Use custom blob format for combining shader code for multiple backends into a single file

CMakeLists.txt

@@ -8,8 +8,8 @@ option(BUILD_SHARED_LIBS "Build shared library" ON)
 
 # Version
 SET(LIB_MAJOR_VERSION "1")
-SET(LIB_MINOR_VERSION "15")
-SET(LIB_REVISION "4")
+SET(LIB_MINOR_VERSION "10")
+SET(LIB_REVISION "0")
 SET(LIB_VERSION "${LIB_MAJOR_VERSION}.${LIB_MINOR_VERSION}.${LIB_REVISION}")
 
 # Build Type
@@ -60,11 +60,6 @@ if(NOT MSVC)
 	set_property(TARGET Refresh PROPERTY COMPILE_FLAGS "-std=gnu99 -Wall -Wno-strict-aliasing -pedantic")
 endif()
 
-# Windows is silly and we need to manually include the Vulkan SDK
-if(MSVC)
-	target_include_directories(Refresh PUBLIC $ENV{VULKAN_SDK}/include)
-endif()
-
 # Refresh folders as includes, for other targets to consume
 target_include_directories(Refresh PUBLIC
 	$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/src>

include/Refresh.h

@@ -55,8 +55,8 @@ extern "C" {
 /* Version API */
 
 #define REFRESH_MAJOR_VERSION   1
-#define REFRESH_MINOR_VERSION	15
-#define REFRESH_PATCH_VERSION	4
+#define REFRESH_MINOR_VERSION	10
+#define REFRESH_PATCH_VERSION	0
 
 #define REFRESH_COMPILED_VERSION ( \
 	(REFRESH_MAJOR_VERSION * 100 * 100) + \
@@ -76,7 +76,6 @@ typedef struct Refresh_ShaderModule Refresh_ShaderModule;
 typedef struct Refresh_ComputePipeline Refresh_ComputePipeline;
 typedef struct Refresh_GraphicsPipeline Refresh_GraphicsPipeline;
 typedef struct Refresh_CommandBuffer Refresh_CommandBuffer;
-typedef struct Refresh_Fence Refresh_Fence;
 
 typedef enum Refresh_PresentMode
 {
@@ -464,7 +463,6 @@ typedef struct Refresh_TextureCreateInfo
 	uint32_t depth;
 	uint8_t isCube;
 	uint32_t levelCount;
-	Refresh_SampleCount sampleCount;
 	Refresh_TextureFormat format;
 	Refresh_TextureUsageFlags usageFlags;
 } Refresh_TextureCreateInfo;
@@ -553,6 +551,7 @@ typedef struct Refresh_ColorAttachmentInfo
 	uint32_t depth;
 	uint32_t layer;
 	uint32_t level;
+	Refresh_SampleCount sampleCount;
 	Refresh_Vec4 clearColor; /* Can be ignored by RenderPass */
 	Refresh_LoadOp loadOp;
 	Refresh_StoreOp storeOp;
@@ -786,13 +785,8 @@ REFRESHAPI void Refresh_SetTextureData(
  * yHeight:	The height of the Y plane.
  * uvWidth:	The width of the U/V planes.
  * uvHeight:	The height of the U/V planes.
- * yData:        A pointer to the raw Y image data.
- * uData:        A pointer to the raw U image data.
- * vData:        A pointer to the raw V image data.
- * yDataLength:  The size of the Y image data in bytes.
- * uvDataLength: The size of the UV image data in bytes.
- * yStride:      The length of a Y image data row in bytes.
- * uvStride:     The length of a UV image data row in bytes.
+ * data:	A pointer to the raw YUV image data.
+ * dataLength:	The size of the image data in bytes.
  */
 REFRESHAPI void Refresh_SetTextureDataYUV(
 	Refresh_Device *driverData,
@@ -804,13 +798,8 @@ REFRESHAPI void Refresh_SetTextureDataYUV(
 	uint32_t yHeight,
 	uint32_t uvWidth,
 	uint32_t uvHeight,
-	void *yDataPtr,
-	void *uDataPtr,
-	void *vDataPtr,
-	uint32_t yDataLength,
-	uint32_t uvDataLength,
-	uint32_t yStride,
-	uint32_t uvStride
+	void* data,
+	uint32_t dataLength
 );
 
