// Copyright 2016 The SwiftShader Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//    http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "Blitter.hpp"

#include "Pipeline/ShaderCore.hpp"
#include "Reactor/Reactor.hpp"
#include "System/Half.hpp"
#include "System/Memory.hpp"
#include "Vulkan/VkDebug.hpp"
#include "Vulkan/VkImage.hpp"
#include "Vulkan/VkBuffer.hpp"

#include <utility>

namespace sw
{
	Blitter::Blitter() :
		blitMutex(),
		blitCache(1024),
		cornerUpdateMutex(),
		cornerUpdateCache(64) // We only need one of these per format
	{
	}

	Blitter::~Blitter()
	{
	}

	void Blitter::clear(void *pixel, vk::Format format, vk::Image *dest, const vk::Format& viewFormat, const VkImageSubresourceRange& subresourceRange, const VkRect2D* renderArea)
	{
		VkImageAspectFlagBits aspect = static_cast<VkImageAspectFlagBits>(subresourceRange.aspectMask);
		vk::Format dstFormat = viewFormat.getAspectFormat(aspect);
		if(dstFormat == VK_FORMAT_UNDEFINED)
		{
			return;
		}

		if(fastClear(pixel, format, dest, dstFormat, subresourceRange, renderArea))
		{
			return;
		}

		State state(format, dstFormat, 1, dest->getSampleCountFlagBits(), { 0xF });
		auto blitRoutine = getBlitRoutine(state);
		if(!blitRoutine)
		{
			return;
		}

		void(*blitFunction)(const BlitData *data) = (void(*)(const BlitData*))blitRoutine->getEntry();

		VkImageSubresourceLayers subresLayers =
		{
			subresourceRange.aspectMask,
			subresourceRange.baseMipLevel,
			subresourceRange.baseArrayLayer,
			1
		};

		uint32_t lastMipLevel = dest->getLastMipLevel(subresourceRange);
		uint32_t lastLayer = dest->getLastLayerIndex(subresourceRange);

		VkRect2D area = { { 0, 0 }, { 0, 0 } };
		if(renderArea)
		{
			ASSERT(subresourceRange.levelCount == 1);
			area = *renderArea;
		}

		for(; subresLayers.mipLevel <= lastMipLevel; subresLayers.mipLevel++)
		{
			VkExtent3D extent = dest->getMipLevelExtent(aspect, subresLayers.mipLevel);
			if(!renderArea)
			{
				area.extent.width = extent.width;
				area.extent.height = extent.height;
			}

			BlitData data =
			{
				pixel, nullptr, // source, dest

				format.bytes(),                                       // sPitchB
				dest->rowPitchBytes(aspect, subresLayers.mipLevel),   // dPitchB
				0,                                                    // sSliceB (unused in clear operations)
				dest->slicePitchBytes(aspect, subresLayers.mipLevel), // dSliceB

				0.5f, 0.5f, 0.0f, 0.0f, // x0, y0, w, h

				area.offset.y, static_cast<int>(area.offset.y + area.extent.height), // y0d, y1d
				area.offset.x, static_cast<int>(area.offset.x + area.extent.width),  // x0d, x1d

				0, 0, // sWidth, sHeight
			};

			if (renderArea && dest->is3DSlice())
			{
				// Reinterpret layers as depth slices
				subresLayers.baseArrayLayer = 0;
				subresLayers.layerCount = 1;
				for (uint32_t depth = subresourceRange.baseArrayLayer; depth <= lastLayer; depth++)
				{
					data.dest = dest->getTexelPointer({0, 0, static_cast<int32_t>(depth)}, subresLayers);
					blitFunction(&data);
				}
			}
			else
			{
				for(subresLayers.baseArrayLayer = subresourceRange.baseArrayLayer; subresLayers.baseArrayLayer <= lastLayer; subresLayers.baseArrayLayer++)
				{
					for(uint32_t depth = 0; depth < extent.depth; depth++)
					{
						data.dest = dest->getTexelPointer({ 0, 0, static_cast<int32_t>(depth) }, subresLayers);

						blitFunction(&data);
					}
				}
			}
		}
	}

	bool Blitter::fastClear(void *pixel, vk::Format format, vk::Image *dest, const vk::Format& viewFormat, const VkImageSubresourceRange& subresourceRange, const VkRect2D* renderArea)
	{
		if(format != VK_FORMAT_R32G32B32A32_SFLOAT)
		{
			return false;
		}

		float *color = (float*)pixel;
		float r = color[0];
		float g = color[1];
		float b = color[2];
		float a = color[3];

		uint32_t packed;

		VkImageAspectFlagBits aspect = static_cast<VkImageAspectFlagBits>(subresourceRange.aspectMask);
		switch(viewFormat)
		{
		case VK_FORMAT_R5G6B5_UNORM_PACK16:
			packed = ((uint16_t)(31 * b + 0.5f) << 0) |
			         ((uint16_t)(63 * g + 0.5f) << 5) |
			         ((uint16_t)(31 * r + 0.5f) << 11);
			break;
		case VK_FORMAT_B5G6R5_UNORM_PACK16:
			packed = ((uint16_t)(31 * r + 0.5f) << 0) |
			         ((uint16_t)(63 * g + 0.5f) << 5) |
			         ((uint16_t)(31 * b + 0.5f) << 11);
			break;
		case VK_FORMAT_A8B8G8R8_UINT_PACK32:
		case VK_FORMAT_A8B8G8R8_UNORM_PACK32:
		case VK_FORMAT_R8G8B8A8_UNORM:
			packed = ((uint32_t)(255 * a + 0.5f) << 24) |
			         ((uint32_t)(255 * b + 0.5f) << 16) |
			         ((uint32_t)(255 * g + 0.5f) << 8) |
			         ((uint32_t)(255 * r + 0.5f) << 0);
			break;
		case VK_FORMAT_B8G8R8A8_UNORM:
			packed = ((uint32_t)(255 * a + 0.5f) << 24) |
			         ((uint32_t)(255 * r + 0.5f) << 16) |
			         ((uint32_t)(255 * g + 0.5f) << 8) |
			         ((uint32_t)(255 * b + 0.5f) << 0);
			break;
		case VK_FORMAT_B10G11R11_UFLOAT_PACK32:
			packed = R11G11B10F(color);
			break;
		case VK_FORMAT_E5B9G9R9_UFLOAT_PACK32:
			packed = RGB9E5(color);
			break;
		default:
			return false;
		}

		VkImageSubresourceLayers subresLayers =
		{
			subresourceRange.aspectMask,
			subresourceRange.baseMipLevel,
			subresourceRange.baseArrayLayer,
			1
		};
		uint32_t lastMipLevel = dest->getLastMipLevel(subresourceRange);
		uint32_t lastLayer = dest->getLastLayerIndex(subresourceRange);

		VkRect2D area = { { 0, 0 }, { 0, 0 } };
		if(renderArea)
		{
			ASSERT(subresourceRange.levelCount == 1);
			area = *renderArea;
		}

		for(; subresLayers.mipLevel <= lastMipLevel; subresLayers.mipLevel++)
		{
			int rowPitchBytes = dest->rowPitchBytes(aspect, subresLayers.mipLevel);
			int slicePitchBytes = dest->slicePitchBytes(aspect, subresLayers.mipLevel);
			VkExtent3D extent = dest->getMipLevelExtent(aspect, subresLayers.mipLevel);
			if(!renderArea)
			{
				area.extent.width = extent.width;
				area.extent.height = extent.height;
			}
			if(dest->is3DSlice())
			{
				extent.depth = 1; // The 3D image is instead interpreted as a 2D image with layers
			}

			for(subresLayers.baseArrayLayer = subresourceRange.baseArrayLayer; subresLayers.baseArrayLayer <= lastLayer; subresLayers.baseArrayLayer++)
			{
				for(uint32_t depth = 0; depth < extent.depth; depth++)
				{
					uint8_t *slice = (uint8_t*)dest->getTexelPointer(
						{ area.offset.x, area.offset.y, static_cast<int32_t>(depth) }, subresLayers);

					for(int j = 0; j < dest->getSampleCountFlagBits(); j++)
					{
						uint8_t *d = slice;

						switch(viewFormat.bytes())
						{
						case 2:
							for(uint32_t i = 0; i < area.extent.height; i++)
							{
								ASSERT(d < dest->end());
								sw::clear((uint16_t*)d, static_cast<uint16_t>(packed), area.extent.width);
								d += rowPitchBytes;
							}
							break;
						case 4:
							for(uint32_t i = 0; i < area.extent.height; i++)
							{
								ASSERT(d < dest->end());
								sw::clear((uint32_t*)d, packed, area.extent.width);
								d += rowPitchBytes;
							}
							break;
						default:
							assert(false);
						}

						slice += slicePitchBytes;
					}
				}
			}
		}

		return true;
	}

	bool Blitter::read(Float4 &c, Pointer<Byte> element, const State &state)
	{
		c = Float4(0.0f, 0.0f, 0.0f, 1.0f);

