Prep for PS5Driver

CMakeLists.txt

@@ -8,7 +8,7 @@ option(BUILD_SHARED_LIBS "Build shared library" ON)
 
 # Version
 SET(LIB_MAJOR_VERSION "1")
-SET(LIB_MINOR_VERSION "4")
+SET(LIB_MINOR_VERSION "5")
 SET(LIB_REVISION "0")
 SET(LIB_VERSION "${LIB_MAJOR_VERSION}.${LIB_MINOR_VERSION}.${LIB_REVISION}")
 

include/Refresh.h

@@ -55,7 +55,7 @@ extern "C" {
 /* Version API */
 
 #define REFRESH_MAJOR_VERSION   1
-#define REFRESH_MINOR_VERSION	4
+#define REFRESH_MINOR_VERSION	5
 #define REFRESH_PATCH_VERSION	0
 
 #define REFRESH_COMPILED_VERSION ( \
@@ -124,6 +124,7 @@ typedef enum Refresh_TextureFormat
 	REFRESH_TEXTUREFORMAT_BC1,
 	REFRESH_TEXTUREFORMAT_BC2,
 	REFRESH_TEXTUREFORMAT_BC3,
+	REFRESH_TEXTUREFORMAT_BC7,
 	REFRESH_TEXTUREFORMAT_R8G8_SNORM,
 	REFRESH_TEXTUREFORMAT_R8G8B8A8_SNORM,
 	REFRESH_TEXTUREFORMAT_A2R10G10B10,

src/Refresh.c

@@ -34,7 +34,12 @@
 /* Drivers */
 
 static const Refresh_Driver *drivers[] = {
+#ifdef REFRESH_DRIVER_VULKAN
 	&VulkanDriver,
+#endif
+#ifdef REFRESH_DRIVER_PS5
+	&PS5Driver,
+#endif
 	NULL
 };
 

src/Refresh_Driver.h

@@ -36,9 +36,15 @@
 
 /* Logging */
 
-extern void Refresh_LogInfo(const char *fmt, ...);
+#ifdef __cplusplus
-extern void Refresh_LogWarn(const char *fmt, ...);
+extern "C" {
-extern void Refresh_LogError(const char *fmt, ...);
+#endif
+void Refresh_LogInfo(const char *fmt, ...);
+void Refresh_LogWarn(const char *fmt, ...);
+void Refresh_LogError(const char *fmt, ...);
+#ifdef __cplusplus
+}
+#endif
 
 /* Internal Helper Utilities */
 
@@ -51,6 +57,7 @@ static inline uint32_t Texture_GetFormatSize(
 			return 8;
 		case REFRESH_TEXTUREFORMAT_BC2:
 		case REFRESH_TEXTUREFORMAT_BC3:
+		case REFRESH_TEXTUREFORMAT_BC7:
 			return 16;
 		case REFRESH_TEXTUREFORMAT_R8:
 			return 1;
@@ -117,7 +124,8 @@ static inline uint32_t BytesPerRow(
 
 	if (	format == REFRESH_TEXTUREFORMAT_BC1 ||
 		format == REFRESH_TEXTUREFORMAT_BC2 ||
-		format == REFRESH_TEXTUREFORMAT_BC3	)
+		format == REFRESH_TEXTUREFORMAT_BC3 ||
+		format == REFRESH_TEXTUREFORMAT_BC7	)
 	{
 		blocksPerRow = (width + 3) / 4;
 	}
@@ -135,7 +143,8 @@ static inline int32_t BytesPerImage(
 
 	if (	format == REFRESH_TEXTUREFORMAT_BC1 ||
 		format == REFRESH_TEXTUREFORMAT_BC2 ||
-		format == REFRESH_TEXTUREFORMAT_BC3	)
+		format == REFRESH_TEXTUREFORMAT_BC3 ||
+		format == REFRESH_TEXTUREFORMAT_BC7	)
 	{
 		blocksPerRow = (width + 3) / 4;
 		blocksPerColumn = (height + 3) / 4;
@@ -522,6 +531,7 @@ typedef struct Refresh_Driver
 } Refresh_Driver;
 
 extern Refresh_Driver VulkanDriver;
+extern Refresh_Driver PS5Driver;
 