 /* Performs an asynchronous texture-to-texture copy.
@@ -1207,50 +1196,15 @@ REFRESHAPI Refresh_Texture* Refresh_AcquireSwapchainTexture(
 /* Submits all of the enqueued commands. */
 REFRESHAPI void Refresh_Submit(
 	Refresh_Device* device,
-	Refresh_CommandBuffer *commandBuffer
+	uint32_t commandBufferCount,
+	Refresh_CommandBuffer **pCommandBuffers
 );
 
-/* Submits a command buffer and acquires a fence.
- * You can use the fence to check if or wait until the command buffer has finished processing.
- * You are responsible for releasing this fence when you are done using it.
- */
-REFRESHAPI Refresh_Fence* Refresh_SubmitAndAcquireFence(
-	Refresh_Device* device,
-	Refresh_CommandBuffer *commandBuffer
-);
-
-/* Waits for the device to become idle. */
+/* Waits for all submissions to complete. */
 REFRESHAPI void Refresh_Wait(
 	Refresh_Device *device
 );
 
-/* Waits for given fences to be signaled.
- *
- * waitAll: If 0, waits for any fence to be signaled. If 1, waits for all fences to be signaled.
- * fenceCount: The number of fences being submitted.
- * pFences: An array of fences to be waited on.
- */
-REFRESHAPI void Refresh_WaitForFences(
-	Refresh_Device *device,
-	uint8_t waitAll,
-	uint32_t fenceCount,
-	Refresh_Fence **pFences
-);
-
-/* Check the status of a fence. 1 means the fence is signaled. */
-REFRESHAPI int Refresh_QueryFence(
-	Refresh_Device *device,
-	Refresh_Fence *fence
-);
-
-/* Allows the fence to be reused by future command buffer submissions.
- * If you do not release fences after acquiring them, you will cause unbounded resource growth.
- */
-REFRESHAPI void Refresh_ReleaseFence(
-	Refresh_Device *device,
-	Refresh_Fence *fence
-);
-
 #ifdef __cplusplus
 }
 #endif /* __cplusplus */

include/Refresh_Image.h

@@ -44,45 +44,44 @@
 extern "C" {
 #endif /* __cplusplus */
 
-/* Image Read API */
-
-/* Decodes image data into raw RGBA8 texture data.
+/* Decodes PNG data into raw RGBA8 texture data.
  *
  * w:		    Filled with the width of the image.
  * h:		    Filled with the height of the image.
- * len:			Filled with the length of pixel data in bytes.
+ * numChannels: Filled with the number of channels in the image.
  *
  * Returns a block of memory suitable for use with Refresh_SetTextureData2D.
  * Be sure to free the memory with Refresh_Image_Free after use!
  */
 REFRESHAPI uint8_t* Refresh_Image_Load(
-	uint8_t *bufferPtr,
-	int32_t bufferLength,
+	char const *filename,
 	int32_t *w,
 	int32_t *h,
-	int32_t *len
+	int32_t *numChannels
 );
 
-/* Frees memory returned by Refresh_Image_Load. Do NOT free the memory yourself!
+/* Frees memory returned by Refresh_Image_Load. (Do NOT free the memory yourself!)
  *
- * mem: A pointer previously returned by Refresh_Image_LoadPNG.
+ * mem: A pointer previously returned by Refresh_Image_Load.
  */
 REFRESHAPI void Refresh_Image_Free(uint8_t *mem);
 
