[gl canvas] balance dispatch of raster shader along 2 dimensions to avoid hitting the max workgroup count per dimension

build.bat

@@ -4,7 +4,7 @@ setlocal EnableDelayedExpansion
 
 if not exist bin mkdir bin
 
-set glsl_shaders=src\glsl_shaders\common.glsl src\glsl_shaders\blit_vertex.glsl src\glsl_shaders\blit_fragment.glsl src\glsl_shaders\path_setup.glsl src\glsl_shaders\segment_setup.glsl src\glsl_shaders\backprop.glsl src\glsl_shaders\merge.glsl src\glsl_shaders\raster.glsl
+set glsl_shaders=src\glsl_shaders\common.glsl src\glsl_shaders\blit_vertex.glsl src\glsl_shaders\blit_fragment.glsl src\glsl_shaders\path_setup.glsl src\glsl_shaders\segment_setup.glsl src\glsl_shaders\backprop.glsl src\glsl_shaders\merge.glsl src\glsl_shaders\raster.glsl src\glsl_shaders\balance_workgroups.glsl
 
 call python3 scripts\embed_text.py %glsl_shaders% --prefix=glsl_ --output src\glsl_shaders.h
 

src/gl_canvas.c

@@ -149,6 +149,7 @@ typedef struct mg_gl_canvas_backend
 	GLuint segmentSetup;
 	GLuint backprop;
 	GLuint merge;
+	GLuint balanceWorkgroups;
 	GLuint raster;
 	GLuint blit;
 
@@ -167,6 +168,7 @@ typedef struct mg_gl_canvas_backend
 	GLuint tileOpBuffer;
 	GLuint tileOpCountBuffer;
 	GLuint screenTilesBuffer;
+	GLuint screenTilesCountBuffer;
 	GLuint rasterDispatchBuffer;
 	GLuint dummyVertexBuffer;
 
@@ -1095,6 +1097,9 @@ void mg_gl_render_batch(mg_gl_canvas_backend* backend,
 	glBindBuffer(GL_SHADER_STORAGE_BUFFER, backend->rasterDispatchBuffer);
 	glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(mg_gl_dispatch_indirect_command), &zero, GL_DYNAMIC_COPY);
 
+	glBindBuffer(GL_SHADER_STORAGE_BUFFER, backend->screenTilesCountBuffer);
+	glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(int), &zero, GL_DYNAMIC_COPY);
+
 	glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
 
 	int err = glGetError();
@@ -1210,7 +1215,7 @@ void mg_gl_render_batch(mg_gl_canvas_backend* backend,
 	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, backend->tileOpCountBuffer);
 	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 4, backend->tileOpBuffer);
 	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 5, backend->screenTilesBuffer);
-	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 6, backend->rasterDispatchBuffer);
+	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 6, backend->screenTilesCountBuffer);
 
 	glUniform1i(0, tileSize);
 	glUniform1f(1, scale);
@@ -1239,6 +1244,17 @@ void mg_gl_render_batch(mg_gl_canvas_backend* backend,
 			log_error("gl error %i\n", err);
 		}
 	}
+
+	//NOTE: balance work groups
+	glUseProgram(backend->balanceWorkgroups);
+
+	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 0, backend->screenTilesCountBuffer);
+	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, backend->rasterDispatchBuffer);
+	glUniform1ui(0, maxWorkGroupCount);
+
+	glDispatchCompute(1, 1, 1);
+	glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
+
 	//NOTE: raster pass
 	glUseProgram(backend->raster);
 
@@ -1246,6 +1262,7 @@ void mg_gl_render_batch(mg_gl_canvas_backend* backend,
 	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 1, backend->segmentBuffer);
 	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 2, backend->tileOpBuffer);
 	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 3, backend->screenTilesBuffer);
+	glBindBufferBase(GL_SHADER_STORAGE_BUFFER, 4, backend->screenTilesCountBuffer);
 
 	glUniform1f(0, scale);
 	glUniform1i(1, backend->msaaCount);
@@ -1265,6 +1282,7 @@ void mg_gl_render_batch(mg_gl_canvas_backend* backend,
 	}
 
 	glUniform1i(3, backend->pathBatchStart);
+	glUniform1ui(4, maxWorkGroupCount);
 
 	glBindBuffer(GL_DISPATCH_INDIRECT_BUFFER, backend->rasterDispatchBuffer);
 	glDispatchComputeIndirect(0);
@@ -1659,6 +1677,7 @@ mg_canvas_backend* gl_canvas_backend_create(mg_wgl_surface* surface)
 		err |= mg_gl_canvas_compile_compute_program(glsl_segment_setup, &backend->segmentSetup);
 		err |= mg_gl_canvas_compile_compute_program(glsl_backprop, &backend->backprop);
 		err |= mg_gl_canvas_compile_compute_program(glsl_merge, &backend->merge);
+		err |= mg_gl_canvas_compile_compute_program(glsl_balance_workgroups, &backend->balanceWorkgroups);
 		err |= mg_gl_canvas_compile_compute_program(glsl_raster, &backend->raster);
 		err |= mg_gl_canvas_compile_render_program("blit", glsl_blit_vertex, glsl_blit_fragment, &backend->blit);
 
