Uniform buffer binding rework for clarity + minor cleanup

src/Refresh_Driver_D3D11.c

@@ -443,6 +443,11 @@ typedef struct D3D11UniformBuffer
 	uint32_t offset;
 } D3D11UniformBuffer;
 
+typedef struct D3D11Fence
+{
+	ID3D11Query *handle;
+} D3D11Fence;
+
 typedef struct D3D11CommandBuffer
 {
 	/* D3D11 Object References */
@@ -459,7 +464,7 @@ typedef struct D3D11CommandBuffer
 
 	/* State */
 	SDL_threadID threadID;
-	ID3D11Query *completionQuery;
+	D3D11Fence *fence;
 
 	/* Uniforms */
 	D3D11UniformBuffer *vertexUniformBuffer;
@@ -509,9 +514,14 @@ typedef struct D3D11Renderer
 	uint32_t availableUniformBufferCount;
 	uint32_t availableUniformBufferCapacity;
 
+	D3D11Fence **availableFences;
+	uint32_t availableFenceCount;
+	uint32_t availableFenceCapacity;
+
 	SDL_mutex *contextLock;
 	SDL_mutex *acquireCommandBufferLock;
 	SDL_mutex *uniformBufferLock;
+	SDL_mutex *acquireFenceLock;
 } D3D11Renderer;
 
 /* Logging */
@@ -572,7 +582,7 @@ static void D3D11_INTERNAL_LogError(
 	Refresh_LogError("%s! Error Code: %s (0x%08X)", msg, wszMsgBuff, res);
 }
 
-/* Subresources */
+/* Helper Functions */
 
 static inline uint32_t D3D11_INTERNAL_CalcSubresource(
 	uint32_t mipLevel,
@@ -582,6 +592,13 @@ static inline uint32_t D3D11_INTERNAL_CalcSubresource(
 	return mipLevel + (arraySlice * numLevels);
 }
 
+static inline uint32_t D3D11_INTERNAL_NextHighestAlignment(
+	uint32_t n,
+	uint32_t align
+) {
+	return align * ((n + align - 1) / align);
+}
+
 /* Quit */
 
 static void D3D11_DestroyDevice(
@@ -606,7 +623,6 @@ static void D3D11_DestroyDevice(
 	for (uint32_t i = 0; i < renderer->availableCommandBufferCount; i += 1)
 	{
 		D3D11CommandBuffer *commandBuffer = renderer->availableCommandBuffers[i];
-		ID3D11Query_Release(commandBuffer->completionQuery);
 		ID3D11DeviceContext_Release(commandBuffer->context);
 		SDL_free(commandBuffer);
 	}
@@ -623,10 +639,20 @@ static void D3D11_DestroyDevice(
 	}
 	SDL_free(renderer->availableUniformBuffers);
 
+	/* Release fence infrastructure */
+	for (uint32_t i = 0; i < renderer->availableFenceCount; i += 1)
+	{
+		D3D11Fence *fence = renderer->availableFences[i];
+		ID3D11Query_Release(fence->handle);
+		SDL_free(fence);
+	}
+	SDL_free(renderer->availableFences);
+
 	/* Release the mutexes */
 	SDL_DestroyMutex(renderer->acquireCommandBufferLock);
 	SDL_DestroyMutex(renderer->contextLock);
 	SDL_DestroyMutex(renderer->uniformBufferLock);
+	SDL_DestroyMutex(renderer->acquireFenceLock);
 
 	/* Release the device and associated objects */
 	ID3D11DeviceContext_Release(renderer->immediateContext);
@@ -877,7 +903,6 @@ static ID3D11BlendState* D3D11_INTERNAL_FetchBlendState(
 ) {
 	ID3D11BlendState *result;
 	D3D11_BLEND_DESC blendDesc;
-	uint32_t i;
 	HRESULT res;
 
 	/* Create a new blend state.
@@ -888,7 +913,7 @@ static ID3D11BlendState* D3D11_INTERNAL_FetchBlendState(
 	blendDesc.AlphaToCoverageEnable = FALSE;
 	blendDesc.IndependentBlendEnable = TRUE;
 