 #endif /* REFRESH_DRIVER_H */
 

src/Refresh_Driver_Vulkan.c

@@ -190,6 +190,7 @@ static VkFormat RefreshToVK_SurfaceFormat[] =
 	VK_FORMAT_BC1_RGBA_UNORM_BLOCK,		/* BC1 */
 	VK_FORMAT_BC2_UNORM_BLOCK,		/* BC3 */
 	VK_FORMAT_BC3_UNORM_BLOCK,		/* BC5 */
+	VK_FORMAT_BC7_UNORM_BLOCK,		/* BC7 */
 	VK_FORMAT_R8G8_SNORM,			/* R8G8_SNORM */
 	VK_FORMAT_R8G8B8A8_SNORM,		/* R8G8B8A8_SNORM */
 	VK_FORMAT_A2R10G10B10_UNORM_PACK32,	/* A2R10G10B10 */
@@ -912,6 +913,7 @@ static inline void DescriptorSetLayoutHashTable_Insert(
 
 typedef struct RenderPassColorTargetDescription
 {
+	VkFormat format;
 	Refresh_Vec4 clearColor;
 	Refresh_LoadOp loadOp;
 	Refresh_StoreOp storeOp;
@@ -919,6 +921,7 @@ typedef struct RenderPassColorTargetDescription
 
 typedef struct RenderPassDepthStencilTargetDescription
 {
+	VkFormat format;
 	Refresh_LoadOp loadOp;
 	Refresh_StoreOp storeOp;
 	Refresh_LoadOp stencilLoadOp;
@@ -958,6 +961,11 @@ static inline uint8_t RenderPassHash_Compare(
 
 	for (i = 0; i < a->colorAttachmentCount; i += 1)
 	{
+		if (a->colorTargetDescriptions[i].format != b->colorTargetDescriptions[i].format)
+		{
+			return 0;
+		}
+
 		if (	a->colorTargetDescriptions[i].clearColor.x != b->colorTargetDescriptions[i].clearColor.x ||
 			a->colorTargetDescriptions[i].clearColor.y != b->colorTargetDescriptions[i].clearColor.y ||
 			a->colorTargetDescriptions[i].clearColor.z != b->colorTargetDescriptions[i].clearColor.z ||
@@ -977,6 +985,11 @@ static inline uint8_t RenderPassHash_Compare(
 		}
 	}
 
+	if (a->depthStencilTargetDescription.format != b->depthStencilTargetDescription.format)
+	{
+		return 0;
+	}
+
 	if (a->depthStencilTargetDescription.loadOp != b->depthStencilTargetDescription.loadOp)
 	{
 		return 0;
@@ -1674,6 +1687,10 @@ typedef struct VulkanRenderer
 
 	VkDeviceSize minUBOAlignment;
 
+	/* Some drivers don't support D16 for some reason. Fun! */
+	VkFormat D16Format;
+	VkFormat D16S8Format;
+
 	VulkanTexture **texturesToDestroy;
 	uint32_t texturesToDestroyCount;
 	uint32_t texturesToDestroyCapacity;
@@ -1782,6 +1799,40 @@ static inline void LogVulkanResultAsWarn(
 
 /* Utility */
 
+static inline VkFormat RefreshToVK_DepthFormat(
+	VulkanRenderer* renderer,
+	Refresh_TextureFormat format
+) {
+	switch (format)
+	{
+		case REFRESH_TEXTUREFORMAT_D16_UNORM:
+			return renderer->D16Format;
+		case REFRESH_TEXTUREFORMAT_D16_UNORM_S8_UINT:
+			return renderer->D16S8Format;
+		case REFRESH_TEXTUREFORMAT_D32_SFLOAT:
+			return VK_FORMAT_D32_SFLOAT;
+		case REFRESH_TEXTUREFORMAT_D32_SFLOAT_S8_UINT:
+			return VK_FORMAT_D32_SFLOAT_S8_UINT;
+		default:
+			return VK_FORMAT_UNDEFINED;
+	}
+}
+
+static inline uint8_t IsRefreshDepthFormat(Refresh_TextureFormat format)
+{
+	switch (format)
+	{
+		case REFRESH_TEXTUREFORMAT_D16_UNORM:
+		case REFRESH_TEXTUREFORMAT_D32_SFLOAT:
+		case REFRESH_TEXTUREFORMAT_D16_UNORM_S8_UINT:
+		case REFRESH_TEXTUREFORMAT_D32_SFLOAT_S8_UINT:
+			return 1;
+
+		default:
+			return 0;
+	}
+}
+
 static inline uint8_t IsDepthFormat(VkFormat format)
 {
 	switch(format)
@@ -1817,6 +1868,7 @@ static inline uint32_t VULKAN_INTERNAL_BytesPerPixel(VkFormat format)
 		case VK_FORMAT_R32G32B32A32_SFLOAT:
 		case VK_FORMAT_BC2_UNORM_BLOCK:
 		case VK_FORMAT_BC3_UNORM_BLOCK:
+		case VK_FORMAT_BC7_UNORM_BLOCK:
 			return 16;
 