 /* Image Write API */
 
-/* Returns a buffer of PNG encoded from RGBA8 color data.
+/* Encodes 32-bit color data into PNG data.
  *
+ * filename:    The filename that the image will be written to.
+ * w:	        The width of the PNG data.
+ * h:	        The height of the PNG data.
+ * bgra:		Whether the data is in BGRA8 format. Otherwise will assume RBGA8.
  * data:	    The raw color data.
- * w:		The width of the color data.
- * h:		The height of the color data.
- * len:		Filled with the length of PNG data in bytes.
  */
 REFRESHAPI void Refresh_Image_SavePNG(
-	const char* filename,
-	uint8_t* data,
+	char const *filename,
 	int32_t w,
-	int32_t h
+	int32_t h,
+	uint8_t bgra,
+	uint8_t *data
 );
 
 #ifdef __cplusplus

shadercompiler/Program.cs

@@ -186,7 +186,7 @@ partial class Program
 		}
 
 		// Create the output blob file
-		string outputFilepath = Path.Combine(data.outputDir, $"{shaderName}{shaderType}.refresh");
+		string outputFilepath = Path.Combine(data.outputDir, $"{shaderName}.refresh");
 		using (FileStream fs = File.Create(outputFilepath))
 		{
 			using (BinaryWriter writer = new BinaryWriter(fs))
@@ -238,7 +238,7 @@ partial class Program
 	{
 		Process glslc = Process.Start(
 			"glslc",
-			$"\"{glslPath}\" -o \"{outputPath}\""
+			$"{glslPath} -o {outputPath}"
 		);
 		glslc.WaitForExit();
 		if (glslc.ExitCode != 0)
@@ -254,7 +254,7 @@ partial class Program
 	{
 		Process spirvcross = Process.Start(
 			"spirv-cross",
-			$"\"{spirvPath}\" --hlsl --shader-model 50 --output \"{outputPath}\""
+			$"{spirvPath} --hlsl --shader-model 50 --output {outputPath}"
 		);
 		spirvcross.WaitForExit();
 		if (spirvcross.ExitCode != 0)

shadercompiler/refreshc.csproj

@@ -4,7 +4,6 @@
     <OutputType>Exe</OutputType>
     <TargetFramework>net7.0</TargetFramework>
     <TargetName>refreshc</TargetName>
-	<PublishAot>true</PublishAot>
   </PropertyGroup>
 
 </Project>

src/Refresh.c

@@ -438,13 +438,8 @@ void Refresh_SetTextureDataYUV(
 	uint32_t yHeight,
 	uint32_t uvWidth,
 	uint32_t uvHeight,
-	void *yDataPtr,
-	void *uDataPtr,
-	void *vDataPtr,
-	uint32_t yDataLength,
-	uint32_t uvDataLength,
-	uint32_t yStride,
-	uint32_t uvStride
+	void* data,
+	uint32_t dataLength
 ) {
 	NULL_RETURN(device);
 	device->SetTextureDataYUV(
@@ -457,13 +452,8 @@ void Refresh_SetTextureDataYUV(
 		yHeight,
 		uvWidth,
 		uvHeight,
-		yDataPtr,
-		uDataPtr,
-		vDataPtr,
-		yDataLength,
-		uvDataLength,
-		yStride,
-		uvStride
+		data,
+		dataLength
 	);
 }
 
@@ -892,23 +882,14 @@ void Refresh_SetSwapchainPresentMode(
 
 void Refresh_Submit(
 	Refresh_Device *device,
-	Refresh_CommandBuffer *commandBuffer
+	uint32_t commandBufferCount,
+	Refresh_CommandBuffer **pCommandBuffers
 ) {
 	NULL_RETURN(device);
 	device->Submit(
 		device->driverData,
-		commandBuffer
-	);
-}
-
-Refresh_Fence* Refresh_SubmitAndAcquireFence(
-	Refresh_Device *device,
-	Refresh_CommandBuffer *commandBuffer
-) {
-	NULL_RETURN_NULL(device);
-	return device->SubmitAndAcquireFence(
-		device->driverData,
-		commandBuffer
+		commandBufferCount,
+		pCommandBuffers
 	);
 }
 