		switch(state.sourceFormat)
		{
		case VK_FORMAT_B4G4R4A4_UNORM_PACK16:
			c.w = Float(Int(*Pointer<Byte>(element)) & Int(0xF));
			c.x = Float((Int(*Pointer<Byte>(element)) >> 4) & Int(0xF));
			c.y = Float(Int(*Pointer<Byte>(element + 1)) & Int(0xF));
			c.z = Float((Int(*Pointer<Byte>(element + 1)) >> 4) & Int(0xF));
			break;
		case VK_FORMAT_R8_SINT:
		case VK_FORMAT_R8_SNORM:
			c.x = Float(Int(*Pointer<SByte>(element)));
			c.w = float(0x7F);
			break;
		case VK_FORMAT_R8_UNORM:
		case VK_FORMAT_R8_UINT:
		case VK_FORMAT_R8_SRGB:
			c.x = Float(Int(*Pointer<Byte>(element)));
			c.w = float(0xFF);
			break;
		case VK_FORMAT_R16_SINT:
			c.x = Float(Int(*Pointer<Short>(element)));
			c.w = float(0x7FFF);
			break;
		case VK_FORMAT_R16_UINT:
			c.x = Float(Int(*Pointer<UShort>(element)));
			c.w = float(0xFFFF);
			break;
		case VK_FORMAT_R32_SINT:
			c.x = Float(*Pointer<Int>(element));
			c.w = float(0x7FFFFFFF);
			break;
		case VK_FORMAT_R32_UINT:
			c.x = Float(*Pointer<UInt>(element));
			c.w = float(0xFFFFFFFF);
			break;
		case VK_FORMAT_B8G8R8A8_SRGB:
		case VK_FORMAT_B8G8R8A8_UNORM:
			c = Float4(*Pointer<Byte4>(element)).zyxw;
			break;
		case VK_FORMAT_A8B8G8R8_SINT_PACK32:
		case VK_FORMAT_R8G8B8A8_SINT:
		case VK_FORMAT_A8B8G8R8_SNORM_PACK32:
		case VK_FORMAT_R8G8B8A8_SNORM:
			c = Float4(*Pointer<SByte4>(element));
			break;
		case VK_FORMAT_A8B8G8R8_UINT_PACK32:
		case VK_FORMAT_A8B8G8R8_UNORM_PACK32:
		case VK_FORMAT_R8G8B8A8_UNORM:
		case VK_FORMAT_R8G8B8A8_UINT:
		case VK_FORMAT_A8B8G8R8_SRGB_PACK32:
		case VK_FORMAT_R8G8B8A8_SRGB:
			c = Float4(*Pointer<Byte4>(element));
			break;
		case VK_FORMAT_R16G16B16A16_SINT:
			c = Float4(*Pointer<Short4>(element));
			break;
		case VK_FORMAT_R16G16B16A16_UNORM:
		case VK_FORMAT_R16G16B16A16_UINT:
			c = Float4(*Pointer<UShort4>(element));
			break;
		case VK_FORMAT_R32G32B32A32_SINT:
			c = Float4(*Pointer<Int4>(element));
			break;
		case VK_FORMAT_R32G32B32A32_UINT:
			c = Float4(*Pointer<UInt4>(element));
			break;
		case VK_FORMAT_R8G8_SINT:
		case VK_FORMAT_R8G8_SNORM:
			c.x = Float(Int(*Pointer<SByte>(element + 0)));
			c.y = Float(Int(*Pointer<SByte>(element + 1)));
			c.w = float(0x7F);
			break;
		case VK_FORMAT_R8G8_UNORM:
		case VK_FORMAT_R8G8_UINT:
		case VK_FORMAT_R8G8_SRGB:
			c.x = Float(Int(*Pointer<Byte>(element + 0)));
			c.y = Float(Int(*Pointer<Byte>(element + 1)));
			c.w = float(0xFF);
			break;
		case VK_FORMAT_R16G16_SINT:
			c.x = Float(Int(*Pointer<Short>(element + 0)));
			c.y = Float(Int(*Pointer<Short>(element + 2)));
			c.w = float(0x7FFF);
			break;
		case VK_FORMAT_R16G16_UNORM:
		case VK_FORMAT_R16G16_UINT:
			c.x = Float(Int(*Pointer<UShort>(element + 0)));
			c.y = Float(Int(*Pointer<UShort>(element + 2)));
			c.w = float(0xFFFF);
			break;
		case VK_FORMAT_R32G32_SINT:
			c.x = Float(*Pointer<Int>(element + 0));
			c.y = Float(*Pointer<Int>(element + 4));
			c.w = float(0x7FFFFFFF);
			break;
		case VK_FORMAT_R32G32_UINT:
			c.x = Float(*Pointer<UInt>(element + 0));
			c.y = Float(*Pointer<UInt>(element + 4));
			c.w = float(0xFFFFFFFF);
			break;
		case VK_FORMAT_R32G32B32A32_SFLOAT:
			c = *Pointer<Float4>(element);
			break;
		case VK_FORMAT_R32G32_SFLOAT:
			c.x = *Pointer<Float>(element + 0);
			c.y = *Pointer<Float>(element + 4);
			break;
		case VK_FORMAT_R32_SFLOAT:
			c.x = *Pointer<Float>(element);
			break;
		case VK_FORMAT_R16G16B16A16_SFLOAT:
			c.w = Float(*Pointer<Half>(element + 6));
		case VK_FORMAT_R16G16B16_SFLOAT:
			c.z = Float(*Pointer<Half>(element + 4));
		case VK_FORMAT_R16G16_SFLOAT:
			c.y = Float(*Pointer<Half>(element + 2));
		case VK_FORMAT_R16_SFLOAT:
			c.x = Float(*Pointer<Half>(element));
			break;
		case VK_FORMAT_B10G11R11_UFLOAT_PACK32:
			// 10 (or 11) bit float formats are unsigned formats with a 5 bit exponent and a 5 (or 6) bit mantissa.
			// Since the Half float format also has a 5 bit exponent, we can convert these formats to half by
			// copy/pasting the bits so the the exponent bits and top mantissa bits are aligned to the half format.
			// In this case, we have:
			//              B B B B B B B B B B G G G G G G G G G G G R R R R R R R R R R R
			// 1st Short:                                  |xxxxxxxxxx---------------------|
			// 2nd Short:                  |xxxx---------------------xxxxxx|
			// 3rd Short: |--------------------xxxxxxxxxxxx|
			// These memory reads overlap, but each of them contains an entire channel, so we can read this without
			// any int -> short conversion.
			c.x = Float(As<Half>((*Pointer<UShort>(element + 0) & UShort(0x07FF)) << UShort(4)));
			c.y = Float(As<Half>((*Pointer<UShort>(element + 1) & UShort(0x3FF8)) << UShort(1)));
			c.z = Float(As<Half>((*Pointer<UShort>(element + 2) & UShort(0xFFC0)) >> UShort(1)));
			break;
		case VK_FORMAT_E5B9G9R9_UFLOAT_PACK32:
			// This type contains a common 5 bit exponent (E) and a 9 bit the mantissa for R, G and B.
			c.x = Float(*Pointer<UInt>(element) & UInt(0x000001FF));         // R's mantissa (bits 0-8)
			c.y = Float((*Pointer<UInt>(element) & UInt(0x0003FE00)) >> 9);  // G's mantissa (bits 9-17)
			c.z = Float((*Pointer<UInt>(element) & UInt(0x07FC0000)) >> 18); // B's mantissa (bits 18-26)
			c *= Float4(
				// 2^E, using the exponent (bits 27-31) and treating it as an unsigned integer value
				Float(UInt(1) << ((*Pointer<UInt>(element) & UInt(0xF8000000)) >> 27)) *
				// Since the 9 bit mantissa values currently stored in RGB were converted straight
				// from int to float (in the [0, 1<<9] range instead of the [0, 1] range), they
				// are (1 << 9) times too high.
				// Also, the exponent has 5 bits and we compute the exponent bias of floating point
				// formats using "2^(k-1) - 1", so, in this case, the exponent bias is 2^(5-1)-1 = 15
				// Exponent bias (15) + number of mantissa bits per component (9) = 24
				Float(1.0f / (1 << 24)));
			c.w = 1.0f;
			break;
		case VK_FORMAT_R5G6B5_UNORM_PACK16:
			c.x = Float(Int((*Pointer<UShort>(element) & UShort(0xF800)) >> UShort(11)));
			c.y = Float(Int((*Pointer<UShort>(element) & UShort(0x07E0)) >> UShort(5)));
			c.z = Float(Int(*Pointer<UShort>(element) & UShort(0x001F)));
			break;
		case VK_FORMAT_A1R5G5B5_UNORM_PACK16:
			c.w = Float(Int((*Pointer<UShort>(element) & UShort(0x8000)) >> UShort(15)));
			c.x = Float(Int((*Pointer<UShort>(element) & UShort(0x7C00)) >> UShort(10)));
			c.y = Float(Int((*Pointer<UShort>(element) & UShort(0x03E0)) >> UShort(5)));
			c.z = Float(Int(*Pointer<UShort>(element) & UShort(0x001F)));
			break;
		case VK_FORMAT_A2B10G10R10_UNORM_PACK32:
		case VK_FORMAT_A2B10G10R10_UINT_PACK32:
			c.x = Float(Int((*Pointer<UInt>(element) & UInt(0x000003FF))));
			c.y = Float(Int((*Pointer<UInt>(element) & UInt(0x000FFC00)) >> 10));
			c.z = Float(Int((*Pointer<UInt>(element) & UInt(0x3FF00000)) >> 20));
			c.w = Float(Int((*Pointer<UInt>(element) & UInt(0xC0000000)) >> 30));
			break;
		case VK_FORMAT_D16_UNORM:
			c.x = Float(Int((*Pointer<UShort>(element))));
			break;
		case VK_FORMAT_D24_UNORM_S8_UINT:
		case VK_FORMAT_X8_D24_UNORM_PACK32:
			c.x = Float(Int((*Pointer<UInt>(element) & UInt(0xFFFFFF00)) >> 8));
			break;
		case VK_FORMAT_D32_SFLOAT:
		case VK_FORMAT_D32_SFLOAT_S8_UINT:
			c.x = *Pointer<Float>(element);
			break;
		case VK_FORMAT_S8_UINT:
			c.x = Float(Int(*Pointer<Byte>(element)));
			break;
		default:
			return false;
		}

		return true;
	}

	bool Blitter::write(Float4 &c, Pointer<Byte> element, const State &state)
	{
		bool writeR = state.writeRed;
		bool writeG = state.writeGreen;
		bool writeB = state.writeBlue;
		bool writeA = state.writeAlpha;
		bool writeRGBA = writeR && writeG && writeB && writeA;

		switch(state.destFormat)
		{
		case VK_FORMAT_R4G4_UNORM_PACK8:
			if(writeR | writeG)
			{
				if(!writeR)
				{
					*Pointer<Byte>(element) = (Byte(RoundInt(Float(c.y))) & Byte(0xF)) |
				                              (*Pointer<Byte>(element) & Byte(0xF0));
				}
				else if(!writeG)
				{
					*Pointer<Byte>(element) = (*Pointer<Byte>(element) & Byte(0xF)) |
				                              (Byte(RoundInt(Float(c.x))) << Byte(4));
				}
				else
				{
					*Pointer<Byte>(element) = (Byte(RoundInt(Float(c.y))) & Byte(0xF)) |
				                              (Byte(RoundInt(Float(c.x))) << Byte(4));
				}
			}
			break;
		case VK_FORMAT_R4G4B4A4_UNORM_PACK16:
			if(writeR || writeG || writeB || writeA)
			{
				*Pointer<UShort>(element) = (writeR ? ((UShort(RoundInt(Float(c.x))) & UShort(0xF)) << UShort(12)) :
				                                      (*Pointer<UShort>(element) & UShort(0x000F))) |
				                            (writeG ? ((UShort(RoundInt(Float(c.y))) & UShort(0xF)) << UShort(8)) :
				                                      (*Pointer<UShort>(element) & UShort(0x00F0))) |
				                            (writeB ? ((UShort(RoundInt(Float(c.z))) & UShort(0xF)) << UShort(4)) :
			                                          (*Pointer<UShort>(element) & UShort(0x0F00))) |
			                                (writeA ? (UShort(RoundInt(Float(c.w))) & UShort(0xF)) :
			                                          (*Pointer<UShort>(element) & UShort(0xF000)));
			}
			break;
		case VK_FORMAT_B4G4R4A4_UNORM_PACK16:
			if(writeRGBA)
			{
				*Pointer<UShort>(element) = UShort(RoundInt(Float(c.w)) & Int(0xF)) |
				                            UShort((RoundInt(Float(c.x)) & Int(0xF)) << 4) |
				                            UShort((RoundInt(Float(c.y)) & Int(0xF)) << 8) |
				                            UShort((RoundInt(Float(c.z)) & Int(0xF)) << 12);
			}
			else
			{
				unsigned short mask = (writeA ? 0x000F : 0x0000) |
				                      (writeR ? 0x00F0 : 0x0000) |
				                      (writeG ? 0x0F00 : 0x0000) |
				                      (writeB ? 0xF000 : 0x0000);
				unsigned short unmask = ~mask;
				*Pointer<UShort>(element) = (*Pointer<UShort>(element) & UShort(unmask)) |
				                            ((UShort(RoundInt(Float(c.w)) & Int(0xF)) |
				                              UShort((RoundInt(Float(c.x)) & Int(0xF)) << 4) |
				                              UShort((RoundInt(Float(c.y)) & Int(0xF)) << 8) |
				                              UShort((RoundInt(Float(c.z)) & Int(0xF)) << 12)) & UShort(mask));
			}
			break;
		case VK_FORMAT_B8G8R8A8_SRGB:
		case VK_FORMAT_B8G8R8A8_UNORM:
			if(writeRGBA)
			{
				Short4 c0 = RoundShort4(c.zyxw);
				*Pointer<Byte4>(element) = Byte4(PackUnsigned(c0, c0));
			}
			else
			{
				if(writeB) { *Pointer<Byte>(element + 0) = Byte(RoundInt(Float(c.z))); }
				if(writeG) { *Pointer<Byte>(element + 1) = Byte(RoundInt(Float(c.y))); }
				if(writeR) { *Pointer<Byte>(element + 2) = Byte(RoundInt(Float(c.x))); }
				if(writeA) { *Pointer<Byte>(element + 3) = Byte(RoundInt(Float(c.w))); }
			}
			break;
		case VK_FORMAT_B8G8R8_SNORM:
			if(writeB) { *Pointer<SByte>(element + 0) = SByte(RoundInt(Float(c.z))); }
			if(writeG) { *Pointer<SByte>(element + 1) = SByte(RoundInt(Float(c.y))); }
			if(writeR) { *Pointer<SByte>(element + 2) = SByte(RoundInt(Float(c.x))); }
			break;
		case VK_FORMAT_B8G8R8_UNORM:
		case VK_FORMAT_B8G8R8_SRGB:
			if(writeB) { *Pointer<Byte>(element + 0) = Byte(RoundInt(Float(c.z))); }
			if(writeG) { *Pointer<Byte>(element + 1) = Byte(RoundInt(Float(c.y))); }
			if(writeR) { *Pointer<Byte>(element + 2) = Byte(RoundInt(Float(c.x))); }
			break;
		case VK_FORMAT_A8B8G8R8_UNORM_PACK32:
		case VK_FORMAT_R8G8B8A8_UNORM:
		case VK_FORMAT_A8B8G8R8_SRGB_PACK32:
		case VK_FORMAT_R8G8B8A8_SRGB:
		case VK_FORMAT_A8B8G8R8_UINT_PACK32:
		case VK_FORMAT_R8G8B8A8_UINT:
		case VK_FORMAT_R8G8B8A8_USCALED:
		case VK_FORMAT_A8B8G8R8_USCALED_PACK32:
			if(writeRGBA)
			{
				Short4 c0 = RoundShort4(c);
				*Pointer<Byte4>(element) = Byte4(PackUnsigned(c0, c0));
			}
			else
			{
				if(writeR) { *Pointer<Byte>(element + 0) = Byte(RoundInt(Float(c.x))); }
				if(writeG) { *Pointer<Byte>(element + 1) = Byte(RoundInt(Float(c.y))); }
				if(writeB) { *Pointer<Byte>(element + 2) = Byte(RoundInt(Float(c.z))); }
				if(writeA) { *Pointer<Byte>(element + 3) = Byte(RoundInt(Float(c.w))); }
			}
			break;
		case VK_FORMAT_R32G32B32A32_SFLOAT:
			if(writeRGBA)
			{
				*Pointer<Float4>(element) = c;
			}
			else
			{
				if(writeR) { *Pointer<Float>(element) = c.x; }
				if(writeG) { *Pointer<Float>(element + 4) = c.y; }
				if(writeB) { *Pointer<Float>(element + 8) = c.z; }
				if(writeA) { *Pointer<Float>(element + 12) = c.w; }
			}
			break;
		case VK_FORMAT_R32G32B32_SFLOAT:
			if(writeR) { *Pointer<Float>(element) = c.x; }
			if(writeG) { *Pointer<Float>(element + 4) = c.y; }
			if(writeB) { *Pointer<Float>(element + 8) = c.z; }
			break;
		case VK_FORMAT_R32G32_SFLOAT:
			if(writeR && writeG)
			{
				*Pointer<Float2>(element) = Float2(c);
			}
			else
			{
				if(writeR) { *Pointer<Float>(element) = c.x; }
				if(writeG) { *Pointer<Float>(element + 4) = c.y; }
			}
			break;
		case VK_FORMAT_R32_SFLOAT:
			if(writeR) { *Pointer<Float>(element) = c.x; }
			break;
		case VK_FORMAT_R16G16B16A16_SFLOAT:
			if(writeA) { *Pointer<Half>(element + 6) = Half(c.w); }
		case VK_FORMAT_R16G16B16_SFLOAT:
			if(writeB) { *Pointer<Half>(element + 4) = Half(c.z); }
		case VK_FORMAT_R16G16_SFLOAT:
			if(writeG) { *Pointer<Half>(element + 2) = Half(c.y); }
		case VK_FORMAT_R16_SFLOAT:
			if(writeR) { *Pointer<Half>(element) = Half(c.x); }
			break;
		case VK_FORMAT_B10G11R11_UFLOAT_PACK32:
			{
				// 10 (or 11) bit float formats are unsigned formats with a 5 bit exponent and a 5 (or 6) bit mantissa.
				// Since the 16-bit half-precision float format also has a 5 bit exponent, we can extract these minifloats from them.

