/**
 ******************************************************************************
 * Xenia : Xbox 360 Emulator Research Project                                 *
 ******************************************************************************
 * Copyright 2022 Ben Vanik. All rights reserved.                             *
 * Released under the BSD license - see LICENSE in the root for more details. *
 ******************************************************************************
 */

#include "xenia/gpu/vulkan/vulkan_render_target_cache.h"

#include <cstddef>
#include <cstdint>
#include <cstring>

#include "third_party/glslang/SPIRV/GLSL.std.450.h"
#include "xenia/base/assert.h"
#include "xenia/base/cvar.h"
#include "xenia/base/logging.h"
#include "xenia/base/math.h"
#include "xenia/gpu/draw_util.h"
#include "xenia/gpu/registers.h"
#include "xenia/gpu/spirv_builder.h"
#include "xenia/gpu/spirv_shader_translator.h"
#include "xenia/gpu/texture_cache.h"
#include "xenia/gpu/vulkan/deferred_command_buffer.h"
#include "xenia/gpu/vulkan/vulkan_command_processor.h"
#include "xenia/gpu/xenos.h"
#include "xenia/ui/vulkan/vulkan_util.h"

DEFINE_string(
    render_target_path_vulkan, "",
    "Render target emulation path to use on Vulkan.\n"
    "Use: [any, fbo, fsi]\n"
    " fbo:\n"
    "  Host framebuffers and fixed-function blending and depth / stencil "
    "testing, copying between render targets when needed.\n"
    "  Lower accuracy (limited pixel format support).\n"
    "  Performance limited primarily by render target layout changes requiring "
    "copying, but generally higher.\n"
    " fsi:\n"
    "  Manual pixel packing, blending and depth / stencil testing, with free "
    "render target layout changes.\n"
    "  Requires a GPU supporting fragment shader interlock.\n"
    "  Highest accuracy (all pixel formats handled in software).\n"
    "  Performance limited primarily by overdraw.\n"
    " Any other value:\n"
    "  Choose what is considered the most optimal for the system (currently "
    "always FB because the FSI path is much slower now).",
    "GPU");

namespace xe {
namespace gpu {
namespace vulkan {

// Generated with `xb buildshaders`.
namespace shaders {
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/host_depth_store_1xmsaa_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/host_depth_store_2xmsaa_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/host_depth_store_4xmsaa_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/passthrough_position_xy_vs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_clear_32bpp_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_clear_32bpp_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_clear_64bpp_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_clear_64bpp_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_fast_32bpp_1x2xmsaa_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_fast_32bpp_1x2xmsaa_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_fast_32bpp_4xmsaa_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_fast_32bpp_4xmsaa_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_fast_64bpp_1x2xmsaa_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_fast_64bpp_1x2xmsaa_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_fast_64bpp_4xmsaa_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_fast_64bpp_4xmsaa_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_128bpp_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_128bpp_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_16bpp_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_16bpp_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_32bpp_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_32bpp_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_64bpp_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_64bpp_scaled_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_8bpp_cs.h"
#include "xenia/gpu/shaders/bytecode/vulkan_spirv/resolve_full_8bpp_scaled_cs.h"
}  // namespace shaders

const VulkanRenderTargetCache::ResolveCopyShaderCode
    VulkanRenderTargetCache::kResolveCopyShaders[size_t(
        draw_util::ResolveCopyShaderIndex::kCount)] = {
        {shaders::resolve_fast_32bpp_1x2xmsaa_cs,
         sizeof(shaders::resolve_fast_32bpp_1x2xmsaa_cs),
         shaders::resolve_fast_32bpp_1x2xmsaa_scaled_cs,
         sizeof(shaders::resolve_fast_32bpp_1x2xmsaa_scaled_cs)},
        {shaders::resolve_fast_32bpp_4xmsaa_cs,
         sizeof(shaders::resolve_fast_32bpp_4xmsaa_cs),
         shaders::resolve_fast_32bpp_4xmsaa_scaled_cs,
         sizeof(shaders::resolve_fast_32bpp_4xmsaa_scaled_cs)},
        {shaders::resolve_fast_64bpp_1x2xmsaa_cs,
         sizeof(shaders::resolve_fast_64bpp_1x2xmsaa_cs),
         shaders::resolve_fast_64bpp_1x2xmsaa_scaled_cs,
         sizeof(shaders::resolve_fast_64bpp_1x2xmsaa_scaled_cs)},
        {shaders::resolve_fast_64bpp_4xmsaa_cs,
         sizeof(shaders::resolve_fast_64bpp_4xmsaa_cs),
         shaders::resolve_fast_64bpp_4xmsaa_scaled_cs,
         sizeof(shaders::resolve_fast_64bpp_4xmsaa_scaled_cs)},
        {shaders::resolve_full_8bpp_cs, sizeof(shaders::resolve_full_8bpp_cs),
         shaders::resolve_full_8bpp_scaled_cs,
         sizeof(shaders::resolve_full_8bpp_scaled_cs)},
        {shaders::resolve_full_16bpp_cs, sizeof(shaders::resolve_full_16bpp_cs),
         shaders::resolve_full_16bpp_scaled_cs,
         sizeof(shaders::resolve_full_16bpp_scaled_cs)},
        {shaders::resolve_full_32bpp_cs, sizeof(shaders::resolve_full_32bpp_cs),
         shaders::resolve_full_32bpp_scaled_cs,
         sizeof(shaders::resolve_full_32bpp_scaled_cs)},
        {shaders::resolve_full_64bpp_cs, sizeof(shaders::resolve_full_64bpp_cs),
         shaders::resolve_full_64bpp_scaled_cs,
         sizeof(shaders::resolve_full_64bpp_scaled_cs)},
        {shaders::resolve_full_128bpp_cs,
         sizeof(shaders::resolve_full_128bpp_cs),
         shaders::resolve_full_128bpp_scaled_cs,
         sizeof(shaders::resolve_full_128bpp_scaled_cs)},
};

const VulkanRenderTargetCache::TransferPipelineLayoutInfo
    VulkanRenderTargetCache::kTransferPipelineLayoutInfos[size_t(
        TransferPipelineLayoutIndex::kCount)] = {
        // kColor
        {kTransferUsedDescriptorSetColorTextureBit,
         kTransferUsedPushConstantDwordAddressBit},
        // kDepth
        {kTransferUsedDescriptorSetDepthStencilTexturesBit,
         kTransferUsedPushConstantDwordAddressBit},
        // kColorToStencilBit
        {kTransferUsedDescriptorSetColorTextureBit,
         kTransferUsedPushConstantDwordAddressBit |
             kTransferUsedPushConstantDwordStencilMaskBit},
        // kDepthToStencilBit
        {kTransferUsedDescriptorSetDepthStencilTexturesBit,
         kTransferUsedPushConstantDwordAddressBit |
             kTransferUsedPushConstantDwordStencilMaskBit},
        // kColorAndHostDepthTexture
        {kTransferUsedDescriptorSetHostDepthStencilTexturesBit |
             kTransferUsedDescriptorSetColorTextureBit,
         kTransferUsedPushConstantDwordHostDepthAddressBit |
             kTransferUsedPushConstantDwordAddressBit},
        // kColorAndHostDepthBuffer
        {kTransferUsedDescriptorSetHostDepthBufferBit |
             kTransferUsedDescriptorSetColorTextureBit,
         kTransferUsedPushConstantDwordHostDepthAddressBit |
             kTransferUsedPushConstantDwordAddressBit},
        // kDepthAndHostDepthTexture
        {kTransferUsedDescriptorSetHostDepthStencilTexturesBit |
             kTransferUsedDescriptorSetDepthStencilTexturesBit,
         kTransferUsedPushConstantDwordHostDepthAddressBit |
             kTransferUsedPushConstantDwordAddressBit},
        // kDepthAndHostDepthBuffer
        {kTransferUsedDescriptorSetHostDepthBufferBit |
             kTransferUsedDescriptorSetDepthStencilTexturesBit,
         kTransferUsedPushConstantDwordHostDepthAddressBit |
             kTransferUsedPushConstantDwordAddressBit},
};

const VulkanRenderTargetCache::TransferModeInfo
    VulkanRenderTargetCache::kTransferModes[size_t(TransferMode::kCount)] = {
        // kColorToDepth
        {TransferOutput::kDepth, TransferPipelineLayoutIndex::kColor},
        // kColorToColor
        {TransferOutput::kColor, TransferPipelineLayoutIndex::kColor},
        // kDepthToDepth
        {TransferOutput::kDepth, TransferPipelineLayoutIndex::kDepth},
        // kDepthToColor
        {TransferOutput::kColor, TransferPipelineLayoutIndex::kDepth},
        // kColorToStencilBit
        {TransferOutput::kStencilBit,
         TransferPipelineLayoutIndex::kColorToStencilBit},
        // kDepthToStencilBit
        {TransferOutput::kStencilBit,
         TransferPipelineLayoutIndex::kDepthToStencilBit},
        // kColorAndHostDepthToDepth
        {TransferOutput::kDepth,
         TransferPipelineLayoutIndex::kColorAndHostDepthTexture},
        // kDepthAndHostDepthToDepth
        {TransferOutput::kDepth,
         TransferPipelineLayoutIndex::kDepthAndHostDepthTexture},
        // kColorAndHostDepthCopyToDepth
        {TransferOutput::kDepth,
         TransferPipelineLayoutIndex::kColorAndHostDepthBuffer},
        // kDepthAndHostDepthCopyToDepth
        {TransferOutput::kDepth,
         TransferPipelineLayoutIndex::kDepthAndHostDepthBuffer},
};

VulkanRenderTargetCache::VulkanRenderTargetCache(
    const RegisterFile& register_file, const Memory& memory,
    TraceWriter& trace_writer, uint32_t draw_resolution_scale_x,
    uint32_t draw_resolution_scale_y, VulkanCommandProcessor& command_processor)
    : RenderTargetCache(register_file, memory, &trace_writer,
                        draw_resolution_scale_x, draw_resolution_scale_y),
      command_processor_(command_processor),
      trace_writer_(trace_writer) {}

VulkanRenderTargetCache::~VulkanRenderTargetCache() { Shutdown(true); }

bool VulkanRenderTargetCache::Initialize(uint32_t shared_memory_binding_count) {
  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanInstance::Functions& ifn =
      vulkan_device->vulkan_instance()->functions();
  const VkPhysicalDevice physical_device = vulkan_device->physical_device();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();
  const ui::vulkan::VulkanDevice::Properties& device_properties =
      vulkan_device->properties();

  if (cvars::render_target_path_vulkan == "fsi") {
    path_ = Path::kPixelShaderInterlock;
  } else {
    path_ = Path::kHostRenderTargets;
  }
  // Fragment shader interlock is a feature implemented by pretty advanced GPUs,
  // closer to Direct3D 11 / OpenGL ES 3.2 level mainly, not Direct3D 10 /
  // OpenGL ES 3.1. Thus, it's fine to demand a wide range of other optional
  // features for the fragment shader interlock backend to work.
  if (path_ == Path::kPixelShaderInterlock) {
    // Interlocking between fragments with common sample coverage is enough, but
    // interlocking more is acceptable too (fragmentShaderShadingRateInterlock
    // would be okay too, but it's unlikely that an implementation would
    // advertise only it and not any other ones, as it's a very specific feature
    // interacting with another optional feature that is variable shading rate,
    // so there's no need to overcomplicate the checks and the shader execution
    // mode setting).
    // Sample-rate shading is required by certain SPIR-V revisions to access the
    // sample mask fragment shader input.
    // Stanard sample locations are needed for calculating the depth at the
    // samples.
    // It's unlikely that a device exposing fragment shader interlock won't have
    // a large enough storage buffer range and a sufficient SSBO slot count for
    // all the shared memory buffers and the EDRAM buffer - an in a conflict
    // between, for instance, the ability to vfetch and memexport in fragment
    // shaders, and the usage of fragment shader interlock, prefer the former
    // for simplicity.
    if (!(device_properties.fragmentShaderSampleInterlock ||
          device_properties.fragmentShaderPixelInterlock) ||
        !device_properties.fragmentStoresAndAtomics ||
        !device_properties.sampleRateShading ||
        !device_properties.standardSampleLocations ||
        shared_memory_binding_count >=
            device_properties.maxPerStageDescriptorStorageBuffers) {
      path_ = Path::kHostRenderTargets;
    }
  }

  // Format support.
  constexpr VkFormatFeatureFlags kUsedDepthFormatFeatures =
      VK_FORMAT_FEATURE_SAMPLED_IMAGE_BIT |
      VK_FORMAT_FEATURE_DEPTH_STENCIL_ATTACHMENT_BIT;
  VkFormatProperties depth_unorm24_properties;
  ifn.vkGetPhysicalDeviceFormatProperties(
      physical_device, VK_FORMAT_D24_UNORM_S8_UINT, &depth_unorm24_properties);
  depth_unorm24_vulkan_format_supported_ =
      (depth_unorm24_properties.optimalTilingFeatures &
       kUsedDepthFormatFeatures) == kUsedDepthFormatFeatures;

  // 2x MSAA support.
  // TODO(Triang3l): Handle sampledImageIntegerSampleCounts 4 not supported in
  // transfers.
  if (cvars::native_2x_msaa) {
    // Multisampled integer sampled images are optional in Vulkan and in Xenia.
    msaa_2x_attachments_supported_ =
        (device_properties.framebufferColorSampleCounts &
         device_properties.framebufferDepthSampleCounts &
         device_properties.framebufferStencilSampleCounts &
         device_properties.sampledImageColorSampleCounts &
         device_properties.sampledImageDepthSampleCounts &
         device_properties.sampledImageStencilSampleCounts &
         VK_SAMPLE_COUNT_2_BIT) &&
        (device_properties.sampledImageIntegerSampleCounts &
         (VK_SAMPLE_COUNT_2_BIT | VK_SAMPLE_COUNT_4_BIT)) !=
            VK_SAMPLE_COUNT_4_BIT;
    msaa_2x_no_attachments_supported_ =
        (device_properties.framebufferNoAttachmentsSampleCounts &
         VK_SAMPLE_COUNT_2_BIT) != 0;
  } else {
    msaa_2x_attachments_supported_ = false;
    msaa_2x_no_attachments_supported_ = false;
  }

  // Descriptor set layouts.
  VkDescriptorSetLayoutBinding descriptor_set_layout_bindings[2];
  descriptor_set_layout_bindings[0].binding = 0;
  descriptor_set_layout_bindings[0].descriptorType =
      VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
  descriptor_set_layout_bindings[0].descriptorCount = 1;
  descriptor_set_layout_bindings[0].stageFlags =
      VK_SHADER_STAGE_FRAGMENT_BIT | VK_SHADER_STAGE_COMPUTE_BIT;
  descriptor_set_layout_bindings[0].pImmutableSamplers = nullptr;
  VkDescriptorSetLayoutCreateInfo descriptor_set_layout_create_info;
  descriptor_set_layout_create_info.sType =
      VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO;
  descriptor_set_layout_create_info.pNext = nullptr;
  descriptor_set_layout_create_info.flags = 0;
  descriptor_set_layout_create_info.bindingCount = 1;
  descriptor_set_layout_create_info.pBindings = descriptor_set_layout_bindings;
  if (dfn.vkCreateDescriptorSetLayout(
          device, &descriptor_set_layout_create_info, nullptr,
          &descriptor_set_layout_storage_buffer_) != VK_SUCCESS) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to create the descriptor set layout "
        "with one storage buffer");
    Shutdown();
    return false;
  }
  descriptor_set_layout_bindings[0].descriptorType =
      VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE;
  if (dfn.vkCreateDescriptorSetLayout(
          device, &descriptor_set_layout_create_info, nullptr,
          &descriptor_set_layout_sampled_image_) != VK_SUCCESS) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to create the descriptor set layout "
        "with one sampled image");
    Shutdown();
    return false;
  }
  descriptor_set_layout_bindings[1].binding = 1;
  descriptor_set_layout_bindings[1].descriptorType =
      VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE;
  descriptor_set_layout_bindings[1].descriptorCount = 1;
  descriptor_set_layout_bindings[1].stageFlags =
      descriptor_set_layout_bindings[0].stageFlags;
  descriptor_set_layout_bindings[1].pImmutableSamplers = nullptr;
  descriptor_set_layout_create_info.bindingCount = 2;
  if (dfn.vkCreateDescriptorSetLayout(
          device, &descriptor_set_layout_create_info, nullptr,
          &descriptor_set_layout_sampled_image_x2_) != VK_SUCCESS) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to create the descriptor set layout "
        "with two sampled images");
    Shutdown();
    return false;
  }

  // Descriptor set pools.
  // The pool sizes were chosen without a specific reason.
  VkDescriptorPoolSize descriptor_set_layout_size;
  descriptor_set_layout_size.type = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE;
  descriptor_set_layout_size.descriptorCount = 1;
  descriptor_set_pool_sampled_image_ =
      std::make_unique<ui::vulkan::SingleLayoutDescriptorSetPool>(
          vulkan_device, 256, 1, &descriptor_set_layout_size,
          descriptor_set_layout_sampled_image_);
  descriptor_set_layout_size.descriptorCount = 2;
  descriptor_set_pool_sampled_image_x2_ =
      std::make_unique<ui::vulkan::SingleLayoutDescriptorSetPool>(
          vulkan_device, 256, 1, &descriptor_set_layout_size,
          descriptor_set_layout_sampled_image_x2_);

  // EDRAM contents reinterpretation buffer.
  // 90 MB with 9x resolution scaling - within the minimum
  // maxStorageBufferRange.
  if (!ui::vulkan::util::CreateDedicatedAllocationBuffer(
          vulkan_device,
          VkDeviceSize(xenos::kEdramSizeBytes *
                       (draw_resolution_scale_x() * draw_resolution_scale_y())),
          VK_BUFFER_USAGE_TRANSFER_SRC_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT |
              VK_BUFFER_USAGE_STORAGE_BUFFER_BIT,
          ui::vulkan::util::MemoryPurpose::kDeviceLocal, edram_buffer_,
          edram_buffer_memory_)) {
    XELOGE("VulkanRenderTargetCache: Failed to create the EDRAM buffer");
    Shutdown();
    return false;
  }
  if (GetPath() == Path::kPixelShaderInterlock) {
    // The first operation will likely be drawing.
    edram_buffer_usage_ = EdramBufferUsage::kFragmentReadWrite;
  } else {
    // The first operation will likely be depth self-comparison.
    edram_buffer_usage_ = EdramBufferUsage::kFragmentRead;
  }
  edram_buffer_modification_status_ =
      EdramBufferModificationStatus::kUnmodified;
  VkDescriptorPoolSize edram_storage_buffer_descriptor_pool_size;
  edram_storage_buffer_descriptor_pool_size.type =
      VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
  edram_storage_buffer_descriptor_pool_size.descriptorCount = 1;
  VkDescriptorPoolCreateInfo edram_storage_buffer_descriptor_pool_create_info;
  edram_storage_buffer_descriptor_pool_create_info.sType =
      VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO;
  edram_storage_buffer_descriptor_pool_create_info.pNext = nullptr;
  edram_storage_buffer_descriptor_pool_create_info.flags = 0;
  edram_storage_buffer_descriptor_pool_create_info.maxSets = 1;
  edram_storage_buffer_descriptor_pool_create_info.poolSizeCount = 1;
  edram_storage_buffer_descriptor_pool_create_info.pPoolSizes =
      &edram_storage_buffer_descriptor_pool_size;
  if (dfn.vkCreateDescriptorPool(
          device, &edram_storage_buffer_descriptor_pool_create_info, nullptr,
          &edram_storage_buffer_descriptor_pool_) != VK_SUCCESS) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to create the EDRAM buffer storage "
        "buffer descriptor pool");
    Shutdown();
    return false;
  }
  VkDescriptorSetAllocateInfo edram_storage_buffer_descriptor_set_allocate_info;
  edram_storage_buffer_descriptor_set_allocate_info.sType =
      VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO;
  edram_storage_buffer_descriptor_set_allocate_info.pNext = nullptr;
  edram_storage_buffer_descriptor_set_allocate_info.descriptorPool =
      edram_storage_buffer_descriptor_pool_;
  edram_storage_buffer_descriptor_set_allocate_info.descriptorSetCount = 1;
  edram_storage_buffer_descriptor_set_allocate_info.pSetLayouts =
      &descriptor_set_layout_storage_buffer_;
  if (dfn.vkAllocateDescriptorSets(
          device, &edram_storage_buffer_descriptor_set_allocate_info,
          &edram_storage_buffer_descriptor_set_) != VK_SUCCESS) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to allocate the EDRAM buffer storage "
        "buffer descriptor set");
    Shutdown();
    return false;
  }
  VkDescriptorBufferInfo edram_storage_buffer_descriptor_buffer_info;
  edram_storage_buffer_descriptor_buffer_info.buffer = edram_buffer_;
  edram_storage_buffer_descriptor_buffer_info.offset = 0;
  edram_storage_buffer_descriptor_buffer_info.range = VK_WHOLE_SIZE;
  VkWriteDescriptorSet edram_storage_buffer_descriptor_write;
  edram_storage_buffer_descriptor_write.sType =
      VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
  edram_storage_buffer_descriptor_write.pNext = nullptr;
  edram_storage_buffer_descriptor_write.dstSet =
      edram_storage_buffer_descriptor_set_;
  edram_storage_buffer_descriptor_write.dstBinding = 0;
  edram_storage_buffer_descriptor_write.dstArrayElement = 0;
  edram_storage_buffer_descriptor_write.descriptorCount = 1;
  edram_storage_buffer_descriptor_write.descriptorType =
      VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
  edram_storage_buffer_descriptor_write.pImageInfo = nullptr;
  edram_storage_buffer_descriptor_write.pBufferInfo =
      &edram_storage_buffer_descriptor_buffer_info;
  edram_storage_buffer_descriptor_write.pTexelBufferView = nullptr;
  dfn.vkUpdateDescriptorSets(device, 1, &edram_storage_buffer_descriptor_write,
                             0, nullptr);

  bool draw_resolution_scaled = IsDrawResolutionScaled();

  // Resolve copy pipeline layout.
  VkDescriptorSetLayout
      resolve_copy_descriptor_set_layouts[kResolveCopyDescriptorSetCount] = {};
  resolve_copy_descriptor_set_layouts[kResolveCopyDescriptorSetEdram] =
      descriptor_set_layout_storage_buffer_;
  resolve_copy_descriptor_set_layouts[kResolveCopyDescriptorSetDest] =
      command_processor_.GetSingleTransientDescriptorLayout(
          VulkanCommandProcessor::SingleTransientDescriptorLayout ::
              kStorageBufferCompute);
  VkPushConstantRange resolve_copy_push_constant_range;
  resolve_copy_push_constant_range.stageFlags = VK_SHADER_STAGE_COMPUTE_BIT;
  resolve_copy_push_constant_range.offset = 0;
  // Potentially binding all of the shared memory at 1x resolution, but only
  // portions with scaled resolution.
  resolve_copy_push_constant_range.size =
      draw_resolution_scaled
          ? sizeof(draw_util::ResolveCopyShaderConstants::DestRelative)
          : sizeof(draw_util::ResolveCopyShaderConstants);
  VkPipelineLayoutCreateInfo resolve_copy_pipeline_layout_create_info;
  resolve_copy_pipeline_layout_create_info.sType =
      VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
  resolve_copy_pipeline_layout_create_info.pNext = nullptr;
  resolve_copy_pipeline_layout_create_info.flags = 0;
  resolve_copy_pipeline_layout_create_info.setLayoutCount =
      kResolveCopyDescriptorSetCount;
  resolve_copy_pipeline_layout_create_info.pSetLayouts =
      resolve_copy_descriptor_set_layouts;
  resolve_copy_pipeline_layout_create_info.pushConstantRangeCount = 1;
  resolve_copy_pipeline_layout_create_info.pPushConstantRanges =
      &resolve_copy_push_constant_range;
  if (dfn.vkCreatePipelineLayout(
          device, &resolve_copy_pipeline_layout_create_info, nullptr,
          &resolve_copy_pipeline_layout_) != VK_SUCCESS) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to create the resolve copy pipeline "
        "layout");
    Shutdown();
    return false;
  }

  // Resolve copy pipelines.
  for (size_t i = 0; i < size_t(draw_util::ResolveCopyShaderIndex::kCount);
       ++i) {
    const draw_util::ResolveCopyShaderInfo& resolve_copy_shader_info =
        draw_util::resolve_copy_shader_info[i];
    const ResolveCopyShaderCode& resolve_copy_shader_code =
        kResolveCopyShaders[i];
    // Somewhat verification whether resolve_copy_shaders_ is up to date.
    assert_true(resolve_copy_shader_code.unscaled &&
                resolve_copy_shader_code.unscaled_size_bytes &&
                resolve_copy_shader_code.scaled &&
                resolve_copy_shader_code.scaled_size_bytes);
    VkPipeline resolve_copy_pipeline = ui::vulkan::util::CreateComputePipeline(
        vulkan_device, resolve_copy_pipeline_layout_,
        draw_resolution_scaled ? resolve_copy_shader_code.scaled
                               : resolve_copy_shader_code.unscaled,
        draw_resolution_scaled ? resolve_copy_shader_code.scaled_size_bytes
                               : resolve_copy_shader_code.unscaled_size_bytes);
    if (resolve_copy_pipeline == VK_NULL_HANDLE) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the resolve copy "
          "pipeline {}",
          resolve_copy_shader_info.debug_name);
      Shutdown();
      return false;
    }
    vulkan_device->SetObjectName(VK_OBJECT_TYPE_PIPELINE, resolve_copy_pipeline,
                                 resolve_copy_shader_info.debug_name);
    resolve_copy_pipelines_[i] = resolve_copy_pipeline;
  }