@@ -921,44 +902,4 @@ void Refresh_Wait(
 	);
 }
 
-void Refresh_WaitForFences(
-	Refresh_Device *device,
-	uint8_t waitAll,
-	uint32_t fenceCount,
-	Refresh_Fence **pFences
-) {
-	NULL_RETURN(device);
-	device->WaitForFences(
-		device->driverData,
-		waitAll,
-		fenceCount,
-		pFences
-	);
-}
-
-int Refresh_QueryFence(
-	Refresh_Device *device,
-	Refresh_Fence *fence
-) {
-	if (device == NULL) {
-		return 0;
-	}
-
-	return device->QueryFence(
-		device->driverData,
-		fence
-	);
-}
-
-void Refresh_ReleaseFence(
-	Refresh_Device *device,
-	Refresh_Fence *fence
-) {
-	NULL_RETURN(device);
-	device->ReleaseFence(
-		device->driverData,
-		fence
-	);
-}
-
 /* vim: set noexpandtab shiftwidth=8 tabstop=8: */

src/Refresh_Driver.h

@@ -286,13 +286,8 @@ struct Refresh_Device
 		uint32_t yHeight,
 		uint32_t uvWidth,
 		uint32_t uvHeight,
-		void *yDataPtr,
-		void *uDataPtr,
-		void *vDataPtr,
-		uint32_t yDataLength,
-		uint32_t uvDataLength,
-		uint32_t yStride,
-		uint32_t uvStride
+		void* data,
+		uint32_t dataLength
 	);
 
 	void(*CopyTextureToTexture)(
@@ -499,35 +494,14 @@ struct Refresh_Device
 
 	void(*Submit)(
 		Refresh_Renderer *driverData,
-		Refresh_CommandBuffer *commandBuffer
-	);
-
-	Refresh_Fence* (*SubmitAndAcquireFence)(
-		Refresh_Renderer *driverData,
-		Refresh_CommandBuffer *commandBuffer
+		uint32_t commandBufferCount,
+		Refresh_CommandBuffer **pCommandBuffers
 	);
 
 	void(*Wait)(
 		Refresh_Renderer *driverData
 	);
 
-	void (*WaitForFences)(
-		Refresh_Renderer *driverData,
-		uint8_t waitAll,
-		uint32_t fenceCount,
-		Refresh_Fence **pFences
-	);
-
-	int (*QueryFence)(
-		Refresh_Renderer *driverData,
-		Refresh_Fence *fence
-	);
-
-	void (*ReleaseFence)(
-		Refresh_Renderer *driverData,
-		Refresh_Fence *fence
-	);
-
 	/* Opaque pointer for the Driver */
 	Refresh_Renderer *driverData;
 };
@@ -581,11 +555,7 @@ struct Refresh_Device
 	ASSIGN_DRIVER_FUNC(GetSwapchainFormat, name) \
 	ASSIGN_DRIVER_FUNC(SetSwapchainPresentMode, name) \
 	ASSIGN_DRIVER_FUNC(Submit, name) \
-	ASSIGN_DRIVER_FUNC(SubmitAndAcquireFence, name) \
-	ASSIGN_DRIVER_FUNC(Wait, name) \
-	ASSIGN_DRIVER_FUNC(WaitForFences, name) \
-	ASSIGN_DRIVER_FUNC(QueryFence, name) \
-	ASSIGN_DRIVER_FUNC(ReleaseFence, name)
+	ASSIGN_DRIVER_FUNC(Wait, name)
 