				// FIXME(b/138944025): Handle negative values, Inf, and NaN.
				// FIXME(b/138944025): Perform rounding before truncating the mantissa.
				UInt r = (UInt(As<UShort>(Half(c.x))) & 0x00007FF0) >> 4;
				UInt g = (UInt(As<UShort>(Half(c.y))) & 0x00007FF0) << 7;
				UInt b = (UInt(As<UShort>(Half(c.z))) & 0x00007FE0) << 17;

				UInt rgb = r | g | b;

				UInt old = *Pointer<UInt>(element);

				unsigned int mask = (writeR ? 0x000007FF : 0) |
				                    (writeG ? 0x003FF800 : 0) |
				                    (writeB ? 0xFFC00000 : 0);

				*Pointer<UInt>(element) = (rgb & mask) | (old & ~mask);
			}
			break;
		case VK_FORMAT_E5B9G9R9_UFLOAT_PACK32:
			{
				ASSERT(writeRGBA);  // Can't sensibly write just part of this format.

				// Vulkan 1.1.117 section 15.2.1 RGB to Shared Exponent Conversion

				constexpr int N = 9;       // number of mantissa bits per component
				constexpr int B = 15;      // exponent bias
				constexpr int E_max = 31;  // maximum possible biased exponent value

				// Maximum representable value.
				constexpr float sharedexp_max = ((static_cast<float>(1 << N) - 1) / static_cast<float>(1 << N)) * static_cast<float>(1 << (E_max - B));

				// Clamp components to valid range.
				Float red_c = Max(0, Min(sharedexp_max, c.x));
				Float green_c = Max(0, Min(sharedexp_max, c.y));
				Float blue_c = Max(0, Min(sharedexp_max, c.z));

				// We're reducing the mantissa to 9 bits, so we must round up if the next
				// bit is 1. In other words add 0.5 to the new mantissa's position and
				// allow overflow into the exponent so we can scale correctly.
				constexpr int half = 1 << (23 - N);
				Float red_r = As<Float>(As<Int>(red_c) + half);
				Float green_r = As<Float>(As<Int>(green_c) + half);
				Float blue_r = As<Float>(As<Int>(blue_c) + half);

				// The largest component determines the shared exponent. It can't be lower
				// than 0 (after bias subtraction) so also limit to the mimimum representable.
				constexpr float min_s = 0.5f / (1 << B);
				Float max_s = Max(Max(red_r, green_r), Max(blue_r, min_s));

				// Obtain the reciprocal of the shared exponent by inverting the bits,
				// and scale by the new mantissa's size. Note that the IEEE-754 single-precision
				// format has an implicit leading 1, but this shared component format does not.
				Float scale = As<Float>((As<Int>(max_s) & 0x7F800000) ^ 0x7F800000) * (1 << (N - 2));

				UInt R9 = RoundInt(red_c * scale);
				UInt G9 = UInt(RoundInt(green_c * scale));
				UInt B9 = UInt(RoundInt(blue_c * scale));
				UInt E5 = (As<UInt>(max_s) >> 23) - 127 + 15 + 1;

				UInt E5B9G9R9 = (E5 << 27) | (B9 << 18) | (G9 << 9) | R9;