  // TODO(Triang3l): All paths (FSI).

  if (path_ == Path::kHostRenderTargets) {
    // Host render targets.

    depth_float24_round_ = cvars::depth_float24_round;

    // Host depth storing pipeline layout.
    VkDescriptorSetLayout host_depth_store_descriptor_set_layouts[] = {
        // Destination EDRAM storage buffer.
        descriptor_set_layout_storage_buffer_,
        // Source depth / stencil texture (only depth is used).
        descriptor_set_layout_sampled_image_x2_,
    };
    VkPushConstantRange host_depth_store_push_constant_range;
    host_depth_store_push_constant_range.stageFlags =
        VK_SHADER_STAGE_COMPUTE_BIT;
    host_depth_store_push_constant_range.offset = 0;
    host_depth_store_push_constant_range.size = sizeof(HostDepthStoreConstants);
    VkPipelineLayoutCreateInfo host_depth_store_pipeline_layout_create_info;
    host_depth_store_pipeline_layout_create_info.sType =
        VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
    host_depth_store_pipeline_layout_create_info.pNext = nullptr;
    host_depth_store_pipeline_layout_create_info.flags = 0;
    host_depth_store_pipeline_layout_create_info.setLayoutCount =
        uint32_t(xe::countof(host_depth_store_descriptor_set_layouts));
    host_depth_store_pipeline_layout_create_info.pSetLayouts =
        host_depth_store_descriptor_set_layouts;
    host_depth_store_pipeline_layout_create_info.pushConstantRangeCount = 1;
    host_depth_store_pipeline_layout_create_info.pPushConstantRanges =
        &host_depth_store_push_constant_range;
    if (dfn.vkCreatePipelineLayout(
            device, &host_depth_store_pipeline_layout_create_info, nullptr,
            &host_depth_store_pipeline_layout_) != VK_SUCCESS) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the host depth storing "
          "pipeline layout");
      Shutdown();
      return false;
    }
    constexpr std::pair<const uint32_t*, size_t> host_depth_store_shaders[] = {
        {shaders::host_depth_store_1xmsaa_cs,
         sizeof(shaders::host_depth_store_1xmsaa_cs)},
        {shaders::host_depth_store_2xmsaa_cs,
         sizeof(shaders::host_depth_store_2xmsaa_cs)},
        {shaders::host_depth_store_4xmsaa_cs,
         sizeof(shaders::host_depth_store_4xmsaa_cs)},
    };
    for (size_t i = 0; i < xe::countof(host_depth_store_shaders); ++i) {
      const std::pair<const uint32_t*, size_t> host_depth_store_shader =
          host_depth_store_shaders[i];
      VkPipeline host_depth_store_pipeline =
          ui::vulkan::util::CreateComputePipeline(
              vulkan_device, host_depth_store_pipeline_layout_,
              host_depth_store_shader.first, host_depth_store_shader.second);
      if (host_depth_store_pipeline == VK_NULL_HANDLE) {
        XELOGE(
            "VulkanRenderTargetCache: Failed to create the {}-sample host "
            "depth storing pipeline",
            uint32_t(1) << i);
        Shutdown();
        return false;
      }
      host_depth_store_pipelines_[i] = host_depth_store_pipeline;
    }

    // Transfer and clear vertex buffer, for quads of up to tile granularity.
    transfer_vertex_buffer_pool_ =
        std::make_unique<ui::vulkan::VulkanUploadBufferPool>(
            vulkan_device, VK_BUFFER_USAGE_VERTEX_BUFFER_BIT,
            std::max(ui::vulkan::VulkanUploadBufferPool::kDefaultPageSize,
                     sizeof(float) * 2 * 6 *
                         Transfer::kMaxCutoutBorderRectangles *
                         xenos::kEdramTileCount));

    // Transfer vertex shader.
    transfer_passthrough_vertex_shader_ = ui::vulkan::util::CreateShaderModule(
        vulkan_device, shaders::passthrough_position_xy_vs,
        sizeof(shaders::passthrough_position_xy_vs));
    if (transfer_passthrough_vertex_shader_ == VK_NULL_HANDLE) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the render target "
          "ownership transfer vertex shader");
      Shutdown();
      return false;
    }

    // Transfer pipeline layouts.
    VkDescriptorSetLayout transfer_pipeline_layout_descriptor_set_layouts
        [kTransferUsedDescriptorSetCount];
    VkPushConstantRange transfer_pipeline_layout_push_constant_range;
    transfer_pipeline_layout_push_constant_range.stageFlags =
        VK_SHADER_STAGE_FRAGMENT_BIT;
    transfer_pipeline_layout_push_constant_range.offset = 0;
    VkPipelineLayoutCreateInfo transfer_pipeline_layout_create_info;
    transfer_pipeline_layout_create_info.sType =
        VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
    transfer_pipeline_layout_create_info.pNext = nullptr;
    transfer_pipeline_layout_create_info.flags = 0;
    transfer_pipeline_layout_create_info.pSetLayouts =
        transfer_pipeline_layout_descriptor_set_layouts;
    transfer_pipeline_layout_create_info.pPushConstantRanges =
        &transfer_pipeline_layout_push_constant_range;
    for (size_t i = 0; i < size_t(TransferPipelineLayoutIndex::kCount); ++i) {
      const TransferPipelineLayoutInfo& transfer_pipeline_layout_info =
          kTransferPipelineLayoutInfos[i];
      transfer_pipeline_layout_create_info.setLayoutCount = 0;
      uint32_t transfer_pipeline_layout_descriptor_sets_remaining =
          transfer_pipeline_layout_info.used_descriptor_sets;
      uint32_t transfer_pipeline_layout_descriptor_set_index;
      while (xe::bit_scan_forward(
          transfer_pipeline_layout_descriptor_sets_remaining,
          &transfer_pipeline_layout_descriptor_set_index)) {
        transfer_pipeline_layout_descriptor_sets_remaining &=
            ~(uint32_t(1) << transfer_pipeline_layout_descriptor_set_index);
        VkDescriptorSetLayout transfer_pipeline_layout_descriptor_set_layout =
            VK_NULL_HANDLE;
        switch (TransferUsedDescriptorSet(
            transfer_pipeline_layout_descriptor_set_index)) {
          case kTransferUsedDescriptorSetHostDepthBuffer:
            transfer_pipeline_layout_descriptor_set_layout =
                descriptor_set_layout_storage_buffer_;
            break;
          case kTransferUsedDescriptorSetHostDepthStencilTextures:
          case kTransferUsedDescriptorSetDepthStencilTextures:
            transfer_pipeline_layout_descriptor_set_layout =
                descriptor_set_layout_sampled_image_x2_;
            break;
          case kTransferUsedDescriptorSetColorTexture:
            transfer_pipeline_layout_descriptor_set_layout =
                descriptor_set_layout_sampled_image_;
            break;
          default:
            assert_unhandled_case(TransferUsedDescriptorSet(
                transfer_pipeline_layout_descriptor_set_index));
        }
        transfer_pipeline_layout_descriptor_set_layouts
            [transfer_pipeline_layout_create_info.setLayoutCount++] =
                transfer_pipeline_layout_descriptor_set_layout;
      }
      transfer_pipeline_layout_push_constant_range.size = uint32_t(
          sizeof(uint32_t) *
          xe::bit_count(
              transfer_pipeline_layout_info.used_push_constant_dwords));
      transfer_pipeline_layout_create_info.pushConstantRangeCount =
          transfer_pipeline_layout_info.used_push_constant_dwords ? 1 : 0;
      if (dfn.vkCreatePipelineLayout(
              device, &transfer_pipeline_layout_create_info, nullptr,
              &transfer_pipeline_layouts_[i]) != VK_SUCCESS) {
        XELOGE(
            "VulkanRenderTargetCache: Failed to create the render target "
            "ownership transfer pipeline layout {}",
            i);
        Shutdown();
        return false;
      }
    }

    // Dump pipeline layouts.
    VkDescriptorSetLayout
        dump_pipeline_layout_descriptor_set_layouts[kDumpDescriptorSetCount];
    dump_pipeline_layout_descriptor_set_layouts[kDumpDescriptorSetEdram] =
        descriptor_set_layout_storage_buffer_;
    dump_pipeline_layout_descriptor_set_layouts[kDumpDescriptorSetSource] =
        descriptor_set_layout_sampled_image_;
    VkPushConstantRange dump_pipeline_layout_push_constant_range;
    dump_pipeline_layout_push_constant_range.stageFlags =
        VK_SHADER_STAGE_COMPUTE_BIT;
    dump_pipeline_layout_push_constant_range.offset = 0;
    dump_pipeline_layout_push_constant_range.size =
        sizeof(uint32_t) * kDumpPushConstantCount;
    VkPipelineLayoutCreateInfo dump_pipeline_layout_create_info;
    dump_pipeline_layout_create_info.sType =
        VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
    dump_pipeline_layout_create_info.pNext = nullptr;
    dump_pipeline_layout_create_info.flags = 0;
    dump_pipeline_layout_create_info.setLayoutCount =
        uint32_t(xe::countof(dump_pipeline_layout_descriptor_set_layouts));
    dump_pipeline_layout_create_info.pSetLayouts =
        dump_pipeline_layout_descriptor_set_layouts;
    dump_pipeline_layout_create_info.pushConstantRangeCount = 1;
    dump_pipeline_layout_create_info.pPushConstantRanges =
        &dump_pipeline_layout_push_constant_range;
    if (dfn.vkCreatePipelineLayout(device, &dump_pipeline_layout_create_info,
                                   nullptr, &dump_pipeline_layout_color_) !=
        VK_SUCCESS) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the color render target "
          "dumping pipeline layout");
      Shutdown();
      return false;
    }
    dump_pipeline_layout_descriptor_set_layouts[kDumpDescriptorSetSource] =
        descriptor_set_layout_sampled_image_x2_;
    if (dfn.vkCreatePipelineLayout(device, &dump_pipeline_layout_create_info,
                                   nullptr, &dump_pipeline_layout_depth_) !=
        VK_SUCCESS) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the depth render target "
          "dumping pipeline layout");
      Shutdown();
      return false;
    }
  } else if (path_ == Path::kPixelShaderInterlock) {
    // Pixel (fragment) shader interlock.

    // Blending is done in linear space directly in shaders.
    gamma_render_target_as_srgb_ = false;

    // Always true float24 depth rounded to the nearest even.
    depth_float24_round_ = true;

    // The pipeline layout and the pipelines for clearing the EDRAM buffer in
    // resolves.
    VkPushConstantRange resolve_fsi_clear_push_constant_range;
    resolve_fsi_clear_push_constant_range.stageFlags =
        VK_SHADER_STAGE_COMPUTE_BIT;
    resolve_fsi_clear_push_constant_range.offset = 0;
    resolve_fsi_clear_push_constant_range.size =
        sizeof(draw_util::ResolveClearShaderConstants);
    VkPipelineLayoutCreateInfo resolve_fsi_clear_pipeline_layout_create_info;
    resolve_fsi_clear_pipeline_layout_create_info.sType =
        VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO;
    resolve_fsi_clear_pipeline_layout_create_info.pNext = nullptr;
    resolve_fsi_clear_pipeline_layout_create_info.flags = 0;
    resolve_fsi_clear_pipeline_layout_create_info.setLayoutCount = 1;
    resolve_fsi_clear_pipeline_layout_create_info.pSetLayouts =
        &descriptor_set_layout_storage_buffer_;
    resolve_fsi_clear_pipeline_layout_create_info.pushConstantRangeCount = 1;
    resolve_fsi_clear_pipeline_layout_create_info.pPushConstantRanges =
        &resolve_fsi_clear_push_constant_range;
    if (dfn.vkCreatePipelineLayout(
            device, &resolve_fsi_clear_pipeline_layout_create_info, nullptr,
            &resolve_fsi_clear_pipeline_layout_) != VK_SUCCESS) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the resolve EDRAM buffer "
          "clear pipeline layout");
      Shutdown();
      return false;
    }
    resolve_fsi_clear_32bpp_pipeline_ = ui::vulkan::util::CreateComputePipeline(
        vulkan_device, resolve_fsi_clear_pipeline_layout_,
        draw_resolution_scaled ? shaders::resolve_clear_32bpp_scaled_cs
                               : shaders::resolve_clear_32bpp_cs,
        draw_resolution_scaled ? sizeof(shaders::resolve_clear_32bpp_scaled_cs)
                               : sizeof(shaders::resolve_clear_32bpp_cs));
    if (resolve_fsi_clear_32bpp_pipeline_ == VK_NULL_HANDLE) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the 32bpp resolve EDRAM "
          "buffer clear pipeline");
      Shutdown();
      return false;
    }
    resolve_fsi_clear_64bpp_pipeline_ = ui::vulkan::util::CreateComputePipeline(
        vulkan_device, resolve_fsi_clear_pipeline_layout_,
        draw_resolution_scaled ? shaders::resolve_clear_64bpp_scaled_cs
                               : shaders::resolve_clear_64bpp_cs,
        draw_resolution_scaled ? sizeof(shaders::resolve_clear_64bpp_scaled_cs)
                               : sizeof(shaders::resolve_clear_64bpp_cs));
    if (resolve_fsi_clear_64bpp_pipeline_ == VK_NULL_HANDLE) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the 64bpp resolve EDRAM "
          "buffer clear pipeline");
      Shutdown();
      return false;
    }

    // Common render pass.
    VkSubpassDescription fsi_subpass = {};
    fsi_subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS;
    // Fragment shader interlock provides synchronization and ordering within a
    // subpass, create an external by-region dependency to maintain interlocking
    // between passes. Framebuffer-global dependencies will be made with
    // explicit barriers when the addressing of the EDRAM buffer relatively to
    // the fragment coordinates is changed.
    VkSubpassDependency fsi_subpass_dependencies[2];
    fsi_subpass_dependencies[0].srcSubpass = VK_SUBPASS_EXTERNAL;
    fsi_subpass_dependencies[0].dstSubpass = 0;
    fsi_subpass_dependencies[0].srcStageMask =
        VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
    fsi_subpass_dependencies[0].dstStageMask =
        VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
    fsi_subpass_dependencies[0].srcAccessMask =
        VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
    fsi_subpass_dependencies[0].dstAccessMask =
        VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
    fsi_subpass_dependencies[0].dependencyFlags = VK_DEPENDENCY_BY_REGION_BIT;
    fsi_subpass_dependencies[1] = fsi_subpass_dependencies[0];
    std::swap(fsi_subpass_dependencies[1].srcSubpass,
              fsi_subpass_dependencies[1].dstSubpass);
    VkRenderPassCreateInfo fsi_render_pass_create_info;
    fsi_render_pass_create_info.sType =
        VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO;
    fsi_render_pass_create_info.pNext = nullptr;
    fsi_render_pass_create_info.flags = 0;
    fsi_render_pass_create_info.attachmentCount = 0;
    fsi_render_pass_create_info.pAttachments = nullptr;
    fsi_render_pass_create_info.subpassCount = 1;
    fsi_render_pass_create_info.pSubpasses = &fsi_subpass;
    fsi_render_pass_create_info.dependencyCount =
        uint32_t(xe::countof(fsi_subpass_dependencies));
    fsi_render_pass_create_info.pDependencies = fsi_subpass_dependencies;
    if (dfn.vkCreateRenderPass(device, &fsi_render_pass_create_info, nullptr,
                               &fsi_render_pass_) != VK_SUCCESS) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the fragment shader "
          "interlock render backend render pass");
      Shutdown();
      return false;
    }

    // Common framebuffer.
    VkFramebufferCreateInfo fsi_framebuffer_create_info;
    fsi_framebuffer_create_info.sType =
        VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO;
    fsi_framebuffer_create_info.pNext = nullptr;
    fsi_framebuffer_create_info.flags = 0;
    fsi_framebuffer_create_info.renderPass = fsi_render_pass_;
    fsi_framebuffer_create_info.attachmentCount = 0;
    fsi_framebuffer_create_info.pAttachments = nullptr;
    fsi_framebuffer_create_info.width = std::min(
        xenos::kTexture2DCubeMaxWidthHeight * draw_resolution_scale_x(),
        device_properties.maxFramebufferWidth);
    fsi_framebuffer_create_info.height = std::min(
        xenos::kTexture2DCubeMaxWidthHeight * draw_resolution_scale_y(),
        device_properties.maxFramebufferHeight);
    fsi_framebuffer_create_info.layers = 1;
    if (dfn.vkCreateFramebuffer(device, &fsi_framebuffer_create_info, nullptr,
                                &fsi_framebuffer_.framebuffer) != VK_SUCCESS) {
      XELOGE(
          "VulkanRenderTargetCache: Failed to create the fragment shader "
          "interlock render backend framebuffer");
      Shutdown();
      return false;
    }
    fsi_framebuffer_.host_extent.width = fsi_framebuffer_create_info.width;
    fsi_framebuffer_.host_extent.height = fsi_framebuffer_create_info.height;
  } else {
    assert_unhandled_case(path_);
    Shutdown();
    return false;
  }

  // Reset the last update structures, to keep the defaults consistent between
  // paths regardless of whether the update for the path actually modifies them.
  last_update_render_pass_key_ = RenderPassKey();
  last_update_render_pass_ = VK_NULL_HANDLE;
  last_update_framebuffer_pitch_tiles_at_32bpp_ = 0;
  std::memset(last_update_framebuffer_attachments_, 0,
              sizeof(last_update_framebuffer_attachments_));
  last_update_framebuffer_ = VK_NULL_HANDLE;

  InitializeCommon();
  return true;
}

void VulkanRenderTargetCache::Shutdown(bool from_destructor) {
  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();

  // Destroy all render targets before the descriptor set pool is destroyed -
  // may happen if shutting down the VulkanRenderTargetCache by destroying it,
  // so ShutdownCommon is called by the RenderTargetCache destructor, when it's
  // already too late.
  DestroyAllRenderTargets(true);

  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipeline, device,
                                         resolve_fsi_clear_64bpp_pipeline_);
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipeline, device,
                                         resolve_fsi_clear_32bpp_pipeline_);
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipelineLayout, device,
                                         resolve_fsi_clear_pipeline_layout_);

  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyFramebuffer, device,
                                         fsi_framebuffer_.framebuffer);
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyRenderPass, device,
                                         fsi_render_pass_);

  for (const auto& dump_pipeline_pair : dump_pipelines_) {
    // May be null to prevent recreation attempts.
    if (dump_pipeline_pair.second != VK_NULL_HANDLE) {
      dfn.vkDestroyPipeline(device, dump_pipeline_pair.second, nullptr);
    }
  }
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipelineLayout, device,
                                         dump_pipeline_layout_depth_);
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipelineLayout, device,
                                         dump_pipeline_layout_color_);

  for (const auto& transfer_pipeline_array_pair : transfer_pipelines_) {
    for (VkPipeline transfer_pipeline : transfer_pipeline_array_pair.second) {
      // May be null to prevent recreation attempts.
      if (transfer_pipeline != VK_NULL_HANDLE) {
        dfn.vkDestroyPipeline(device, transfer_pipeline, nullptr);
      }
    }
  }
  transfer_pipelines_.clear();
  for (const auto& transfer_shader_pair : transfer_shaders_) {
    if (transfer_shader_pair.second != VK_NULL_HANDLE) {
      dfn.vkDestroyShaderModule(device, transfer_shader_pair.second, nullptr);
    }
  }
  transfer_shaders_.clear();
  for (size_t i = 0; i < size_t(TransferPipelineLayoutIndex::kCount); ++i) {
    ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipelineLayout, device,
                                           transfer_pipeline_layouts_[i]);
  }
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyShaderModule, device,
                                         transfer_passthrough_vertex_shader_);
  transfer_vertex_buffer_pool_.reset();

  for (size_t i = 0; i < xe::countof(host_depth_store_pipelines_); ++i) {
    ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipeline, device,
                                           host_depth_store_pipelines_[i]);
  }
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipelineLayout, device,
                                         host_depth_store_pipeline_layout_);

  last_update_framebuffer_ = VK_NULL_HANDLE;
  for (const auto& framebuffer_pair : framebuffers_) {
    dfn.vkDestroyFramebuffer(device, framebuffer_pair.second.framebuffer,
                             nullptr);
  }
  framebuffers_.clear();

  last_update_render_pass_ = VK_NULL_HANDLE;
  for (const auto& render_pass_pair : render_passes_) {
    if (render_pass_pair.second != VK_NULL_HANDLE) {
      dfn.vkDestroyRenderPass(device, render_pass_pair.second, nullptr);
    }
  }
  render_passes_.clear();

  for (VkPipeline& resolve_copy_pipeline : resolve_copy_pipelines_) {
    ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipeline, device,
                                           resolve_copy_pipeline);
  }
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyPipelineLayout, device,
                                         resolve_copy_pipeline_layout_);

  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyDescriptorPool, device,
                                         edram_storage_buffer_descriptor_pool_);
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyBuffer, device,
                                         edram_buffer_);
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkFreeMemory, device,
                                         edram_buffer_memory_);

  descriptor_set_pool_sampled_image_x2_.reset();
  descriptor_set_pool_sampled_image_.reset();

  ui::vulkan::util::DestroyAndNullHandle(
      dfn.vkDestroyDescriptorSetLayout, device,
      descriptor_set_layout_sampled_image_x2_);
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyDescriptorSetLayout,
                                         device,
                                         descriptor_set_layout_sampled_image_);
  ui::vulkan::util::DestroyAndNullHandle(dfn.vkDestroyDescriptorSetLayout,
                                         device,
                                         descriptor_set_layout_storage_buffer_);

  if (!from_destructor) {
    ShutdownCommon();
  }
}

void VulkanRenderTargetCache::ClearCache() {
  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();

  // Framebuffer objects must be destroyed because they reference views of
  // attachment images, which may be removed by the common ClearCache.
  last_update_framebuffer_ = VK_NULL_HANDLE;
  for (const auto& framebuffer_pair : framebuffers_) {
    dfn.vkDestroyFramebuffer(device, framebuffer_pair.second.framebuffer,
                             nullptr);
  }
  framebuffers_.clear();

  last_update_render_pass_ = VK_NULL_HANDLE;
  for (const auto& render_pass_pair : render_passes_) {
    dfn.vkDestroyRenderPass(device, render_pass_pair.second, nullptr);
  }
  render_passes_.clear();

  RenderTargetCache::ClearCache();
}

void VulkanRenderTargetCache::CompletedSubmissionUpdated() {
  if (transfer_vertex_buffer_pool_) {
    transfer_vertex_buffer_pool_->Reclaim(
        command_processor_.GetCompletedSubmission());
  }
}

void VulkanRenderTargetCache::EndSubmission() {
  if (transfer_vertex_buffer_pool_) {
    transfer_vertex_buffer_pool_->FlushWrites();
  }
}

bool VulkanRenderTargetCache::Resolve(const Memory& memory,
                                      VulkanSharedMemory& shared_memory,
                                      VulkanTextureCache& texture_cache,
                                      uint32_t& written_address_out,
                                      uint32_t& written_length_out) {
  written_address_out = 0;
  written_length_out = 0;

  bool draw_resolution_scaled = IsDrawResolutionScaled();

  draw_util::ResolveInfo resolve_info;
  if (!draw_util::GetResolveInfo(
          register_file(), memory, trace_writer_, draw_resolution_scale_x(),
          draw_resolution_scale_y(), IsFixedRG16TruncatedToMinus1To1(),
          IsFixedRGBA16TruncatedToMinus1To1(), resolve_info)) {
    return false;
  }

  // Nothing to copy/clear.
  if (!resolve_info.coordinate_info.width_div_8 || !resolve_info.height_div_8) {
    return true;
  }

  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();
  DeferredCommandBuffer& command_buffer =
      command_processor_.deferred_command_buffer();