-	for (i = 0; i < numColorAttachments; i += 1)
+	for (uint32_t i = 0; i < numColorAttachments; i += 1)
 	{
 		blendDesc.RenderTarget[i].BlendEnable = colorAttachments[i].blendState.blendEnable;
 		blendDesc.RenderTarget[i].BlendOp = RefreshToD3D11_BlendOp[
@@ -1009,8 +1034,7 @@ static uint32_t D3D11_INTERNAL_FindIndexOfVertexBinding(
 	const Refresh_VertexBinding *bindings,
 	uint32_t numBindings
 ) {
-	uint32_t i;
+	for (uint32_t i = 0; i < numBindings; i += 1)
-	for (i = 0; i < numBindings; i += 1)
 	{
 		if (bindings[i].binding == targetBinding)
 		{
@@ -1030,7 +1054,7 @@ static ID3D11InputLayout* D3D11_INTERNAL_FetchInputLayout(
 ) {
 	ID3D11InputLayout *result = NULL;
 	D3D11_INPUT_ELEMENT_DESC *elementDescs;
-	uint32_t i, bindingIndex;
+	uint32_t bindingIndex;
 	HRESULT res;
 
 	/* Don't bother creating/fetching an input layout if there are no attributes. */
@@ -1046,7 +1070,7 @@ static ID3D11InputLayout* D3D11_INTERNAL_FetchInputLayout(
 	);
 
 	/* Create the array of input elements */
-	for (i = 0; i < inputState.vertexAttributeCount; i += 1)
+	for (uint32_t i = 0; i < inputState.vertexAttributeCount; i += 1)
 	{
 		elementDescs[i].AlignedByteOffset = inputState.vertexAttributes[i].offset;
 		elementDescs[i].Format = RefreshToD3D11_VertexFormat[
@@ -1099,7 +1123,6 @@ static Refresh_ComputePipeline* D3D11_CreateComputePipeline(
 	Refresh_Renderer *driverData,
 	Refresh_ComputeShaderInfo *computeShaderInfo
 ) {
-	D3D11Renderer* renderer = (D3D11Renderer*) driverData;
 	D3D11ComputePipeline* pipeline = (D3D11ComputePipeline*) SDL_malloc(sizeof(D3D11ComputePipeline));
 	D3D11ShaderModule* shaderModule = (D3D11ShaderModule*) computeShaderInfo->shaderModule;
 
@@ -1111,13 +1134,6 @@ static Refresh_ComputePipeline* D3D11_CreateComputePipeline(
 	return (Refresh_ComputePipeline*) pipeline;
 }
 
-static inline uint32_t D3D11_INTERNAL_NextHighestAlignment(
-	uint32_t n,
-	uint32_t align
-) {
-	return align * ((n + align - 1) / align);
-}
-
 static Refresh_GraphicsPipeline* D3D11_CreateGraphicsPipeline(
 	Refresh_Renderer *driverData,
 	Refresh_GraphicsPipelineCreateInfo *pipelineCreateInfo
@@ -1129,7 +1145,7 @@ static Refresh_GraphicsPipeline* D3D11_CreateGraphicsPipeline(
 	ID3D10Blob *errorBlob;
 	HRESULT res;
 
-	/* Color */
+	/* Blend */
 
 	pipeline->colorAttachmentBlendState = D3D11_INTERNAL_FetchBlendState(
 		renderer,
@@ -1150,9 +1166,11 @@ static Refresh_GraphicsPipeline* D3D11_CreateGraphicsPipeline(
 	pipeline->blendConstants[2] = pipelineCreateInfo->blendConstants[2];
 	pipeline->blendConstants[3] = pipelineCreateInfo->blendConstants[3];
 
+	/* Multisample */
+
 	pipeline->multisampleState = pipelineCreateInfo->multisampleState;
 