				*Pointer<UInt>(element) = E5B9G9R9;
			}
			break;
		case VK_FORMAT_B8G8R8A8_SNORM:
			if(writeB) { *Pointer<SByte>(element) = SByte(RoundInt(Float(c.z))); }
			if(writeG) { *Pointer<SByte>(element + 1) = SByte(RoundInt(Float(c.y))); }
			if(writeR) { *Pointer<SByte>(element + 2) = SByte(RoundInt(Float(c.x))); }
			if(writeA) { *Pointer<SByte>(element + 3) = SByte(RoundInt(Float(c.w))); }
			break;
		case VK_FORMAT_A8B8G8R8_SINT_PACK32:
		case VK_FORMAT_R8G8B8A8_SINT:
		case VK_FORMAT_A8B8G8R8_SNORM_PACK32:
		case VK_FORMAT_R8G8B8A8_SNORM:
		case VK_FORMAT_R8G8B8A8_SSCALED:
		case VK_FORMAT_A8B8G8R8_SSCALED_PACK32:
			if(writeA) { *Pointer<SByte>(element + 3) = SByte(RoundInt(Float(c.w))); }
		case VK_FORMAT_R8G8B8_SINT:
		case VK_FORMAT_R8G8B8_SNORM:
		case VK_FORMAT_R8G8B8_SSCALED:
			if(writeB) { *Pointer<SByte>(element + 2) = SByte(RoundInt(Float(c.z))); }
		case VK_FORMAT_R8G8_SINT:
		case VK_FORMAT_R8G8_SNORM:
		case VK_FORMAT_R8G8_SSCALED:
			if(writeG) { *Pointer<SByte>(element + 1) = SByte(RoundInt(Float(c.y))); }
		case VK_FORMAT_R8_SINT:
		case VK_FORMAT_R8_SNORM:
		case VK_FORMAT_R8_SSCALED:
			if(writeR) { *Pointer<SByte>(element) = SByte(RoundInt(Float(c.x))); }
			break;
		case VK_FORMAT_R8G8B8_UINT:
		case VK_FORMAT_R8G8B8_UNORM:
		case VK_FORMAT_R8G8B8_USCALED:
		case VK_FORMAT_R8G8B8_SRGB:
			if(writeB) { *Pointer<Byte>(element + 2) = Byte(RoundInt(Float(c.z))); }
		case VK_FORMAT_R8G8_UINT:
		case VK_FORMAT_R8G8_UNORM:
		case VK_FORMAT_R8G8_USCALED:
		case VK_FORMAT_R8G8_SRGB:
			if(writeG) { *Pointer<Byte>(element + 1) = Byte(RoundInt(Float(c.y))); }
		case VK_FORMAT_R8_UINT:
		case VK_FORMAT_R8_UNORM:
		case VK_FORMAT_R8_USCALED:
		case VK_FORMAT_R8_SRGB:
			if(writeR) { *Pointer<Byte>(element) = Byte(RoundInt(Float(c.x))); }
			break;
		case VK_FORMAT_R16G16B16A16_SINT:
		case VK_FORMAT_R16G16B16A16_SNORM:
		case VK_FORMAT_R16G16B16A16_SSCALED:
			if(writeRGBA)
			{
				*Pointer<Short4>(element) = Short4(RoundInt(c));
			}
			else
			{
				if(writeR) { *Pointer<Short>(element) = Short(RoundInt(Float(c.x))); }
				if(writeG) { *Pointer<Short>(element + 2) = Short(RoundInt(Float(c.y))); }
				if(writeB) { *Pointer<Short>(element + 4) = Short(RoundInt(Float(c.z))); }
				if(writeA) { *Pointer<Short>(element + 6) = Short(RoundInt(Float(c.w))); }
			}
			break;
		case VK_FORMAT_R16G16B16_SINT:
		case VK_FORMAT_R16G16B16_SNORM:
		case VK_FORMAT_R16G16B16_SSCALED:
			if(writeR) { *Pointer<Short>(element) = Short(RoundInt(Float(c.x))); }
			if(writeG) { *Pointer<Short>(element + 2) = Short(RoundInt(Float(c.y))); }
			if(writeB) { *Pointer<Short>(element + 4) = Short(RoundInt(Float(c.z))); }
			break;
		case VK_FORMAT_R16G16_SINT:
		case VK_FORMAT_R16G16_SNORM:
		case VK_FORMAT_R16G16_SSCALED:
			if(writeR && writeG)
			{
				*Pointer<Short2>(element) = Short2(Short4(RoundInt(c)));
			}
			else
			{
				if(writeR) { *Pointer<Short>(element) = Short(RoundInt(Float(c.x))); }
				if(writeG) { *Pointer<Short>(element + 2) = Short(RoundInt(Float(c.y))); }
			}
			break;
		case VK_FORMAT_R16_SINT:
		case VK_FORMAT_R16_SNORM:
		case VK_FORMAT_R16_SSCALED:
			if(writeR) { *Pointer<Short>(element) = Short(RoundInt(Float(c.x))); }
			break;
		case VK_FORMAT_R16G16B16A16_UINT:
		case VK_FORMAT_R16G16B16A16_UNORM:
		case VK_FORMAT_R16G16B16A16_USCALED:
			if(writeRGBA)
			{
				*Pointer<UShort4>(element) = UShort4(RoundInt(c));
			}
			else
			{
				if(writeR) { *Pointer<UShort>(element) = UShort(RoundInt(Float(c.x))); }
				if(writeG) { *Pointer<UShort>(element + 2) = UShort(RoundInt(Float(c.y))); }
				if(writeB) { *Pointer<UShort>(element + 4) = UShort(RoundInt(Float(c.z))); }
				if(writeA) { *Pointer<UShort>(element + 6) = UShort(RoundInt(Float(c.w))); }
			}
			break;
		case VK_FORMAT_R16G16B16_UINT:
		case VK_FORMAT_R16G16B16_UNORM:
		case VK_FORMAT_R16G16B16_USCALED:
			if(writeR) { *Pointer<UShort>(element) = UShort(RoundInt(Float(c.x))); }
			if(writeG) { *Pointer<UShort>(element + 2) = UShort(RoundInt(Float(c.y))); }
			if(writeB) { *Pointer<UShort>(element + 4) = UShort(RoundInt(Float(c.z))); }
			break;
		case VK_FORMAT_R16G16_UINT:
		case VK_FORMAT_R16G16_UNORM:
		case VK_FORMAT_R16G16_USCALED:
			if(writeR && writeG)
			{
				*Pointer<UShort2>(element) = UShort2(UShort4(RoundInt(c)));
			}
			else
			{
				if(writeR) { *Pointer<UShort>(element) = UShort(RoundInt(Float(c.x))); }
				if(writeG) { *Pointer<UShort>(element + 2) = UShort(RoundInt(Float(c.y))); }
			}
			break;
		case VK_FORMAT_R16_UINT:
		case VK_FORMAT_R16_UNORM:
		case VK_FORMAT_R16_USCALED:
			if(writeR) { *Pointer<UShort>(element) = UShort(RoundInt(Float(c.x))); }
			break;
		case VK_FORMAT_R32G32B32A32_SINT:
			if(writeRGBA)
			{
				*Pointer<Int4>(element) = RoundInt(c);
			}
			else
			{
				if(writeR) { *Pointer<Int>(element) = RoundInt(Float(c.x)); }
				if(writeG) { *Pointer<Int>(element + 4) = RoundInt(Float(c.y)); }
				if(writeB) { *Pointer<Int>(element + 8) = RoundInt(Float(c.z)); }
				if(writeA) { *Pointer<Int>(element + 12) = RoundInt(Float(c.w)); }
			}
			break;
		case VK_FORMAT_R32G32B32_SINT:
			if(writeB) { *Pointer<Int>(element + 8) = RoundInt(Float(c.z)); }
		case VK_FORMAT_R32G32_SINT:
			if(writeG) { *Pointer<Int>(element + 4) = RoundInt(Float(c.y)); }
		case VK_FORMAT_R32_SINT:
			if(writeR) { *Pointer<Int>(element) = RoundInt(Float(c.x)); }
			break;
		case VK_FORMAT_R32G32B32A32_UINT:
			if(writeRGBA)
			{
				*Pointer<UInt4>(element) = UInt4(RoundInt(c));
			}
			else
			{
				if(writeR) { *Pointer<UInt>(element) = As<UInt>(RoundInt(Float(c.x))); }
				if(writeG) { *Pointer<UInt>(element + 4) = As<UInt>(RoundInt(Float(c.y))); }
				if(writeB) { *Pointer<UInt>(element + 8) = As<UInt>(RoundInt(Float(c.z))); }
				if(writeA) { *Pointer<UInt>(element + 12) = As<UInt>(RoundInt(Float(c.w))); }
			}
			break;
		case VK_FORMAT_R32G32B32_UINT:
			if(writeB) { *Pointer<UInt>(element + 8) = As<UInt>(RoundInt(Float(c.z))); }
		case VK_FORMAT_R32G32_UINT:
			if(writeG) { *Pointer<UInt>(element + 4) = As<UInt>(RoundInt(Float(c.y))); }
		case VK_FORMAT_R32_UINT:
			if(writeR) { *Pointer<UInt>(element) = As<UInt>(RoundInt(Float(c.x))); }
			break;
		case VK_FORMAT_R5G6B5_UNORM_PACK16:
			if(writeR && writeG && writeB)
			{
				*Pointer<UShort>(element) = UShort(RoundInt(Float(c.z)) |
				                                  (RoundInt(Float(c.y)) << Int(5)) |
				                                  (RoundInt(Float(c.x)) << Int(11)));
			}
			else
			{
				unsigned short mask = (writeB ? 0x001F : 0x0000) | (writeG ? 0x07E0 : 0x0000) | (writeR ? 0xF800 : 0x0000);
				unsigned short unmask = ~mask;
				*Pointer<UShort>(element) = (*Pointer<UShort>(element) & UShort(unmask)) |
				                            (UShort(RoundInt(Float(c.z)) |
				                                   (RoundInt(Float(c.y)) << Int(5)) |
				                                   (RoundInt(Float(c.x)) << Int(11))) & UShort(mask));
			}
			break;
		case VK_FORMAT_R5G5B5A1_UNORM_PACK16:
			if(writeRGBA)
			{
				*Pointer<UShort>(element) = UShort(RoundInt(Float(c.w)) |
				                                  (RoundInt(Float(c.z)) << Int(1)) |
				                                  (RoundInt(Float(c.y)) << Int(6)) |
				                                  (RoundInt(Float(c.x)) << Int(11)));
			}
			else
			{
				unsigned short mask = (writeA ? 0x8000 : 0x0000) |
				                      (writeR ? 0x7C00 : 0x0000) |
				                      (writeG ? 0x03E0 : 0x0000) |
				                      (writeB ? 0x001F : 0x0000);
				unsigned short unmask = ~mask;
				*Pointer<UShort>(element) = (*Pointer<UShort>(element) & UShort(unmask)) |
				                            (UShort(RoundInt(Float(c.w)) |
				                                   (RoundInt(Float(c.z)) << Int(1)) |
				                                   (RoundInt(Float(c.y)) << Int(6)) |
				                                   (RoundInt(Float(c.x)) << Int(11))) & UShort(mask));
			}
			break;
		case VK_FORMAT_B5G5R5A1_UNORM_PACK16:
			if(writeRGBA)
			{
				*Pointer<UShort>(element) = UShort(RoundInt(Float(c.w)) |
				                                  (RoundInt(Float(c.x)) << Int(1)) |
				                                  (RoundInt(Float(c.y)) << Int(6)) |
				                                  (RoundInt(Float(c.z)) << Int(11)));
			}
			else
			{
				unsigned short mask = (writeA ? 0x8000 : 0x0000) |
				                      (writeR ? 0x7C00 : 0x0000) |
				                      (writeG ? 0x03E0 : 0x0000) |
				                      (writeB ? 0x001F : 0x0000);
				unsigned short unmask = ~mask;
				*Pointer<UShort>(element) = (*Pointer<UShort>(element) & UShort(unmask)) |
				                            (UShort(RoundInt(Float(c.w)) |
				                                   (RoundInt(Float(c.x)) << Int(1)) |
				                                   (RoundInt(Float(c.y)) << Int(6)) |
				                                   (RoundInt(Float(c.z)) << Int(11))) & UShort(mask));
			}
			break;
		case VK_FORMAT_A1R5G5B5_UNORM_PACK16:
			if(writeRGBA)
			{
				*Pointer<UShort>(element) = UShort(RoundInt(Float(c.z)) |
				                                  (RoundInt(Float(c.y)) << Int(5)) |
				                                  (RoundInt(Float(c.x)) << Int(10)) |
				                                  (RoundInt(Float(c.w)) << Int(15)));
			}
			else
			{
				unsigned short mask = (writeA ? 0x8000 : 0x0000) |
				                      (writeR ? 0x7C00 : 0x0000) |
				                      (writeG ? 0x03E0 : 0x0000) |
				                      (writeB ? 0x001F : 0x0000);
				unsigned short unmask = ~mask;
				*Pointer<UShort>(element) = (*Pointer<UShort>(element) & UShort(unmask)) |
				                            (UShort(RoundInt(Float(c.z)) |
				                                   (RoundInt(Float(c.y)) << Int(5)) |
				                                   (RoundInt(Float(c.x)) << Int(10)) |
				                                   (RoundInt(Float(c.w)) << Int(15))) & UShort(mask));
			}
			break;
		case VK_FORMAT_A2B10G10R10_UNORM_PACK32:
		case VK_FORMAT_A2B10G10R10_UINT_PACK32:
		case VK_FORMAT_A2B10G10R10_SNORM_PACK32:
			if(writeRGBA)
			{
				*Pointer<UInt>(element) = UInt(RoundInt(Float(c.x)) |
				                              (RoundInt(Float(c.y)) << 10) |
				                              (RoundInt(Float(c.z)) << 20) |
				                              (RoundInt(Float(c.w)) << 30));
			}
			else
			{
				unsigned int mask = (writeA ? 0xC0000000 : 0x0000) |
				                    (writeB ? 0x3FF00000 : 0x0000) |
				                    (writeG ? 0x000FFC00 : 0x0000) |
				                    (writeR ? 0x000003FF : 0x0000);
				unsigned int unmask = ~mask;
				*Pointer<UInt>(element) = (*Pointer<UInt>(element) & UInt(unmask)) |
				                            (UInt(RoundInt(Float(c.x)) |
				                                 (RoundInt(Float(c.y)) << 10) |
				                                 (RoundInt(Float(c.z)) << 20) |
				                                 (RoundInt(Float(c.w)) << 30)) & UInt(mask));
			}
			break;
		case VK_FORMAT_A2R10G10B10_UNORM_PACK32:
		case VK_FORMAT_A2R10G10B10_UINT_PACK32:
		case VK_FORMAT_A2R10G10B10_SNORM_PACK32:
			if(writeRGBA)
			{
				*Pointer<UInt>(element) = UInt(RoundInt(Float(c.z)) |
				                              (RoundInt(Float(c.y)) << 10) |
				                              (RoundInt(Float(c.x)) << 20) |
				                              (RoundInt(Float(c.w)) << 30));
			}
			else
			{
				unsigned int mask = (writeA ? 0xC0000000 : 0x0000) |
				                    (writeR ? 0x3FF00000 : 0x0000) |
				                    (writeG ? 0x000FFC00 : 0x0000) |
				                    (writeB ? 0x000003FF : 0x0000);
				unsigned int unmask = ~mask;
				*Pointer<UInt>(element) = (*Pointer<UInt>(element) & UInt(unmask)) |
				                            (UInt(RoundInt(Float(c.z)) |
				                                 (RoundInt(Float(c.y)) << 10) |
				                                 (RoundInt(Float(c.x)) << 20) |
				                                 (RoundInt(Float(c.w)) << 30)) & UInt(mask));
			}
			break;
		case VK_FORMAT_D16_UNORM:
			*Pointer<UShort>(element) = UShort(RoundInt(Float(c.x)));
			break;
		case VK_FORMAT_D24_UNORM_S8_UINT:
		case VK_FORMAT_X8_D24_UNORM_PACK32:
			*Pointer<UInt>(element) = UInt(RoundInt(Float(c.x)) << 8);
			break;
		case VK_FORMAT_D32_SFLOAT:
		case VK_FORMAT_D32_SFLOAT_S8_UINT:
			*Pointer<Float>(element) = c.x;
			break;
		case VK_FORMAT_S8_UINT:
			*Pointer<Byte>(element) = Byte(RoundInt(Float(c.x)));
			break;
		default:
			return false;
		}
		return true;
	}

	bool Blitter::read(Int4 &c, Pointer<Byte> element, const State &state)
	{
		c = Int4(0, 0, 0, 1);