  // Copying.
  bool copied = false;
  if (resolve_info.copy_dest_extent_length) {
    if (GetPath() == Path::kHostRenderTargets) {
      // Dump the current contents of the render targets owning the affected
      // range to edram_buffer_.
      // TODO(Triang3l): Direct host render target -> shared memory resolve
      // shaders for non-converting cases.
      uint32_t dump_base;
      uint32_t dump_row_length_used;
      uint32_t dump_rows;
      uint32_t dump_pitch;
      resolve_info.GetCopyEdramTileSpan(dump_base, dump_row_length_used,
                                        dump_rows, dump_pitch);
      // Scale tile parameters for resolution scaling to match resolve shader
      // expectations
      if (IsDrawResolutionScaled()) {
        dump_row_length_used *= draw_resolution_scale_x();
        dump_rows *= draw_resolution_scale_y();
        dump_pitch *= draw_resolution_scale_x();
      }
      DumpRenderTargets(dump_base, dump_row_length_used, dump_rows, dump_pitch);
    }

    draw_util::ResolveCopyShaderConstants copy_shader_constants;
    uint32_t copy_group_count_x, copy_group_count_y;
    draw_util::ResolveCopyShaderIndex copy_shader = resolve_info.GetCopyShader(
        draw_resolution_scale_x(), draw_resolution_scale_y(),
        copy_shader_constants, copy_group_count_x, copy_group_count_y);
    assert_true(copy_group_count_x && copy_group_count_y);
    if (copy_shader != draw_util::ResolveCopyShaderIndex::kUnknown) {
      const draw_util::ResolveCopyShaderInfo& copy_shader_info =
          draw_util::resolve_copy_shader_info[size_t(copy_shader)];

      // Make sure there is memory to write to.
      bool copy_dest_committed;
      // TODO(Triang3l): Resolution-scaled buffer committing.
      copy_dest_committed =
          shared_memory.RequestRange(resolve_info.copy_dest_extent_start,
                                     resolve_info.copy_dest_extent_length);
      if (!copy_dest_committed) {
        XELOGE(
            "VulkanRenderTargetCache: Failed to obtain the resolve destination "
            "memory region");
      } else {
        // TODO(Triang3l): Switching between descriptors if exceeding
        // maxStorageBufferRange.
        // TODO(Triang3l): Use a single 512 MB shared memory binding if
        // possible.
        VkDescriptorSet descriptor_set_dest =
            command_processor_.AllocateSingleTransientDescriptor(
                VulkanCommandProcessor::SingleTransientDescriptorLayout ::
                    kStorageBufferCompute);
        if (descriptor_set_dest != VK_NULL_HANDLE) {
          // Write the destination descriptor.
          VkDescriptorBufferInfo write_descriptor_set_dest_buffer_info;

          bool scaled_buffer_ready = false;
          if (draw_resolution_scaled) {
            // For scaled resolve, ensure the scaled buffer exists and bind to
            // it
            uint32_t dest_address = resolve_info.copy_dest_base;
            uint32_t dest_length = resolve_info.copy_dest_extent_start -
                                   resolve_info.copy_dest_base +
                                   resolve_info.copy_dest_extent_length;

            // Ensure scaled resolve memory is committed
            scaled_buffer_ready = true;
            if (!texture_cache.EnsureScaledResolveMemoryCommittedPublic(
                    dest_address, dest_length)) {
              XELOGE(
                  "Failed to commit scaled resolve memory for resolve dest at "
                  "0x{:08X}",
                  dest_address);
              scaled_buffer_ready = false;
            }

            // Make the range current to get the buffer
            if (scaled_buffer_ready &&
                !texture_cache.MakeScaledResolveRangeCurrent(dest_address,
                                                             dest_length)) {
              XELOGE(
                  "Failed to make scaled resolve range current for resolve "
                  "dest at 0x{:08X}",
                  dest_address);
              scaled_buffer_ready = false;
            }

            // Get the current scaled buffer
            VkBuffer scaled_buffer = VK_NULL_HANDLE;
            if (scaled_buffer_ready) {
              scaled_buffer = texture_cache.GetCurrentScaledResolveBuffer();
              if (scaled_buffer == VK_NULL_HANDLE) {
                XELOGE(
                    "No current scaled resolve buffer for resolve dest at "
                    "0x{:08X}",
                    dest_address);
                scaled_buffer_ready = false;
              }
            }

            if (scaled_buffer_ready) {
              // Calculate offset within the scaled buffer
              uint32_t draw_resolution_scale_area =
                  draw_resolution_scale_x() * draw_resolution_scale_y();
              uint64_t scaled_offset =
                  uint64_t(dest_address) * draw_resolution_scale_area;

              // Get the buffer's base offset to calculate relative offset
              uint64_t buffer_relative_offset = 0;
              size_t buffer_index =
                  texture_cache.GetScaledResolveCurrentBufferIndex();
              auto* buffer_info =
                  texture_cache.GetScaledResolveBufferInfo(buffer_index);
              if (buffer_info) {
                buffer_relative_offset =
                    scaled_offset - buffer_info->range_start_scaled;
              }

              write_descriptor_set_dest_buffer_info.buffer = scaled_buffer;
              write_descriptor_set_dest_buffer_info.offset =
                  buffer_relative_offset;
              write_descriptor_set_dest_buffer_info.range =
                  dest_length * draw_resolution_scale_area;
            }
          }

          if (!scaled_buffer_ready) {
            // Regular unscaled resolve - write to shared memory
            if (draw_resolution_scaled) {
              XELOGW(
                  "Falling back to unscaled resolve at 0x{:08X} - scaled "
                  "buffer not available",
                  resolve_info.copy_dest_base);
            }
            write_descriptor_set_dest_buffer_info.buffer =
                shared_memory.buffer();
            write_descriptor_set_dest_buffer_info.offset =
                resolve_info.copy_dest_base;
            write_descriptor_set_dest_buffer_info.range =
                resolve_info.copy_dest_extent_start -
                resolve_info.copy_dest_base +
                resolve_info.copy_dest_extent_length;
          }
          VkWriteDescriptorSet write_descriptor_set_dest;
          write_descriptor_set_dest.sType =
              VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
          write_descriptor_set_dest.pNext = nullptr;
          write_descriptor_set_dest.dstSet = descriptor_set_dest;
          write_descriptor_set_dest.dstBinding = 0;
          write_descriptor_set_dest.dstArrayElement = 0;
          write_descriptor_set_dest.descriptorCount = 1;
          write_descriptor_set_dest.descriptorType =
              VK_DESCRIPTOR_TYPE_STORAGE_BUFFER;
          write_descriptor_set_dest.pImageInfo = nullptr;
          write_descriptor_set_dest.pBufferInfo =
              &write_descriptor_set_dest_buffer_info;
          write_descriptor_set_dest.pTexelBufferView = nullptr;
          dfn.vkUpdateDescriptorSets(device, 1, &write_descriptor_set_dest, 0,
                                     nullptr);

          // Submit the resolve.
          if (!scaled_buffer_ready) {
            // Regular unscaled - transition shared memory for write
            shared_memory.Use(VulkanSharedMemory::Usage::kComputeWrite,
                              std::pair<uint32_t, uint32_t>(
                                  resolve_info.copy_dest_extent_start,
                                  resolve_info.copy_dest_extent_length));
          } else {
            // Scaled - add barrier for the scaled resolve buffer
            // The buffer transitions from compute shader read (texture loading)
            // to compute shader write
            VkBuffer scaled_buffer =
                texture_cache.GetCurrentScaledResolveBuffer();
            if (scaled_buffer != VK_NULL_HANDLE) {
              VkBufferMemoryBarrier buffer_barrier = {};
              buffer_barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
              // More specific: previous compute shader reads to compute shader
              // write
              buffer_barrier.srcAccessMask = VK_ACCESS_SHADER_READ_BIT;
              buffer_barrier.dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
              buffer_barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
              buffer_barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
              buffer_barrier.buffer = scaled_buffer;
              buffer_barrier.offset = 0;
              buffer_barrier.size = VK_WHOLE_SIZE;

              command_buffer.CmdVkPipelineBarrier(
                  VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,  // From compute shader
                  VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,  // To compute shader
                  0, 0, nullptr, 1, &buffer_barrier, 0, nullptr);
            }
          }
          UseEdramBuffer(EdramBufferUsage::kComputeRead);
          command_processor_.BindExternalComputePipeline(
              resolve_copy_pipelines_[size_t(copy_shader)]);
          VkDescriptorSet descriptor_sets[kResolveCopyDescriptorSetCount] = {};
          descriptor_sets[kResolveCopyDescriptorSetEdram] =
              edram_storage_buffer_descriptor_set_;
          descriptor_sets[kResolveCopyDescriptorSetDest] = descriptor_set_dest;
          command_buffer.CmdVkBindDescriptorSets(
              VK_PIPELINE_BIND_POINT_COMPUTE, resolve_copy_pipeline_layout_, 0,
              uint32_t(xe::countof(descriptor_sets)), descriptor_sets, 0,
              nullptr);
          if (draw_resolution_scaled) {
            command_buffer.CmdVkPushConstants(
                resolve_copy_pipeline_layout_, VK_SHADER_STAGE_COMPUTE_BIT, 0,
                sizeof(copy_shader_constants.dest_relative),
                &copy_shader_constants.dest_relative);
          } else {
            // TODO(Triang3l): Proper dest_base in case of one 512 MB shared
            // memory binding, or multiple shared memory bindings in case of
            // splitting due to maxStorageBufferRange overflow.
            copy_shader_constants.dest_base -=
                uint32_t(write_descriptor_set_dest_buffer_info.offset);
            command_buffer.CmdVkPushConstants(
                resolve_copy_pipeline_layout_, VK_SHADER_STAGE_COMPUTE_BIT, 0,
                sizeof(copy_shader_constants), &copy_shader_constants);
          }
          command_processor_.SubmitBarriers(true);
          command_buffer.CmdVkDispatch(copy_group_count_x, copy_group_count_y,
                                       1);

          // Add barrier after writing to scaled resolve buffer
          if (scaled_buffer_ready) {
            VkBuffer scaled_buffer =
                texture_cache.GetCurrentScaledResolveBuffer();
            if (scaled_buffer != VK_NULL_HANDLE) {
              VkBufferMemoryBarrier buffer_barrier = {};
              buffer_barrier.sType = VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER;
              buffer_barrier.srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT;
              buffer_barrier.dstAccessMask = VK_ACCESS_SHADER_READ_BIT;
              buffer_barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
              buffer_barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
              buffer_barrier.buffer = scaled_buffer;
              buffer_barrier.offset = 0;
              buffer_barrier.size = VK_WHOLE_SIZE;

              command_buffer.CmdVkPipelineBarrier(
                  VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
                  VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, 0, nullptr, 1,
                  &buffer_barrier, 0, nullptr);
            }
          }

          // Invalidate textures and mark the range as scaled if needed.
          texture_cache.MarkRangeAsResolved(
              resolve_info.copy_dest_extent_start,
              resolve_info.copy_dest_extent_length);
          written_address_out = resolve_info.copy_dest_extent_start;
          written_length_out = resolve_info.copy_dest_extent_length;
          copied = true;
        }
      }
    }
  } else {
    copied = true;
  }

  // Clearing.
  bool cleared = false;
  bool clear_depth = resolve_info.IsClearingDepth();
  bool clear_color = resolve_info.IsClearingColor();
  if (clear_depth || clear_color) {
    switch (GetPath()) {
      case Path::kHostRenderTargets: {
        Transfer::Rectangle clear_rectangle;
        RenderTarget* clear_render_targets[2];
        // If PrepareHostRenderTargetsResolveClear returns false, may be just an
        // empty region (success) or an error - don't care.
        if (PrepareHostRenderTargetsResolveClear(
                resolve_info, clear_rectangle, clear_render_targets[0],
                clear_transfers_[0], clear_render_targets[1],
                clear_transfers_[1])) {
          uint64_t clear_values[2];
          clear_values[0] = resolve_info.rb_depth_clear;
          clear_values[1] = resolve_info.rb_color_clear |
                            (uint64_t(resolve_info.rb_color_clear_lo) << 32);
          PerformTransfersAndResolveClears(2, clear_render_targets,
                                           clear_transfers_, clear_values,
                                           &clear_rectangle);
        }
        cleared = true;
      } break;
      case Path::kPixelShaderInterlock: {
        UseEdramBuffer(EdramBufferUsage::kComputeWrite);
        // Should be safe to only commit once (if was accessed as unordered or
        // with fragment shader interlock previously - if there was nothing to
        // copy, only to clear, for some reason, for instance), overlap of the
        // depth and the color ranges is highly unlikely.
        CommitEdramBufferShaderWrites();
        command_buffer.CmdVkBindDescriptorSets(
            VK_PIPELINE_BIND_POINT_COMPUTE, resolve_fsi_clear_pipeline_layout_,
            0, 1, &edram_storage_buffer_descriptor_set_, 0, nullptr);
        std::pair<uint32_t, uint32_t> clear_group_count =
            resolve_info.GetClearShaderGroupCount(draw_resolution_scale_x(),
                                                  draw_resolution_scale_y());
        assert_true(clear_group_count.first && clear_group_count.second);
        if (clear_depth) {
          command_processor_.BindExternalComputePipeline(
              resolve_fsi_clear_32bpp_pipeline_);
          draw_util::ResolveClearShaderConstants depth_clear_constants;
          resolve_info.GetDepthClearShaderConstants(depth_clear_constants);
          command_buffer.CmdVkPushConstants(
              resolve_fsi_clear_pipeline_layout_, VK_SHADER_STAGE_COMPUTE_BIT,
              0, sizeof(depth_clear_constants), &depth_clear_constants);
          command_processor_.SubmitBarriers(true);
          command_buffer.CmdVkDispatch(clear_group_count.first,
                                       clear_group_count.second, 1);
        }
        if (clear_color) {
          command_processor_.BindExternalComputePipeline(
              resolve_info.color_edram_info.format_is_64bpp
                  ? resolve_fsi_clear_64bpp_pipeline_
                  : resolve_fsi_clear_32bpp_pipeline_);
          draw_util::ResolveClearShaderConstants color_clear_constants;
          resolve_info.GetColorClearShaderConstants(color_clear_constants);
          if (clear_depth) {
            // Non-RT-specific constants have already been set.
            command_buffer.CmdVkPushConstants(
                resolve_fsi_clear_pipeline_layout_, VK_SHADER_STAGE_COMPUTE_BIT,
                uint32_t(offsetof(draw_util::ResolveClearShaderConstants,
                                  rt_specific)),
                sizeof(color_clear_constants.rt_specific),
                &color_clear_constants.rt_specific);
          } else {
            command_buffer.CmdVkPushConstants(
                resolve_fsi_clear_pipeline_layout_, VK_SHADER_STAGE_COMPUTE_BIT,
                0, sizeof(color_clear_constants), &color_clear_constants);
          }
          command_processor_.SubmitBarriers(true);
          command_buffer.CmdVkDispatch(clear_group_count.first,
                                       clear_group_count.second, 1);
        }
        MarkEdramBufferModified();
        cleared = true;
      } break;
      default:
        assert_unhandled_case(GetPath());
    }
  } else {
    cleared = true;
  }

  return copied && cleared;
}

bool VulkanRenderTargetCache::Update(
    bool is_rasterization_done, reg::RB_DEPTHCONTROL normalized_depth_control,
    uint32_t normalized_color_mask, const Shader& vertex_shader) {
  if (!RenderTargetCache::Update(is_rasterization_done,
                                 normalized_depth_control,
                                 normalized_color_mask, vertex_shader)) {
    return false;
  }

  auto rb_surface_info = register_file().Get<reg::RB_SURFACE_INFO>();

  RenderPassKey render_pass_key;
  // Needed even with the fragment shader interlock render backend for passing
  // the sample count to the pipeline cache.
  render_pass_key.msaa_samples = rb_surface_info.msaa_samples;

  switch (GetPath()) {
    case Path::kHostRenderTargets: {
      RenderTarget* const* depth_and_color_render_targets =
          last_update_accumulated_render_targets();

      PerformTransfersAndResolveClears(1 + xenos::kMaxColorRenderTargets,
                                       depth_and_color_render_targets,
                                       last_update_transfers());

      uint32_t render_targets_are_srgb =
          gamma_render_target_as_srgb_
              ? last_update_accumulated_color_targets_are_gamma()
              : 0;

      if (depth_and_color_render_targets[0]) {
        render_pass_key.depth_and_color_used |= 1 << 0;
        render_pass_key.depth_format =
            depth_and_color_render_targets[0]->key().GetDepthFormat();
      }
      if (depth_and_color_render_targets[1]) {
        render_pass_key.depth_and_color_used |= 1 << 1;
        render_pass_key.color_0_view_format =
            (render_targets_are_srgb & (1 << 0))
                ? xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA
                : depth_and_color_render_targets[1]->key().GetColorFormat();
      }
      if (depth_and_color_render_targets[2]) {
        render_pass_key.depth_and_color_used |= 1 << 2;
        render_pass_key.color_1_view_format =
            (render_targets_are_srgb & (1 << 1))
                ? xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA
                : depth_and_color_render_targets[2]->key().GetColorFormat();
      }
      if (depth_and_color_render_targets[3]) {
        render_pass_key.depth_and_color_used |= 1 << 3;
        render_pass_key.color_2_view_format =
            (render_targets_are_srgb & (1 << 2))
                ? xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA
                : depth_and_color_render_targets[3]->key().GetColorFormat();
      }
      if (depth_and_color_render_targets[4]) {
        render_pass_key.depth_and_color_used |= 1 << 4;
        render_pass_key.color_3_view_format =
            (render_targets_are_srgb & (1 << 3))
                ? xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA
                : depth_and_color_render_targets[4]->key().GetColorFormat();
      }

      const Framebuffer* framebuffer = last_update_framebuffer_;
      VkRenderPass render_pass = last_update_render_pass_key_ == render_pass_key
                                     ? last_update_render_pass_
                                     : VK_NULL_HANDLE;
      if (render_pass == VK_NULL_HANDLE) {
        render_pass = GetHostRenderTargetsRenderPass(render_pass_key);
        if (render_pass == VK_NULL_HANDLE) {
          return false;
        }
        // Framebuffer for a different render pass needed now.
        framebuffer = nullptr;
      }

      uint32_t pitch_tiles_at_32bpp =
          ((rb_surface_info.surface_pitch << uint32_t(
                rb_surface_info.msaa_samples >= xenos::MsaaSamples::k4X)) +
           (xenos::kEdramTileWidthSamples - 1)) /
          xenos::kEdramTileWidthSamples;
      if (framebuffer) {
        if (last_update_framebuffer_pitch_tiles_at_32bpp_ !=
                pitch_tiles_at_32bpp ||
            std::memcmp(last_update_framebuffer_attachments_,
                        depth_and_color_render_targets,
                        sizeof(last_update_framebuffer_attachments_))) {
          framebuffer = nullptr;
        }
      }
      if (!framebuffer) {
        framebuffer = GetHostRenderTargetsFramebuffer(
            render_pass_key, pitch_tiles_at_32bpp,
            depth_and_color_render_targets);
        if (!framebuffer) {
          return false;
        }
      }

      // Successful update - write the new configuration.
      last_update_render_pass_key_ = render_pass_key;
      last_update_render_pass_ = render_pass;
      last_update_framebuffer_pitch_tiles_at_32bpp_ = pitch_tiles_at_32bpp;
      std::memcpy(last_update_framebuffer_attachments_,
                  depth_and_color_render_targets,
                  sizeof(last_update_framebuffer_attachments_));
      last_update_framebuffer_ = framebuffer;

      // Transition the used render targets.
      for (uint32_t i = 0; i < 1 + xenos::kMaxColorRenderTargets; ++i) {
        RenderTarget* rt = depth_and_color_render_targets[i];
        if (!rt) {
          continue;
        }
        auto& vulkan_rt = *static_cast<VulkanRenderTarget*>(rt);
        VkPipelineStageFlags rt_dst_stage_mask;
        VkAccessFlags rt_dst_access_mask;
        VkImageLayout rt_new_layout;
        VulkanRenderTarget::GetDrawUsage(i == 0, &rt_dst_stage_mask,
                                         &rt_dst_access_mask, &rt_new_layout);
        command_processor_.PushImageMemoryBarrier(
            vulkan_rt.image(),
            ui::vulkan::util::InitializeSubresourceRange(
                i ? VK_IMAGE_ASPECT_COLOR_BIT
                  : (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)),
            vulkan_rt.current_stage_mask(), rt_dst_stage_mask,
            vulkan_rt.current_access_mask(), rt_dst_access_mask,
            vulkan_rt.current_layout(), rt_new_layout);
        vulkan_rt.SetUsage(rt_dst_stage_mask, rt_dst_access_mask,
                           rt_new_layout);
      }
    } break;

    case Path::kPixelShaderInterlock: {
      // For FSI, only the barrier is needed - already scheduled if required.
      // But the buffer will be used for FSI drawing now.
      UseEdramBuffer(EdramBufferUsage::kFragmentReadWrite);
      // Commit preceding unordered (but not FSI) writes like clears as they
      // aren't synchronized with FSI accesses.
      CommitEdramBufferShaderWrites(
          EdramBufferModificationStatus::kViaUnordered);
      // TODO(Triang3l): Check if this draw call modifies color or depth /
      // stencil, at least coarsely, to prevent useless barriers.
      MarkEdramBufferModified(
          EdramBufferModificationStatus::kViaFragmentShaderInterlock);
      last_update_render_pass_key_ = render_pass_key;
      last_update_render_pass_ = fsi_render_pass_;
      last_update_framebuffer_ = &fsi_framebuffer_;
    } break;

    default:
      assert_unhandled_case(GetPath());
      return false;
  }

  return true;
}

VkRenderPass VulkanRenderTargetCache::GetHostRenderTargetsRenderPass(
    RenderPassKey key) {
  assert_true(GetPath() == Path::kHostRenderTargets);

  auto it = render_passes_.find(key);
  if (it != render_passes_.end()) {
    return it->second;
  }

  VkSampleCountFlagBits samples;
  switch (key.msaa_samples) {
    case xenos::MsaaSamples::k1X:
      samples = VK_SAMPLE_COUNT_1_BIT;
      break;
    case xenos::MsaaSamples::k2X:
      samples = IsMsaa2xSupported(key.depth_and_color_used != 0)
                    ? VK_SAMPLE_COUNT_2_BIT
                    : VK_SAMPLE_COUNT_4_BIT;
      break;
    case xenos::MsaaSamples::k4X:
      samples = VK_SAMPLE_COUNT_4_BIT;
      break;
    default:
      return VK_NULL_HANDLE;
  }

  VkAttachmentDescription attachments[1 + xenos::kMaxColorRenderTargets];
  if (key.depth_and_color_used & 0b1) {
    VkAttachmentDescription& attachment = attachments[0];
    attachment.flags = 0;
    attachment.format = GetDepthVulkanFormat(key.depth_format);
    attachment.samples = samples;
    attachment.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
    attachment.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
    attachment.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
    attachment.stencilStoreOp = VK_ATTACHMENT_STORE_OP_STORE;
    attachment.initialLayout = VulkanRenderTarget::kDepthDrawLayout;
    attachment.finalLayout = VulkanRenderTarget::kDepthDrawLayout;
  }
  VkAttachmentReference color_attachments[xenos::kMaxColorRenderTargets];
  xenos::ColorRenderTargetFormat color_formats[] = {
      key.color_0_view_format,
      key.color_1_view_format,
      key.color_2_view_format,
      key.color_3_view_format,
  };
  for (uint32_t i = 0; i < xenos::kMaxColorRenderTargets; ++i) {
    VkAttachmentReference& color_attachment = color_attachments[i];
    color_attachment.layout = VulkanRenderTarget::kColorDrawLayout;
    uint32_t attachment_bit = uint32_t(1) << (1 + i);
    if (!(key.depth_and_color_used & attachment_bit)) {
      color_attachment.attachment = VK_ATTACHMENT_UNUSED;
      continue;
    }
    uint32_t attachment_index =
        xe::bit_count(key.depth_and_color_used & (attachment_bit - 1));
    color_attachment.attachment = attachment_index;
    VkAttachmentDescription& attachment = attachments[attachment_index];
    attachment.flags = 0;
    xenos::ColorRenderTargetFormat color_format = color_formats[i];
    attachment.format =
        key.color_rts_use_transfer_formats
            ? GetColorOwnershipTransferVulkanFormat(color_format)
            : GetColorVulkanFormat(color_format);
    attachment.samples = samples;
    attachment.loadOp = VK_ATTACHMENT_LOAD_OP_LOAD;
    attachment.storeOp = VK_ATTACHMENT_STORE_OP_STORE;
    attachment.stencilLoadOp = VK_ATTACHMENT_LOAD_OP_DONT_CARE;
    attachment.stencilStoreOp = VK_ATTACHMENT_STORE_OP_DONT_CARE;
    attachment.initialLayout = VulkanRenderTarget::kColorDrawLayout;
    attachment.finalLayout = VulkanRenderTarget::kColorDrawLayout;
  }

  VkAttachmentReference depth_stencil_attachment;
  depth_stencil_attachment.attachment =
      (key.depth_and_color_used & 0b1) ? 0 : VK_ATTACHMENT_UNUSED;
  depth_stencil_attachment.layout = VulkanRenderTarget::kDepthDrawLayout;