@@ -1744,11 +1763,14 @@ mg_canvas_backend* gl_canvas_backend_create(mg_wgl_surface* surface)
 		glBindBuffer(GL_SHADER_STORAGE_BUFFER, backend->screenTilesBuffer);
 		glBufferData(GL_SHADER_STORAGE_BUFFER, nTilesX*nTilesY*MG_GL_SCREEN_TILE_SIZE, 0, GL_DYNAMIC_COPY);
 
+		glGenBuffers(1, &backend->screenTilesCountBuffer);
+		glBindBuffer(GL_SHADER_STORAGE_BUFFER, backend->screenTilesCountBuffer);
+		glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(int), 0, GL_DYNAMIC_COPY);
+
 		glGenBuffers(1, &backend->rasterDispatchBuffer);
 		glBindBuffer(GL_SHADER_STORAGE_BUFFER, backend->rasterDispatchBuffer);
 		glBufferData(GL_SHADER_STORAGE_BUFFER, sizeof(mg_gl_dispatch_indirect_command), 0, GL_DYNAMIC_COPY);
 
-
 		if(err)
 		{
 			mg_gl_canvas_destroy((mg_canvas_backend*)backend);

src/glsl_shaders/balance_workgroups.glsl

@@ -0,0 +1,27 @@
+
+layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+
+precision mediump float;
+layout(std430) buffer;
+
+layout(binding = 0) coherent restrict readonly buffer screenTilesCountBufferSSBO
+{
+	int elements[];
+} screenTilesCountBuffer;
+
+layout(binding = 1) coherent restrict writeonly buffer dispatchBufferSSBO
+{
+	mg_gl_dispatch_indirect_command elements[];
+} dispatchBuffer;
+
+
+layout(location = 0) uniform uint maxWorkGroupCount;
+
+void main()
+{
+	uint totalWorkGroupCount = screenTilesCountBuffer.elements[0];
+
+	dispatchBuffer.elements[0].num_groups_x = totalWorkGroupCount > maxWorkGroupCount ? maxWorkGroupCount : totalWorkGroupCount;
+	dispatchBuffer.elements[0].num_groups_y = (totalWorkGroupCount + maxWorkGroupCount - 1) / maxWorkGroupCount;
+	dispatchBuffer.elements[0].num_groups_z = 1;
+}

src/glsl_shaders/merge.glsl

@@ -34,10 +34,10 @@ layout(binding = 5) restrict writeonly buffer screenTilesBufferSSBO
 	mg_gl_screen_tile elements[];
 } screenTilesBuffer;
 
-layout(binding = 6) coherent restrict buffer dispatchBufferSSBO
+layout(binding = 6) coherent restrict buffer screenTilesCountBufferSSBO
 {
-	mg_gl_dispatch_indirect_command elements[];
+	int elements[];
-} dispatchBuffer;
+} screenTilesCountBuffer;
 
 
 layout(location = 0) uniform int tileSize;
@@ -53,9 +53,6 @@ void main()
 
 	int lastOpIndex = -1;
 
-	dispatchBuffer.elements[0].num_groups_y = 1;
-	dispatchBuffer.elements[0].num_groups_z = 1;
-
 	for(int pathIndex = 0; pathIndex < pathCount; pathIndex++)
 	{
 		mg_gl_path_queue pathQueue = pathQueueBuffer.elements[pathIndex];
@@ -75,7 +72,7 @@ void main()
 		{
 			if(tileIndex < 0)
 			{
-				tileIndex = int(atomicAdd(dispatchBuffer.elements[0].num_groups_x, 1));
+				tileIndex = int(atomicAdd(screenTilesCountBuffer.elements[0], 1));
 				screenTilesBuffer.elements[tileIndex].tileCoord = uvec2(tileCoord);
 				screenTilesBuffer.elements[tileIndex].first = -1;
 			}

src/glsl_shaders/raster.glsl

@@ -24,17 +24,30 @@ layout(binding = 3) restrict readonly buffer screenTilesBufferSSBO
 	mg_gl_screen_tile elements[];
 } screenTilesBuffer;
 
+layout(binding = 4) restrict readonly buffer screenTilesCountBufferSSBO
+{
+	int elements[];
+} screenTilesCountBuffer;
+
+
 layout(location = 0) uniform float scale;
 layout(location = 1) uniform int msaaSampleCount;
 layout(location = 2) uniform uint useTexture;
 layout(location = 3) uniform int pathBufferStart;
+layout(location = 4) uniform uint maxWorkGroupCount;
 
 layout(rgba8, binding = 0) uniform restrict writeonly image2D outTexture;
 layout(binding = 1) uniform sampler2D srcTexture;
 
 void main()
 {
-	uint tileIndex = gl_WorkGroupID.x;
+	uint tileIndex = gl_WorkGroupID.y * maxWorkGroupCount + gl_WorkGroupID.x;
+
+	if(tileIndex >= screenTilesCountBuffer.elements[0])
+	{
+		return;
+	}
+
 	uvec2 tileCoord = screenTilesBuffer.elements[tileIndex].tileCoord;
 	ivec2 pixelCoord = ivec2(tileCoord * gl_WorkGroupSize.x + gl_LocalInvocationID.xy);