		switch(state.sourceFormat)
		{
		case VK_FORMAT_A8B8G8R8_SINT_PACK32:
		case VK_FORMAT_R8G8B8A8_SINT:
			c = Insert(c, Int(*Pointer<SByte>(element + 3)), 3);
			c = Insert(c, Int(*Pointer<SByte>(element + 2)), 2);
		case VK_FORMAT_R8G8_SINT:
			c = Insert(c, Int(*Pointer<SByte>(element + 1)), 1);
		case VK_FORMAT_R8_SINT:
			c = Insert(c, Int(*Pointer<SByte>(element)), 0);
			break;
		case VK_FORMAT_A2B10G10R10_UINT_PACK32:
			c = Insert(c, Int((*Pointer<UInt>(element) & UInt(0x000003FF))), 0);
			c = Insert(c, Int((*Pointer<UInt>(element) & UInt(0x000FFC00)) >> 10), 1);
			c = Insert(c, Int((*Pointer<UInt>(element) & UInt(0x3FF00000)) >> 20), 2);
			c = Insert(c, Int((*Pointer<UInt>(element) & UInt(0xC0000000)) >> 30), 3);
			break;
		case VK_FORMAT_A8B8G8R8_UINT_PACK32:
		case VK_FORMAT_R8G8B8A8_UINT:
			c = Insert(c, Int(*Pointer<Byte>(element + 3)), 3);
			c = Insert(c, Int(*Pointer<Byte>(element + 2)), 2);
		case VK_FORMAT_R8G8_UINT:
			c = Insert(c, Int(*Pointer<Byte>(element + 1)), 1);
		case VK_FORMAT_R8_UINT:
			c = Insert(c, Int(*Pointer<Byte>(element)), 0);
			break;
		case VK_FORMAT_R16G16B16A16_SINT:
			c = Insert(c, Int(*Pointer<Short>(element + 6)), 3);
			c = Insert(c, Int(*Pointer<Short>(element + 4)), 2);
		case VK_FORMAT_R16G16_SINT:
			c = Insert(c, Int(*Pointer<Short>(element + 2)), 1);
		case VK_FORMAT_R16_SINT:
			c = Insert(c, Int(*Pointer<Short>(element)), 0);
			break;
		case VK_FORMAT_R16G16B16A16_UINT:
			c = Insert(c, Int(*Pointer<UShort>(element + 6)), 3);
			c = Insert(c, Int(*Pointer<UShort>(element + 4)), 2);
		case VK_FORMAT_R16G16_UINT:
			c = Insert(c, Int(*Pointer<UShort>(element + 2)), 1);
		case VK_FORMAT_R16_UINT:
			c = Insert(c, Int(*Pointer<UShort>(element)), 0);
			break;
		case VK_FORMAT_R32G32B32A32_SINT:
		case VK_FORMAT_R32G32B32A32_UINT:
			c = *Pointer<Int4>(element);
			break;
		case VK_FORMAT_R32G32_SINT:
		case VK_FORMAT_R32G32_UINT:
			c = Insert(c, *Pointer<Int>(element + 4), 1);
		case VK_FORMAT_R32_SINT:
		case VK_FORMAT_R32_UINT:
			c = Insert(c, *Pointer<Int>(element), 0);
			break;
		default:
			return false;
		}

		return true;
	}

	bool Blitter::write(Int4 &c, Pointer<Byte> element, const State &state)
	{
		bool writeR = state.writeRed;
		bool writeG = state.writeGreen;
		bool writeB = state.writeBlue;
		bool writeA = state.writeAlpha;
		bool writeRGBA = writeR && writeG && writeB && writeA;

		switch(state.destFormat)
		{
		case VK_FORMAT_A2B10G10R10_UINT_PACK32:
			c = Min(As<UInt4>(c), UInt4(0x03FF, 0x03FF, 0x03FF, 0x0003));
			break;
		case VK_FORMAT_A8B8G8R8_UINT_PACK32:
		case VK_FORMAT_R8G8B8A8_UINT:
		case VK_FORMAT_R8G8B8_UINT:
		case VK_FORMAT_R8G8_UINT:
		case VK_FORMAT_R8_UINT:
		case VK_FORMAT_R8G8B8A8_USCALED:
		case VK_FORMAT_R8G8B8_USCALED:
		case VK_FORMAT_R8G8_USCALED:
		case VK_FORMAT_R8_USCALED:
			c = Min(As<UInt4>(c), UInt4(0xFF));
			break;
		case VK_FORMAT_R16G16B16A16_UINT:
		case VK_FORMAT_R16G16B16_UINT:
		case VK_FORMAT_R16G16_UINT:
		case VK_FORMAT_R16_UINT:
		case VK_FORMAT_R16G16B16A16_USCALED:
		case VK_FORMAT_R16G16B16_USCALED:
		case VK_FORMAT_R16G16_USCALED:
		case VK_FORMAT_R16_USCALED:
			c = Min(As<UInt4>(c), UInt4(0xFFFF));
			break;
		case VK_FORMAT_A8B8G8R8_SINT_PACK32:
		case VK_FORMAT_R8G8B8A8_SINT:
		case VK_FORMAT_R8G8_SINT:
		case VK_FORMAT_R8_SINT:
		case VK_FORMAT_R8G8B8A8_SSCALED:
		case VK_FORMAT_R8G8B8_SSCALED:
		case VK_FORMAT_R8G8_SSCALED:
		case VK_FORMAT_R8_SSCALED:
			c = Min(Max(c, Int4(-0x80)), Int4(0x7F));
			break;
		case VK_FORMAT_R16G16B16A16_SINT:
		case VK_FORMAT_R16G16B16_SINT:
		case VK_FORMAT_R16G16_SINT:
		case VK_FORMAT_R16_SINT:
		case VK_FORMAT_R16G16B16A16_SSCALED:
		case VK_FORMAT_R16G16B16_SSCALED:
		case VK_FORMAT_R16G16_SSCALED:
		case VK_FORMAT_R16_SSCALED:
			c = Min(Max(c, Int4(-0x8000)), Int4(0x7FFF));
			break;
		default:
			break;
		}

		switch(state.destFormat)
		{
		case VK_FORMAT_B8G8R8A8_SINT:
		case VK_FORMAT_B8G8R8A8_SSCALED:
			if(writeA) { *Pointer<SByte>(element + 3) = SByte(Extract(c, 3)); }
		case VK_FORMAT_B8G8R8_SINT:
		case VK_FORMAT_B8G8R8_SSCALED:
			if(writeB) { *Pointer<SByte>(element) = SByte(Extract(c, 2)); }
			if(writeG) { *Pointer<SByte>(element + 1) = SByte(Extract(c, 1)); }
			if(writeR) { *Pointer<SByte>(element + 2) = SByte(Extract(c, 0)); }
			break;
		case VK_FORMAT_A8B8G8R8_SINT_PACK32:
		case VK_FORMAT_R8G8B8A8_SINT:
		case VK_FORMAT_R8G8B8A8_SSCALED:
		case VK_FORMAT_A8B8G8R8_SSCALED_PACK32:
			if(writeA) { *Pointer<SByte>(element + 3) = SByte(Extract(c, 3)); }
		case VK_FORMAT_R8G8B8_SINT:
		case VK_FORMAT_R8G8B8_SSCALED:
			if(writeB) { *Pointer<SByte>(element + 2) = SByte(Extract(c, 2)); }
		case VK_FORMAT_R8G8_SINT:
		case VK_FORMAT_R8G8_SSCALED:
			if(writeG) { *Pointer<SByte>(element + 1) = SByte(Extract(c, 1)); }
		case VK_FORMAT_R8_SINT:
		case VK_FORMAT_R8_SSCALED:
			if(writeR) { *Pointer<SByte>(element) = SByte(Extract(c, 0)); }
			break;
		case VK_FORMAT_A2B10G10R10_UINT_PACK32:
		case VK_FORMAT_A2B10G10R10_SINT_PACK32:
		case VK_FORMAT_A2B10G10R10_USCALED_PACK32:
		case VK_FORMAT_A2B10G10R10_SSCALED_PACK32:
			if(writeRGBA)
			{
				*Pointer<UInt>(element) =
					UInt((Extract(c, 0)) | (Extract(c, 1) << 10) | (Extract(c, 2) << 20) | (Extract(c, 3) << 30));
			}
			else
			{
				unsigned int mask = (writeA ? 0xC0000000 : 0x0000) |
				                    (writeB ? 0x3FF00000 : 0x0000) |
				                    (writeG ? 0x000FFC00 : 0x0000) |
				                    (writeR ? 0x000003FF : 0x0000);
				unsigned int unmask = ~mask;
				*Pointer<UInt>(element) = (*Pointer<UInt>(element) & UInt(unmask)) |
					(UInt(Extract(c, 0) | (Extract(c, 1) << 10) | (Extract(c, 2) << 20) | (Extract(c, 3) << 30)) & UInt(mask));
			}
			break;
		case VK_FORMAT_A2R10G10B10_UINT_PACK32:
		case VK_FORMAT_A2R10G10B10_SINT_PACK32:
		case VK_FORMAT_A2R10G10B10_USCALED_PACK32:
		case VK_FORMAT_A2R10G10B10_SSCALED_PACK32:
			if(writeRGBA)
			{
				*Pointer<UInt>(element) =
					UInt((Extract(c, 2)) | (Extract(c, 1) << 10) | (Extract(c, 0) << 20) | (Extract(c, 3) << 30));
			}
			else
			{
				unsigned int mask = (writeA ? 0xC0000000 : 0x0000) |
				                    (writeR ? 0x3FF00000 : 0x0000) |
				                    (writeG ? 0x000FFC00 : 0x0000) |
				                    (writeB ? 0x000003FF : 0x0000);
				unsigned int unmask = ~mask;
				*Pointer<UInt>(element) = (*Pointer<UInt>(element) & UInt(unmask)) |
					(UInt(Extract(c, 2) | (Extract(c, 1) << 10) | (Extract(c, 0) << 20) | (Extract(c, 3) << 30)) & UInt(mask));
			}
			break;
		case VK_FORMAT_B8G8R8A8_UINT:
		case VK_FORMAT_B8G8R8A8_USCALED:
			if(writeA) { *Pointer<Byte>(element + 3) = Byte(Extract(c, 3)); }
		case VK_FORMAT_B8G8R8_UINT:
		case VK_FORMAT_B8G8R8_USCALED:
		case VK_FORMAT_B8G8R8_SRGB:
			if(writeB) { *Pointer<Byte>(element) = Byte(Extract(c, 2)); }
			if(writeG) { *Pointer<Byte>(element + 1) = Byte(Extract(c, 1)); }
			if(writeR) { *Pointer<Byte>(element + 2) = Byte(Extract(c, 0)); }
			break;
		case VK_FORMAT_A8B8G8R8_UINT_PACK32:
		case VK_FORMAT_R8G8B8A8_UINT:
		case VK_FORMAT_R8G8B8A8_USCALED:
		case VK_FORMAT_A8B8G8R8_USCALED_PACK32:
			if(writeA) { *Pointer<Byte>(element + 3) = Byte(Extract(c, 3)); }
		case VK_FORMAT_R8G8B8_UINT:
		case VK_FORMAT_R8G8B8_USCALED:
			if(writeB) { *Pointer<Byte>(element + 2) = Byte(Extract(c, 2)); }
		case VK_FORMAT_R8G8_UINT:
		case VK_FORMAT_R8G8_USCALED:
			if(writeG) { *Pointer<Byte>(element + 1) = Byte(Extract(c, 1)); }
		case VK_FORMAT_R8_UINT:
		case VK_FORMAT_R8_USCALED:
			if(writeR) { *Pointer<Byte>(element) = Byte(Extract(c, 0)); }
			break;
		case VK_FORMAT_R16G16B16A16_SINT:
		case VK_FORMAT_R16G16B16A16_SSCALED:
			if(writeA) { *Pointer<Short>(element + 6) = Short(Extract(c, 3)); }
		case VK_FORMAT_R16G16B16_SINT:
		case VK_FORMAT_R16G16B16_SSCALED:
			if(writeB) { *Pointer<Short>(element + 4) = Short(Extract(c, 2)); }
		case VK_FORMAT_R16G16_SINT:
		case VK_FORMAT_R16G16_SSCALED:
			if(writeG) { *Pointer<Short>(element + 2) = Short(Extract(c, 1)); }
		case VK_FORMAT_R16_SINT:
		case VK_FORMAT_R16_SSCALED:
			if(writeR) { *Pointer<Short>(element) = Short(Extract(c, 0)); }
			break;
		case VK_FORMAT_R16G16B16A16_UINT:
		case VK_FORMAT_R16G16B16A16_USCALED:
			if(writeA) { *Pointer<UShort>(element + 6) = UShort(Extract(c, 3)); }
		case VK_FORMAT_R16G16B16_UINT:
		case VK_FORMAT_R16G16B16_USCALED:
			if(writeB) { *Pointer<UShort>(element + 4) = UShort(Extract(c, 2)); }
		case VK_FORMAT_R16G16_UINT:
		case VK_FORMAT_R16G16_USCALED:
			if(writeG) { *Pointer<UShort>(element + 2) = UShort(Extract(c, 1)); }
		case VK_FORMAT_R16_UINT:
		case VK_FORMAT_R16_USCALED:
			if(writeR) { *Pointer<UShort>(element) = UShort(Extract(c, 0)); }
			break;
		case VK_FORMAT_R32G32B32A32_SINT:
			if(writeRGBA)
			{
				*Pointer<Int4>(element) = c;
			}
			else
			{
				if(writeR) { *Pointer<Int>(element) = Extract(c, 0); }
				if(writeG) { *Pointer<Int>(element + 4) = Extract(c, 1); }
				if(writeB) { *Pointer<Int>(element + 8) = Extract(c, 2); }
				if(writeA) { *Pointer<Int>(element + 12) = Extract(c, 3); }
			}
			break;
		case VK_FORMAT_R32G32B32_SINT:
			if(writeR) { *Pointer<Int>(element) = Extract(c, 0); }
			if(writeG) { *Pointer<Int>(element + 4) = Extract(c, 1); }
			if(writeB) { *Pointer<Int>(element + 8) = Extract(c, 2); }
			break;
		case VK_FORMAT_R32G32_SINT:
			if(writeR) { *Pointer<Int>(element) = Extract(c, 0); }
			if(writeG) { *Pointer<Int>(element + 4) = Extract(c, 1); }
			break;
		case VK_FORMAT_R32_SINT:
			if(writeR) { *Pointer<Int>(element) = Extract(c, 0); }
			break;
		case VK_FORMAT_R32G32B32A32_UINT:
			if(writeRGBA)
			{
				*Pointer<UInt4>(element) = As<UInt4>(c);
			}
			else
			{
				if(writeR) { *Pointer<UInt>(element) = As<UInt>(Extract(c, 0)); }
				if(writeG) { *Pointer<UInt>(element + 4) = As<UInt>(Extract(c, 1)); }
				if(writeB) { *Pointer<UInt>(element + 8) = As<UInt>(Extract(c, 2)); }
				if(writeA) { *Pointer<UInt>(element + 12) = As<UInt>(Extract(c, 3)); }
			}
			break;
		case VK_FORMAT_R32G32B32_UINT:
			if(writeB) { *Pointer<UInt>(element + 8) = As<UInt>(Extract(c, 2)); }
		case VK_FORMAT_R32G32_UINT:
			if(writeG) { *Pointer<UInt>(element + 4) = As<UInt>(Extract(c, 1)); }
		case VK_FORMAT_R32_UINT:
			if(writeR) { *Pointer<UInt>(element) = As<UInt>(Extract(c, 0)); }
			break;
		default:
			return false;
		}