 		case VK_FORMAT_R8G8B8A8_UNORM:
@@ -1857,6 +1909,40 @@ static inline uint32_t VULKAN_INTERNAL_BytesPerPixel(VkFormat format)
 	}
 }
 
+static inline uint32_t VULKAN_INTERNAL_GetTextureBlockSize(
+	VkFormat format
+) {
+	switch (format)
+	{
+	case VK_FORMAT_BC1_RGBA_UNORM_BLOCK:
+	case VK_FORMAT_BC2_UNORM_BLOCK:
+	case VK_FORMAT_BC3_UNORM_BLOCK:
+	case VK_FORMAT_BC7_UNORM_BLOCK:
+		return 4;
+	case VK_FORMAT_R8G8B8A8_UNORM:
+	case VK_FORMAT_B8G8R8A8_UNORM:
+	case VK_FORMAT_R5G6B5_UNORM_PACK16:
+	case VK_FORMAT_A1R5G5B5_UNORM_PACK16:
+	case VK_FORMAT_B4G4R4A4_UNORM_PACK16:
+	case VK_FORMAT_R8G8_SNORM:
+	case VK_FORMAT_R8G8B8A8_SNORM:
+	case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
+	case VK_FORMAT_R16G16_UNORM:
+	case VK_FORMAT_R16G16B16A16_UNORM:
+	case VK_FORMAT_R8_UNORM:
+	case VK_FORMAT_R32_SFLOAT:
+	case VK_FORMAT_R32G32_SFLOAT:
+	case VK_FORMAT_R32G32B32A32_SFLOAT:
+	case VK_FORMAT_R16_SFLOAT:
+	case VK_FORMAT_R16G16_SFLOAT:
+	case VK_FORMAT_R16G16B16A16_SFLOAT:
+		return 1;
+	default:
+		Refresh_LogError("Unrecognized texture format!");
+		return 0;
+	}
+}
+
 static inline VkDeviceSize VULKAN_INTERNAL_BytesPerImage(
 	uint32_t width,
 	uint32_t height,
@@ -1864,12 +1950,12 @@ static inline VkDeviceSize VULKAN_INTERNAL_BytesPerImage(
 ) {
 	uint32_t blocksPerRow = width;
 	uint32_t blocksPerColumn = height;
+	uint32_t blockSize = VULKAN_INTERNAL_GetTextureBlockSize(format);
 
-	if (format == VK_FORMAT_BC1_RGBA_UNORM_BLOCK ||
+	if (blockSize > 1)
-		format == VK_FORMAT_BC3_UNORM_BLOCK ||
-		format == VK_FORMAT_BC5_UNORM_BLOCK)
 	{
-		blocksPerRow = (width + 3) / 4;
+		blocksPerRow = (width + blockSize - 1) / blockSize;
+		blocksPerColumn = (height + blockSize - 1) / blockSize;
 	}
 