 typedef struct Refresh_Driver
 {

src/Refresh_Driver_Template.txt

@@ -268,19 +268,6 @@ static void TEMPLATE_DrawPrimitives(
 	NOT_IMPLEMENTED
 }
 
-static void TEMPLATE_DrawPrimitivesIndirect(
-	Refresh_Renderer *driverData,
-	Refresh_CommandBuffer *commandBuffer,
-	Refresh_Buffer *buffer,
-	uint32_t offsetInBytes,
-	uint32_t drawCount,
-	uint32_t stride,
-	uint32_t vertexParamOffset,
-	uint32_t fragmentParamOffset
-) {
-	NOT_IMPLEMENTED
-}
-
 static void TEMPLATE_DispatchCompute(
 	Refresh_Renderer *device,
 	Refresh_CommandBuffer *commandBuffer,
@@ -360,13 +347,8 @@ static void TEMPLATE_SetTextureDataYUV(
 	uint32_t yHeight,
 	uint32_t uvWidth,
 	uint32_t uvHeight,
-	void *yDataPtr,
-	void *uDataPtr,
-	void *vDataPtr,
-	uint32_t yDataLength,
-	uint32_t uvDataLength,
-	uint32_t yStride,
-	uint32_t uvStride
+	void* data,
+	uint32_t dataLength
 ) {
 	NOT_IMPLEMENTED
 }
@@ -503,15 +485,10 @@ static void TEMPLATE_QueueDestroyGraphicsPipeline(
 
 /* Graphics State */
 
-static Refresh_CommandBuffer* TEMPLATE_AcquireCommandBuffer(
-	Refresh_Renderer *driverData
-) {
-	NOT_IMPLEMENTED
-}
-
 static void TEMPLATE_BeginRenderPass(
 	Refresh_Renderer *driverData,
 	Refresh_CommandBuffer *commandBuffer,
+	Refresh_Rect *renderArea,
 	Refresh_ColorAttachmentInfo *colorAttachmentInfos,
 	uint32_t colorAttachmentCount,
 	Refresh_DepthStencilAttachmentInfo *depthStencilAttachmentInfo
@@ -571,8 +548,6 @@ static void TEMPLATE_BindIndexBuffer(
 	NOT_IMPLEMENTED
 }
 
-/* Compute State */
-
 static void TEMPLATE_BindComputePipeline(
 	Refresh_Renderer *driverData,
 	Refresh_CommandBuffer *commandBuffer,
@@ -597,24 +572,14 @@ static void TEMPLATE_BindComputeTextures(
 	NOT_IMPLEMENTED
 }
 
-/* Window and Swapchain Management */
-
-static uint8_t TEMPLATE_ClaimWindow(
+static Refresh_CommandBuffer* TEMPLATE_AcquireCommandBuffer(
 	Refresh_Renderer *driverData,
-	void *windowHandle,
-	Refresh_PresentMode presentMode
+	uint8_t fixed
 ) {
 	NOT_IMPLEMENTED
 }
 
-static void TEMPLATE_UnclaimWindow(
-	Refresh_Renderer *driverData,
-	void *windowHandle
-) {
-	NOT_IMPLEMENTED
-}
-
-static Refresh_Texture* TEMPLATE_AcquireSwapchainTexture(
+Refresh_Texture* TEMPLATE_AcquireSwapchainTexture(
 	Refresh_Renderer *driverData,
 	Refresh_CommandBuffer *commandBuffer,
 	void *windowHandle,
@@ -624,33 +589,17 @@ static Refresh_Texture* TEMPLATE_AcquireSwapchainTexture(
 	NOT_IMPLEMENTED
 }
 
-static Refresh_TextureFormat TEMPLATE_GetSwapchainFormat(
+Refresh_TextureFormat TEMPLATE_GetSwapchainFormat(
 	Refresh_Renderer *driverData,
 	void *windowHandle
 ) {
 	NOT_IMPLEMENTED
 }
 
-static void TEMPLATE_SetSwapchainPresentMode(
-	Refresh_Renderer *driverData,
-	void *windowHandle,
-	Refresh_PresentMode presentMode
-) {
-	NOT_IMPLEMENTED
-}
-
-/* Submission and Fences */
-
 static void TEMPLATE_Submit(
 	Refresh_Renderer *driverData,
-	Refresh_CommandBuffer *commandBuffer
-) {
-	NOT_IMPLEMENTED
-}
-
-static Refresh_Fence* TEMPLATE_SubmitAndAcquireFence(
-	Refresh_Renderer *driverData,
-	Refresh_CommandBuffer *commandBuffer
+	uint32_t commandBufferCount,
+	Refresh_CommandBuffer **pCommandBuffers
 ) {
 	NOT_IMPLEMENTED
 }
@@ -661,38 +610,8 @@ static void TEMPLATE_Wait(
 	NOT_IMPLEMENTED
 }
 