		return true;
	}

	bool Blitter::ApplyScaleAndClamp(Float4 &value, const State &state, bool preScaled)
	{
		float4 scale, unscale;
		if(state.clearOperation &&
		   state.sourceFormat.isNonNormalizedInteger() &&
		   !state.destFormat.isNonNormalizedInteger())
		{
			// If we're clearing a buffer from an int or uint color into a normalized color,
			// then the whole range of the int or uint color must be scaled between 0 and 1.
			switch(state.sourceFormat)
			{
			case VK_FORMAT_R32G32B32A32_SINT:
				unscale = replicate(static_cast<float>(0x7FFFFFFF));
				break;
			case VK_FORMAT_R32G32B32A32_UINT:
				unscale = replicate(static_cast<float>(0xFFFFFFFF));
				break;
			default:
				return false;
			}
		}
		else if(!state.sourceFormat.getScale(unscale))
		{
			return false;
		}

		if(!state.destFormat.getScale(scale))
		{
			return false;
		}

		bool srcSRGB = state.sourceFormat.isSRGBformat();
		bool dstSRGB = state.destFormat.isSRGBformat();

		if(state.convertSRGB && ((srcSRGB && !preScaled) || dstSRGB))   // One of the formats is sRGB encoded.
		{
			value *= preScaled ? Float4(1.0f / scale.x, 1.0f / scale.y, 1.0f / scale.z, 1.0f / scale.w) : // Unapply scale
			                     Float4(1.0f / unscale.x, 1.0f / unscale.y, 1.0f / unscale.z, 1.0f / unscale.w); // Apply unscale
			value = (srcSRGB && !preScaled) ? sRGBtoLinear(value) : LinearToSRGB(value);
			value *= Float4(scale.x, scale.y, scale.z, scale.w); // Apply scale
		}
		else if(unscale != scale)
		{
			value *= Float4(scale.x / unscale.x, scale.y / unscale.y, scale.z / unscale.z, scale.w / unscale.w);
		}

		if(state.sourceFormat.isFloatFormat() && !state.destFormat.isFloatFormat())
		{
			value = Min(value, Float4(scale.x, scale.y, scale.z, scale.w));

			value = Max(value, Float4(state.destFormat.isUnsignedComponent(0) ? 0.0f : -scale.x,
			                          state.destFormat.isUnsignedComponent(1) ? 0.0f : -scale.y,
			                          state.destFormat.isUnsignedComponent(2) ? 0.0f : -scale.z,
			                          state.destFormat.isUnsignedComponent(3) ? 0.0f : -scale.w));
		}

		return true;
	}

	Int Blitter::ComputeOffset(Int &x, Int &y, Int &pitchB, int bytes, bool quadLayout)
	{
		if(!quadLayout)
		{
			return y * pitchB + x * bytes;
		}
		else
		{
			// (x & ~1) * 2 + (x & 1) == (x - (x & 1)) * 2 + (x & 1) == x * 2 - (x & 1) * 2 + (x & 1) == x * 2 - (x & 1)
			return (y & Int(~1)) * pitchB +
			       ((y & Int(1)) * 2 + x * 2 - (x & Int(1))) * bytes;
		}
	}

	Float4 Blitter::LinearToSRGB(Float4 &c)
	{
		Float4 lc = Min(c, Float4(0.0031308f)) * Float4(12.92f);
		Float4 ec = Float4(1.055f) * power(c, Float4(1.0f / 2.4f)) - Float4(0.055f);

		Float4 s = c;
		s.xyz = Max(lc, ec);

		return s;
	}

	Float4 Blitter::sRGBtoLinear(Float4 &c)
	{
		Float4 lc = c * Float4(1.0f / 12.92f);
		Float4 ec = power((c + Float4(0.055f)) * Float4(1.0f / 1.055f), Float4(2.4f));

		Int4 linear = CmpLT(c, Float4(0.04045f));

		Float4 s = c;
		s.xyz = As<Float4>((linear & As<Int4>(lc)) | (~linear & As<Int4>(ec)));   // TODO: IfThenElse()

		return s;
	}

	std::shared_ptr<Routine> Blitter::generate(const State &state)
	{
		Function<Void(Pointer<Byte>)> function;
		{
			Pointer<Byte> blit(function.Arg<0>());

			Pointer<Byte> source = *Pointer<Pointer<Byte>>(blit + OFFSET(BlitData,source));
			Pointer<Byte> dest = *Pointer<Pointer<Byte>>(blit + OFFSET(BlitData,dest));
			Int sPitchB = *Pointer<Int>(blit + OFFSET(BlitData,sPitchB));
			Int dPitchB = *Pointer<Int>(blit + OFFSET(BlitData,dPitchB));

			Float x0 = *Pointer<Float>(blit + OFFSET(BlitData,x0));
			Float y0 = *Pointer<Float>(blit + OFFSET(BlitData,y0));
			Float w = *Pointer<Float>(blit + OFFSET(BlitData,w));
			Float h = *Pointer<Float>(blit + OFFSET(BlitData,h));

			Int x0d = *Pointer<Int>(blit + OFFSET(BlitData,x0d));
			Int x1d = *Pointer<Int>(blit + OFFSET(BlitData,x1d));
			Int y0d = *Pointer<Int>(blit + OFFSET(BlitData,y0d));
			Int y1d = *Pointer<Int>(blit + OFFSET(BlitData,y1d));

			Int sWidth = *Pointer<Int>(blit + OFFSET(BlitData,sWidth));
			Int sHeight = *Pointer<Int>(blit + OFFSET(BlitData,sHeight));

			bool intSrc = state.sourceFormat.isNonNormalizedInteger();
			bool intDst = state.destFormat.isNonNormalizedInteger();
			bool intBoth = intSrc && intDst;
			bool srcQuadLayout = state.sourceFormat.hasQuadLayout();
			bool dstQuadLayout = state.destFormat.hasQuadLayout();
			int srcBytes = state.sourceFormat.bytes();
			int dstBytes = state.destFormat.bytes();

			bool hasConstantColorI = false;
			Int4 constantColorI;
			bool hasConstantColorF = false;
			Float4 constantColorF;
			if(state.clearOperation)
			{
				if(intBoth) // Integer types
				{
					if(!read(constantColorI, source, state))
					{
						return nullptr;
					}
					hasConstantColorI = true;
				}
				else
				{
					if(!read(constantColorF, source, state))
					{
						return nullptr;
					}
					hasConstantColorF = true;

					if(!ApplyScaleAndClamp(constantColorF, state))
					{
						return nullptr;
					}
				}
			}

			For(Int j = y0d, j < y1d, j++)
			{
				Float y = state.clearOperation ? RValue<Float>(y0) : y0 + Float(j) * h;
				Pointer<Byte> destLine = dest + (dstQuadLayout ? j & Int(~1) : RValue<Int>(j)) * dPitchB;

				For(Int i = x0d, i < x1d, i++)
				{
					Float x = state.clearOperation ? RValue<Float>(x0) : x0 + Float(i) * w;
					Pointer<Byte> d = destLine + (dstQuadLayout ? (((j & Int(1)) << 1) + (i * 2) - (i & Int(1))) : RValue<Int>(i)) * dstBytes;

					if(hasConstantColorI)
					{
						if(!write(constantColorI, d, state))
						{
							return nullptr;
						}
					}
					else if(hasConstantColorF)
					{
						for(int s = 0; s < state.destSamples; s++)
						{
							if(!write(constantColorF, d, state))
							{
								return nullptr;
							}

							d += *Pointer<Int>(blit + OFFSET(BlitData, dSliceB));
						}
					}
					else if(intBoth) // Integer types do not support filtering
					{
						Int4 color; // When both formats are true integer types, we don't go to float to avoid losing precision
						Int X = Int(x);
						Int Y = Int(y);

						if(state.clampToEdge)
						{
							X = Clamp(X, 0, sWidth - 1);
							Y = Clamp(Y, 0, sHeight - 1);
						}

						Pointer<Byte> s = source + ComputeOffset(X, Y, sPitchB, srcBytes, srcQuadLayout);

						if(!read(color, s, state))
						{
							return nullptr;
						}

						if(!write(color, d, state))
						{
							return nullptr;
						}
					}
					else
					{
						Float4 color;

						bool preScaled = false;
						if(!state.filter || intSrc)
						{
							Int X = Int(x);
							Int Y = Int(y);

							if(state.clampToEdge)
							{
								X = Clamp(X, 0, sWidth - 1);
								Y = Clamp(Y, 0, sHeight - 1);
							}

							Pointer<Byte> s = source + ComputeOffset(X, Y, sPitchB, srcBytes, srcQuadLayout);

							if(!read(color, s, state))
							{
								return nullptr;
							}

							if(state.srcSamples > 1) // Resolve multisampled source
							{
								if(state.convertSRGB && state.sourceFormat.isSRGBformat()) // sRGB -> RGB
								{
									if(!ApplyScaleAndClamp(color, state)) return nullptr;
									preScaled = true;
								}
								Float4 accum = color;
								for(int sample = 1; sample < state.srcSamples; sample++)
								{
									s += *Pointer<Int>(blit + OFFSET(BlitData, sSliceB));
									if(!read(color, s, state))
									{
										return nullptr;
									}
									if(state.convertSRGB && state.sourceFormat.isSRGBformat()) // sRGB -> RGB
									{
										if(!ApplyScaleAndClamp(color, state)) return nullptr;
										preScaled = true;
									}
									accum += color;
								}
								color = accum * Float4(1.0f / static_cast<float>(state.srcSamples));
							}
						}
						else   // Bilinear filtering
						{
							Float X = x;
							Float Y = y;

							if(state.clampToEdge)
							{
								X = Min(Max(x, 0.5f), Float(sWidth) - 0.5f);
								Y = Min(Max(y, 0.5f), Float(sHeight) - 0.5f);
							}