 	return blocksPerRow * blocksPerColumn * VULKAN_INTERNAL_BytesPerPixel(format);
@@ -5614,9 +5700,10 @@ static VkRenderPass VULKAN_INTERNAL_CreateTransientRenderPass(
 	if (attachmentInfo.hasDepthStencilAttachment)
 	{
 		attachmentDescriptions[attachmentDescriptionCount].flags = 0;
-		attachmentDescriptions[attachmentDescriptionCount].format = RefreshToVK_SurfaceFormat[
+		attachmentDescriptions[attachmentDescriptionCount].format = RefreshToVK_DepthFormat(
+			renderer,
 			attachmentInfo.depthStencilFormat
-		];
+		);
 		attachmentDescriptions[attachmentDescriptionCount].samples =
 			VK_SAMPLE_COUNT_1_BIT; /* FIXME: do these take multisamples? */
 		attachmentDescriptions[attachmentDescriptionCount].loadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
@@ -6313,7 +6400,16 @@ static Refresh_Texture* VULKAN_CreateTexture(
 		VK_IMAGE_USAGE_TRANSFER_SRC_BIT
 	);
 	VkImageAspectFlags imageAspectFlags;
-	VkFormat format = RefreshToVK_SurfaceFormat[textureCreateInfo->format];
+	VkFormat format;
+
+	if (IsRefreshDepthFormat(textureCreateInfo->format))
+	{
+		format = RefreshToVK_DepthFormat(renderer, textureCreateInfo->format);
+	}
+	else
+	{
+		format = RefreshToVK_SurfaceFormat[textureCreateInfo->format];
+	}
 
 	if (textureCreateInfo->usageFlags & REFRESH_TEXTUREUSAGE_SAMPLER_BIT)
 	{
@@ -6507,6 +6603,9 @@ static void VULKAN_SetTextureData(
 	VulkanTransferBuffer *transferBuffer;
 	VkBufferImageCopy imageCopy;
 	uint8_t *stagingBufferPointer;
+	uint32_t blockSize = VULKAN_INTERNAL_GetTextureBlockSize(vulkanTexture->format);
+	uint32_t bufferRowLength;
+	uint32_t bufferImageHeight;
 
 	if (vulkanCommandBuffer->renderPassInProgress)
 	{
@@ -6555,6 +6654,9 @@ static void VULKAN_SetTextureData(
 		&vulkanTexture->resourceAccessType
 	);
 
+	bufferRowLength = SDL_max(blockSize, textureSlice->rectangle.w);
+	bufferImageHeight = SDL_max(blockSize, textureSlice->rectangle.h);
+
 	imageCopy.imageExtent.width = textureSlice->rectangle.w;
 	imageCopy.imageExtent.height = textureSlice->rectangle.h;
 	imageCopy.imageExtent.depth = 1;
@@ -6566,8 +6668,8 @@ static void VULKAN_SetTextureData(
 	imageCopy.imageSubresource.layerCount = 1;
 	imageCopy.imageSubresource.mipLevel = textureSlice->level;
 	imageCopy.bufferOffset = transferBuffer->offset;
-	imageCopy.bufferRowLength = 0;
+	imageCopy.bufferRowLength = bufferRowLength;
-	imageCopy.bufferImageHeight = 0;
+	imageCopy.bufferImageHeight = bufferImageHeight;
 
 	renderer->vkCmdCopyBufferToImage(
 		vulkanCommandBuffer->commandBuffer,
@@ -6692,6 +6794,24 @@ static void VULKAN_SetTextureDataYUV(
 		&imageCopy
 	);
 
+	if (tex->usageFlags & VK_IMAGE_USAGE_SAMPLED_BIT)
+	{
+		/* TODO: is it worth it to only transition the specific subresource? */
+		VULKAN_INTERNAL_ImageMemoryBarrier(
+			renderer,
+			vulkanCommandBuffer->commandBuffer,
+			RESOURCE_ACCESS_ANY_SHADER_READ_SAMPLED_IMAGE,
+			VK_IMAGE_ASPECT_COLOR_BIT,
+			0,
+			tex->layerCount,
+			0,
+			tex->levelCount,
+			0,
+			tex->image,
+			&tex->resourceAccessType
+		);
+	}
+
 	VULKAN_INTERNAL_TrackTexture(renderer, vulkanCommandBuffer, tex);
 
 	/* These apply to both U and V */
@@ -6736,6 +6856,24 @@ static void VULKAN_SetTextureDataYUV(
 		&imageCopy
 	);
 