-	/* Depth stencil */
+	/* Depth-Stencil */
 
 	pipeline->depthStencilState = D3D11_INTERNAL_FetchDepthStencilState(
 		renderer,
@@ -1173,7 +1191,7 @@ static Refresh_GraphicsPipeline* D3D11_CreateGraphicsPipeline(
 		pipelineCreateInfo->rasterizerState
 	);
 
-	/* Vertex shader */
+	/* Vertex Shader */
 
 	if (vertShaderModule->shader == NULL)
 	{
@@ -1325,6 +1343,7 @@ static Refresh_Sampler* D3D11_CreateSampler(
 	);
 	ERROR_CHECK_RETURN("Could not create sampler state", NULL);
 
+	/* FIXME: Is there even a point to having a D3D11Sampler struct if it's just a wrapper? */
 	d3d11Sampler = (D3D11Sampler*) SDL_malloc(sizeof(D3D11Sampler));
 	d3d11Sampler->handle = samplerStateHandle;
 
@@ -1659,7 +1678,7 @@ static Refresh_Buffer* D3D11_CreateBuffer(
 	return (Refresh_Buffer*) d3d11Buffer;
 }
 
-/* Setters */
+/* Texture Data */
 
 static void D3D11_SetTextureData(
 	Refresh_Renderer *driverData,
@@ -1746,6 +1765,8 @@ static void D3D11_CopyTextureToBuffer(
 	NOT_IMPLEMENTED
 }
 
+/* Buffer Data */
+
 static void D3D11_INTERNAL_SetBufferData(
 	D3D11Renderer *renderer,
 	D3D11CommandBuffer *commandBuffer,
@@ -1824,23 +1845,17 @@ static void D3D11_SetBufferData(
 	);
 }
 
-static void D3D11_INTERNAL_BindUniformBuffer(
+static void D3D11_GetBufferData(
-	D3D11CommandBuffer *commandBuffer,
+	Refresh_Renderer *driverData,
-	D3D11UniformBuffer *uniformBuffer
+	Refresh_Buffer *buffer,
+	void* data,
+	uint32_t dataLengthInBytes
 ) {
-	if (commandBuffer->boundUniformBufferCount >= commandBuffer->boundUniformBufferCapacity)
+	NOT_IMPLEMENTED
-	{
-		commandBuffer->boundUniformBufferCapacity *= 2;
-		commandBuffer->boundUniformBuffers = SDL_realloc(
-			commandBuffer->boundUniformBuffers,
-			sizeof(D3D11UniformBuffer*) * commandBuffer->boundUniformBufferCapacity
-		);
-	}
-
-	commandBuffer->boundUniformBuffers[commandBuffer->boundUniformBufferCount] = uniformBuffer;
-	commandBuffer->boundUniformBufferCount += 1;
 }
 
+/* Uniforms */
+
 static uint8_t D3D11_INTERNAL_CreateUniformBuffer(
 	D3D11Renderer *renderer
 ) {
@@ -1869,6 +1884,8 @@ static uint8_t D3D11_INTERNAL_CreateUniformBuffer(
 	uniformBuffer->d3d11Buffer = SDL_malloc(sizeof(D3D11Buffer));
 	uniformBuffer->d3d11Buffer->handle = bufferHandle;
 	uniformBuffer->d3d11Buffer->size = UBO_BUFFER_SIZE;
+	uniformBuffer->d3d11Buffer->isDynamic = 1;
+	uniformBuffer->d3d11Buffer->uav = NULL;
 
 	/* Add it to the available pool */
 	if (renderer->availableUniformBufferCount >= renderer->availableUniformBufferCapacity)
@@ -1887,12 +1904,15 @@ static uint8_t D3D11_INTERNAL_CreateUniformBuffer(
 	return 1;
 }
 
-static D3D11UniformBuffer* D3D11_INTERNAL_AcquireUniformBufferFromPool(
+static uint8_t D3D11_INTERNAL_AcquireUniformBuffer(
 	D3D11Renderer *renderer,
+	D3D11CommandBuffer *commandBuffer,
+	D3D11UniformBuffer **uniformBufferToBind,
 	uint64_t blockSize
 ) {
 	D3D11UniformBuffer *uniformBuffer;
 