							Float x0 = X - 0.5f;
							Float y0 = Y - 0.5f;

							Int X0 = Max(Int(x0), 0);
							Int Y0 = Max(Int(y0), 0);

							Int X1 = X0 + 1;
							Int Y1 = Y0 + 1;
							X1 = IfThenElse(X1 >= sWidth, X0, X1);
							Y1 = IfThenElse(Y1 >= sHeight, Y0, Y1);

							Pointer<Byte> s00 = source + ComputeOffset(X0, Y0, sPitchB, srcBytes, srcQuadLayout);
							Pointer<Byte> s01 = source + ComputeOffset(X1, Y0, sPitchB, srcBytes, srcQuadLayout);
							Pointer<Byte> s10 = source + ComputeOffset(X0, Y1, sPitchB, srcBytes, srcQuadLayout);
							Pointer<Byte> s11 = source + ComputeOffset(X1, Y1, sPitchB, srcBytes, srcQuadLayout);

							Float4 c00; if(!read(c00, s00, state)) return nullptr;
							Float4 c01; if(!read(c01, s01, state)) return nullptr;
							Float4 c10; if(!read(c10, s10, state)) return nullptr;
							Float4 c11; if(!read(c11, s11, state)) return nullptr;

							if(state.convertSRGB && state.sourceFormat.isSRGBformat()) // sRGB -> RGB
							{
								if(!ApplyScaleAndClamp(c00, state)) return nullptr;
								if(!ApplyScaleAndClamp(c01, state)) return nullptr;
								if(!ApplyScaleAndClamp(c10, state)) return nullptr;
								if(!ApplyScaleAndClamp(c11, state)) return nullptr;
								preScaled = true;
							}

							Float4 fx = Float4(x0 - Float(X0));
							Float4 fy = Float4(y0 - Float(Y0));
							Float4 ix = Float4(1.0f) - fx;
							Float4 iy = Float4(1.0f) - fy;

							color = (c00 * ix + c01 * fx) * iy +
							        (c10 * ix + c11 * fx) * fy;
						}

						if(!ApplyScaleAndClamp(color, state, preScaled))
						{
							return nullptr;
						}

						for(int s = 0; s < state.destSamples; s++)
						{
							if(!write(color, d, state))
							{
								return nullptr;
							}

							d += *Pointer<Int>(blit + OFFSET(BlitData,dSliceB));
						}
					}
				}
			}
		}

		return function("BlitRoutine");
	}

	std::shared_ptr<Routine> Blitter::getBlitRoutine(const State &state)
	{
		std::unique_lock<std::mutex> lock(blitMutex);
		auto blitRoutine = blitCache.query(state);

		if(!blitRoutine)
		{
			blitRoutine = generate(state);

			if(!blitRoutine)
			{
				UNIMPLEMENTED("blitRoutine");
				return nullptr;
			}

			blitCache.add(state, blitRoutine);
		}

		return blitRoutine;
	}

	std::shared_ptr<Routine> Blitter::getCornerUpdateRoutine(const State &state)
	{
		std::unique_lock<std::mutex> lock(cornerUpdateMutex);
		auto cornerUpdateRoutine = cornerUpdateCache.query(state);

		if(!cornerUpdateRoutine)
		{
			cornerUpdateRoutine = generateCornerUpdate(state);

			if(!cornerUpdateRoutine)
			{
				UNIMPLEMENTED("cornerUpdateRoutine");
				return nullptr;
			}

			cornerUpdateCache.add(state, cornerUpdateRoutine);
		}

		return cornerUpdateRoutine;
	}

	void Blitter::blitToBuffer(const vk::Image *src, VkImageSubresourceLayers subresource, VkOffset3D offset, VkExtent3D extent, uint8_t *dst, int bufferRowPitch, int bufferSlicePitch)
	{
		auto aspect = static_cast<VkImageAspectFlagBits>(subresource.aspectMask);
		auto format = src->getFormat(aspect);
		State state(format, format.getNonQuadLayoutFormat(), VK_SAMPLE_COUNT_1_BIT, VK_SAMPLE_COUNT_1_BIT,
					{false, false});

		auto blitRoutine = getBlitRoutine(state);
		if(!blitRoutine)
		{
			return;
		}

		void(*blitFunction)(const BlitData *data) = (void(*)(const BlitData*))blitRoutine->getEntry();

		BlitData data =
		{
			nullptr, // source
			dst, // dest
			src->rowPitchBytes(aspect, subresource.mipLevel),   // sPitchB
			bufferRowPitch,   // dPitchB
			src->slicePitchBytes(aspect, subresource.mipLevel), // sSliceB
			bufferSlicePitch, // dSliceB

			0, 0, 1, 1,

			0, // y0d
			static_cast<int>(extent.height), // y1d
			0, // x0d
			static_cast<int>(extent.width), // x1d

			static_cast<int>(extent.width), // sWidth
			static_cast<int>(extent.height) // sHeight;
		};

		VkOffset3D srcOffset = { 0, 0, offset.z };

		VkImageSubresourceLayers srcSubresLayers = subresource;
		srcSubresLayers.layerCount = 1;

		VkImageSubresourceRange srcSubresRange =
		{
			subresource.aspectMask,
			subresource.mipLevel,
			1,
			subresource.baseArrayLayer,
			subresource.layerCount
		};

		uint32_t lastLayer = src->getLastLayerIndex(srcSubresRange);

		for(; srcSubresLayers.baseArrayLayer <= lastLayer; srcSubresLayers.baseArrayLayer++)
		{
			srcOffset.z = offset.z;

			for(auto i = 0u; i < extent.depth; i++)
			{
				data.source = src->getTexelPointer(srcOffset, srcSubresLayers);
				ASSERT(data.source < src->end());
				blitFunction(&data);
				srcOffset.z++;
				data.dest = (dst += bufferSlicePitch);
			}
		}
	}

	void Blitter::blitFromBuffer(const vk::Image *dst, VkImageSubresourceLayers subresource, VkOffset3D offset, VkExtent3D extent, uint8_t *src, int bufferRowPitch, int bufferSlicePitch)
	{
		auto aspect = static_cast<VkImageAspectFlagBits>(subresource.aspectMask);
		auto format = dst->getFormat(aspect);
		State state(format.getNonQuadLayoutFormat(), format, VK_SAMPLE_COUNT_1_BIT, VK_SAMPLE_COUNT_1_BIT,
					{false, false});

		auto blitRoutine = getBlitRoutine(state);
		if(!blitRoutine)
		{
			return;
		}

		void(*blitFunction)(const BlitData *data) = (void(*)(const BlitData*))blitRoutine->getEntry();

		BlitData data =
		{
			src, // source
			nullptr, // dest
			bufferRowPitch,   // sPitchB
			dst->rowPitchBytes(aspect, subresource.mipLevel),   // dPitchB
			bufferSlicePitch, // sSliceB
			dst->slicePitchBytes(aspect, subresource.mipLevel), // dSliceB

			0, 0, 1, 1,

			offset.y, // y0d
			static_cast<int>(offset.y + extent.height), // y1d
			offset.x, // x0d
			static_cast<int>(offset.x + extent.width), // x1d

			static_cast<int>(extent.width), // sWidth
			static_cast<int>(extent.height) // sHeight;
		};

		VkOffset3D dstOffset = { 0, 0, offset.z };

		VkImageSubresourceLayers dstSubresLayers = subresource;
		dstSubresLayers.layerCount = 1;

		VkImageSubresourceRange dstSubresRange =
		{
			subresource.aspectMask,
			subresource.mipLevel,
			1,
			subresource.baseArrayLayer,
			subresource.layerCount
		};

		uint32_t lastLayer = dst->getLastLayerIndex(dstSubresRange);

		for(; dstSubresLayers.baseArrayLayer <= lastLayer; dstSubresLayers.baseArrayLayer++)
		{
			dstOffset.z = offset.z;

			for(auto i = 0u; i < extent.depth; i++)
			{
				data.dest = dst->getTexelPointer(dstOffset, dstSubresLayers);
				ASSERT(data.dest < dst->end());
				blitFunction(&data);
				dstOffset.z++;
				data.source = (src += bufferSlicePitch);
			}
		}
	}

	void Blitter::blit(const vk::Image *src, vk::Image *dst, VkImageBlit region, VkFilter filter)
	{
		if(dst->getFormat() == VK_FORMAT_UNDEFINED)
		{
			return;
		}

		if((region.srcSubresource.layerCount != region.dstSubresource.layerCount) ||
		   (region.srcSubresource.aspectMask != region.dstSubresource.aspectMask))
		{
			UNIMPLEMENTED("region");
		}

		if(region.dstOffsets[0].x > region.dstOffsets[1].x)
		{
			std::swap(region.srcOffsets[0].x, region.srcOffsets[1].x);
			std::swap(region.dstOffsets[0].x, region.dstOffsets[1].x);
		}

		if(region.dstOffsets[0].y > region.dstOffsets[1].y)
		{
			std::swap(region.srcOffsets[0].y, region.srcOffsets[1].y);
			std::swap(region.dstOffsets[0].y, region.dstOffsets[1].y);
		}

		VkImageAspectFlagBits srcAspect = static_cast<VkImageAspectFlagBits>(region.srcSubresource.aspectMask);
		VkImageAspectFlagBits dstAspect = static_cast<VkImageAspectFlagBits>(region.dstSubresource.aspectMask);
		VkExtent3D srcExtent = src->getMipLevelExtent(srcAspect, region.srcSubresource.mipLevel);

		int32_t numSlices = (region.srcOffsets[1].z - region.srcOffsets[0].z);
		ASSERT(numSlices == (region.dstOffsets[1].z - region.dstOffsets[0].z));

		float widthRatio = static_cast<float>(region.srcOffsets[1].x - region.srcOffsets[0].x) /
		                   static_cast<float>(region.dstOffsets[1].x - region.dstOffsets[0].x);
		float heightRatio = static_cast<float>(region.srcOffsets[1].y - region.srcOffsets[0].y) /
		                    static_cast<float>(region.dstOffsets[1].y - region.dstOffsets[0].y);
		float x0 = region.srcOffsets[0].x + (0.5f - region.dstOffsets[0].x) * widthRatio;
		float y0 = region.srcOffsets[0].y + (0.5f - region.dstOffsets[0].y) * heightRatio;

		bool doFilter = (filter != VK_FILTER_NEAREST);
		State state(src->getFormat(srcAspect), dst->getFormat(dstAspect), src->getSampleCountFlagBits(), dst->getSampleCountFlagBits(),
		            { doFilter, doFilter || (src->getSampleCountFlagBits() > 1) });
		state.clampToEdge = (region.srcOffsets[0].x < 0) ||
		                    (region.srcOffsets[0].y < 0) ||
		                    (static_cast<uint32_t>(region.srcOffsets[1].x) > srcExtent.width) ||
		                    (static_cast<uint32_t>(region.srcOffsets[1].y) > srcExtent.height) ||
		                    (doFilter && ((x0 < 0.5f) || (y0 < 0.5f)));

		auto blitRoutine = getBlitRoutine(state);
		if(!blitRoutine)
		{
			return;
		}

		void(*blitFunction)(const BlitData *data) = (void(*)(const BlitData*))blitRoutine->getEntry();

		BlitData data =
		{
			nullptr, // source
			nullptr, // dest
			src->rowPitchBytes(srcAspect, region.srcSubresource.mipLevel),   // sPitchB
			dst->rowPitchBytes(dstAspect, region.dstSubresource.mipLevel),   // dPitchB
			src->slicePitchBytes(srcAspect, region.srcSubresource.mipLevel), // sSliceB
			dst->slicePitchBytes(dstAspect, region.dstSubresource.mipLevel), // dSliceB

			x0,
			y0,
			widthRatio,
			heightRatio,

			region.dstOffsets[0].y, // y0d
			region.dstOffsets[1].y, // y1d
			region.dstOffsets[0].x, // x0d
			region.dstOffsets[1].x, // x1d

			static_cast<int>(srcExtent.width), // sWidth
			static_cast<int>(srcExtent.height) // sHeight;
		};