-static void TEMPLATE_WaitForFences(
-	Refresh_Renderer *driverData,
-	uint8_t waitAll,
-	uint32_t fenceCount,
-	Refresh_Fence **pFences
-) {
-	NOT_IMPLEMENTED
-}
-
-static int TEMPLATE_QueryFence(
-	Refresh_Renderer *driverData,
-	Refresh_Fence *fence
-) {
-	NOT_IMPLEMENTED
-}
-
-static void TEMPLATE_ReleaseFence(
-	Refresh_Renderer *driverData,
-	Refresh_Fence *fence
-) {
-	NOT_IMPLEMENTED
-}
-
-/* Device Creation */
-
-static uint8_t TEMPLATE_PrepareDriver(
-	uint32_t *flags
-) {
-	NOT_IMPLEMENTED
-}
-
 static Refresh_Device* TEMPLATE_CreateDevice(
+	Refresh_PresentationParameters *presentationParameters,
 	uint8_t debugMode
 ) {
 	NOT_IMPLEMENTED
@@ -700,7 +619,6 @@ static Refresh_Device* TEMPLATE_CreateDevice(
 
 Refresh_Driver TEMPLATEDriver = {
 	"TEMPLATE",
-	TEMPLATE_PrepareDriver,
 	TEMPLATE_CreateDevice
 };
 

src/Refresh_Driver_Vulkan_vkfuncs.h

@@ -89,7 +89,6 @@ VULKAN_DEVICE_FUNCTION(BaseVK, void, vkCmdClearDepthStencilImage, (VkCommandBuff
 VULKAN_DEVICE_FUNCTION(BaseVK, void, vkCmdCopyBuffer, (VkCommandBuffer commandBuffer, VkBuffer srcBuffer, VkBuffer dstBuffer, uint32_t regionCount, const VkBufferCopy* pRegions))
 VULKAN_DEVICE_FUNCTION(BaseVK, void, vkCmdCopyBufferToImage, (VkCommandBuffer commandBuffer, VkBuffer srcBuffer, VkImage dstImage, VkImageLayout dstImageLayout, uint32_t regionCount, const VkBufferImageCopy *pRegions))
 VULKAN_DEVICE_FUNCTION(BaseVK, void, vkCmdCopyImageToBuffer, (VkCommandBuffer commandBuffer, VkImage srcImage, VkImageLayout srcImageLayout, VkBuffer dstBuffer, uint32_t regionCount, const VkBufferImageCopy *pRegions))
-VULKAN_DEVICE_FUNCTION(BaseVK, void, vkCmdCopyImage, (VkCommandBuffer commandBuffer, VkImage srcImage, VkImageLayout srcImageLayout, VkImage dstImage, VkImageLayout dstImageLayout, uint32_t regionCount, const VkImageCopy* pRegions))
 VULKAN_DEVICE_FUNCTION(BaseVK, void, vkCmdDispatch, (VkCommandBuffer commandBuffer, uint32_t groupCountX, uint32_t groupCountY, uint32_t groupCountZ))
 VULKAN_DEVICE_FUNCTION(BaseVK, void, vkCmdDraw, (VkCommandBuffer commandBuffer, uint32_t vertexCount, uint32_t instanceCount, uint32_t firstVertex, uint32_t firstInstance))
 VULKAN_DEVICE_FUNCTION(BaseVK, void, vkCmdDrawIndexed, (VkCommandBuffer commandBuffer, uint32_t indexCount, uint32_t instanceCount, uint32_t firstIndex, int32_t vertexOffset, uint32_t firstInstance))

src/Refresh_Image.c

@@ -48,6 +48,7 @@
 #define floorf SDL_floorf
 #define ldexp SDL_scalbn
 #define pow SDL_pow
+#define strtol SDL_strtol
 