+	/* Acquire a uniform buffer from the pool */
 	SDL_LockMutex(renderer->uniformBufferLock);
 
 	if (renderer->availableUniformBufferCount == 0)
@@ -1901,7 +1921,7 @@ static D3D11UniformBuffer* D3D11_INTERNAL_AcquireUniformBufferFromPool(
 		{
 			SDL_UnlockMutex(renderer->uniformBufferLock);
 			Refresh_LogError("Failed to create uniform buffer!");
-			return NULL;
+			return 0;
 		}
 	}
 
@@ -1910,9 +1930,24 @@ static D3D11UniformBuffer* D3D11_INTERNAL_AcquireUniformBufferFromPool(
 
 	SDL_UnlockMutex(renderer->uniformBufferLock);
 
+	/* Reset the uniform buffer */
 	uniformBuffer->offset = 0;
 
-	return uniformBuffer;
+	/* Bind the uniform buffer to the command buffer */
+	if (commandBuffer->boundUniformBufferCount >= commandBuffer->boundUniformBufferCapacity)
+	{
+		commandBuffer->boundUniformBufferCapacity *= 2;
+		commandBuffer->boundUniformBuffers = SDL_realloc(
+			commandBuffer->boundUniformBuffers,
+			sizeof(D3D11UniformBuffer*) * commandBuffer->boundUniformBufferCapacity
+		);
+	}
+	commandBuffer->boundUniformBuffers[commandBuffer->boundUniformBufferCount] = uniformBuffer;
+	commandBuffer->boundUniformBufferCount += 1;
+
+	*uniformBufferToBind = uniformBuffer;
+
+	return 1;
 }
 
 static uint32_t D3D11_PushVertexShaderUniforms(
@@ -1928,13 +1963,11 @@ static uint32_t D3D11_PushVertexShaderUniforms(
 
 	if (d3d11CommandBuffer->vertexUniformBuffer->offset + graphicsPipeline->vertexUniformBlockSize >= UBO_BUFFER_SIZE)
 	{
-		/* We're out of space in this buffer, bind the old one and acquire a new one */
+		/* Out of space! Get a new uniform buffer. */
-		D3D11_INTERNAL_BindUniformBuffer(
+		D3D11_INTERNAL_AcquireUniformBuffer(
-			d3d11CommandBuffer,
-			d3d11CommandBuffer->vertexUniformBuffer
-		);
-		d3d11CommandBuffer->vertexUniformBuffer = D3D11_INTERNAL_AcquireUniformBufferFromPool(
 			renderer,
+			d3d11CommandBuffer,
+			&d3d11CommandBuffer->vertexUniformBuffer,
 			graphicsPipeline->vertexUniformBlockSize
 		);
 	}
@@ -1969,13 +2002,11 @@ static uint32_t D3D11_PushFragmentShaderUniforms(
 
 	if (d3d11CommandBuffer->fragmentUniformBuffer->offset + graphicsPipeline->fragmentUniformBlockSize >= UBO_BUFFER_SIZE)
 	{
-		/* We're out of space in this buffer, bind the old one and acquire a new one */
+		/* Out of space! Get a new uniform buffer. */
-		D3D11_INTERNAL_BindUniformBuffer(
+		D3D11_INTERNAL_AcquireUniformBuffer(
-			d3d11CommandBuffer,
-			d3d11CommandBuffer->fragmentUniformBuffer
-		);
-		d3d11CommandBuffer->fragmentUniformBuffer = D3D11_INTERNAL_AcquireUniformBufferFromPool(
 			renderer,
+			d3d11CommandBuffer,
+			&d3d11CommandBuffer->fragmentUniformBuffer,
 			graphicsPipeline->fragmentUniformBlockSize
 		);
 	}
@@ -2007,6 +2038,8 @@ static uint32_t D3D11_PushComputeShaderUniforms(
 	return 0;
 }
 