+	if (tex->usageFlags & VK_IMAGE_USAGE_SAMPLED_BIT)
+	{
+		/* TODO: is it worth it to only transition the specific subresource? */
+		VULKAN_INTERNAL_ImageMemoryBarrier(
+			renderer,
+			vulkanCommandBuffer->commandBuffer,
+			RESOURCE_ACCESS_ANY_SHADER_READ_SAMPLED_IMAGE,
+			VK_IMAGE_ASPECT_COLOR_BIT,
+			0,
+			tex->layerCount,
+			0,
+			tex->levelCount,
+			0,
+			tex->image,
+			&tex->resourceAccessType
+		);
+	}
+
 	VULKAN_INTERNAL_TrackTexture(renderer, vulkanCommandBuffer, tex);
 
 	/* V */
@@ -6775,11 +6913,9 @@ static void VULKAN_SetTextureDataYUV(
 
 	transferBuffer->offset += yDataLength + uvDataLength;
 
-	VULKAN_INTERNAL_TrackTexture(renderer, vulkanCommandBuffer, tex);
-
-	/* FIXME: don't we have to do this for every image? */
 	if (tex->usageFlags & VK_IMAGE_USAGE_SAMPLED_BIT)
 	{
+		/* TODO: is it worth it to only transition the specific subresource? */
 		VULKAN_INTERNAL_ImageMemoryBarrier(
 			renderer,
 			vulkanCommandBuffer->commandBuffer,
@@ -6794,6 +6930,8 @@ static void VULKAN_SetTextureDataYUV(
 			&tex->resourceAccessType
 		);
 	}
+
+	VULKAN_INTERNAL_TrackTexture(renderer, vulkanCommandBuffer, tex);
 }
 
 static void VULKAN_INTERNAL_BlitImage(
@@ -6858,8 +6996,8 @@ static void VULKAN_INTERNAL_BlitImage(
 	blit.dstOffsets[1].y = destinationTextureSlice->rectangle.y + destinationTextureSlice->rectangle.h;
 	blit.dstOffsets[1].z = 1;
 
-	blit.dstSubresource.mipLevel = sourceTextureSlice->level;
+	blit.dstSubresource.mipLevel = destinationTextureSlice->level;
-	blit.dstSubresource.baseArrayLayer = sourceTextureSlice->layer;
+	blit.dstSubresource.baseArrayLayer = destinationTextureSlice->layer;
 	blit.dstSubresource.layerCount = 1;
 	blit.dstSubresource.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
 
@@ -7656,6 +7794,7 @@ static VkRenderPass VULKAN_INTERNAL_FetchRenderPass(
 
 	for (i = 0; i < colorAttachmentCount; i += 1)
 	{
+		hash.colorTargetDescriptions[i].format = ((VulkanTexture*) colorAttachmentInfos[i].texture)->format;
 		hash.colorTargetDescriptions[i].clearColor = colorAttachmentInfos[i].clearColor;
 		hash.colorTargetDescriptions[i].loadOp = colorAttachmentInfos[i].loadOp;
 		hash.colorTargetDescriptions[i].storeOp = colorAttachmentInfos[i].storeOp;
@@ -7665,6 +7804,7 @@ static VkRenderPass VULKAN_INTERNAL_FetchRenderPass(
 
 	if (depthStencilAttachmentInfo == NULL)
 	{
+		hash.depthStencilTargetDescription.format = 0;
 		hash.depthStencilTargetDescription.loadOp = REFRESH_LOADOP_DONT_CARE;
 		hash.depthStencilTargetDescription.storeOp = REFRESH_STOREOP_DONT_CARE;
 		hash.depthStencilTargetDescription.stencilLoadOp = REFRESH_LOADOP_DONT_CARE;
@@ -7672,6 +7812,7 @@ static VkRenderPass VULKAN_INTERNAL_FetchRenderPass(
 	}
 	else
 	{
+		hash.depthStencilTargetDescription.format = ((VulkanTexture*) depthStencilAttachmentInfo->texture)->format;
 		hash.depthStencilTargetDescription.loadOp = depthStencilAttachmentInfo->loadOp;
 		hash.depthStencilTargetDescription.storeOp = depthStencilAttachmentInfo->storeOp;
 		hash.depthStencilTargetDescription.stencilLoadOp = depthStencilAttachmentInfo->stencilLoadOp;
@@ -8000,6 +8141,8 @@ static void VULKAN_BeginRenderPass(
 
 	if (depthStencilAttachmentInfo != NULL)
 	{
+		texture = (VulkanTexture*) depthStencilAttachmentInfo->texture;
+
 		if (texture->dimensions.width < framebufferWidth)
 		{
 			framebufferWidth = texture->dimensions.width;
@@ -9363,9 +9506,9 @@ static void VULKAN_Submit(
 