  VkSubpassDescription subpass;
  subpass.flags = 0;
  subpass.pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS;
  subpass.inputAttachmentCount = 0;
  subpass.pInputAttachments = nullptr;
  subpass.colorAttachmentCount =
      32 - xe::lzcnt(uint32_t(key.depth_and_color_used >> 1));
  subpass.pColorAttachments = color_attachments;
  subpass.pResolveAttachments = nullptr;
  subpass.pDepthStencilAttachment =
      (key.depth_and_color_used & 0b1) ? &depth_stencil_attachment : nullptr;
  subpass.preserveAttachmentCount = 0;
  subpass.pPreserveAttachments = nullptr;

  VkPipelineStageFlags dependency_stage_mask = 0;
  VkAccessFlags dependency_access_mask = 0;
  if (key.depth_and_color_used & 0b1) {
    dependency_stage_mask |= VulkanRenderTarget::kDepthDrawStageMask;
    dependency_access_mask |= VulkanRenderTarget::kDepthDrawAccessMask;
  }
  if (key.depth_and_color_used >> 1) {
    dependency_stage_mask |= VulkanRenderTarget::kColorDrawStageMask;
    dependency_access_mask |= VulkanRenderTarget::kColorDrawAccessMask;
  }
  VkSubpassDependency subpass_dependencies[2];
  subpass_dependencies[0].srcSubpass = VK_SUBPASS_EXTERNAL;
  subpass_dependencies[0].dstSubpass = 0;
  subpass_dependencies[0].srcStageMask = dependency_stage_mask;
  subpass_dependencies[0].dstStageMask = dependency_stage_mask;
  subpass_dependencies[0].srcAccessMask = dependency_access_mask;
  subpass_dependencies[0].dstAccessMask = dependency_access_mask;
  subpass_dependencies[0].dependencyFlags = VK_DEPENDENCY_BY_REGION_BIT;
  subpass_dependencies[1].srcSubpass = 0;
  subpass_dependencies[1].dstSubpass = VK_SUBPASS_EXTERNAL;
  subpass_dependencies[1].srcStageMask = dependency_stage_mask;
  subpass_dependencies[1].dstStageMask = dependency_stage_mask;
  subpass_dependencies[1].srcAccessMask = dependency_access_mask;
  subpass_dependencies[1].dstAccessMask = dependency_access_mask;
  subpass_dependencies[1].dependencyFlags = VK_DEPENDENCY_BY_REGION_BIT;

  VkRenderPassCreateInfo render_pass_create_info;
  render_pass_create_info.sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO;
  render_pass_create_info.pNext = nullptr;
  render_pass_create_info.flags = 0;
  render_pass_create_info.attachmentCount =
      xe::bit_count(key.depth_and_color_used);
  render_pass_create_info.pAttachments = attachments;
  render_pass_create_info.subpassCount = 1;
  render_pass_create_info.pSubpasses = &subpass;
  render_pass_create_info.dependencyCount =
      key.depth_and_color_used ? uint32_t(xe::countof(subpass_dependencies))
                               : 0;
  render_pass_create_info.pDependencies = subpass_dependencies;

  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();
  VkRenderPass render_pass;
  if (dfn.vkCreateRenderPass(device, &render_pass_create_info, nullptr,
                             &render_pass) != VK_SUCCESS) {
    XELOGE("VulkanRenderTargetCache: Failed to create a render pass");
    render_passes_.emplace(key, VK_NULL_HANDLE);
    return VK_NULL_HANDLE;
  }
  render_passes_.emplace(key, render_pass);
  return render_pass;
}

VkFormat VulkanRenderTargetCache::GetDepthVulkanFormat(
    xenos::DepthRenderTargetFormat format) const {
  if (format == xenos::DepthRenderTargetFormat::kD24S8 &&
      depth_unorm24_vulkan_format_supported()) {
    return VK_FORMAT_D24_UNORM_S8_UINT;
  }
  return VK_FORMAT_D32_SFLOAT_S8_UINT;
}

VkFormat VulkanRenderTargetCache::GetColorVulkanFormat(
    xenos::ColorRenderTargetFormat format) const {
  switch (format) {
    case xenos::ColorRenderTargetFormat::k_8_8_8_8:
      return VK_FORMAT_R8G8B8A8_UNORM;
    case xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA:
      return gamma_render_target_as_srgb_ ? VK_FORMAT_R8G8B8A8_SRGB
                                          : VK_FORMAT_R8G8B8A8_UNORM;
    case xenos::ColorRenderTargetFormat::k_2_10_10_10:
    case xenos::ColorRenderTargetFormat::k_2_10_10_10_AS_10_10_10_10:
      return VK_FORMAT_A8B8G8R8_UNORM_PACK32;
    case xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT:
    case xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT_AS_16_16_16_16:
      return VK_FORMAT_R16G16B16A16_SFLOAT;
    case xenos::ColorRenderTargetFormat::k_16_16:
      // TODO(Triang3l): Fallback to float16 (disregarding clearing correctness
      // likely) - possibly on render target gathering, treating them entirely
      // as float16.
      return VK_FORMAT_R16G16_SNORM;
    case xenos::ColorRenderTargetFormat::k_16_16_16_16:
      // TODO(Triang3l): Fallback to float16 (disregarding clearing correctness
      // likely) - possibly on render target gathering, treating them entirely
      // as float16.
      return VK_FORMAT_R16G16B16A16_SNORM;
    case xenos::ColorRenderTargetFormat::k_16_16_FLOAT:
      return VK_FORMAT_R16G16_SFLOAT;
    case xenos::ColorRenderTargetFormat::k_16_16_16_16_FLOAT:
      return VK_FORMAT_R16G16B16A16_SFLOAT;
    case xenos::ColorRenderTargetFormat::k_32_FLOAT:
      return VK_FORMAT_R32_SFLOAT;
    case xenos::ColorRenderTargetFormat::k_32_32_FLOAT:
      return VK_FORMAT_R32G32_SFLOAT;
    default:
      assert_unhandled_case(format);
      return VK_FORMAT_UNDEFINED;
  }
}

VkFormat VulkanRenderTargetCache::GetColorOwnershipTransferVulkanFormat(
    xenos::ColorRenderTargetFormat format, bool* is_integer_out) const {
  if (is_integer_out) {
    *is_integer_out = true;
  }
  // Floating-point numbers have NaNs that need to be propagated without
  // modifications to the bit representation, and SNORM has two representations
  // of -1.
  switch (format) {
    case xenos::ColorRenderTargetFormat::k_16_16:
    case xenos::ColorRenderTargetFormat::k_16_16_FLOAT:
      return VK_FORMAT_R16G16_UINT;
    case xenos::ColorRenderTargetFormat::k_16_16_16_16:
    case xenos::ColorRenderTargetFormat::k_16_16_16_16_FLOAT:
      return VK_FORMAT_R16G16B16A16_UINT;
    case xenos::ColorRenderTargetFormat::k_32_FLOAT:
      return VK_FORMAT_R32_UINT;
    case xenos::ColorRenderTargetFormat::k_32_32_FLOAT:
      return VK_FORMAT_R32G32_UINT;
    default:
      if (is_integer_out) {
        *is_integer_out = false;
      }
      return GetColorVulkanFormat(format);
  }
}

VulkanRenderTargetCache::VulkanRenderTarget::~VulkanRenderTarget() {
  const ui::vulkan::VulkanDevice* const vulkan_device =
      render_target_cache_.command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();
  ui::vulkan::SingleLayoutDescriptorSetPool& descriptor_set_pool =
      key().is_depth
          ? *render_target_cache_.descriptor_set_pool_sampled_image_x2_
          : *render_target_cache_.descriptor_set_pool_sampled_image_;
  descriptor_set_pool.Free(descriptor_set_index_transfer_source_);
  if (view_color_transfer_separate_ != VK_NULL_HANDLE) {
    dfn.vkDestroyImageView(device, view_color_transfer_separate_, nullptr);
  }
  if (view_srgb_ != VK_NULL_HANDLE) {
    dfn.vkDestroyImageView(device, view_srgb_, nullptr);
  }
  if (view_stencil_ != VK_NULL_HANDLE) {
    dfn.vkDestroyImageView(device, view_stencil_, nullptr);
  }
  if (view_depth_stencil_ != VK_NULL_HANDLE) {
    dfn.vkDestroyImageView(device, view_depth_stencil_, nullptr);
  }
  dfn.vkDestroyImageView(device, view_depth_color_, nullptr);
  dfn.vkDestroyImage(device, image_, nullptr);
  dfn.vkFreeMemory(device, memory_, nullptr);
}

uint32_t VulkanRenderTargetCache::GetMaxRenderTargetWidth() const {
  const ui::vulkan::VulkanDevice::Properties& device_properties =
      command_processor_.GetVulkanDevice()->properties();
  return std::min(device_properties.maxFramebufferWidth,
                  device_properties.maxImageDimension2D);
}

uint32_t VulkanRenderTargetCache::GetMaxRenderTargetHeight() const {
  const ui::vulkan::VulkanDevice::Properties& device_properties =
      command_processor_.GetVulkanDevice()->properties();
  return std::min(device_properties.maxFramebufferHeight,
                  device_properties.maxImageDimension2D);
}

RenderTargetCache::RenderTarget* VulkanRenderTargetCache::CreateRenderTarget(
    RenderTargetKey key) {
  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();

  // Create the image.

  VkImageCreateInfo image_create_info;
  image_create_info.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
  image_create_info.pNext = nullptr;
  image_create_info.flags = 0;
  image_create_info.imageType = VK_IMAGE_TYPE_2D;
  image_create_info.extent.width = key.GetWidth() * draw_resolution_scale_x();
  image_create_info.extent.height =
      GetRenderTargetHeight(key.pitch_tiles_at_32bpp, key.msaa_samples) *
      draw_resolution_scale_y();
  image_create_info.extent.depth = 1;
  image_create_info.mipLevels = 1;
  image_create_info.arrayLayers = 1;
  if (key.msaa_samples == xenos::MsaaSamples::k2X &&
      !msaa_2x_attachments_supported_) {
    image_create_info.samples = VK_SAMPLE_COUNT_4_BIT;
  } else {
    image_create_info.samples =
        VkSampleCountFlagBits(uint32_t(1) << uint32_t(key.msaa_samples));
  }
  image_create_info.tiling = VK_IMAGE_TILING_OPTIMAL;
  image_create_info.usage = VK_IMAGE_USAGE_SAMPLED_BIT;
  image_create_info.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
  image_create_info.queueFamilyIndexCount = 0;
  image_create_info.pQueueFamilyIndices = nullptr;
  image_create_info.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
  VkFormat transfer_format;
  bool is_srgb_view_needed = false;
  if (key.is_depth) {
    image_create_info.format = GetDepthVulkanFormat(key.GetDepthFormat());
    transfer_format = image_create_info.format;
    image_create_info.usage |= VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT;
  } else {
    xenos::ColorRenderTargetFormat color_format = key.GetColorFormat();
    image_create_info.format = GetColorVulkanFormat(color_format);
    transfer_format = GetColorOwnershipTransferVulkanFormat(color_format);
    is_srgb_view_needed =
        gamma_render_target_as_srgb_ &&
        (color_format == xenos::ColorRenderTargetFormat::k_8_8_8_8 ||
         color_format == xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA);
    if (image_create_info.format != transfer_format || is_srgb_view_needed) {
      image_create_info.flags |= VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT;
    }
    image_create_info.usage |= VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
  }
  if (image_create_info.format == VK_FORMAT_UNDEFINED) {
    XELOGE("VulkanRenderTargetCache: Unknown {} render target format {}",
           key.is_depth ? "depth" : "color", key.resource_format);
    return nullptr;
  }
  VkImage image;
  VkDeviceMemory memory;
  if (!ui::vulkan::util::CreateDedicatedAllocationImage(
          vulkan_device, image_create_info,
          ui::vulkan::util::MemoryPurpose::kDeviceLocal, image, memory)) {
    XELOGE(
        "VulkanRenderTarget: Failed to create a {}x{} {}xMSAA {} render target "
        "image",
        image_create_info.extent.width, image_create_info.extent.height,
        uint32_t(1) << uint32_t(key.msaa_samples), key.GetFormatName());
    return nullptr;
  }

  // Create the image views.

  VkImageViewCreateInfo view_create_info;
  view_create_info.sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO;
  view_create_info.pNext = nullptr;
  view_create_info.flags = 0;
  view_create_info.image = image;
  view_create_info.viewType = VK_IMAGE_VIEW_TYPE_2D;
  view_create_info.format = image_create_info.format;
  view_create_info.components.r = VK_COMPONENT_SWIZZLE_IDENTITY;
  view_create_info.components.g = VK_COMPONENT_SWIZZLE_IDENTITY;
  view_create_info.components.b = VK_COMPONENT_SWIZZLE_IDENTITY;
  view_create_info.components.a = VK_COMPONENT_SWIZZLE_IDENTITY;
  view_create_info.subresourceRange =
      ui::vulkan::util::InitializeSubresourceRange(
          key.is_depth ? VK_IMAGE_ASPECT_DEPTH_BIT : VK_IMAGE_ASPECT_COLOR_BIT);
  VkImageView view_depth_color;
  if (dfn.vkCreateImageView(device, &view_create_info, nullptr,
                            &view_depth_color) != VK_SUCCESS) {
    XELOGE(
        "VulkanRenderTarget: Failed to create a {} view for a {}x{} {}xMSAA {} "
        "render target",
        key.is_depth ? "depth" : "color", image_create_info.extent.width,
        image_create_info.extent.height,
        uint32_t(1) << uint32_t(key.msaa_samples), key.GetFormatName());
    dfn.vkDestroyImage(device, image, nullptr);
    dfn.vkFreeMemory(device, memory, nullptr);
    return nullptr;
  }
  VkImageView view_depth_stencil = VK_NULL_HANDLE;
  VkImageView view_stencil = VK_NULL_HANDLE;
  VkImageView view_srgb = VK_NULL_HANDLE;
  VkImageView view_color_transfer_separate = VK_NULL_HANDLE;
  if (key.is_depth) {
    view_create_info.subresourceRange.aspectMask =
        VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT;
    if (dfn.vkCreateImageView(device, &view_create_info, nullptr,
                              &view_depth_stencil) != VK_SUCCESS) {
      XELOGE(
          "VulkanRenderTarget: Failed to create a depth / stencil view for a "
          "{}x{} {}xMSAA {} render target",
          image_create_info.extent.width, image_create_info.extent.height,
          uint32_t(1) << uint32_t(key.msaa_samples),
          xenos::GetDepthRenderTargetFormatName(key.GetDepthFormat()));
      dfn.vkDestroyImageView(device, view_depth_color, nullptr);
      dfn.vkDestroyImage(device, image, nullptr);
      dfn.vkFreeMemory(device, memory, nullptr);
      return nullptr;
    }
    view_create_info.subresourceRange.aspectMask = VK_IMAGE_ASPECT_STENCIL_BIT;
    if (dfn.vkCreateImageView(device, &view_create_info, nullptr,
                              &view_stencil) != VK_SUCCESS) {
      XELOGE(
          "VulkanRenderTarget: Failed to create a stencil view for a {}x{} "
          "{}xMSAA render target",
          image_create_info.extent.width, image_create_info.extent.height,
          uint32_t(1) << uint32_t(key.msaa_samples),
          xenos::GetDepthRenderTargetFormatName(key.GetDepthFormat()));
      dfn.vkDestroyImageView(device, view_depth_stencil, nullptr);
      dfn.vkDestroyImageView(device, view_depth_color, nullptr);
      dfn.vkDestroyImage(device, image, nullptr);
      dfn.vkFreeMemory(device, memory, nullptr);
      return nullptr;
    }
  } else {
    if (is_srgb_view_needed) {
      view_create_info.format = VK_FORMAT_R8G8B8A8_SRGB;
      if (dfn.vkCreateImageView(device, &view_create_info, nullptr,
                                &view_srgb) != VK_SUCCESS) {
        XELOGE(
            "VulkanRenderTarget: Failed to create an sRGB view for a {}x{} "
            "{}xMSAA render target",
            image_create_info.extent.width, image_create_info.extent.height,
            uint32_t(1) << uint32_t(key.msaa_samples),
            xenos::GetColorRenderTargetFormatName(key.GetColorFormat()));
        dfn.vkDestroyImageView(device, view_depth_color, nullptr);
        dfn.vkDestroyImage(device, image, nullptr);
        dfn.vkFreeMemory(device, memory, nullptr);
        return nullptr;
      }
    }
    if (transfer_format != image_create_info.format) {
      view_create_info.format = transfer_format;
      if (dfn.vkCreateImageView(device, &view_create_info, nullptr,
                                &view_color_transfer_separate) != VK_SUCCESS) {
        XELOGE(
            "VulkanRenderTarget: Failed to create a transfer view for a {}x{} "
            "{}xMSAA {} render target",
            image_create_info.extent.width, image_create_info.extent.height,
            uint32_t(1) << uint32_t(key.msaa_samples), key.GetFormatName());
        if (view_srgb != VK_NULL_HANDLE) {
          dfn.vkDestroyImageView(device, view_srgb, nullptr);
        }
        dfn.vkDestroyImageView(device, view_depth_color, nullptr);
        dfn.vkDestroyImage(device, image, nullptr);
        dfn.vkFreeMemory(device, memory, nullptr);
        return nullptr;
      }
    }
  }

  ui::vulkan::SingleLayoutDescriptorSetPool& descriptor_set_pool =
      key.is_depth ? *descriptor_set_pool_sampled_image_x2_
                   : *descriptor_set_pool_sampled_image_;
  size_t descriptor_set_index_transfer_source = descriptor_set_pool.Allocate();
  if (descriptor_set_index_transfer_source == SIZE_MAX) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to allocate sampled image descriptors "
        "for a {} render target",
        key.is_depth ? "depth/stencil" : "color");
    if (view_color_transfer_separate != VK_NULL_HANDLE) {
      dfn.vkDestroyImageView(device, view_color_transfer_separate, nullptr);
    }
    if (view_srgb != VK_NULL_HANDLE) {
      dfn.vkDestroyImageView(device, view_srgb, nullptr);
    }
    dfn.vkDestroyImageView(device, view_depth_color, nullptr);
    dfn.vkDestroyImage(device, image, nullptr);
    dfn.vkFreeMemory(device, memory, nullptr);
    return nullptr;
  }
  VkDescriptorSet descriptor_set_transfer_source =
      descriptor_set_pool.Get(descriptor_set_index_transfer_source);
  VkWriteDescriptorSet descriptor_set_write[2];
  VkDescriptorImageInfo descriptor_set_write_depth_color;
  descriptor_set_write_depth_color.sampler = VK_NULL_HANDLE;
  descriptor_set_write_depth_color.imageView =
      view_color_transfer_separate != VK_NULL_HANDLE
          ? view_color_transfer_separate
          : view_depth_color;
  descriptor_set_write_depth_color.imageLayout =
      VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
  descriptor_set_write[0].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
  descriptor_set_write[0].pNext = nullptr;
  descriptor_set_write[0].dstSet = descriptor_set_transfer_source;
  descriptor_set_write[0].dstBinding = 0;
  descriptor_set_write[0].dstArrayElement = 0;
  descriptor_set_write[0].descriptorCount = 1;
  descriptor_set_write[0].descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE;
  descriptor_set_write[0].pImageInfo = &descriptor_set_write_depth_color;
  descriptor_set_write[0].pBufferInfo = nullptr;
  descriptor_set_write[0].pTexelBufferView = nullptr;
  VkDescriptorImageInfo descriptor_set_write_stencil;
  if (key.is_depth) {
    descriptor_set_write_stencil.sampler = VK_NULL_HANDLE;
    descriptor_set_write_stencil.imageView = view_stencil;
    descriptor_set_write_stencil.imageLayout =
        VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
    descriptor_set_write[1].sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET;
    descriptor_set_write[1].pNext = nullptr;
    descriptor_set_write[1].dstSet = descriptor_set_transfer_source;
    descriptor_set_write[1].dstBinding = 1;
    descriptor_set_write[1].dstArrayElement = 0;
    descriptor_set_write[1].descriptorCount = 1;
    descriptor_set_write[1].descriptorType = VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE;
    descriptor_set_write[1].pImageInfo = &descriptor_set_write_stencil;
    descriptor_set_write[1].pBufferInfo = nullptr;
    descriptor_set_write[1].pTexelBufferView = nullptr;
  }
  dfn.vkUpdateDescriptorSets(device, key.is_depth ? 2 : 1, descriptor_set_write,
                             0, nullptr);

  return new VulkanRenderTarget(key, *this, image, memory, view_depth_color,
                                view_depth_stencil, view_stencil, view_srgb,
                                view_color_transfer_separate,
                                descriptor_set_index_transfer_source);
}

bool VulkanRenderTargetCache::IsHostDepthEncodingDifferent(
    xenos::DepthRenderTargetFormat format) const {
  // TODO(Triang3l): Conversion directly in shaders.
  switch (format) {
    case xenos::DepthRenderTargetFormat::kD24S8:
      return !depth_unorm24_vulkan_format_supported();
    case xenos::DepthRenderTargetFormat::kD24FS8:
      return true;
  }
  return false;
}

void VulkanRenderTargetCache::RequestPixelShaderInterlockBarrier() {
  if (edram_buffer_usage_ == EdramBufferUsage::kFragmentReadWrite) {
    CommitEdramBufferShaderWrites();
  }
}

void VulkanRenderTargetCache::GetEdramBufferUsageMasks(
    EdramBufferUsage usage, VkPipelineStageFlags& stage_mask_out,
    VkAccessFlags& access_mask_out) {
  switch (usage) {
    case EdramBufferUsage::kFragmentRead:
      stage_mask_out = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
      access_mask_out = VK_ACCESS_SHADER_READ_BIT;
      break;
    case EdramBufferUsage::kFragmentReadWrite:
      stage_mask_out = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
      access_mask_out = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT;
      break;
    case EdramBufferUsage::kComputeRead:
      stage_mask_out = VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
      access_mask_out = VK_ACCESS_SHADER_READ_BIT;
      break;
    case EdramBufferUsage::kComputeWrite:
      stage_mask_out = VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
      access_mask_out = VK_ACCESS_SHADER_WRITE_BIT;
      break;
    case EdramBufferUsage::kTransferRead:
      stage_mask_out = VK_PIPELINE_STAGE_TRANSFER_BIT;
      access_mask_out = VK_ACCESS_TRANSFER_READ_BIT;
      break;
    case EdramBufferUsage::kTransferWrite:
      stage_mask_out = VK_PIPELINE_STAGE_TRANSFER_BIT;
      access_mask_out = VK_ACCESS_TRANSFER_WRITE_BIT;
      break;
    default:
      assert_unhandled_case(usage);
  }
}

void VulkanRenderTargetCache::UseEdramBuffer(EdramBufferUsage new_usage) {
  if (edram_buffer_usage_ == new_usage) {
    return;
  }
  VkPipelineStageFlags src_stage_mask, dst_stage_mask;
  VkAccessFlags src_access_mask, dst_access_mask;
  GetEdramBufferUsageMasks(edram_buffer_usage_, src_stage_mask,
                           src_access_mask);
  GetEdramBufferUsageMasks(new_usage, dst_stage_mask, dst_access_mask);
  if (command_processor_.PushBufferMemoryBarrier(
          edram_buffer_, 0, VK_WHOLE_SIZE, src_stage_mask, dst_stage_mask,
          src_access_mask, dst_access_mask)) {
    // Resetting edram_buffer_modification_status_ only if the barrier has been
    // truly inserted.
    edram_buffer_modification_status_ =
        EdramBufferModificationStatus::kUnmodified;
  }
  edram_buffer_usage_ = new_usage;
}

void VulkanRenderTargetCache::MarkEdramBufferModified(
    EdramBufferModificationStatus modification_status) {
  assert_true(modification_status !=
              EdramBufferModificationStatus::kUnmodified);
  switch (edram_buffer_usage_) {
    case EdramBufferUsage::kFragmentReadWrite:
      // max because being modified via unordered access requires stricter
      // synchronization than via fragment shader interlocks.
      edram_buffer_modification_status_ =
          std::max(edram_buffer_modification_status_, modification_status);
      break;
    case EdramBufferUsage::kComputeWrite:
      assert_true(modification_status ==
                  EdramBufferModificationStatus::kViaUnordered);
      modification_status = EdramBufferModificationStatus::kViaUnordered;
      break;
    default:
      assert_always(
          "While changing the usage of the EDRAM buffer before marking it as "
          "modified is handled safely (but will cause spurious marking as "
          "modified after the changes have been implicitly committed by the "
          "usage switch), normally that shouldn't be done and is an "
          "indication of architectural mistakes. Alternatively, this may "
          "indicate that the usage switch has been forgotten before writing, "
          "which is a clearly invalid situation.");
  }
}

void VulkanRenderTargetCache::CommitEdramBufferShaderWrites(
    EdramBufferModificationStatus commit_status) {
  assert_true(commit_status != EdramBufferModificationStatus::kUnmodified);
  if (edram_buffer_modification_status_ < commit_status) {
    return;
  }
  VkPipelineStageFlags stage_mask;
  VkAccessFlags access_mask;
  GetEdramBufferUsageMasks(edram_buffer_usage_, stage_mask, access_mask);
  assert_not_zero(access_mask & VK_ACCESS_SHADER_WRITE_BIT);
  command_processor_.PushBufferMemoryBarrier(
      edram_buffer_, 0, VK_WHOLE_SIZE, stage_mask, stage_mask, access_mask,
      access_mask, VK_QUEUE_FAMILY_IGNORED, VK_QUEUE_FAMILY_IGNORED, false);
  edram_buffer_modification_status_ =
      EdramBufferModificationStatus::kUnmodified;
  PixelShaderInterlockFullEdramBarrierPlaced();
}

const VulkanRenderTargetCache::Framebuffer*
VulkanRenderTargetCache::GetHostRenderTargetsFramebuffer(
    RenderPassKey render_pass_key, uint32_t pitch_tiles_at_32bpp,
    const RenderTarget* const* depth_and_color_render_targets) {
  FramebufferKey key;
  key.render_pass_key = render_pass_key;
  key.pitch_tiles_at_32bpp = pitch_tiles_at_32bpp;
  if (render_pass_key.depth_and_color_used & (1 << 0)) {
    key.depth_base_tiles = depth_and_color_render_targets[0]->key().base_tiles;
  }
  if (render_pass_key.depth_and_color_used & (1 << 1)) {
    key.color_0_base_tiles =
        depth_and_color_render_targets[1]->key().base_tiles;
  }
  if (render_pass_key.depth_and_color_used & (1 << 2)) {
    key.color_1_base_tiles =
        depth_and_color_render_targets[2]->key().base_tiles;
  }
  if (render_pass_key.depth_and_color_used & (1 << 3)) {
    key.color_2_base_tiles =
        depth_and_color_render_targets[3]->key().base_tiles;
  }
  if (render_pass_key.depth_and_color_used & (1 << 4)) {
    key.color_3_base_tiles =
        depth_and_color_render_targets[4]->key().base_tiles;
  }
  auto it = framebuffers_.find(key);
  if (it != framebuffers_.end()) {
    return &it->second;
  }