+/* Samplers */
+
 static void D3D11_BindVertexSamplers(
 	Refresh_Renderer *driverData,
 	Refresh_CommandBuffer *commandBuffer,
@@ -2075,17 +2108,6 @@ static void D3D11_BindFragmentSamplers(
 	);
 }
 
-/* Getters */
-
-static void D3D11_GetBufferData(
-	Refresh_Renderer *driverData,
-	Refresh_Buffer *buffer,
-	void *data,
-	uint32_t dataLengthInBytes
-) {
-	NOT_IMPLEMENTED
-}
-
 /* Disposal */
 
 static void D3D11_QueueDestroyTexture(
@@ -2193,7 +2215,6 @@ static void D3D11_INTERNAL_AllocateCommandBuffers(
 	uint32_t allocateCount
 ) {
 	D3D11CommandBuffer *commandBuffer;
-	D3D11_QUERY_DESC queryDesc;
 	HRESULT res;
 
 	renderer->availableCommandBufferCapacity += allocateCount;
@@ -2215,16 +2236,6 @@ static void D3D11_INTERNAL_AllocateCommandBuffers(
 		);
 		ERROR_CHECK("Could not create deferred context");
 
-		/* Fence Query */
-		queryDesc.Query = D3D11_QUERY_EVENT;
-		queryDesc.MiscFlags = 0;
-		res = ID3D11Device_CreateQuery(
-			renderer->device,
-			&queryDesc,
-			&commandBuffer->completionQuery
-		);
-		ERROR_CHECK("Could not create query");
-
 		/* Bound Uniform Buffers */
 		commandBuffer->boundUniformBufferCapacity = 16;
 		commandBuffer->boundUniformBufferCount = 0;
@@ -2232,8 +2243,6 @@ static void D3D11_INTERNAL_AllocateCommandBuffers(
 			commandBuffer->boundUniformBufferCapacity * sizeof(D3D11UniformBuffer*)
 		);
 
-		/* FIXME: Resource tracking? */
-
 		renderer->availableCommandBuffers[renderer->availableCommandBufferCount] = commandBuffer;
 		renderer->availableCommandBufferCount += 1;
 	}
@@ -2258,17 +2267,69 @@ static D3D11CommandBuffer* D3D11_INTERNAL_GetInactiveCommandBufferFromPool(
 	return commandBuffer;
 }
 
+static void D3D11_INTERNAL_AllocateFences(
+	D3D11Renderer *renderer,
+	uint32_t allocateCount
+) {
+	D3D11Fence *fence;
+	D3D11_QUERY_DESC queryDesc;
+	HRESULT res;
+
+	renderer->availableFenceCapacity += allocateCount;
+
+	renderer->availableFences = SDL_realloc(
+		renderer->availableFences,
+		sizeof(D3D11Fence*) * renderer->availableFenceCapacity
+	);
+
+	for (uint32_t i = 0; i < allocateCount; i += 1)
+	{
+		fence = SDL_malloc(sizeof(D3D11Fence));
+
+		queryDesc.Query = D3D11_QUERY_EVENT;
+		queryDesc.MiscFlags = 0;
+		res = ID3D11Device_CreateQuery(
+			renderer->device,
+			&queryDesc,
+			&fence->handle
+		);
+		ERROR_CHECK("Could not create query"); /* FIXME: Should this return an error code? */
+
+		renderer->availableFences[renderer->availableFenceCount] = fence;
+		renderer->availableFenceCount += 1;
+	}
+}
+
+static D3D11Fence* D3D11_INTERNAL_GetFenceFromPool(
+	D3D11Renderer* renderer
+) {
+	D3D11Fence *fence;
+
+	SDL_LockMutex(renderer->acquireFenceLock);
+
+	if (renderer->availableFenceCount == 0)
+	{
+		D3D11_INTERNAL_AllocateFences(renderer, renderer->availableFenceCapacity);
+	}
+
+	fence = renderer->availableFences[renderer->availableFenceCount - 1];
+	renderer->availableFenceCount -= 1;
+
+	SDL_UnlockMutex(renderer->acquireFenceLock);
+
+	return fence;
+}
+
 static Refresh_CommandBuffer* D3D11_AcquireCommandBuffer(
 	Refresh_Renderer *driverData
 ) {
 	D3D11Renderer *renderer = (D3D11Renderer*) driverData;
 	D3D11CommandBuffer *commandBuffer;
-	uint32_t i;
 
