to tiles on row/column 0.
@@ -233,6 +237,7 @@ kernel void TriangleKernel(constant mg_vertex* vertexBuffer [[buffer(0)]],
 
 				device mg_tile_elt* elt = &eltBuffer[eltIndex];
 				elt->triangleIndex = gid;
+
 				elt->next = atomic_exchange_explicit(&tile->firstElt, eltIndex, memory_order_relaxed);
 
 				atomic_fetch_add_explicit(&tile->eltCount, 1, memory_order_relaxed);
@@ -254,6 +259,7 @@ kernel void TriangleKernel(constant mg_vertex* vertexBuffer [[buffer(0)]],
 
 				if(allLeftFromEdge0 && allLeftFromEdge1 && allLeftFromEdge2 && triangleFull)
 				{
+					elt->triangleIndex |= mg_cmd_flip;
 					atomic_fetch_add_explicit(&tile->flipCount, 1, memory_order_relaxed);
 				}
 				else
@@ -310,7 +316,7 @@ kernel void GatherKernel(const device mg_shape_queue* shapeQueueBuffer [[buffer(
 						const device mg_tile_elt* elt = &eltBuffer[firstEltIndex];
 
 						count = 0;
-						tileArray[count] = mg_cmd_color | elt->triangleIndex;
+						tileArray[count] = mg_cmd_color | (elt->triangleIndex & ~MG_TILE_CMD_MASK);
 						count++;
 						continue;
 					}
@@ -458,7 +464,9 @@ kernel void RenderKernel(const device uint* tileCounters [[buffer(0)]],
 			currentShapeIndex = triangle->shapeIndex;
 		}
 
-		if(cmdKind == mg_cmd_color)
+		switch(cmdKind)
+		{
+			case mg_cmd_color:
 			{
 				for(int sampleIndex=0; sampleIndex<sampleCount; sampleIndex++)
 				{
@@ -467,8 +475,17 @@ kernel void RenderKernel(const device uint* tileCounters [[buffer(0)]],
 					sampleColor[sampleIndex] = nextColor;
 					flipCount[sampleIndex] = 0;
 				}
+			} break;
+
+			case mg_cmd_flip:
+			{
+				for(int sampleIndex=0; sampleIndex<sampleCount; sampleIndex++)
+				{
+					flipCount[sampleIndex]++;
 				}
-		else
+			} break;
+
+			case mg_cmd_triangle:
 			{
 				int2 p0 = triangle->p0;
 				int2 p1 = triangle->p1;
@@ -515,6 +532,7 @@ kernel void RenderKernel(const device uint* tileCounters [[buffer(0)]],
 						}
 					}
 				}
+			} break;
 		}
     }