  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();
  const ui::vulkan::VulkanDevice::Properties& device_properties =
      vulkan_device->properties();

  VkRenderPass render_pass = GetHostRenderTargetsRenderPass(render_pass_key);
  if (render_pass == VK_NULL_HANDLE) {
    return nullptr;
  }

  VkImageView attachments[1 + xenos::kMaxColorRenderTargets];
  uint32_t attachment_count = 0;
  uint32_t depth_and_color_rts_remaining = render_pass_key.depth_and_color_used;
  uint32_t rt_index;
  while (xe::bit_scan_forward(depth_and_color_rts_remaining, &rt_index)) {
    depth_and_color_rts_remaining &= ~(uint32_t(1) << rt_index);
    const auto& vulkan_rt = *static_cast<const VulkanRenderTarget*>(
        depth_and_color_render_targets[rt_index]);
    VkImageView attachment;
    if (rt_index) {
      attachment = render_pass_key.color_rts_use_transfer_formats
                       ? vulkan_rt.view_color_transfer()
                       : vulkan_rt.view_depth_color();
    } else {
      attachment = vulkan_rt.view_depth_stencil();
    }
    attachments[attachment_count++] = attachment;
  }

  VkFramebufferCreateInfo framebuffer_create_info;
  framebuffer_create_info.sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO;
  framebuffer_create_info.pNext = nullptr;
  framebuffer_create_info.flags = 0;
  framebuffer_create_info.renderPass = render_pass;
  framebuffer_create_info.attachmentCount = attachment_count;
  framebuffer_create_info.pAttachments = attachments;
  VkExtent2D host_extent;
  if (pitch_tiles_at_32bpp) {
    host_extent.width = RenderTargetKey::GetWidth(pitch_tiles_at_32bpp,
                                                  render_pass_key.msaa_samples);
    host_extent.height = GetRenderTargetHeight(pitch_tiles_at_32bpp,
                                               render_pass_key.msaa_samples);
  } else {
    assert_zero(render_pass_key.depth_and_color_used);
    // Still needed for occlusion queries.
    host_extent.width = xenos::kTexture2DCubeMaxWidthHeight;
    host_extent.height = xenos::kTexture2DCubeMaxWidthHeight;
  }
  // Limiting to the device limit for the case of no attachments, for which
  // there's no limit imposed by the sizes of the attachments that have been
  // created successfully.
  host_extent.width = std::min(host_extent.width * draw_resolution_scale_x(),
                               device_properties.maxFramebufferWidth);
  host_extent.height = std::min(host_extent.height * draw_resolution_scale_y(),
                                device_properties.maxFramebufferHeight);
  framebuffer_create_info.width = host_extent.width;
  framebuffer_create_info.height = host_extent.height;
  framebuffer_create_info.layers = 1;
  VkFramebuffer framebuffer;
  if (dfn.vkCreateFramebuffer(device, &framebuffer_create_info, nullptr,
                              &framebuffer) != VK_SUCCESS) {
    return nullptr;
  }
  // Creates at a persistent location - safe to use pointers.
  return &framebuffers_
              .emplace(std::piecewise_construct, std::forward_as_tuple(key),
                       std::forward_as_tuple(framebuffer, host_extent))
              .first->second;
}

VkShaderModule VulkanRenderTargetCache::GetTransferShader(
    TransferShaderKey key) {
  auto shader_it = transfer_shaders_.find(key);
  if (shader_it != transfer_shaders_.end()) {
    return shader_it->second;
  }

  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Properties& device_properties =
      vulkan_device->properties();

  std::vector<spv::Id> id_vector_temp;
  std::vector<unsigned int> uint_vector_temp;

  SpirvBuilder builder(spv::Spv_1_0,
                       (SpirvShaderTranslator::kSpirvMagicToolId << 16) | 1,
                       nullptr);
  spv::Id ext_inst_glsl_std_450 = builder.import("GLSL.std.450");
  builder.addCapability(spv::CapabilityShader);
  builder.setMemoryModel(spv::AddressingModelLogical, spv::MemoryModelGLSL450);
  builder.setSource(spv::SourceLanguageUnknown, 0);

  spv::Id type_void = builder.makeVoidType();
  spv::Id type_bool = builder.makeBoolType();
  spv::Id type_int = builder.makeIntType(32);
  spv::Id type_int2 = builder.makeVectorType(type_int, 2);
  spv::Id type_uint = builder.makeUintType(32);
  spv::Id type_uint2 = builder.makeVectorType(type_uint, 2);
  spv::Id type_uint4 = builder.makeVectorType(type_uint, 4);
  spv::Id type_float = builder.makeFloatType(32);
  spv::Id type_float2 = builder.makeVectorType(type_float, 2);
  spv::Id type_float4 = builder.makeVectorType(type_float, 4);

  const TransferModeInfo& mode = kTransferModes[size_t(key.mode)];
  const TransferPipelineLayoutInfo& pipeline_layout_info =
      kTransferPipelineLayoutInfos[size_t(mode.pipeline_layout)];

  // If not dest_is_color, it's depth, or stencil bit - 40-sample columns are
  // swapped as opposed to color source.
  bool dest_is_color = (mode.output == TransferOutput::kColor);
  xenos::ColorRenderTargetFormat dest_color_format =
      xenos::ColorRenderTargetFormat(key.dest_resource_format);
  xenos::DepthRenderTargetFormat dest_depth_format =
      xenos::DepthRenderTargetFormat(key.dest_resource_format);
  bool dest_is_64bpp =
      dest_is_color && xenos::IsColorRenderTargetFormat64bpp(dest_color_format);

  xenos::ColorRenderTargetFormat source_color_format =
      xenos::ColorRenderTargetFormat(key.source_resource_format);
  xenos::DepthRenderTargetFormat source_depth_format =
      xenos::DepthRenderTargetFormat(key.source_resource_format);
  // If not source_is_color, it's depth / stencil - 40-sample columns are
  // swapped as opposed to color destination.
  bool source_is_color = (pipeline_layout_info.used_descriptor_sets &
                          kTransferUsedDescriptorSetColorTextureBit) != 0;
  bool source_is_64bpp;
  uint32_t source_color_format_component_count;
  uint32_t source_color_texture_component_mask;
  bool source_color_is_uint;
  spv::Id source_color_component_type;
  if (source_is_color) {
    assert_zero(pipeline_layout_info.used_descriptor_sets &
                kTransferUsedDescriptorSetDepthStencilTexturesBit);
    source_is_64bpp =
        xenos::IsColorRenderTargetFormat64bpp(source_color_format);
    source_color_format_component_count =
        xenos::GetColorRenderTargetFormatComponentCount(source_color_format);
    if (mode.output == TransferOutput::kStencilBit) {
      if (source_is_64bpp && !dest_is_64bpp) {
        // Need one component, but choosing from the two 32bpp halves of the
        // 64bpp sample.
        source_color_texture_component_mask =
            0b1 | (0b1 << (source_color_format_component_count >> 1));
      } else {
        // Red is at least 8 bits per component in all formats.
        source_color_texture_component_mask = 0b1;
      }
    } else {
      source_color_texture_component_mask =
          (uint32_t(1) << source_color_format_component_count) - 1;
    }
    GetColorOwnershipTransferVulkanFormat(source_color_format,
                                          &source_color_is_uint);
    source_color_component_type = source_color_is_uint ? type_uint : type_float;
  } else {
    source_is_64bpp = false;
    source_color_format_component_count = 0;
    source_color_texture_component_mask = 0;
    source_color_is_uint = false;
    source_color_component_type = spv::NoType;
  }

  std::vector<spv::Id> main_interface;

  // Outputs.
  bool shader_uses_stencil_reference_output =
      mode.output == TransferOutput::kDepth &&
      vulkan_device->extensions().ext_EXT_shader_stencil_export;
  bool dest_color_is_uint = false;
  uint32_t dest_color_component_count = 0;
  spv::Id type_fragment_data_component = spv::NoResult;
  spv::Id type_fragment_data = spv::NoResult;
  spv::Id output_fragment_data = spv::NoResult;
  spv::Id output_fragment_depth = spv::NoResult;
  spv::Id output_fragment_stencil_ref = spv::NoResult;
  switch (mode.output) {
    case TransferOutput::kColor:
      GetColorOwnershipTransferVulkanFormat(dest_color_format,
                                            &dest_color_is_uint);
      dest_color_component_count =
          xenos::GetColorRenderTargetFormatComponentCount(dest_color_format);
      type_fragment_data_component =
          dest_color_is_uint ? type_uint : type_float;
      type_fragment_data =
          dest_color_component_count > 1
              ? builder.makeVectorType(type_fragment_data_component,
                                       dest_color_component_count)
              : type_fragment_data_component;
      output_fragment_data = builder.createVariable(
          spv::NoPrecision, spv::StorageClassOutput, type_fragment_data,
          "xe_transfer_fragment_data");
      builder.addDecoration(output_fragment_data, spv::DecorationLocation,
                            key.dest_color_rt_index);
      main_interface.push_back(output_fragment_data);
      break;
    case TransferOutput::kDepth:
      output_fragment_depth =
          builder.createVariable(spv::NoPrecision, spv::StorageClassOutput,
                                 type_float, "gl_FragDepth");
      builder.addDecoration(output_fragment_depth, spv::DecorationBuiltIn,
                            spv::BuiltInFragDepth);
      main_interface.push_back(output_fragment_depth);
      if (shader_uses_stencil_reference_output) {
        builder.addExtension("SPV_EXT_shader_stencil_export");
        builder.addCapability(spv::CapabilityStencilExportEXT);
        output_fragment_stencil_ref =
            builder.createVariable(spv::NoPrecision, spv::StorageClassOutput,
                                   type_int, "gl_FragStencilRefARB");
        builder.addDecoration(output_fragment_stencil_ref,
                              spv::DecorationBuiltIn,
                              spv::BuiltInFragStencilRefEXT);
        main_interface.push_back(output_fragment_stencil_ref);
      }
      break;
    default:
      break;
  }

  // Bindings.
  // Generating SPIR-V 1.0, no need to add bindings to the entry point's
  // interface until SPIR-V 1.4.
  // Color source.
  bool source_is_multisampled =
      key.source_msaa_samples != xenos::MsaaSamples::k1X;
  spv::Id source_color_texture = spv::NoResult;
  if (pipeline_layout_info.used_descriptor_sets &
      kTransferUsedDescriptorSetColorTextureBit) {
    source_color_texture = builder.createVariable(
        spv::NoPrecision, spv::StorageClassUniformConstant,
        builder.makeImageType(source_color_component_type, spv::Dim2D, false,
                              false, source_is_multisampled, 1,
                              spv::ImageFormatUnknown),
        "xe_transfer_color");
    builder.addDecoration(
        source_color_texture, spv::DecorationDescriptorSet,
        xe::bit_count(pipeline_layout_info.used_descriptor_sets &
                      (kTransferUsedDescriptorSetColorTextureBit - 1)));
    builder.addDecoration(source_color_texture, spv::DecorationBinding, 0);
  }
  // Depth / stencil source.
  spv::Id source_depth_texture = spv::NoResult;
  spv::Id source_stencil_texture = spv::NoResult;
  if (pipeline_layout_info.used_descriptor_sets &
      kTransferUsedDescriptorSetDepthStencilTexturesBit) {
    uint32_t source_depth_stencil_descriptor_set =
        xe::bit_count(pipeline_layout_info.used_descriptor_sets &
                      (kTransferUsedDescriptorSetDepthStencilTexturesBit - 1));
    // Using `depth == false` in makeImageType because comparisons are not
    // required, and other values of `depth` are causing issues in drivers.
    // https://github.com/microsoft/DirectXShaderCompiler/issues/1107
    if (mode.output != TransferOutput::kStencilBit) {
      source_depth_texture = builder.createVariable(
          spv::NoPrecision, spv::StorageClassUniformConstant,
          builder.makeImageType(type_float, spv::Dim2D, false, false,
                                source_is_multisampled, 1,
                                spv::ImageFormatUnknown),
          "xe_transfer_depth");
      builder.addDecoration(source_depth_texture, spv::DecorationDescriptorSet,
                            source_depth_stencil_descriptor_set);
      builder.addDecoration(source_depth_texture, spv::DecorationBinding, 0);
    }
    if (mode.output != TransferOutput::kDepth ||
        shader_uses_stencil_reference_output) {
      source_stencil_texture = builder.createVariable(
          spv::NoPrecision, spv::StorageClassUniformConstant,
          builder.makeImageType(type_uint, spv::Dim2D, false, false,
                                source_is_multisampled, 1,
                                spv::ImageFormatUnknown),
          "xe_transfer_stencil");
      builder.addDecoration(source_stencil_texture,
                            spv::DecorationDescriptorSet,
                            source_depth_stencil_descriptor_set);
      builder.addDecoration(source_stencil_texture, spv::DecorationBinding, 1);
    }
  }
  // Host depth source buffer.
  spv::Id host_depth_source_buffer = spv::NoResult;
  if (pipeline_layout_info.used_descriptor_sets &
      kTransferUsedDescriptorSetHostDepthBufferBit) {
    id_vector_temp.clear();
    id_vector_temp.push_back(builder.makeRuntimeArray(type_uint));
    // Storage buffers have std430 packing, no padding to 4-component vectors.
    builder.addDecoration(id_vector_temp.back(), spv::DecorationArrayStride,
                          sizeof(uint32_t));
    spv::Id type_host_depth_source_buffer =
        builder.makeStructType(id_vector_temp, "XeTransferHostDepthBuffer");
    builder.addMemberName(type_host_depth_source_buffer, 0, "host_depth");
    builder.addMemberDecoration(type_host_depth_source_buffer, 0,
                                spv::DecorationNonWritable);
    builder.addMemberDecoration(type_host_depth_source_buffer, 0,
                                spv::DecorationOffset, 0);
    // Block since SPIR-V 1.3, but since SPIR-V 1.0 is generated, it's
    // BufferBlock.
    builder.addDecoration(type_host_depth_source_buffer,
                          spv::DecorationBufferBlock);
    // StorageBuffer since SPIR-V 1.3, but since SPIR-V 1.0 is generated, it's
    // Uniform.
    host_depth_source_buffer = builder.createVariable(
        spv::NoPrecision, spv::StorageClassUniform,
        type_host_depth_source_buffer, "xe_transfer_host_depth_buffer");
    builder.addDecoration(
        host_depth_source_buffer, spv::DecorationDescriptorSet,
        xe::bit_count(pipeline_layout_info.used_descriptor_sets &
                      (kTransferUsedDescriptorSetHostDepthBufferBit - 1)));
    builder.addDecoration(host_depth_source_buffer, spv::DecorationBinding, 0);
  }
  // Host depth source texture (the depth / stencil descriptor set is reused,
  // but stencil is not needed).
  spv::Id host_depth_source_texture = spv::NoResult;
  if (pipeline_layout_info.used_descriptor_sets &
      kTransferUsedDescriptorSetHostDepthStencilTexturesBit) {
    host_depth_source_texture = builder.createVariable(
        spv::NoPrecision, spv::StorageClassUniformConstant,
        builder.makeImageType(
            type_float, spv::Dim2D, false, false,
            key.host_depth_source_msaa_samples != xenos::MsaaSamples::k1X, 1,
            spv::ImageFormatUnknown),
        "xe_transfer_host_depth");
    builder.addDecoration(
        host_depth_source_texture, spv::DecorationDescriptorSet,
        xe::bit_count(
            pipeline_layout_info.used_descriptor_sets &
            (kTransferUsedDescriptorSetHostDepthStencilTexturesBit - 1)));
    builder.addDecoration(host_depth_source_texture, spv::DecorationBinding, 0);
  }
  // Push constants.
  id_vector_temp.clear();
  uint32_t push_constants_member_host_depth_address = UINT32_MAX;
  if (pipeline_layout_info.used_push_constant_dwords &
      kTransferUsedPushConstantDwordHostDepthAddressBit) {
    push_constants_member_host_depth_address = uint32_t(id_vector_temp.size());
    id_vector_temp.push_back(type_uint);
  }
  uint32_t push_constants_member_address = UINT32_MAX;
  if (pipeline_layout_info.used_push_constant_dwords &
      kTransferUsedPushConstantDwordAddressBit) {
    push_constants_member_address = uint32_t(id_vector_temp.size());
    id_vector_temp.push_back(type_uint);
  }
  uint32_t push_constants_member_stencil_mask = UINT32_MAX;
  if (pipeline_layout_info.used_push_constant_dwords &
      kTransferUsedPushConstantDwordStencilMaskBit) {
    push_constants_member_stencil_mask = uint32_t(id_vector_temp.size());
    id_vector_temp.push_back(type_uint);
  }
  spv::Id push_constants = spv::NoResult;
  if (!id_vector_temp.empty()) {
    spv::Id type_push_constants =
        builder.makeStructType(id_vector_temp, "XeTransferPushConstants");
    if (pipeline_layout_info.used_push_constant_dwords &
        kTransferUsedPushConstantDwordHostDepthAddressBit) {
      assert_true(push_constants_member_host_depth_address != UINT32_MAX);
      builder.addMemberName(type_push_constants,
                            push_constants_member_host_depth_address,
                            "host_depth_address");
      builder.addMemberDecoration(
          type_push_constants, push_constants_member_host_depth_address,
          spv::DecorationOffset,
          sizeof(uint32_t) *
              xe::bit_count(
                  pipeline_layout_info.used_push_constant_dwords &
                  (kTransferUsedPushConstantDwordHostDepthAddressBit - 1)));
    }
    if (pipeline_layout_info.used_push_constant_dwords &
        kTransferUsedPushConstantDwordAddressBit) {
      assert_true(push_constants_member_address != UINT32_MAX);
      builder.addMemberName(type_push_constants, push_constants_member_address,
                            "address");
      builder.addMemberDecoration(
          type_push_constants, push_constants_member_address,
          spv::DecorationOffset,
          sizeof(uint32_t) *
              xe::bit_count(pipeline_layout_info.used_push_constant_dwords &
                            (kTransferUsedPushConstantDwordAddressBit - 1)));
    }
    if (pipeline_layout_info.used_push_constant_dwords &
        kTransferUsedPushConstantDwordStencilMaskBit) {
      assert_true(push_constants_member_stencil_mask != UINT32_MAX);
      builder.addMemberName(type_push_constants,
                            push_constants_member_stencil_mask, "stencil_mask");
      builder.addMemberDecoration(
          type_push_constants, push_constants_member_stencil_mask,
          spv::DecorationOffset,
          sizeof(uint32_t) *
              xe::bit_count(
                  pipeline_layout_info.used_push_constant_dwords &
                  (kTransferUsedPushConstantDwordStencilMaskBit - 1)));
    }
    builder.addDecoration(type_push_constants, spv::DecorationBlock);
    push_constants = builder.createVariable(
        spv::NoPrecision, spv::StorageClassPushConstant, type_push_constants,
        "xe_transfer_push_constants");
  }

  // Coordinate inputs.
  spv::Id input_fragment_coord = builder.createVariable(
      spv::NoPrecision, spv::StorageClassInput, type_float4, "gl_FragCoord");
  builder.addDecoration(input_fragment_coord, spv::DecorationBuiltIn,
                        spv::BuiltInFragCoord);
  main_interface.push_back(input_fragment_coord);
  spv::Id input_sample_id = spv::NoResult;
  spv::Id spec_const_sample_id = spv::NoResult;
  if (key.dest_msaa_samples != xenos::MsaaSamples::k1X) {
    if (device_properties.sampleRateShading) {
      // One draw for all samples.
      builder.addCapability(spv::CapabilitySampleRateShading);
      input_sample_id = builder.createVariable(
          spv::NoPrecision, spv::StorageClassInput, type_int, "gl_SampleID");
      builder.addDecoration(input_sample_id, spv::DecorationFlat);
      builder.addDecoration(input_sample_id, spv::DecorationBuiltIn,
                            spv::BuiltInSampleId);
      main_interface.push_back(input_sample_id);
    } else {
      // One sample per draw, with different sample masks.
      spec_const_sample_id = builder.makeUintConstant(0, true);
      builder.addName(spec_const_sample_id, "xe_transfer_sample_id");
      builder.addDecoration(spec_const_sample_id, spv::DecorationSpecId, 0);
    }
  }

  // Begin the main function.
  std::vector<spv::Id> main_param_types;
  std::vector<std::vector<spv::Decoration>> main_precisions;
  spv::Block* main_entry;
  spv::Function* main_function =
      builder.makeFunctionEntry(spv::NoPrecision, type_void, "main",
                                main_param_types, main_precisions, &main_entry);

  // Working with unsigned numbers for simplicity now, bitcasting to signed will
  // be done at texture fetch.

  uint32_t tile_width_samples =
      xenos::kEdramTileWidthSamples * draw_resolution_scale_x();
  uint32_t tile_height_samples =
      xenos::kEdramTileHeightSamples * draw_resolution_scale_y();

  // Split the destination pixel index into 32bpp tile and 32bpp-tile-relative
  // pixel index.
  // Note that division by non-power-of-two constants will include a 4-cycle
  // 32*32 multiplication on AMD, even though so many bits are not needed for
  // the pixel position - however, if an OpUnreachable path is inserted for the
  // case when the position has upper bits set, for some reason, the code for it
  // is not eliminated when compiling the shader for AMD via RenderDoc on
  // Windows, as of June 2022.
  uint_vector_temp.clear();
  uint_vector_temp.push_back(0);
  uint_vector_temp.push_back(1);
  spv::Id dest_pixel_coord = builder.createUnaryOp(
      spv::OpConvertFToU, type_uint2,
      builder.createRvalueSwizzle(
          spv::NoPrecision, type_float2,
          builder.createLoad(input_fragment_coord, spv::NoPrecision),
          uint_vector_temp));
  spv::Id dest_pixel_x =
      builder.createCompositeExtract(dest_pixel_coord, type_uint, 0);
  spv::Id const_dest_tile_width_pixels = builder.makeUintConstant(
      tile_width_samples >>
      (uint32_t(dest_is_64bpp) +
       uint32_t(key.dest_msaa_samples >= xenos::MsaaSamples::k4X)));
  spv::Id dest_tile_index_x = builder.createBinOp(
      spv::OpUDiv, type_uint, dest_pixel_x, const_dest_tile_width_pixels);
  spv::Id dest_tile_pixel_x = builder.createBinOp(
      spv::OpUMod, type_uint, dest_pixel_x, const_dest_tile_width_pixels);
  spv::Id dest_pixel_y =
      builder.createCompositeExtract(dest_pixel_coord, type_uint, 1);
  spv::Id const_dest_tile_height_pixels = builder.makeUintConstant(
      tile_height_samples >>
      uint32_t(key.dest_msaa_samples >= xenos::MsaaSamples::k2X));
  spv::Id dest_tile_index_y = builder.createBinOp(
      spv::OpUDiv, type_uint, dest_pixel_y, const_dest_tile_height_pixels);
  spv::Id dest_tile_pixel_y = builder.createBinOp(
      spv::OpUMod, type_uint, dest_pixel_y, const_dest_tile_height_pixels);

  assert_true(push_constants_member_address != UINT32_MAX);
  id_vector_temp.clear();
  id_vector_temp.push_back(
      builder.makeIntConstant(int32_t(push_constants_member_address)));
  spv::Id address_constant = builder.createLoad(
      builder.createAccessChain(spv::StorageClassPushConstant, push_constants,
                                id_vector_temp),
      spv::NoPrecision);

  // Calculate the 32bpp tile index from its X and Y parts.
  spv::Id dest_tile_index = builder.createBinOp(
      spv::OpIAdd, type_uint,
      builder.createBinOp(
          spv::OpIMul, type_uint,
          builder.createTriOp(
              spv::OpBitFieldUExtract, type_uint, address_constant,
              builder.makeUintConstant(0),
              builder.makeUintConstant(xenos::kEdramPitchTilesBits)),
          dest_tile_index_y),
      dest_tile_index_x);

  // Load the destination sample index.
  spv::Id dest_sample_id = spv::NoResult;
  if (key.dest_msaa_samples != xenos::MsaaSamples::k1X) {
    if (device_properties.sampleRateShading) {
      assert_true(input_sample_id != spv::NoResult);
      dest_sample_id = builder.createUnaryOp(
          spv::OpBitcast, type_uint,
          builder.createLoad(input_sample_id, spv::NoPrecision));
    } else {
      assert_true(spec_const_sample_id != spv::NoResult);
      // Already uint.
      dest_sample_id = spec_const_sample_id;
    }
  }

  // Transform the destination framebuffer pixel and sample coordinates into the
  // source texture pixel and sample coordinates.