 	SDL_LockMutex(renderer->acquireCommandBufferLock);
 
-	/* Set up the command buffer */
 	commandBuffer = D3D11_INTERNAL_GetInactiveCommandBufferFromPool(renderer);
+
 	commandBuffer->threadID = SDL_ThreadID();
 	commandBuffer->swapchainData = NULL;
 	commandBuffer->graphicsPipeline = NULL;
@@ -2277,11 +2338,16 @@ static Refresh_CommandBuffer* D3D11_AcquireCommandBuffer(
 	commandBuffer->fragmentUniformBuffer = NULL;
 	commandBuffer->computeUniformBuffer = NULL;
 	commandBuffer->dsView = NULL;
-	for (i = 0; i < MAX_COLOR_TARGET_BINDINGS; i += 1)
+
+	for (uint32_t i = 0; i < MAX_COLOR_TARGET_BINDINGS; i += 1)
 	{
 		commandBuffer->rtViews[i] = NULL;
 	}
 
+	SDL_LockMutex(renderer->acquireFenceLock);
+	commandBuffer->fence = D3D11_INTERNAL_GetFenceFromPool(renderer);
+	SDL_UnlockMutex(renderer->acquireFenceLock);
+
 	SDL_UnlockMutex(renderer->acquireCommandBufferLock);
 
 	return (Refresh_CommandBuffer*) commandBuffer;
@@ -2300,7 +2366,6 @@ static void D3D11_BeginRenderPass(
 	D3D11_CLEAR_FLAG dsClearFlags;
 	D3D11_VIEWPORT viewport;
 	D3D11_RECT scissorRect;
-	uint32_t i;
 
 	/* FIXME:
 	 * We need to unbind the RT textures on the Refresh side
@@ -2308,14 +2373,14 @@ static void D3D11_BeginRenderPass(
 	 */
 
 	/* Clear the bound RTs for the current command buffer */
-	for (i = 0; i < MAX_COLOR_TARGET_BINDINGS; i += 1)
+	for (uint32_t i = 0; i < MAX_COLOR_TARGET_BINDINGS; i += 1)
 	{
 		d3d11CommandBuffer->rtViews[i] = NULL;
 	}
 	d3d11CommandBuffer->dsView = NULL;
 
 	/* Get RTVs for the color attachments */
-	for (i = 0; i < colorAttachmentCount; i += 1)
+	for (uint32_t i = 0; i < colorAttachmentCount; i += 1)
 	{
 		/* FIXME: Cube RTs */
 
@@ -2337,7 +2402,7 @@ static void D3D11_BeginRenderPass(
 	);
 
 	/* Perform load ops on the RTs */
-	for (i = 0; i < colorAttachmentCount; i += 1)
+	for (uint32_t i = 0; i < colorAttachmentCount; i += 1)
 	{
 		if (colorAttachmentInfos[i].loadOp == REFRESH_LOADOP_CLEAR)
 		{
@@ -2410,23 +2475,9 @@ static void D3D11_EndRenderPass(
 ) {
 	D3D11CommandBuffer *d3d11CommandBuffer = (D3D11CommandBuffer*) commandBuffer;
 