 		/* Mark command buffers as submitted */
 
-		if (renderer->submittedCommandBufferCount + commandBufferCount >= renderer->submittedCommandBufferCapacity)
+		if (renderer->submittedCommandBufferCount + 1 >= renderer->submittedCommandBufferCapacity)
 		{
-			renderer->submittedCommandBufferCapacity = renderer->submittedCommandBufferCount + commandBufferCount;
+			renderer->submittedCommandBufferCapacity = renderer->submittedCommandBufferCount + 1;
 
 			renderer->submittedCommandBuffers = SDL_realloc(
 				renderer->submittedCommandBuffers,
@@ -9373,12 +9516,9 @@ static void VULKAN_Submit(
 			);
 		}
 
-		for (i = 0; i < commandBufferCount; i += 1)
-		{
 		((VulkanCommandBuffer*)pCommandBuffers[i])->submitted = 1;
 		renderer->submittedCommandBuffers[renderer->submittedCommandBufferCount] = (VulkanCommandBuffer*) pCommandBuffers[i];
 		renderer->submittedCommandBufferCount += 1;
-		}
 
 		/* Present, if applicable */
 
@@ -10132,6 +10272,9 @@ static Refresh_Device* VULKAN_CreateDevice(
 	VkDescriptorPoolSize poolSizes[4];
 	VkDescriptorSetAllocateInfo descriptorAllocateInfo;
 
+	/* Variables: Image Format Detection */
+	VkImageFormatProperties imageFormatProperties;
+
 	VULKAN_INTERNAL_LoadEntryPoints(renderer);
 
 	renderer->presentMode = presentationParameters->presentMode;
@@ -10571,12 +10714,53 @@ static Refresh_Device* VULKAN_CreateDevice(
 	renderer->renderTargetHashArray.capacity = 0;
 
 	/* Initialize transfer buffer pool */
+
 	renderer->transferBufferPool.lock = SDL_CreateMutex();
 
 	renderer->transferBufferPool.availableBufferCapacity = 4;
 	renderer->transferBufferPool.availableBufferCount = 0;
 	renderer->transferBufferPool.availableBuffers = SDL_malloc(renderer->transferBufferPool.availableBufferCapacity * sizeof(VulkanTransferBuffer*));
 
+	/* Some drivers don't support D16, so we have to fall back to D32. */
+
+	vulkanResult = renderer->vkGetPhysicalDeviceImageFormatProperties(
+		renderer->physicalDevice,
+		VK_FORMAT_D16_UNORM,
+		VK_IMAGE_TYPE_2D,
+		VK_IMAGE_TILING_OPTIMAL,
+		VK_IMAGE_ASPECT_DEPTH_BIT,
+		0,
+		&imageFormatProperties
+	);
+
+	if (vulkanResult == VK_ERROR_FORMAT_NOT_SUPPORTED)
+	{
+		renderer->D16Format = VK_FORMAT_D32_SFLOAT;
+	}
+	else
+	{
+		renderer->D16Format = VK_FORMAT_D16_UNORM;
+	}
+
+	vulkanResult = renderer->vkGetPhysicalDeviceImageFormatProperties(
+		renderer->physicalDevice,
+		VK_FORMAT_D16_UNORM_S8_UINT,
+		VK_IMAGE_TYPE_2D,
+		VK_IMAGE_TILING_OPTIMAL,
+		VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT,
+		0,
+		&imageFormatProperties
+	);
+
+	if (vulkanResult == VK_ERROR_FORMAT_NOT_SUPPORTED)
+	{
+		renderer->D16S8Format = VK_FORMAT_D32_SFLOAT_S8_UINT;
+	}
+	else
+	{
+		renderer->D16S8Format = VK_FORMAT_D16_UNORM_S8_UINT;
+	}
+
 	/* Deferred destroy storage */
 
 	renderer->texturesToDestroyCapacity = 16;