  // First sample bit at 4x with Vulkan standard locations - horizontal sample.
  // Second sample bit at 4x with Vulkan standard locations - vertical sample.
  // At 2x:
  // - Native 2x: top is 1 in Vulkan, bottom is 0.
  // - 2x as 4x: top is 0, bottom is 3.

  spv::Id source_sample_id = dest_sample_id;
  spv::Id source_tile_pixel_x = dest_tile_pixel_x;
  spv::Id source_tile_pixel_y = dest_tile_pixel_y;
  spv::Id source_color_half = spv::NoResult;
  if (!source_is_64bpp && dest_is_64bpp) {
    // 32bpp -> 64bpp, need two samples of the source.
    if (key.source_msaa_samples >= xenos::MsaaSamples::k4X) {
      // 32bpp -> 64bpp, 4x ->.
      // Source has 32bpp halves in two adjacent samples.
      if (key.dest_msaa_samples >= xenos::MsaaSamples::k4X) {
        // 32bpp -> 64bpp, 4x -> 4x.
        // 1 destination horizontal sample = 2 source horizontal samples.
        // D p0,0 s0,0 = S p0,0 s0,0 | S p0,0 s1,0
        // D p0,0 s1,0 = S p1,0 s0,0 | S p1,0 s1,0
        // D p0,0 s0,1 = S p0,0 s0,1 | S p0,0 s1,1
        // D p0,0 s1,1 = S p1,0 s0,1 | S p1,0 s1,1
        // Thus destination horizontal sample -> source horizontal pixel,
        // vertical samples are 1:1.
        source_sample_id =
            builder.createBinOp(spv::OpBitwiseAnd, type_uint, dest_sample_id,
                                builder.makeUintConstant(1 << 1));
        source_tile_pixel_x = builder.createQuadOp(
            spv::OpBitFieldInsert, type_uint, dest_sample_id, dest_tile_pixel_x,
            builder.makeUintConstant(1), builder.makeUintConstant(31));
      } else if (key.dest_msaa_samples == xenos::MsaaSamples::k2X) {
        // 32bpp -> 64bpp, 4x -> 2x.
        // 1 destination horizontal pixel = 2 source horizontal samples.
        // D p0,0 s0 = S p0,0 s0,0 | S p0,0 s1,0
        // D p0,0 s1 = S p0,0 s0,1 | S p0,0 s1,1
        // D p1,0 s0 = S p1,0 s0,0 | S p1,0 s1,0
        // D p1,0 s1 = S p1,0 s0,1 | S p1,0 s1,1
        // Pixel index can be reused. Sample 1 (for native 2x) or 0 (for 2x as
        // 4x) should become samples 01, sample 0 or 3 should become samples 23.
        if (msaa_2x_attachments_supported_) {
          source_sample_id = builder.createBinOp(
              spv::OpShiftLeftLogical, type_uint,
              builder.createBinOp(spv::OpBitwiseXor, type_uint, dest_sample_id,
                                  builder.makeUintConstant(1)),
              builder.makeUintConstant(1));
        } else {
          source_sample_id =
              builder.createBinOp(spv::OpBitwiseAnd, type_uint, dest_sample_id,
                                  builder.makeUintConstant(1 << 1));
        }
      } else {
        // 32bpp -> 64bpp, 4x -> 1x.
        // 1 destination horizontal pixel = 2 source horizontal samples.
        // D p0,0 = S p0,0 s0,0 | S p0,0 s1,0
        // D p0,1 = S p0,0 s0,1 | S p0,0 s1,1
        // Horizontal pixel index can be reused. Vertical pixel 1 should
        // become sample 2.
        source_sample_id = builder.createQuadOp(
            spv::OpBitFieldInsert, type_uint, builder.makeUintConstant(0),
            dest_tile_pixel_y, builder.makeUintConstant(1),
            builder.makeUintConstant(1));
        source_tile_pixel_y =
            builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                dest_tile_pixel_y, builder.makeUintConstant(1));
      }
    } else {
      // 32bpp -> 64bpp, 1x/2x ->.
      // Source has 32bpp halves in two adjacent pixels.
      if (key.dest_msaa_samples >= xenos::MsaaSamples::k4X) {
        // 32bpp -> 64bpp, 1x/2x -> 4x.
        // The X part.
        // 1 destination horizontal sample = 2 source horizontal pixels.
        source_tile_pixel_x = builder.createQuadOp(
            spv::OpBitFieldInsert, type_uint,
            builder.createBinOp(spv::OpShiftLeftLogical, type_uint,
                                dest_tile_pixel_x, builder.makeUintConstant(2)),
            dest_sample_id, builder.makeUintConstant(1),
            builder.makeUintConstant(1));
        // Y is handled by common code.
      } else {
        // 32bpp -> 64bpp, 1x/2x -> 1x/2x.
        // The X part.
        // 1 destination horizontal pixel = 2 source horizontal pixels.
        source_tile_pixel_x =
            builder.createBinOp(spv::OpShiftLeftLogical, type_uint,
                                dest_tile_pixel_x, builder.makeUintConstant(1));
        // Y is handled by common code.
      }
    }
  } else if (source_is_64bpp && !dest_is_64bpp) {
    // 64bpp -> 32bpp, also the half to load.
    if (key.dest_msaa_samples >= xenos::MsaaSamples::k4X) {
      // 64bpp -> 32bpp, -> 4x.
      // The needed half is in the destination horizontal sample index.
      if (key.source_msaa_samples >= xenos::MsaaSamples::k4X) {
        // 64bpp -> 32bpp, 4x -> 4x.
        // D p0,0 s0,0 = S s0,0 low
        // D p0,0 s1,0 = S s0,0 high
        // D p1,0 s0,0 = S s1,0 low
        // D p1,0 s1,0 = S s1,0 high
        // Vertical pixel and sample (second bit) addressing is the same.
        // However, 1 horizontal destination pixel = 1 horizontal source sample.
        source_sample_id = builder.createQuadOp(
            spv::OpBitFieldInsert, type_uint, dest_sample_id, dest_tile_pixel_x,
            builder.makeUintConstant(0), builder.makeUintConstant(1));
        // 2 destination horizontal samples = 1 source horizontal sample, thus
        // 2 destination horizontal pixels = 1 source horizontal pixel.
        source_tile_pixel_x =
            builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                dest_tile_pixel_x, builder.makeUintConstant(1));
      } else {
        // 64bpp -> 32bpp, 1x/2x -> 4x.
        // 2 destination horizontal samples = 1 source horizontal pixel, thus
        // 1 destination horizontal pixel = 1 source horizontal pixel. Can reuse
        // horizontal pixel index.
        // Y is handled by common code.
      }
      // Half from the destination horizontal sample index.
      source_color_half =
          builder.createBinOp(spv::OpBitwiseAnd, type_uint, dest_sample_id,
                              builder.makeUintConstant(1));
    } else {
      // 64bpp -> 32bpp, -> 1x/2x.
      // The needed half is in the destination horizontal pixel index.
      if (key.source_msaa_samples >= xenos::MsaaSamples::k4X) {
        // 64bpp -> 32bpp, 4x -> 1x/2x.
        // (Destination horizontal pixel >> 1) & 1 = source horizontal sample
        // (first bit).
        source_sample_id = builder.createTriOp(
            spv::OpBitFieldUExtract, type_uint, dest_tile_pixel_x,
            builder.makeUintConstant(1), builder.makeUintConstant(1));
        if (key.dest_msaa_samples == xenos::MsaaSamples::k2X) {
          // 64bpp -> 32bpp, 4x -> 2x.
          // Destination vertical samples (1/0 in the first bit for native 2x or
          // 0/1 in the second bit for 2x as 4x) = source vertical samples
          // (second bit).
          if (msaa_2x_attachments_supported_) {
            source_sample_id = builder.createQuadOp(
                spv::OpBitFieldInsert, type_uint, source_sample_id,
                builder.createBinOp(spv::OpBitwiseXor, type_uint,
                                    dest_sample_id,
                                    builder.makeUintConstant(1)),
                builder.makeUintConstant(1), builder.makeUintConstant(1));
          } else {
            source_sample_id = builder.createQuadOp(
                spv::OpBitFieldInsert, type_uint, dest_sample_id,
                source_sample_id, builder.makeUintConstant(0),
                builder.makeUintConstant(1));
          }
        } else {
          // 64bpp -> 32bpp, 4x -> 1x.
          // 1 destination vertical pixel = 1 source vertical sample.
          source_sample_id = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint, source_sample_id,
              source_tile_pixel_y, builder.makeUintConstant(1),
              builder.makeUintConstant(1));
          source_tile_pixel_y = builder.createBinOp(
              spv::OpShiftRightLogical, type_uint, dest_tile_pixel_y,
              builder.makeUintConstant(1));
        }
        // 2 destination horizontal pixels = 1 source horizontal sample.
        // 4 destination horizontal pixels = 1 source horizontal pixel.
        source_tile_pixel_x =
            builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                dest_tile_pixel_x, builder.makeUintConstant(2));
      } else {
        // 64bpp -> 32bpp, 1x/2x -> 1x/2x.
        // The X part.
        // 2 destination horizontal pixels = 1 destination source pixel.
        source_tile_pixel_x =
            builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                dest_tile_pixel_x, builder.makeUintConstant(1));
        // Y is handled by common code.
      }
      // Half from the destination horizontal pixel index.
      source_color_half =
          builder.createBinOp(spv::OpBitwiseAnd, type_uint, dest_tile_pixel_x,
                              builder.makeUintConstant(1));
    }
    assert_true(source_color_half != spv::NoResult);
  } else {
    // Same bit count.
    if (key.source_msaa_samples != key.dest_msaa_samples) {
      if (key.source_msaa_samples >= xenos::MsaaSamples::k4X) {
        // Same BPP, 4x -> 1x/2x.
        if (key.dest_msaa_samples == xenos::MsaaSamples::k2X) {
          // Same BPP, 4x -> 2x.
          // Horizontal pixels to samples. Vertical sample (1/0 in the first bit
          // for native 2x or 0/1 in the second bit for 2x as 4x) to second
          // sample bit.
          if (msaa_2x_attachments_supported_) {
            source_sample_id = builder.createQuadOp(
                spv::OpBitFieldInsert, type_uint, dest_tile_pixel_x,
                builder.createBinOp(spv::OpBitwiseXor, type_uint,
                                    dest_sample_id,
                                    builder.makeUintConstant(1)),
                builder.makeUintConstant(1), builder.makeUintConstant(31));
          } else {
            source_sample_id = builder.createQuadOp(
                spv::OpBitFieldInsert, type_uint, dest_sample_id,
                dest_tile_pixel_x, builder.makeUintConstant(0),
                builder.makeUintConstant(1));
          }
          source_tile_pixel_x = builder.createBinOp(
              spv::OpShiftRightLogical, type_uint, dest_tile_pixel_x,
              builder.makeUintConstant(1));
        } else {
          // Same BPP, 4x -> 1x.
          // Pixels to samples.
          source_sample_id = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint,
              builder.createBinOp(spv::OpBitwiseAnd, type_uint,
                                  dest_tile_pixel_x,
                                  builder.makeUintConstant(1)),
              dest_tile_pixel_y, builder.makeUintConstant(1),
              builder.makeUintConstant(1));
          source_tile_pixel_x = builder.createBinOp(
              spv::OpShiftRightLogical, type_uint, dest_tile_pixel_x,
              builder.makeUintConstant(1));
          source_tile_pixel_y = builder.createBinOp(
              spv::OpShiftRightLogical, type_uint, dest_tile_pixel_y,
              builder.makeUintConstant(1));
        }
      } else {
        // Same BPP, 1x/2x -> 1x/2x/4x (as long as they're different).
        // Only the X part - Y is handled by common code.
        if (key.dest_msaa_samples >= xenos::MsaaSamples::k4X) {
          // Horizontal samples to pixels.
          source_tile_pixel_x = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint, dest_sample_id,
              dest_tile_pixel_x, builder.makeUintConstant(1),
              builder.makeUintConstant(31));
        }
      }
    }
  }
  // Common source Y and sample index for 1x/2x AA sources, independent of bits
  // per sample.
  if (key.source_msaa_samples < xenos::MsaaSamples::k4X &&
      key.source_msaa_samples != key.dest_msaa_samples) {
    if (key.dest_msaa_samples >= xenos::MsaaSamples::k4X) {
      // 1x/2x -> 4x.
      if (key.source_msaa_samples == xenos::MsaaSamples::k2X) {
        // 2x -> 4x.
        // Vertical samples (second bit) of 4x destination to vertical sample
        // (1, 0 for native 2x, or 0, 3 for 2x as 4x) of 2x source.
        source_sample_id =
            builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                dest_sample_id, builder.makeUintConstant(1));
        if (msaa_2x_attachments_supported_) {
          source_sample_id = builder.createBinOp(spv::OpBitwiseXor, type_uint,
                                                 source_sample_id,
                                                 builder.makeUintConstant(1));
        } else {
          source_sample_id = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint, source_sample_id,
              source_sample_id, builder.makeUintConstant(1),
              builder.makeUintConstant(1));
        }
      } else {
        // 1x -> 4x.
        // Vertical samples (second bit) to Y pixels.
        source_tile_pixel_y = builder.createQuadOp(
            spv::OpBitFieldInsert, type_uint,
            builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                dest_sample_id, builder.makeUintConstant(1)),
            dest_tile_pixel_y, builder.makeUintConstant(1),
            builder.makeUintConstant(31));
      }
    } else {
      // 1x/2x -> different 1x/2x.
      if (key.source_msaa_samples == xenos::MsaaSamples::k2X) {
        // 2x -> 1x.
        // Vertical pixels of 2x destination to vertical samples (1, 0 for
        // native 2x, or 0, 3 for 2x as 4x) of 1x source.
        source_sample_id =
            builder.createBinOp(spv::OpBitwiseAnd, type_uint, dest_tile_pixel_y,
                                builder.makeUintConstant(1));
        if (msaa_2x_attachments_supported_) {
          source_sample_id = builder.createBinOp(spv::OpBitwiseXor, type_uint,
                                                 source_sample_id,
                                                 builder.makeUintConstant(1));
        } else {
          source_sample_id = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint, source_sample_id,
              source_sample_id, builder.makeUintConstant(1),
              builder.makeUintConstant(1));
        }
        source_tile_pixel_y =
            builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                dest_tile_pixel_y, builder.makeUintConstant(1));
      } else {
        // 1x -> 2x.
        // Vertical samples (1/0 in the first bit for native 2x or 0/1 in the
        // second bit for 2x as 4x) of 2x destination to vertical pixels of 1x
        // source.
        if (msaa_2x_attachments_supported_) {
          source_tile_pixel_y = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint,
              builder.createBinOp(spv::OpBitwiseXor, type_uint, dest_sample_id,
                                  builder.makeUintConstant(1)),
              dest_tile_pixel_y, builder.makeUintConstant(1),
              builder.makeUintConstant(31));
        } else {
          source_tile_pixel_y = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint,
              builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                  dest_sample_id, builder.makeUintConstant(1)),
              dest_tile_pixel_y, builder.makeUintConstant(1),
              builder.makeUintConstant(31));
        }
      }
    }
  }

  uint32_t source_pixel_width_dwords_log2 =
      uint32_t(key.source_msaa_samples >= xenos::MsaaSamples::k4X) +
      uint32_t(source_is_64bpp);

  if (source_is_color != dest_is_color) {
    // Copying between color and depth / stencil - swap 40-32bpp-sample columns
    // in the pixel index within the source 32bpp tile.
    uint32_t source_32bpp_tile_half_pixels =
        tile_width_samples >> (1 + source_pixel_width_dwords_log2);
    source_tile_pixel_x = builder.createUnaryOp(
        spv::OpBitcast, type_uint,
        builder.createBinOp(
            spv::OpIAdd, type_int,
            builder.createUnaryOp(spv::OpBitcast, type_int,
                                  source_tile_pixel_x),
            builder.createTriOp(
                spv::OpSelect, type_int,
                builder.createBinOp(
                    spv::OpULessThan, builder.makeBoolType(),
                    source_tile_pixel_x,
                    builder.makeUintConstant(source_32bpp_tile_half_pixels)),
                builder.makeIntConstant(int32_t(source_32bpp_tile_half_pixels)),
                builder.makeIntConstant(
                    -int32_t(source_32bpp_tile_half_pixels)))));
  }

  // Transform the destination 32bpp tile index into the source. After the
  // addition, it may be negative - in which case, the transfer is done across
  // EDRAM addressing wrapping, and xenos::kEdramTileCount must be added to it,
  // but `& (xenos::kEdramTileCount - 1)` handles that regardless of the sign.
  spv::Id source_tile_index = builder.createBinOp(
      spv::OpBitwiseAnd, type_uint,
      builder.createUnaryOp(
          spv::OpBitcast, type_uint,
          builder.createBinOp(
              spv::OpIAdd, type_int,
              builder.createUnaryOp(spv::OpBitcast, type_int, dest_tile_index),
              builder.createTriOp(
                  spv::OpBitFieldSExtract, type_int,
                  builder.createUnaryOp(spv::OpBitcast, type_int,
                                        address_constant),
                  builder.makeUintConstant(xenos::kEdramPitchTilesBits * 2),
                  builder.makeUintConstant(xenos::kEdramBaseTilesBits + 1)))),
      builder.makeUintConstant(xenos::kEdramTileCount - 1));
  // Split the source 32bpp tile index into X and Y tile index within the source
  // image.
  spv::Id source_pitch_tiles = builder.createTriOp(
      spv::OpBitFieldUExtract, type_uint, address_constant,
      builder.makeUintConstant(xenos::kEdramPitchTilesBits),
      builder.makeUintConstant(xenos::kEdramPitchTilesBits));
  spv::Id source_tile_index_y = builder.createBinOp(
      spv::OpUDiv, type_uint, source_tile_index, source_pitch_tiles);
  spv::Id source_tile_index_x = builder.createBinOp(
      spv::OpUMod, type_uint, source_tile_index, source_pitch_tiles);
  // Finally calculate the source texture coordinates.
  spv::Id source_pixel_x_int = builder.createUnaryOp(
      spv::OpBitcast, type_int,
      builder.createBinOp(
          spv::OpIAdd, type_uint,
          builder.createBinOp(
              spv::OpIMul, type_uint,
              builder.makeUintConstant(tile_width_samples >>
                                       source_pixel_width_dwords_log2),
              source_tile_index_x),
          source_tile_pixel_x));
  spv::Id source_pixel_y_int = builder.createUnaryOp(
      spv::OpBitcast, type_int,
      builder.createBinOp(
          spv::OpIAdd, type_uint,
          builder.createBinOp(
              spv::OpIMul, type_uint,
              builder.makeUintConstant(
                  tile_height_samples >>
                  uint32_t(key.source_msaa_samples >= xenos::MsaaSamples::k2X)),
              source_tile_index_y),
          source_tile_pixel_y));

  // Load the source.

  spv::Builder::TextureParameters source_texture_parameters = {};
  id_vector_temp.clear();
  id_vector_temp.push_back(source_pixel_x_int);
  id_vector_temp.push_back(source_pixel_y_int);
  spv::Id source_coordinates[2] = {
      builder.createCompositeConstruct(type_int2, id_vector_temp),
  };
  spv::Id source_sample_ids_int[2] = {};
  if (key.source_msaa_samples != xenos::MsaaSamples::k1X) {
    source_sample_ids_int[0] =
        builder.createUnaryOp(spv::OpBitcast, type_int, source_sample_id);
  } else {
    source_texture_parameters.lod = builder.makeIntConstant(0);
  }
  // Go to the next sample or pixel along X if need to load two dwords.
  bool source_load_is_two_32bpp_samples = !source_is_64bpp && dest_is_64bpp;
  if (source_load_is_two_32bpp_samples) {
    if (key.source_msaa_samples >= xenos::MsaaSamples::k4X) {
      source_coordinates[1] = source_coordinates[0];
      source_sample_ids_int[1] = builder.createBinOp(
          spv::OpBitwiseOr, type_int, source_sample_ids_int[0],
          builder.makeIntConstant(1));
    } else {
      id_vector_temp.clear();
      id_vector_temp.push_back(builder.createBinOp(spv::OpBitwiseOr, type_int,
                                                   source_pixel_x_int,
                                                   builder.makeIntConstant(1)));
      id_vector_temp.push_back(source_pixel_y_int);
      source_coordinates[1] =
          builder.createCompositeConstruct(type_int2, id_vector_temp);
      source_sample_ids_int[1] = source_sample_ids_int[0];
    }
  }
  spv::Id source_color[2][4] = {};
  if (source_color_texture != spv::NoResult) {
    source_texture_parameters.sampler =
        builder.createLoad(source_color_texture, spv::NoPrecision);
    assert_true(source_color_component_type != spv::NoType);
    spv::Id source_color_vec4_type =
        builder.makeVectorType(source_color_component_type, 4);
    for (uint32_t i = 0; i <= uint32_t(source_load_is_two_32bpp_samples); ++i) {
      source_texture_parameters.coords = source_coordinates[i];
      source_texture_parameters.sample = source_sample_ids_int[i];
      spv::Id source_color_vec4 = builder.createTextureCall(
          spv::NoPrecision, source_color_vec4_type, false, true, false, false,
          false, source_texture_parameters, spv::ImageOperandsMaskNone);
      uint32_t source_color_components_remaining =
          source_color_texture_component_mask;
      uint32_t source_color_component_index;
      while (xe::bit_scan_forward(source_color_components_remaining,
                                  &source_color_component_index)) {
        source_color_components_remaining &=
            ~(uint32_t(1) << source_color_component_index);
        source_color[i][source_color_component_index] =
            builder.createCompositeExtract(source_color_vec4,
                                           source_color_component_type,
                                           source_color_component_index);
      }
    }
  }
  spv::Id source_depth_float[2] = {};
  if (source_depth_texture != spv::NoResult) {
    source_texture_parameters.sampler =
        builder.createLoad(source_depth_texture, spv::NoPrecision);
    for (uint32_t i = 0; i <= uint32_t(source_load_is_two_32bpp_samples); ++i) {
      source_texture_parameters.coords = source_coordinates[i];
      source_texture_parameters.sample = source_sample_ids_int[i];
      source_depth_float[i] = builder.createCompositeExtract(
          builder.createTextureCall(
              spv::NoPrecision, type_float4, false, true, false, false, false,
              source_texture_parameters, spv::ImageOperandsMaskNone),
          type_float, 0);
    }
  }
  spv::Id source_stencil[2] = {};
  if (source_stencil_texture != spv::NoResult) {
    source_texture_parameters.sampler =
        builder.createLoad(source_stencil_texture, spv::NoPrecision);
    for (uint32_t i = 0; i <= uint32_t(source_load_is_two_32bpp_samples); ++i) {
      source_texture_parameters.coords = source_coordinates[i];
      source_texture_parameters.sample = source_sample_ids_int[i];
      source_stencil[i] = builder.createCompositeExtract(
          builder.createTextureCall(
              spv::NoPrecision, type_uint4, false, true, false, false, false,
              source_texture_parameters, spv::ImageOperandsMaskNone),
          type_uint, 0);
    }
  }