-	if (d3d11CommandBuffer->vertexUniformBuffer != NULL)
-	{
-		D3D11_INTERNAL_BindUniformBuffer(
-			d3d11CommandBuffer,
-			d3d11CommandBuffer->vertexUniformBuffer
-		);
-	}
 	d3d11CommandBuffer->vertexUniformBuffer = NULL;
-
-	if (d3d11CommandBuffer->fragmentUniformBuffer != NULL)
-	{
-		D3D11_INTERNAL_BindUniformBuffer(
-			d3d11CommandBuffer,
-			d3d11CommandBuffer->fragmentUniformBuffer
-		);
-	}
 	d3d11CommandBuffer->fragmentUniformBuffer = NULL;
+	d3d11CommandBuffer->computeUniformBuffer = NULL;
 
 	/* FIXME: Resolve MSAA here! */
 	/* FIXME: Anything else we need to do...? */
@@ -2443,31 +2494,17 @@ static void D3D11_BindGraphicsPipeline(
 
 	d3d11CommandBuffer->graphicsPipeline = pipeline;
 
-	if (d3d11CommandBuffer->vertexUniformBuffer != NULL)
-	{
-		D3D11_INTERNAL_BindUniformBuffer(
-			d3d11CommandBuffer,
-			d3d11CommandBuffer->vertexUniformBuffer
-		);
-	}
-
 	if (pipeline->vertexUniformBlockSize == 0)
 	{
 		d3d11CommandBuffer->vertexUniformBuffer = NULL;
 	}
 	else
 	{
-		d3d11CommandBuffer->vertexUniformBuffer = D3D11_INTERNAL_AcquireUniformBufferFromPool(
+		D3D11_INTERNAL_AcquireUniformBuffer(
 			renderer,
-			pipeline->vertexUniformBlockSize
-		);
-	}
-
-	if (d3d11CommandBuffer->fragmentUniformBuffer != NULL)
-	{
-		D3D11_INTERNAL_BindUniformBuffer(
 			d3d11CommandBuffer,
-			d3d11CommandBuffer->fragmentUniformBuffer
+			&d3d11CommandBuffer->vertexUniformBuffer,
+			pipeline->vertexUniformBlockSize
 		);
 	}
 
@@ -2477,8 +2514,10 @@ static void D3D11_BindGraphicsPipeline(
 	}
 	else
 	{
-		d3d11CommandBuffer->fragmentUniformBuffer = D3D11_INTERNAL_AcquireUniformBufferFromPool(
+		D3D11_INTERNAL_AcquireUniformBuffer(
 			renderer,
+			d3d11CommandBuffer,
+			&d3d11CommandBuffer->fragmentUniformBuffer,
 			pipeline->fragmentUniformBlockSize
 		);
 	}
@@ -2629,22 +2668,16 @@ static void D3D11_BindComputePipeline(
 
 	d3d11CommandBuffer->computePipeline = pipeline;
 
-	if (d3d11CommandBuffer->computeUniformBuffer != NULL)
-	{
-		D3D11_INTERNAL_BindUniformBuffer(
-			d3d11CommandBuffer,
-			d3d11CommandBuffer->computeUniformBuffer
-		);
-	}
-
 	if (pipeline->computeUniformBlockSize == 0)
 	{
 		d3d11CommandBuffer->computeUniformBuffer = NULL;
 	}
 	else
 	{
-		d3d11CommandBuffer->computeUniformBuffer = D3D11_INTERNAL_AcquireUniformBufferFromPool(
+		D3D11_INTERNAL_AcquireUniformBuffer(
 			renderer,
+			d3d11CommandBuffer,
+			&d3d11CommandBuffer->computeUniformBuffer,
 			pipeline->computeUniformBlockSize
 		);
 	}
@@ -2985,7 +3018,6 @@ static void D3D11_UnclaimWindow(
 ) {
 	D3D11Renderer *renderer = (D3D11Renderer*) driverData;
 	D3D11WindowData *windowData = D3D11_INTERNAL_FetchWindowData(windowHandle);
-	uint32_t i;
 
 	if (windowData == NULL)
 	{
@@ -3001,7 +3033,7 @@ static void D3D11_UnclaimWindow(
 		);
 	}
 