 #ifdef memcmp
 #undef memcmp
@@ -73,10 +74,13 @@
 #undef strlen
 #endif
 #define strlen SDL_strlen
+#ifdef strncmp
+#undef strncmp
+#endif
+#define strncmp SDL_strncmp
 
 /* These are per the Texture2D.FromStream spec */
 #define STBI_ONLY_PNG
-#define STBI_ONLY_QOI
 
 /* These are per the Texture2D.SaveAs* spec */
 #define STBIW_ONLY_PNG
@@ -136,7 +140,6 @@ SDL_SIMDRealloc(void *mem, const size_t len)
 #endif
 
 #define STB_IMAGE_STATIC
-#define STBI_NO_HDR
 #define STBI_ASSERT SDL_assert
 #define STBI_MALLOC SDL_SIMDAlloc
 #define STBI_REALLOC SDL_SIMDRealloc
@@ -187,72 +190,51 @@ static unsigned char* dgibson_stbi_zlib_compress(
 /* Image Read API */
 
 uint8_t* Refresh_Image_Load(
-	uint8_t *bufferPtr,
-	int32_t bufferLength,
+	char const *filename,
 	int32_t *w,
 	int32_t *h,
-	int32_t *len
+	int32_t *numChannels
 ) {
-	uint8_t* result;
-	uint8_t* pixels;
-	int32_t format;
-	int32_t i;
-
-	result = stbi_load_from_memory(
-		bufferPtr,
-		bufferLength,
-		w,
-		h,
-		&format,
-		STBI_rgb_alpha
-	);
-
-	if (result == NULL)
-	{
-		SDL_LogWarn(SDL_LOG_CATEGORY_ERROR, "Image loading failed: %s", stbi_failure_reason());
-	}
-
-	/* Ensure that the alpha pixels are... well, actual alpha.
-	 * You think this looks stupid, but be assured: Your paint program is
-	 * almost certainly even stupider.
-	 * -flibit
-	 */
-	pixels = result;
-	*len = (*w) * (*h) *4;
-	for (i = 0; i < *len; i += 4, pixels += 4)
-	{
-		if (pixels[3] == 0)
-		{
-			pixels[0] = 0;
-			pixels[1] = 1;
-			pixels[2] = 2;
-		}
-	}
-
-	return result;
+	return stbi_load(filename, w, h, numChannels, STBI_rgb_alpha);
 }
 
 void Refresh_Image_Free(uint8_t *mem)
 {
-	SDL_SIMDFree(mem);
+	stbi_image_free(mem);
 }
 
 /* Image Write API */
 
 void Refresh_Image_SavePNG(
 	const char *filename,
-	uint8_t* data,
 	int32_t w,
-	int32_t h
+	int32_t h,
+	uint8_t bgra,
+	uint8_t *data
 ) {
-	stbi_write_png(
-		filename,
-		w,
-		h,
-		4,
-		data,
-		w * 4
-	);
+	uint32_t i;
+	uint8_t *bgraData;
+
+	if (bgra)
+	{
+		bgraData = SDL_malloc(w * h * 4);
+
+		for (i = 0; i < w * h * 4; i += 4)
+		{
+			bgraData[i]     = data[i + 2];
+			bgraData[i + 1] = data[i + 1];
+			bgraData[i + 2] = data[i];
+			bgraData[i + 3] = data[i + 3];
+		}
+
+		stbi_write_png(filename, w, h, 4, bgraData, w * 4);
+
+		SDL_free(bgraData);
+	}
+	else
+	{
+		stbi_write_png(filename, w, h, 4, data, w * 4);
+	}
 }
 
 /* vim: set noexpandtab shiftwidth=8 tabstop=8: */