		VkOffset3D srcOffset = { 0, 0, region.srcOffsets[0].z };
		VkOffset3D dstOffset = { 0, 0, region.dstOffsets[0].z };

		VkImageSubresourceLayers srcSubresLayers =
		{
			region.srcSubresource.aspectMask,
			region.srcSubresource.mipLevel,
			region.srcSubresource.baseArrayLayer,
			1
		};

		VkImageSubresourceLayers dstSubresLayers =
		{
			region.dstSubresource.aspectMask,
			region.dstSubresource.mipLevel,
			region.dstSubresource.baseArrayLayer,
			1
		};

		VkImageSubresourceRange srcSubresRange =
		{
			region.srcSubresource.aspectMask,
			region.srcSubresource.mipLevel,
			1,
			region.srcSubresource.baseArrayLayer,
			region.srcSubresource.layerCount
		};

		uint32_t lastLayer = src->getLastLayerIndex(srcSubresRange);

		for(; srcSubresLayers.baseArrayLayer <= lastLayer; srcSubresLayers.baseArrayLayer++, dstSubresLayers.baseArrayLayer++)
		{
			srcOffset.z = region.srcOffsets[0].z;
			dstOffset.z = region.dstOffsets[0].z;

			for(int i = 0; i < numSlices; i++)
			{
				data.source = src->getTexelPointer(srcOffset, srcSubresLayers);
				data.dest = dst->getTexelPointer(dstOffset, dstSubresLayers);

				ASSERT(data.source < src->end());
				ASSERT(data.dest < dst->end());

				blitFunction(&data);
				srcOffset.z++;
				dstOffset.z++;
			}
		}
	}

	void Blitter::computeCubeCorner(Pointer<Byte>& layer, Int& x0, Int& x1, Int& y0, Int& y1, Int& pitchB, const State& state)
	{
		int bytes = state.sourceFormat.bytes();
		bool quadLayout = state.sourceFormat.hasQuadLayout();

		Float4 c0;
		read(c0, layer + ComputeOffset(x0, y1, pitchB, bytes, quadLayout), state);
		Float4 c1;
		read(c1, layer + ComputeOffset(x1, y0, pitchB, bytes, quadLayout), state);
		c0 += c1;
		read(c1, layer + ComputeOffset(x1, y1, pitchB, bytes, quadLayout), state);
		c0 += c1;
		c0 *= Float4(1.0f / 3.0f);
		write(c0, layer + ComputeOffset(x0, y0, pitchB, bytes, quadLayout), state);
	}

	std::shared_ptr<Routine> Blitter::generateCornerUpdate(const State& state)
	{
		// Reading and writing from/to the same image
		ASSERT(state.sourceFormat == state.destFormat);
		ASSERT(state.srcSamples == state.destSamples);

		if(state.srcSamples != 1)
		{
			UNIMPLEMENTED("state.srcSamples %d", state.srcSamples);
		}

		Function<Void(Pointer<Byte>)> function;
		{
			Pointer<Byte> blit(function.Arg<0>());

			Pointer<Byte> layers = *Pointer<Pointer<Byte>>(blit + OFFSET(CubeBorderData, layers));
			Int pitchB = *Pointer<Int>(blit + OFFSET(CubeBorderData, pitchB));
			UInt layerSize = *Pointer<Int>(blit + OFFSET(CubeBorderData, layerSize));
			UInt dim = *Pointer<Int>(blit + OFFSET(CubeBorderData, dim));

			// Low Border, Low Pixel, High Border, High Pixel
			Int LB(-1), LP(0), HB(dim), HP(dim-1);

			for(int face = 0; face < 6; face++)
			{
				computeCubeCorner(layers, LB, LP, LB, LP, pitchB, state);
				computeCubeCorner(layers, LB, LP, HB, HP, pitchB, state);
				computeCubeCorner(layers, HB, HP, LB, LP, pitchB, state);
				computeCubeCorner(layers, HB, HP, HB, HP, pitchB, state);
				layers = layers + layerSize;
			}
		}

		return function("BlitRoutine");
	}

	void Blitter::updateBorders(vk::Image* image, const VkImageSubresourceLayers& subresourceLayers)
	{
		if(image->getArrayLayers() < (subresourceLayers.baseArrayLayer + 6))
		{
			UNIMPLEMENTED("image->getArrayLayers() %d, baseArrayLayer %d",
			              image->getArrayLayers(), subresourceLayers.baseArrayLayer);
		}

		// From Vulkan 1.1 spec, section 11.5. Image Views:
		// "For cube and cube array image views, the layers of the image view starting
		//  at baseArrayLayer correspond to faces in the order +X, -X, +Y, -Y, +Z, -Z."
		VkImageSubresourceLayers posX = subresourceLayers;
		posX.layerCount = 1;
		VkImageSubresourceLayers negX = posX;
		negX.baseArrayLayer++;
		VkImageSubresourceLayers posY = negX;
		posY.baseArrayLayer++;
		VkImageSubresourceLayers negY = posY;
		negY.baseArrayLayer++;
		VkImageSubresourceLayers posZ = negY;
		posZ.baseArrayLayer++;
		VkImageSubresourceLayers negZ = posZ;
		negZ.baseArrayLayer++;

		// Copy top / bottom
		copyCubeEdge(image, posX, BOTTOM, negY, RIGHT);
		copyCubeEdge(image, posY, BOTTOM, posZ, TOP);
		copyCubeEdge(image, posZ, BOTTOM, negY, TOP);
		copyCubeEdge(image, negX, BOTTOM, negY, LEFT);
		copyCubeEdge(image, negY, BOTTOM, negZ, BOTTOM);
		copyCubeEdge(image, negZ, BOTTOM, negY, BOTTOM);

		copyCubeEdge(image, posX, TOP, posY, RIGHT);
		copyCubeEdge(image, posY, TOP, negZ, TOP);
		copyCubeEdge(image, posZ, TOP, posY, BOTTOM);
		copyCubeEdge(image, negX, TOP, posY, LEFT);
		copyCubeEdge(image, negY, TOP, posZ, BOTTOM);
		copyCubeEdge(image, negZ, TOP, posY, TOP);

		// Copy left / right
		copyCubeEdge(image, posX, RIGHT, negZ, LEFT);
		copyCubeEdge(image, posY, RIGHT, posX, TOP);
		copyCubeEdge(image, posZ, RIGHT, posX, LEFT);
		copyCubeEdge(image, negX, RIGHT, posZ, LEFT);
		copyCubeEdge(image, negY, RIGHT, posX, BOTTOM);
		copyCubeEdge(image, negZ, RIGHT, negX, LEFT);

		copyCubeEdge(image, posX, LEFT, posZ, RIGHT);
		copyCubeEdge(image, posY, LEFT, negX, TOP);
		copyCubeEdge(image, posZ, LEFT, negX, RIGHT);
		copyCubeEdge(image, negX, LEFT, negZ, RIGHT);
		copyCubeEdge(image, negY, LEFT, negX, BOTTOM);
		copyCubeEdge(image, negZ, LEFT, posX, RIGHT);

		// Compute corner colors
		VkImageAspectFlagBits aspect = static_cast<VkImageAspectFlagBits>(subresourceLayers.aspectMask);
		vk::Format format = image->getFormat(aspect);
		VkSampleCountFlagBits samples = image->getSampleCountFlagBits();
		State state(format, format, samples, samples, { 0xF });

		if(samples != VK_SAMPLE_COUNT_1_BIT)
		{
			UNIMPLEMENTED("Multi-sampled cube: %d samples", static_cast<int>(samples));
		}

		auto cornerUpdateRoutine = getCornerUpdateRoutine(state);
		if(!cornerUpdateRoutine)
		{
			return;
		}

		void(*cornerUpdateFunction)(const CubeBorderData *data) = (void(*)(const CubeBorderData*))cornerUpdateRoutine->getEntry();

		VkExtent3D extent = image->getMipLevelExtent(aspect, subresourceLayers.mipLevel);
		CubeBorderData data =
		{
			image->getTexelPointer({ 0, 0, 0 }, posX),
			image->rowPitchBytes(aspect, subresourceLayers.mipLevel),
			static_cast<uint32_t>(image->getLayerSize(aspect)),
			extent.width
		};
		cornerUpdateFunction(&data);
	}

	void Blitter::copyCubeEdge(vk::Image* image,
	                           const VkImageSubresourceLayers& dstSubresourceLayers, Edge dstEdge,
	                           const VkImageSubresourceLayers& srcSubresourceLayers, Edge srcEdge)
	{
		ASSERT(srcSubresourceLayers.aspectMask == dstSubresourceLayers.aspectMask);
		ASSERT(srcSubresourceLayers.mipLevel == dstSubresourceLayers.mipLevel);
		ASSERT(srcSubresourceLayers.baseArrayLayer != dstSubresourceLayers.baseArrayLayer);
		ASSERT(srcSubresourceLayers.layerCount == 1);
		ASSERT(dstSubresourceLayers.layerCount == 1);

		// Figure out if the edges to be copied in reverse order respectively from one another
		// The copy should be reversed whenever the same edges are contiguous or if we're
		// copying top <-> right or bottom <-> left. This is explained by the layout, which is:
		//
		//      | +y |
		// | -x | +z | +x | -z |
		//      | -y |

		bool reverse = (srcEdge == dstEdge) ||
		               ((srcEdge == TOP) && (dstEdge == RIGHT)) ||
		               ((srcEdge == RIGHT) && (dstEdge == TOP)) ||
		               ((srcEdge == BOTTOM) && (dstEdge == LEFT)) ||
		               ((srcEdge == LEFT) && (dstEdge == BOTTOM));

		VkImageAspectFlagBits aspect = static_cast<VkImageAspectFlagBits>(srcSubresourceLayers.aspectMask);
		int bytes = image->getFormat(aspect).bytes();
		int pitchB = image->rowPitchBytes(aspect, srcSubresourceLayers.mipLevel);

		VkExtent3D extent = image->getMipLevelExtent(aspect, srcSubresourceLayers.mipLevel);
		int w = extent.width;
		int h = extent.height;
		if(w != h)
		{
			UNIMPLEMENTED("Cube doesn't have square faces : (%d, %d)", w, h);
		}

		// Src is expressed in the regular [0, width-1], [0, height-1] space
		bool srcHorizontal = ((srcEdge == TOP) || (srcEdge == BOTTOM));
		int srcDelta = srcHorizontal ? bytes : pitchB;
		VkOffset3D srcOffset = { (srcEdge == RIGHT) ? (w - 1) : 0, (srcEdge == BOTTOM) ? (h - 1) : 0, 0 };

		// Dst contains borders, so it is expressed in the [-1, width], [-1, height] space
		bool dstHorizontal = ((dstEdge == TOP) || (dstEdge == BOTTOM));
		int dstDelta = (dstHorizontal ? bytes : pitchB) * (reverse ? -1 : 1);
		VkOffset3D dstOffset = { (dstEdge == RIGHT) ? w : -1, (dstEdge == BOTTOM) ? h : -1, 0 };

		// Don't write in the corners
		if(dstHorizontal)
		{
			dstOffset.x += reverse ? w : 1;
		}
		else
		{
			dstOffset.y += reverse ? h : 1;
		}

		const uint8_t* src = static_cast<const uint8_t*>(image->getTexelPointer(srcOffset, srcSubresourceLayers));
		uint8_t *dst = static_cast<uint8_t*>(image->getTexelPointer(dstOffset, dstSubresourceLayers));
		ASSERT((src < image->end()) && ((src + (w * srcDelta)) < image->end()));
		ASSERT((dst < image->end()) && ((dst + (w * dstDelta)) < image->end()));

		for(int i = 0; i < w; ++i, dst += dstDelta, src += srcDelta)
		{
			memcpy(dst, src, bytes);
		}
	}
}