-	for (i = 0; i < renderer->claimedWindowCount; i += 1)
+	for (uint32_t i = 0; i < renderer->claimedWindowCount; i += 1)
 	{
 		if (renderer->claimedWindows[i]->windowHandle == windowHandle)
 		{
@@ -3111,6 +3143,22 @@ static void D3D11_INTERNAL_CleanCommandBuffer(
 
 	commandBuffer->boundUniformBufferCount = 0;
 
+	/* The fence is now available */
+	/* FIXME: Not if auto-release is false! */
+	SDL_LockMutex(renderer->acquireFenceLock);
+	if (renderer->availableFenceCount == renderer->availableFenceCapacity)
+	{
+		renderer->availableFenceCapacity *= 2;
+		renderer->availableFences = SDL_realloc(
+			renderer->availableFences,
+			renderer->availableFenceCapacity * sizeof(D3D11Fence*)
+		);
+	}
+
+	renderer->availableFences[renderer->availableFenceCount] = commandBuffer->fence;
+	renderer->availableFenceCount += 1;
+	SDL_UnlockMutex(renderer->acquireFenceLock);
+
 	/* Return command buffer to pool */
 	SDL_LockMutex(renderer->acquireCommandBufferLock);
 
@@ -3155,7 +3203,7 @@ static void D3D11_Submit(
 	/* Notify the command buffer completion query that we have completed recording */
 	ID3D11DeviceContext_End(
 		renderer->immediateContext,
-		(ID3D11Asynchronous*) d3d11CommandBuffer->completionQuery
+		(ID3D11Asynchronous*) d3d11CommandBuffer->fence->handle
 	);
 
 	/* Serialize the commands into the command list */
@@ -3205,7 +3253,7 @@ static void D3D11_Submit(
 
 		res = ID3D11DeviceContext_GetData(
 			renderer->immediateContext,
-			(ID3D11Asynchronous*) renderer->submittedCommandBuffers[i]->completionQuery,
+			(ID3D11Asynchronous*) renderer->submittedCommandBuffers[i]->fence->handle,
 			&queryData,
 			sizeof(queryData),
 			0
@@ -3245,7 +3293,7 @@ static void D3D11_Wait(
 	{
 		while (S_OK != ID3D11DeviceContext_GetData(
 			renderer->immediateContext,
-			(ID3D11Asynchronous*) renderer->submittedCommandBuffers[i]->completionQuery,
+			(ID3D11Asynchronous*) renderer->submittedCommandBuffers[i]->fence->handle,
 			&queryData,
 			sizeof(queryData),
 			0
@@ -3262,8 +3310,6 @@ static void D3D11_Wait(
 		D3D11_INTERNAL_CleanCommandBuffer(renderer, commandBuffer);
 	}
 
-	/* FIXME: D3D11_INTERNAL_PerformPendingDestroys(renderer); */
-
 	SDL_UnlockMutex(renderer->contextLock);
 }
 
@@ -3615,17 +3661,21 @@ tryCreateDevice:
 	renderer->contextLock = SDL_CreateMutex();
 	renderer->acquireCommandBufferLock = SDL_CreateMutex();
 	renderer->uniformBufferLock = SDL_CreateMutex();
+	renderer->acquireFenceLock = SDL_CreateMutex();
 
 	/* Initialize miscellaneous renderer members */
 	renderer->debugMode = (flags & D3D11_CREATE_DEVICE_DEBUG);
 
-	/* Create command buffers to initialize the pool */
+	/* Create command buffer pool */
 	D3D11_INTERNAL_AllocateCommandBuffers(renderer, 2);
 
 	/* Create uniform buffer pool */
 	renderer->availableUniformBufferCapacity = 2;
 	renderer->availableUniformBuffers = SDL_malloc(sizeof(D3D11UniformBuffer*) * renderer->availableUniformBufferCapacity);
 
+	/* Create fence pool */
+	D3D11_INTERNAL_AllocateFences(renderer, 2);
+
 	/* Create the Refresh Device */
 	result = (Refresh_Device*) SDL_malloc(sizeof(Refresh_Device));
 	ASSIGN_DRIVER(D3D11)