  // Pick the needed 32bpp half of the 64bpp color.
  if (source_is_64bpp && !dest_is_64bpp) {
    uint32_t source_color_half_component_count =
        source_color_format_component_count >> 1;
    assert_true(source_color_half != spv::NoResult);
    spv::Id source_color_is_second_half =
        builder.createBinOp(spv::OpINotEqual, type_bool, source_color_half,
                            builder.makeUintConstant(0));
    if (mode.output == TransferOutput::kStencilBit) {
      source_color[0][0] = builder.createTriOp(
          spv::OpSelect, source_color_component_type,
          source_color_is_second_half,
          source_color[0][source_color_half_component_count],
          source_color[0][0]);
    } else {
      for (uint32_t i = 0; i < source_color_half_component_count; ++i) {
        source_color[0][i] = builder.createTriOp(
            spv::OpSelect, source_color_component_type,
            source_color_is_second_half,
            source_color[0][source_color_half_component_count + i],
            source_color[0][i]);
      }
    }
  }

  if (output_fragment_stencil_ref != spv::NoResult &&
      source_stencil[0] != spv::NoResult) {
    // For the depth -> depth case, write the stencil directly to the output.
    assert_true(mode.output == TransferOutput::kDepth);
    builder.createStore(
        builder.createUnaryOp(spv::OpBitcast, type_int, source_stencil[0]),
        output_fragment_stencil_ref);
  }

  if (dest_is_64bpp) {
    // Construct the 64bpp color from two 32-bit samples or one 64-bit sample.
    // If `packed` (two uints) are created, use the generic path involving
    // unpacking.
    // Otherwise, the fragment data output must be written to directly by the
    // reached control flow path.
    spv::Id packed[2] = {};
    if (source_is_color) {
      switch (source_color_format) {
        case xenos::ColorRenderTargetFormat::k_8_8_8_8:
        case xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA: {
          spv::Id unorm_round_offset = builder.makeFloatConstant(0.5f);
          spv::Id unorm_scale = builder.makeFloatConstant(255.0f);
          spv::Id component_width = builder.makeUintConstant(8);
          for (uint32_t i = 0; i < 2; ++i) {
            packed[i] = builder.createUnaryOp(
                spv::OpConvertFToU, type_uint,
                builder.createBinOp(
                    spv::OpFAdd, type_float,
                    builder.createBinOp(spv::OpFMul, type_float,
                                        source_color[i][0], unorm_scale),
                    unorm_round_offset));
            for (uint32_t j = 1; j < 4; ++j) {
              packed[i] = builder.createQuadOp(
                  spv::OpBitFieldInsert, type_uint, packed[i],
                  builder.createUnaryOp(
                      spv::OpConvertFToU, type_uint,
                      builder.createBinOp(
                          spv::OpFAdd, type_float,
                          builder.createBinOp(spv::OpFMul, type_float,
                                              source_color[i][j], unorm_scale),
                          unorm_round_offset)),
                  builder.makeUintConstant(8 * j), component_width);
            }
          }
        } break;
        case xenos::ColorRenderTargetFormat::k_2_10_10_10:
        case xenos::ColorRenderTargetFormat::k_2_10_10_10_AS_10_10_10_10: {
          spv::Id unorm_round_offset = builder.makeFloatConstant(0.5f);
          spv::Id unorm_scale_rgb = builder.makeFloatConstant(1023.0f);
          spv::Id width_rgb = builder.makeUintConstant(10);
          spv::Id unorm_scale_a = builder.makeFloatConstant(3.0f);
          spv::Id width_a = builder.makeUintConstant(2);
          for (uint32_t i = 0; i < 2; ++i) {
            packed[i] = builder.createUnaryOp(
                spv::OpConvertFToU, type_uint,
                builder.createBinOp(
                    spv::OpFAdd, type_float,
                    builder.createBinOp(spv::OpFMul, type_float,
                                        source_color[i][0], unorm_scale_rgb),
                    unorm_round_offset));
            for (uint32_t j = 1; j < 4; ++j) {
              packed[i] = builder.createQuadOp(
                  spv::OpBitFieldInsert, type_uint, packed[i],
                  builder.createUnaryOp(
                      spv::OpConvertFToU, type_uint,
                      builder.createBinOp(
                          spv::OpFAdd, type_float,
                          builder.createBinOp(
                              spv::OpFMul, type_float, source_color[i][j],
                              j == 3 ? unorm_scale_a : unorm_scale_rgb),
                          unorm_round_offset)),
                  builder.makeUintConstant(10 * j),
                  j == 3 ? width_a : width_rgb);
            }
          }
        } break;
        case xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT:
        case xenos::ColorRenderTargetFormat::
            k_2_10_10_10_FLOAT_AS_16_16_16_16: {
          spv::Id width_rgb = builder.makeUintConstant(10);
          spv::Id float_0 = builder.makeFloatConstant(0.0f);
          spv::Id float_1 = builder.makeFloatConstant(1.0f);
          spv::Id unorm_round_offset = builder.makeFloatConstant(0.5f);
          spv::Id unorm_scale_a = builder.makeFloatConstant(3.0f);
          spv::Id offset_a = builder.makeUintConstant(30);
          spv::Id width_a = builder.makeUintConstant(2);
          for (uint32_t i = 0; i < 2; ++i) {
            // Float16 has a wider range for both color and alpha, also NaNs -
            // clamp and convert.
            packed[i] = SpirvShaderTranslator::UnclampedFloat32To7e3(
                builder, source_color[i][0], ext_inst_glsl_std_450);
            for (uint32_t j = 1; j < 3; ++j) {
              packed[i] = builder.createQuadOp(
                  spv::OpBitFieldInsert, type_uint, packed[i],
                  SpirvShaderTranslator::UnclampedFloat32To7e3(
                      builder, source_color[i][j], ext_inst_glsl_std_450),
                  builder.makeUintConstant(10 * j), width_rgb);
            }
            // Saturate and convert the alpha.
            spv::Id alpha_saturated = builder.createTriBuiltinCall(
                type_float, ext_inst_glsl_std_450, GLSLstd450NClamp,
                source_color[i][3], float_0, float_1);
            packed[i] = builder.createQuadOp(
                spv::OpBitFieldInsert, type_uint, packed[i],
                builder.createUnaryOp(
                    spv::OpConvertFToU, type_uint,
                    builder.createBinOp(
                        spv::OpFAdd, type_float,
                        builder.createBinOp(spv::OpFMul, type_float,
                                            alpha_saturated, unorm_scale_a),
                        unorm_round_offset)),
                offset_a, width_a);
          }
        } break;
        // All 64bpp formats, and all 16 bits per component formats, are
        // represented as integers in ownership transfer for safe handling of
        // NaN encodings and -32768 / -32767.
        // TODO(Triang3l): Handle the case when that's not true (no multisampled
        // sampled images, no 16-bit UNORM, no cross-packing 32bpp aliasing on a
        // portability subset device or a 64bpp format where that wouldn't help
        // anyway).
        case xenos::ColorRenderTargetFormat::k_16_16:
        case xenos::ColorRenderTargetFormat::k_16_16_FLOAT: {
          if (dest_color_format ==
              xenos::ColorRenderTargetFormat::k_32_32_FLOAT) {
            spv::Id component_offset_width = builder.makeUintConstant(16);
            spv::Id color_16_in_32[2];
            for (uint32_t i = 0; i < 2; ++i) {
              color_16_in_32[i] = builder.createQuadOp(
                  spv::OpBitFieldInsert, type_uint, source_color[i][0],
                  source_color[i][1], component_offset_width,
                  component_offset_width);
            }
            id_vector_temp.clear();
            id_vector_temp.push_back(color_16_in_32[0]);
            id_vector_temp.push_back(color_16_in_32[1]);
            builder.createStore(builder.createCompositeConstruct(
                                    type_fragment_data, id_vector_temp),
                                output_fragment_data);
          } else {
            id_vector_temp.clear();
            for (uint32_t i = 0; i < 4; ++i) {
              id_vector_temp.push_back(source_color[i >> 1][i & 1]);
            }
            builder.createStore(builder.createCompositeConstruct(
                                    type_fragment_data, id_vector_temp),
                                output_fragment_data);
          }
        } break;
        case xenos::ColorRenderTargetFormat::k_16_16_16_16:
        case xenos::ColorRenderTargetFormat::k_16_16_16_16_FLOAT: {
          if (dest_color_format ==
              xenos::ColorRenderTargetFormat::k_32_32_FLOAT) {
            spv::Id component_offset_width = builder.makeUintConstant(16);
            spv::Id color_16_in_32[2];
            for (uint32_t i = 0; i < 2; ++i) {
              color_16_in_32[i] = builder.createQuadOp(
                  spv::OpBitFieldInsert, type_uint, source_color[0][i << 1],
                  source_color[0][(i << 1) + 1], component_offset_width,
                  component_offset_width);
            }
            id_vector_temp.clear();
            id_vector_temp.push_back(color_16_in_32[0]);
            id_vector_temp.push_back(color_16_in_32[1]);
            builder.createStore(builder.createCompositeConstruct(
                                    type_fragment_data, id_vector_temp),
                                output_fragment_data);
          } else {
            id_vector_temp.clear();
            for (uint32_t i = 0; i < 4; ++i) {
              id_vector_temp.push_back(source_color[0][i]);
            }
            builder.createStore(builder.createCompositeConstruct(
                                    type_fragment_data, id_vector_temp),
                                output_fragment_data);
          }
        } break;
        // Float32 is transferred as uint32 to preserve NaN encodings. However,
        // multisampled sampled image support is optional in Vulkan.
        case xenos::ColorRenderTargetFormat::k_32_FLOAT: {
          for (uint32_t i = 0; i < 2; ++i) {
            packed[i] = source_color[i][0];
            if (!source_color_is_uint) {
              packed[i] =
                  builder.createUnaryOp(spv::OpBitcast, type_uint, packed[i]);
            }
          }
        } break;
        case xenos::ColorRenderTargetFormat::k_32_32_FLOAT: {
          for (uint32_t i = 0; i < 2; ++i) {
            packed[i] = source_color[0][i];
            if (!source_color_is_uint) {
              packed[i] =
                  builder.createUnaryOp(spv::OpBitcast, type_uint, packed[i]);
            }
          }
        } break;
      }
    } else {
      assert_true(source_depth_texture != spv::NoResult);
      assert_true(source_stencil_texture != spv::NoResult);
      spv::Id depth_offset = builder.makeUintConstant(8);
      spv::Id depth_width = builder.makeUintConstant(24);
      for (uint32_t i = 0; i < 2; ++i) {
        spv::Id depth24 = spv::NoResult;
        switch (source_depth_format) {
          case xenos::DepthRenderTargetFormat::kD24S8: {
            // Round to the nearest even integer. This seems to be the
            // correct conversion, adding +0.5 and rounding towards zero results
            // in red instead of black in the 4D5307E6 clear shader.
            depth24 = builder.createUnaryOp(
                spv::OpConvertFToU, type_uint,
                builder.createUnaryBuiltinCall(
                    type_float, ext_inst_glsl_std_450, GLSLstd450RoundEven,
                    builder.createBinOp(
                        spv::OpFMul, type_float, source_depth_float[i],
                        builder.makeFloatConstant(float(0xFFFFFF)))));
          } break;
          case xenos::DepthRenderTargetFormat::kD24FS8: {
            depth24 = SpirvShaderTranslator::PreClampedDepthTo20e4(
                builder, source_depth_float[i], depth_float24_round(), true,
                ext_inst_glsl_std_450);
          } break;
        }
        // Merge depth and stencil.
        packed[i] = builder.createQuadOp(spv::OpBitFieldInsert, type_uint,
                                         source_stencil[i], depth24,
                                         depth_offset, depth_width);
      }
    }
    // Common path unless there was a specialized one - unpack two packed 32-bit
    // parts.
    if (packed[0] != spv::NoResult) {
      assert_true(packed[1] != spv::NoResult);
      if (dest_color_format == xenos::ColorRenderTargetFormat::k_32_32_FLOAT) {
        id_vector_temp.clear();
        id_vector_temp.push_back(packed[0]);
        id_vector_temp.push_back(packed[1]);
        // Multisampled sampled images are optional in Vulkan, and image views
        // of different formats can't be created separately for sampled image
        // and color attachment usages, so no multisampled integer sampled image
        // support implies no multisampled integer framebuffer attachment
        // support in Xenia.
        if (!dest_color_is_uint) {
          for (spv::Id& float32 : id_vector_temp) {
            float32 =
                builder.createUnaryOp(spv::OpBitcast, type_float, float32);
          }
        }
        builder.createStore(builder.createCompositeConstruct(type_fragment_data,
                                                             id_vector_temp),
                            output_fragment_data);
      } else {
        spv::Id const_uint_0 = builder.makeUintConstant(0);
        spv::Id const_uint_16 = builder.makeUintConstant(16);
        id_vector_temp.clear();
        for (uint32_t i = 0; i < 4; ++i) {
          id_vector_temp.push_back(builder.createTriOp(
              spv::OpBitFieldUExtract, type_uint, packed[i >> 1],
              (i & 1) ? const_uint_16 : const_uint_0, const_uint_16));
        }
        // TODO(Triang3l): Handle the case when that's not true (no multisampled
        // sampled images, no 16-bit UNORM, no cross-packing 32bpp aliasing on a
        // portability subset device or a 64bpp format where that wouldn't help
        // anyway).
        builder.createStore(builder.createCompositeConstruct(type_fragment_data,
                                                             id_vector_temp),
                            output_fragment_data);
      }
    }
  } else {
    // If `packed` is created, use the generic path involving unpacking.
    // - For a color destination, the packed 32bpp color.
    // - For a depth / stencil destination, stencil in 0:7, depth in 8:31
    //   normally, or depth in 0:23 and zeros in 24:31 with packed_only_depth.
    // - For a stencil bit, stencil in 0:7.
    // Otherwise, the fragment data or fragment depth / stencil output must be
    // written to directly by the reached control flow path.
    spv::Id packed = spv::NoResult;
    bool packed_only_depth = false;
    if (source_is_color) {
      switch (source_color_format) {
        case xenos::ColorRenderTargetFormat::k_8_8_8_8:
        case xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA: {
          if (dest_is_color &&
              (dest_color_format == xenos::ColorRenderTargetFormat::k_8_8_8_8 ||
               dest_color_format ==
                   xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA)) {
            // Same format - passthrough.
            id_vector_temp.clear();
            for (uint32_t i = 0; i < 4; ++i) {
              id_vector_temp.push_back(source_color[0][i]);
            }
            builder.createStore(builder.createCompositeConstruct(
                                    type_fragment_data, id_vector_temp),
                                output_fragment_data);
          } else {
            spv::Id unorm_round_offset = builder.makeFloatConstant(0.5f);
            spv::Id unorm_scale = builder.makeFloatConstant(255.0f);
            uint32_t packed_component_offset = 0;
            if (mode.output == TransferOutput::kDepth) {
              // When need only depth, not stencil, skip the red component, and
              // put the depth from GBA directly in the lower bits.
              packed_component_offset = 1;
              packed_only_depth = true;
              if (output_fragment_stencil_ref != spv::NoResult) {
                builder.createStore(
                    builder.createUnaryOp(
                        spv::OpBitcast, type_int,
                        builder.createUnaryOp(
                            spv::OpConvertFToU, type_uint,
                            builder.createBinOp(
                                spv::OpFAdd, type_float,
                                builder.createBinOp(spv::OpFMul, type_float,
                                                    source_color[0][0],
                                                    unorm_scale),
                                unorm_round_offset))),
                    output_fragment_stencil_ref);
              }
            }
            packed = builder.createUnaryOp(
                spv::OpConvertFToU, type_uint,
                builder.createBinOp(
                    spv::OpFAdd, type_float,
                    builder.createBinOp(
                        spv::OpFMul, type_float,
                        source_color[0][packed_component_offset], unorm_scale),
                    unorm_round_offset));
            if (mode.output != TransferOutput::kStencilBit) {
              spv::Id component_width = builder.makeUintConstant(8);
              for (uint32_t i = 1; i < 4 - packed_component_offset; ++i) {
                packed = builder.createQuadOp(
                    spv::OpBitFieldInsert, type_uint, packed,
                    builder.createUnaryOp(
                        spv::OpConvertFToU, type_uint,
                        builder.createBinOp(
                            spv::OpFAdd, type_float,
                            builder.createBinOp(
                                spv::OpFMul, type_float,
                                source_color[0][packed_component_offset + i],
                                unorm_scale),
                            unorm_round_offset)),
                    builder.makeUintConstant(8 * i), component_width);
              }
            }
          }
        } break;
        case xenos::ColorRenderTargetFormat::k_2_10_10_10:
        case xenos::ColorRenderTargetFormat::k_2_10_10_10_AS_10_10_10_10: {
          if (dest_is_color &&
              (dest_color_format ==
                   xenos::ColorRenderTargetFormat::k_2_10_10_10 ||
               dest_color_format == xenos::ColorRenderTargetFormat::
                                        k_2_10_10_10_AS_10_10_10_10)) {
            id_vector_temp.clear();
            for (uint32_t i = 0; i < 4; ++i) {
              id_vector_temp.push_back(source_color[0][i]);
            }
            builder.createStore(builder.createCompositeConstruct(
                                    type_fragment_data, id_vector_temp),
                                output_fragment_data);
          } else {
            spv::Id unorm_round_offset = builder.makeFloatConstant(0.5f);
            spv::Id unorm_scale_rgb = builder.makeFloatConstant(1023.0f);
            packed = builder.createUnaryOp(
                spv::OpConvertFToU, type_uint,
                builder.createBinOp(
                    spv::OpFAdd, type_float,
                    builder.createBinOp(spv::OpFMul, type_float,
                                        source_color[0][0], unorm_scale_rgb),
                    unorm_round_offset));
            if (mode.output != TransferOutput::kStencilBit) {
              spv::Id width_rgb = builder.makeUintConstant(10);
              spv::Id unorm_scale_a = builder.makeFloatConstant(3.0f);
              spv::Id width_a = builder.makeUintConstant(2);
              for (uint32_t i = 1; i < 4; ++i) {
                packed = builder.createQuadOp(
                    spv::OpBitFieldInsert, type_uint, packed,
                    builder.createUnaryOp(
                        spv::OpConvertFToU, type_uint,
                        builder.createBinOp(
                            spv::OpFAdd, type_float,
                            builder.createBinOp(
                                spv::OpFMul, type_float, source_color[0][i],
                                i == 3 ? unorm_scale_a : unorm_scale_rgb),
                            unorm_round_offset)),
                    builder.makeUintConstant(10 * i),
                    i == 3 ? width_a : width_rgb);
              }
            }
          }
        } break;
        case xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT:
        case xenos::ColorRenderTargetFormat::
            k_2_10_10_10_FLOAT_AS_16_16_16_16: {
          if (dest_is_color &&
              (dest_color_format ==
                   xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT ||
               dest_color_format == xenos::ColorRenderTargetFormat::
                                        k_2_10_10_10_FLOAT_AS_16_16_16_16)) {
            id_vector_temp.clear();
            for (uint32_t i = 0; i < 4; ++i) {
              id_vector_temp.push_back(source_color[0][i]);
            }
            builder.createStore(builder.createCompositeConstruct(
                                    type_fragment_data, id_vector_temp),
                                output_fragment_data);
          } else {
            // Float16 has a wider range for both color and alpha, also NaNs -
            // clamp and convert.
            packed = SpirvShaderTranslator::UnclampedFloat32To7e3(
                builder, source_color[0][0], ext_inst_glsl_std_450);
            if (mode.output != TransferOutput::kStencilBit) {
              spv::Id width_rgb = builder.makeUintConstant(10);
              for (uint32_t i = 1; i < 3; ++i) {
                packed = builder.createQuadOp(
                    spv::OpBitFieldInsert, type_uint, packed,
                    SpirvShaderTranslator::UnclampedFloat32To7e3(
                        builder, source_color[0][i], ext_inst_glsl_std_450),
                    builder.makeUintConstant(10 * i), width_rgb);
              }
              // Saturate and convert the alpha.
              spv::Id alpha_saturated = builder.createTriBuiltinCall(
                  type_float, ext_inst_glsl_std_450, GLSLstd450NClamp,
                  source_color[0][3], builder.makeFloatConstant(0.0f),
                  builder.makeFloatConstant(1.0f));
              packed = builder.createQuadOp(
                  spv::OpBitFieldInsert, type_uint, packed,
                  builder.createUnaryOp(
                      spv::OpConvertFToU, type_uint,
                      builder.createBinOp(
                          spv::OpFAdd, type_float,
                          builder.createBinOp(spv::OpFMul, type_float,
                                              alpha_saturated,
                                              builder.makeFloatConstant(3.0f)),
                          builder.makeFloatConstant(0.5f))),
                  builder.makeUintConstant(30), builder.makeUintConstant(2));
            }
          }
        } break;
        case xenos::ColorRenderTargetFormat::k_16_16:
        case xenos::ColorRenderTargetFormat::k_16_16_16_16:
        case xenos::ColorRenderTargetFormat::k_16_16_FLOAT:
        case xenos::ColorRenderTargetFormat::k_16_16_16_16_FLOAT: {
          // All 64bpp formats, and all 16 bits per component formats, are
          // represented as integers in ownership transfer for safe handling of
          // NaN encodings and -32768 / -32767.
          // TODO(Triang3l): Handle the case when that's not true (no
          // multisampled sampled images, no 16-bit UNORM, no cross-packing
          // 32bpp aliasing on a portability subset device or a 64bpp format
          // where that wouldn't help anyway).
          if (dest_is_color &&
              (dest_color_format == xenos::ColorRenderTargetFormat::k_16_16 ||
               dest_color_format ==
                   xenos::ColorRenderTargetFormat::k_16_16_FLOAT)) {
            id_vector_temp.clear();
            for (uint32_t i = 0; i < 2; ++i) {
              id_vector_temp.push_back(source_color[0][i]);
            }
            builder.createStore(builder.createCompositeConstruct(
                                    type_fragment_data, id_vector_temp),
                                output_fragment_data);
          } else {
            packed = source_color[0][0];
            if (mode.output != TransferOutput::kStencilBit) {
              spv::Id component_offset_width = builder.makeUintConstant(16);
              packed = builder.createQuadOp(
                  spv::OpBitFieldInsert, type_uint, packed, source_color[0][1],
                  component_offset_width, component_offset_width);
            }
          }
        } break;
        // Float32 is transferred as uint32 to preserve NaN encodings. However,
        // multisampled sampled image support is optional in Vulkan.
        case xenos::ColorRenderTargetFormat::k_32_FLOAT:
        case xenos::ColorRenderTargetFormat::k_32_32_FLOAT: {
          packed = source_color[0][0];
          if (!source_color_is_uint) {
            packed = builder.createUnaryOp(spv::OpBitcast, type_uint, packed);
          }
        } break;
      }
    } else if (source_depth_float[0] != spv::NoResult) {
      if (mode.output == TransferOutput::kDepth &&
          dest_depth_format == source_depth_format) {
        builder.createStore(source_depth_float[0], output_fragment_depth);
      } else {
        switch (source_depth_format) {
          case xenos::DepthRenderTargetFormat::kD24S8: {
            // Round to the nearest even integer. This seems to be the correct
            // conversion, adding +0.5 and rounding towards zero results in red
            // instead of black in the 4D5307E6 clear shader.
            packed = builder.createUnaryOp(
                spv::OpConvertFToU, type_uint,
                builder.createUnaryBuiltinCall(
                    type_float, ext_inst_glsl_std_450, GLSLstd450RoundEven,
                    builder.createBinOp(
                        spv::OpFMul, type_float, source_depth_float[0],
                        builder.makeFloatConstant(float(0xFFFFFF)))));
          } break;
          case xenos::DepthRenderTargetFormat::kD24FS8: {
            packed = SpirvShaderTranslator::PreClampedDepthTo20e4(
                builder, source_depth_float[0], depth_float24_round(), true,
                ext_inst_glsl_std_450);
          } break;
        }
        if (mode.output == TransferOutput::kDepth) {
          packed_only_depth = true;
        } else {
          // Merge depth and stencil.
          packed = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint, source_stencil[0], packed,
              builder.makeUintConstant(8), builder.makeUintConstant(24));
        }
      }
    }
    switch (mode.output) {
      case TransferOutput::kColor: {
        // Unless a special path was taken, unpack the raw 32bpp value into the
        // 32bpp color output.
        if (packed != spv::NoResult) {
          switch (dest_color_format) {
            case xenos::ColorRenderTargetFormat::k_8_8_8_8:
            case xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA: {
              spv::Id component_width = builder.makeUintConstant(8);
              spv::Id unorm_scale = builder.makeFloatConstant(1.0f / 255.0f);
              id_vector_temp.clear();
              for (uint32_t i = 0; i < 4; ++i) {
                id_vector_temp.push_back(builder.createBinOp(
                    spv::OpFMul, type_float,
                    builder.createUnaryOp(
                        spv::OpConvertUToF, type_float,
                        builder.createTriOp(
                            spv::OpBitFieldUExtract, type_uint, packed,
                            builder.makeUintConstant(8 * i), component_width)),
                    unorm_scale));
              }
              builder.createStore(builder.createCompositeConstruct(
                                      type_fragment_data, id_vector_temp),
                                  output_fragment_data);
            } break;
            case xenos::ColorRenderTargetFormat::k_2_10_10_10:
            case xenos::ColorRenderTargetFormat::k_2_10_10_10_AS_10_10_10_10: {
              spv::Id width_rgb = builder.makeUintConstant(10);
              spv::Id unorm_scale_rgb =
                  builder.makeFloatConstant(1.0f / 1023.0f);
              spv::Id width_a = builder.makeUintConstant(2);
              spv::Id unorm_scale_a = builder.makeFloatConstant(1.0f / 3.0f);
              id_vector_temp.clear();
              for (uint32_t i = 0; i < 4; ++i) {
                id_vector_temp.push_back(builder.createBinOp(
                    spv::OpFMul, type_float,
                    builder.createUnaryOp(
                        spv::OpConvertUToF, type_float,
                        builder.createTriOp(spv::OpBitFieldUExtract, type_uint,
                                            packed,
                                            builder.makeUintConstant(10 * i),
                                            i == 3 ? width_a : width_rgb)),
                    i == 3 ? unorm_scale_a : unorm_scale_rgb));
              }
              builder.createStore(builder.createCompositeConstruct(
                                      type_fragment_data, id_vector_temp),
                                  output_fragment_data);
            } break;
            case xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT:
            case xenos::ColorRenderTargetFormat::
                k_2_10_10_10_FLOAT_AS_16_16_16_16: {
              id_vector_temp.clear();
              // Color.
              spv::Id width_rgb = builder.makeUintConstant(10);
              for (uint32_t i = 0; i < 3; ++i) {
                id_vector_temp.push_back(SpirvShaderTranslator::Float7e3To32(
                    builder, packed, 10 * i, false, ext_inst_glsl_std_450));
              }
              // Alpha.
              id_vector_temp.push_back(builder.createBinOp(
                  spv::OpFMul, type_float,
                  builder.createUnaryOp(
                      spv::OpConvertUToF, type_float,
                      builder.createTriOp(spv::OpBitFieldUExtract, type_uint,
                                          packed, builder.makeUintConstant(30),
                                          builder.makeUintConstant(2))),
                  builder.makeFloatConstant(1.0f / 3.0f)));
              builder.createStore(builder.createCompositeConstruct(
                                      type_fragment_data, id_vector_temp),
                                  output_fragment_data);
            } break;
            case xenos::ColorRenderTargetFormat::k_16_16:
            case xenos::ColorRenderTargetFormat::k_16_16_FLOAT: {
              // All 16 bits per component formats are represented as integers
              // in ownership transfer for safe handling of NaN encodings and
              // -32768 / -32767.
              // TODO(Triang3l): Handle the case when that's not true (no
              // multisampled sampled images, no 16-bit UNORM, no cross-packing
              // 32bpp aliasing on a portability subset device or a 64bpp format
              // where that wouldn't help anyway).
              spv::Id component_offset_width = builder.makeUintConstant(16);
              id_vector_temp.clear();
              for (uint32_t i = 0; i < 2; ++i) {
                id_vector_temp.push_back(builder.createTriOp(
                    spv::OpBitFieldUExtract, type_uint, packed,
                    i ? component_offset_width : builder.makeUintConstant(0),
                    component_offset_width));
              }
              builder.createStore(builder.createCompositeConstruct(
                                      type_fragment_data, id_vector_temp),
                                  output_fragment_data);
            } break;
            case xenos::ColorRenderTargetFormat::k_32_FLOAT: {
              // Float32 is transferred as uint32 to preserve NaN encodings.
              // However, multisampled sampled images are optional in Vulkan,
              // and image views of different formats can't be created
              // separately for sampled image and color attachment usages, so no
              // multisampled integer sampled image support implies no
              // multisampled integer framebuffer attachment support in Xenia.
              spv::Id float32 = packed;
              if (!dest_color_is_uint) {
                float32 =
                    builder.createUnaryOp(spv::OpBitcast, type_float, float32);
              }
              builder.createStore(float32, output_fragment_data);
            } break;
            default:
              // A 64bpp format (handled separately) or an invalid one.
              assert_unhandled_case(dest_color_format);
          }
        }
      } break;
      case TransferOutput::kDepth: {
        if (packed) {
          spv::Id guest_depth24 = packed;
          if (!packed_only_depth) {
            // Extract the depth bits.
            guest_depth24 =
                builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                    guest_depth24, builder.makeUintConstant(8));
          }
          // Load the host float32 depth, check if, when converted to the guest
          // format, it's the same as the guest source, thus up to date, and if
          // it is, write host float32 depth, otherwise do the guest -> host
          // conversion.
          spv::Id host_depth32 = spv::NoResult;
          if (host_depth_source_texture != spv::NoResult) {
            // Convert position and sample index from within the destination
            // tile to within the host depth source tile, like for the guest
            // render target, but for 32bpp -> 32bpp only.
            spv::Id host_depth_source_sample_id = dest_sample_id;
            spv::Id host_depth_source_tile_pixel_x = dest_tile_pixel_x;
            spv::Id host_depth_source_tile_pixel_y = dest_tile_pixel_y;
            if (key.host_depth_source_msaa_samples != key.dest_msaa_samples) {
              if (key.host_depth_source_msaa_samples >=
                  xenos::MsaaSamples::k4X) {
                // 4x -> 1x/2x.
                if (key.dest_msaa_samples == xenos::MsaaSamples::k2X) {
                  // 4x -> 2x.
                  // Horizontal pixels to samples. Vertical sample (1/0 in the
                  // first bit for native 2x or 0/1 in the second bit for 2x as
                  // 4x) to second sample bit.
                  if (msaa_2x_attachments_supported_) {
                    host_depth_source_sample_id = builder.createQuadOp(
                        spv::OpBitFieldInsert, type_uint, dest_tile_pixel_x,
                        builder.createBinOp(spv::OpBitwiseXor, type_uint,
                                            dest_sample_id,
                                            builder.makeUintConstant(1)),
                        builder.makeUintConstant(1),
                        builder.makeUintConstant(31));
                  } else {
                    host_depth_source_sample_id = builder.createQuadOp(
                        spv::OpBitFieldInsert, type_uint, dest_sample_id,
                        dest_tile_pixel_x, builder.makeUintConstant(0),
                        builder.makeUintConstant(1));
                  }
                  host_depth_source_tile_pixel_x = builder.createBinOp(
                      spv::OpShiftRightLogical, type_uint, dest_tile_pixel_x,
                      builder.makeUintConstant(1));
                } else {
                  // 4x -> 1x.
                  // Pixels to samples.
                  host_depth_source_sample_id = builder.createQuadOp(
                      spv::OpBitFieldInsert, type_uint,
                      builder.createBinOp(spv::OpBitwiseAnd, type_uint,
                                          dest_tile_pixel_x,
                                          builder.makeUintConstant(1)),
                      dest_tile_pixel_y, builder.makeUintConstant(1),
                      builder.makeUintConstant(1));
                  host_depth_source_tile_pixel_x = builder.createBinOp(
                      spv::OpShiftRightLogical, type_uint, dest_tile_pixel_x,
                      builder.makeUintConstant(1));
                  host_depth_source_tile_pixel_y = builder.createBinOp(
                      spv::OpShiftRightLogical, type_uint, dest_tile_pixel_y,
                      builder.makeUintConstant(1));
                }
              } else {
                // 1x/2x -> 1x/2x/4x (as long as they're different).
                // Only the X part - Y is handled by common code.
                if (key.dest_msaa_samples >= xenos::MsaaSamples::k4X) {
                  // Horizontal samples to pixels.
                  host_depth_source_tile_pixel_x = builder.createQuadOp(
                      spv::OpBitFieldInsert, type_uint, dest_sample_id,
                      dest_tile_pixel_x, builder.makeUintConstant(1),
                      builder.makeUintConstant(31));
                }
              }
              // Host depth source Y and sample index for 1x/2x AA sources.
              if (key.host_depth_source_msaa_samples <
                  xenos::MsaaSamples::k4X) {
                if (key.dest_msaa_samples >= xenos::MsaaSamples::k4X) {
                  // 1x/2x -> 4x.
                  if (key.host_depth_source_msaa_samples ==
                      xenos::MsaaSamples::k2X) {
                    // 2x -> 4x.
                    // Vertical samples (second bit) of 4x destination to
                    // vertical sample (1, 0 for native 2x, or 0, 3 for 2x as
                    // 4x) of 2x source.
                    host_depth_source_sample_id = builder.createBinOp(
                        spv::OpShiftRightLogical, type_uint, dest_sample_id,
                        builder.makeUintConstant(1));
                    if (msaa_2x_attachments_supported_) {
                      host_depth_source_sample_id =
                          builder.createBinOp(spv::OpBitwiseXor, type_uint,
                                              host_depth_source_sample_id,
                                              builder.makeUintConstant(1));
                    } else {
                      host_depth_source_sample_id =
                          builder.createQuadOp(spv::OpBitFieldInsert, type_uint,
                                               host_depth_source_sample_id,
                                               host_depth_source_sample_id,
                                               builder.makeUintConstant(1),
                                               builder.makeUintConstant(1));
                    }
                  } else {
                    // 1x -> 4x.
                    // Vertical samples (second bit) to Y pixels.
                    host_depth_source_tile_pixel_y = builder.createQuadOp(
                        spv::OpBitFieldInsert, type_uint,
                        builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                            dest_sample_id,
                                            builder.makeUintConstant(1)),
                        dest_tile_pixel_y, builder.makeUintConstant(1),
                        builder.makeUintConstant(31));
                  }
                } else {
                  // 1x/2x -> different 1x/2x.
                  if (key.host_depth_source_msaa_samples ==
                      xenos::MsaaSamples::k2X) {
                    // 2x -> 1x.
                    // Vertical pixels of 2x destination to vertical samples (1,
                    // 0 for native 2x, or 0, 3 for 2x as 4x) of 1x source.
                    host_depth_source_sample_id = builder.createBinOp(
                        spv::OpBitwiseAnd, type_uint, dest_tile_pixel_y,
                        builder.makeUintConstant(1));
                    if (msaa_2x_attachments_supported_) {
                      host_depth_source_sample_id =
                          builder.createBinOp(spv::OpBitwiseXor, type_uint,
                                              host_depth_source_sample_id,
                                              builder.makeUintConstant(1));
                    } else {
                      host_depth_source_sample_id =
                          builder.createQuadOp(spv::OpBitFieldInsert, type_uint,
                                               host_depth_source_sample_id,
                                               host_depth_source_sample_id,
                                               builder.makeUintConstant(1),
                                               builder.makeUintConstant(1));
                    }
                    host_depth_source_tile_pixel_y = builder.createBinOp(
                        spv::OpShiftRightLogical, type_uint, dest_tile_pixel_y,
                        builder.makeUintConstant(1));
                  } else {
                    // 1x -> 2x.
                    // Vertical samples (1/0 in the first bit for native 2x or
                    // 0/1 in the second bit for 2x as 4x) of 2x destination to
                    // vertical pixels of 1x source.
                    if (msaa_2x_attachments_supported_) {
                      host_depth_source_tile_pixel_y = builder.createQuadOp(
                          spv::OpBitFieldInsert, type_uint,
                          builder.createBinOp(spv::OpBitwiseXor, type_uint,
                                              dest_sample_id,
                                              builder.makeUintConstant(1)),
                          dest_tile_pixel_y, builder.makeUintConstant(1),
                          builder.makeUintConstant(31));
                    } else {
                      host_depth_source_tile_pixel_y = builder.createQuadOp(
                          spv::OpBitFieldInsert, type_uint,
                          builder.createBinOp(spv::OpShiftRightLogical,
                                              type_uint, dest_sample_id,
                                              builder.makeUintConstant(1)),
                          dest_tile_pixel_y, builder.makeUintConstant(1),
                          builder.makeUintConstant(31));
                    }
                  }
                }
              }
            }
            assert_true(push_constants_member_host_depth_address != UINT32_MAX);
            id_vector_temp.clear();
            id_vector_temp.push_back(builder.makeIntConstant(
                int32_t(push_constants_member_host_depth_address)));
            spv::Id host_depth_address_constant = builder.createLoad(
                builder.createAccessChain(spv::StorageClassPushConstant,
                                          push_constants, id_vector_temp),
                spv::NoPrecision);
            // Transform the destination tile index into the host depth source.
            // After the addition, it may be negative - in which case, the
            // transfer is done across EDRAM addressing wrapping, and
            // xenos::kEdramTileCount must be added to it, but
            // `& (xenos::kEdramTileCount - 1)` handles that regardless of the
            // sign.
            spv::Id host_depth_source_tile_index = builder.createBinOp(
                spv::OpBitwiseAnd, type_uint,
                builder.createUnaryOp(
                    spv::OpBitcast, type_uint,
                    builder.createBinOp(
                        spv::OpIAdd, type_int,
                        builder.createUnaryOp(spv::OpBitcast, type_int,
                                              dest_tile_index),
                        builder.createTriOp(
                            spv::OpBitFieldSExtract, type_int,
                            builder.createUnaryOp(spv::OpBitcast, type_int,
                                                  host_depth_address_constant),
                            builder.makeUintConstant(
                                xenos::kEdramPitchTilesBits * 2),
                            builder.makeUintConstant(
                                xenos::kEdramBaseTilesBits + 1)))),
                builder.makeUintConstant(xenos::kEdramTileCount - 1));
            // Split the host depth source tile index into X and Y tile index
            // within the source image.
            spv::Id host_depth_source_pitch_tiles = builder.createTriOp(
                spv::OpBitFieldUExtract, type_uint, host_depth_address_constant,
                builder.makeUintConstant(xenos::kEdramPitchTilesBits),
                builder.makeUintConstant(xenos::kEdramPitchTilesBits));
            spv::Id host_depth_source_tile_index_y = builder.createBinOp(
                spv::OpUDiv, type_uint, host_depth_source_tile_index,
                host_depth_source_pitch_tiles);
            spv::Id host_depth_source_tile_index_x = builder.createBinOp(
                spv::OpUMod, type_uint, host_depth_source_tile_index,
                host_depth_source_pitch_tiles);
            // Finally calculate the host depth source texture coordinates.
            spv::Id host_depth_source_pixel_x_int = builder.createUnaryOp(
                spv::OpBitcast, type_int,
                builder.createBinOp(
                    spv::OpIAdd, type_uint,
                    builder.createBinOp(spv::OpIMul, type_uint,
                                        builder.makeUintConstant(
                                            tile_width_samples >>
                                            uint32_t(key.source_msaa_samples >=
                                                     xenos::MsaaSamples::k4X)),
                                        host_depth_source_tile_index_x),
                    host_depth_source_tile_pixel_x));
            spv::Id host_depth_source_pixel_y_int = builder.createUnaryOp(
                spv::OpBitcast, type_int,
                builder.createBinOp(
                    spv::OpIAdd, type_uint,
                    builder.createBinOp(spv::OpIMul, type_uint,
                                        builder.makeUintConstant(
                                            tile_height_samples >>
                                            uint32_t(key.source_msaa_samples >=
                                                     xenos::MsaaSamples::k2X)),
                                        host_depth_source_tile_index_y),
                    host_depth_source_tile_pixel_y));
            // Load the host depth source.
            spv::Builder::TextureParameters
                host_depth_source_texture_parameters = {};
            host_depth_source_texture_parameters.sampler =
                builder.createLoad(host_depth_source_texture, spv::NoPrecision);
            id_vector_temp.clear();
            id_vector_temp.push_back(host_depth_source_pixel_x_int);
            id_vector_temp.push_back(host_depth_source_pixel_y_int);
            host_depth_source_texture_parameters.coords =
                builder.createCompositeConstruct(type_int2, id_vector_temp);
            if (key.host_depth_source_msaa_samples != xenos::MsaaSamples::k1X) {
              host_depth_source_texture_parameters.sample =
                  builder.createUnaryOp(spv::OpBitcast, type_int,
                                        host_depth_source_sample_id);
            } else {
              host_depth_source_texture_parameters.lod =
                  builder.makeIntConstant(0);
            }
            host_depth32 = builder.createCompositeExtract(
                builder.createTextureCall(spv::NoPrecision, type_float4, false,
                                          true, false, false, false,
                                          host_depth_source_texture_parameters,
                                          spv::ImageOperandsMaskNone),
                type_float, 0);
          } else if (host_depth_source_buffer != spv::NoResult) {
            // Get the address in the EDRAM scratch buffer and load from there.
            // The beginning of the buffer is (0, 0) of the destination.
            // 40-sample columns are not swapped for addressing simplicity
            // (because this is used for depth -> depth transfers, where
            // swapping isn't needed).
            // Convert samples to pixels.
            assert_true(key.host_depth_source_msaa_samples ==
                        xenos::MsaaSamples::k1X);
            spv::Id dest_tile_sample_x = dest_tile_pixel_x;
            spv::Id dest_tile_sample_y = dest_tile_pixel_y;
            if (key.dest_msaa_samples >= xenos::MsaaSamples::k2X) {
              if (key.dest_msaa_samples >= xenos::MsaaSamples::k4X) {
                // Horizontal sample index in bit 0.
                dest_tile_sample_x = builder.createQuadOp(
                    spv::OpBitFieldInsert, type_uint, dest_sample_id,
                    dest_tile_pixel_x, builder.makeUintConstant(1),
                    builder.makeUintConstant(31));
              }
              // Vertical sample index as 1 or 0 in bit 0 for true 2x or as 0
              // or 1 in bit 1 for 4x or for 2x emulated as 4x.
              dest_tile_sample_y = builder.createQuadOp(
                  spv::OpBitFieldInsert, type_uint,
                  builder.createBinOp(
                      (key.dest_msaa_samples == xenos::MsaaSamples::k2X &&
                       msaa_2x_attachments_supported_)
                          ? spv::OpBitwiseXor
                          : spv::OpShiftRightLogical,
                      type_uint, dest_sample_id, builder.makeUintConstant(1)),
                  dest_tile_pixel_y, builder.makeUintConstant(1),
                  builder.makeUintConstant(31));
            }
            // Combine the tile sample index and the tile index.
            // The tile index doesn't need to be wrapped, as the host depth is
            // written to the beginning of the buffer, without the base offset.
            spv::Id host_depth_offset = builder.createBinOp(
                spv::OpIAdd, type_uint,
                builder.createBinOp(
                    spv::OpIMul, type_uint,
                    builder.makeUintConstant(tile_width_samples *
                                             tile_height_samples),
                    dest_tile_index),
                builder.createBinOp(
                    spv::OpIAdd, type_uint,
                    builder.createBinOp(
                        spv::OpIMul, type_uint,
                        builder.makeUintConstant(tile_width_samples),
                        dest_tile_sample_y),
                    dest_tile_sample_x));
            id_vector_temp.clear();
            // The only SSBO structure member.
            id_vector_temp.push_back(builder.makeIntConstant(0));
            id_vector_temp.push_back(builder.createUnaryOp(
                spv::OpBitcast, type_int, host_depth_offset));
            // StorageBuffer since SPIR-V 1.3, but since SPIR-V 1.0 is
            // generated, it's Uniform.
            host_depth32 = builder.createUnaryOp(
                spv::OpBitcast, type_float,
                builder.createLoad(
                    builder.createAccessChain(spv::StorageClassUniform,
                                              host_depth_source_buffer,
                                              id_vector_temp),
                    spv::NoPrecision));
          }
          spv::Block* depth24_to_depth32_header = builder.getBuildPoint();
          spv::Id depth24_to_depth32_convert_id = spv::NoResult;
          spv::Block* depth24_to_depth32_merge = nullptr;
          spv::Id host_depth24 = spv::NoResult;
          if (host_depth32 != spv::NoResult) {
            // Convert the host depth value to the guest format and check if it
            // matches the value in the currently owning guest render target.
            switch (dest_depth_format) {
              case xenos::DepthRenderTargetFormat::kD24S8: {
                // Round to the nearest even integer. This seems to be the
                // correct conversion, adding +0.5 and rounding towards zero
                // results in red instead of black in the 4D5307E6 clear shader.
                host_depth24 = builder.createUnaryOp(
                    spv::OpConvertFToU, type_uint,
                    builder.createUnaryBuiltinCall(
                        type_float, ext_inst_glsl_std_450, GLSLstd450RoundEven,
                        builder.createBinOp(
                            spv::OpFMul, type_float, host_depth32,
                            builder.makeFloatConstant(float(0xFFFFFF)))));
              } break;
              case xenos::DepthRenderTargetFormat::kD24FS8: {
                host_depth24 = SpirvShaderTranslator::PreClampedDepthTo20e4(
                    builder, host_depth32, depth_float24_round(), true,
                    ext_inst_glsl_std_450);
              } break;
            }
            assert_true(host_depth24 != spv::NoResult);
            // Update the header block pointer after the conversion (to avoid
            // assuming that the conversion doesn't branch).
            depth24_to_depth32_header = builder.getBuildPoint();
            spv::Id host_depth_outdated = builder.createBinOp(
                spv::OpINotEqual, type_bool, guest_depth24, host_depth24);
            spv::Block& depth24_to_depth32_convert_entry =
                builder.makeNewBlock();
            {
              spv::Block& depth24_to_depth32_merge_block =
                  builder.makeNewBlock();
              depth24_to_depth32_merge = &depth24_to_depth32_merge_block;
            }
            builder.createSelectionMerge(depth24_to_depth32_merge,
                                         spv::SelectionControlMaskNone);
            builder.createConditionalBranch(host_depth_outdated,
                                            &depth24_to_depth32_convert_entry,
                                            depth24_to_depth32_merge);
            builder.setBuildPoint(&depth24_to_depth32_convert_entry);
          }
          // Convert the guest 24-bit depth to float32 (in an open conditional
          // if the host depth is also loaded).
          spv::Id guest_depth32 = spv::NoResult;
          switch (dest_depth_format) {
            case xenos::DepthRenderTargetFormat::kD24S8: {
              // Multiplying by 1.0 / 0xFFFFFF produces an incorrect result (for
              // 0xC00000, for instance - which is 2_10_10_10 clear to 0001) -
              // rescale from 0...0xFFFFFF to 0...0x1000000 doing what true
              // float division followed by multiplication does (on x86-64 MSVC
              // with default SSE rounding) - values starting from 0x800000
              // become bigger by 1; then accurately bias the result's exponent.
              guest_depth32 = builder.createBinOp(
                  spv::OpFMul, type_float,
                  builder.createUnaryOp(
                      spv::OpConvertUToF, type_float,
                      builder.createBinOp(
                          spv::OpIAdd, type_uint, guest_depth24,
                          builder.createBinOp(spv::OpShiftRightLogical,
                                              type_uint, guest_depth24,
                                              builder.makeUintConstant(23)))),
                  builder.makeFloatConstant(1.0f / float(1 << 24)));
            } break;
            case xenos::DepthRenderTargetFormat::kD24FS8: {
              guest_depth32 = SpirvShaderTranslator::Depth20e4To32(
                  builder, guest_depth24, 0, true, false,
                  ext_inst_glsl_std_450);
            } break;
          }
          assert_true(guest_depth32 != spv::NoResult);
          spv::Id fragment_depth32 = guest_depth32;
          if (host_depth32 != spv::NoResult) {
            assert_not_null(depth24_to_depth32_merge);
            spv::Id depth24_to_depth32_result_block_id =
                builder.getBuildPoint()->getId();
            builder.createBranch(depth24_to_depth32_merge);
            builder.setBuildPoint(depth24_to_depth32_merge);
            id_vector_temp.clear();
            id_vector_temp.push_back(guest_depth32);
            id_vector_temp.push_back(depth24_to_depth32_result_block_id);
            id_vector_temp.push_back(host_depth32);
            id_vector_temp.push_back(depth24_to_depth32_header->getId());
            fragment_depth32 =
                builder.createOp(spv::OpPhi, type_float, id_vector_temp);
          }
          builder.createStore(fragment_depth32, output_fragment_depth);
          // Unpack the stencil into the stencil reference output if needed and
          // not already written.
          if (!packed_only_depth &&
              output_fragment_stencil_ref != spv::NoResult) {
            builder.createStore(
                builder.createUnaryOp(
                    spv::OpBitcast, type_int,
                    builder.createBinOp(spv::OpBitwiseAnd, type_uint, packed,
                                        builder.makeUintConstant(UINT8_MAX))),
                output_fragment_stencil_ref);
          }
        }
      } break;
      case TransferOutput::kStencilBit: {
        if (packed) {
          // Kill the sample if the needed stencil bit is not set.
          assert_true(push_constants_member_stencil_mask != UINT32_MAX);
          id_vector_temp.clear();
          id_vector_temp.push_back(builder.makeIntConstant(
              int32_t(push_constants_member_stencil_mask)));
          spv::Id stencil_mask_constant = builder.createLoad(
              builder.createAccessChain(spv::StorageClassPushConstant,
                                        push_constants, id_vector_temp),
              spv::NoPrecision);
          SpirvBuilder::IfBuilder stencil_kill_if(
              builder.createBinOp(
                  spv::OpIEqual, type_bool,
                  builder.createBinOp(spv::OpBitwiseAnd, type_uint, packed,
                                      stencil_mask_constant),
                  builder.makeUintConstant(0)),
              spv::SelectionControlMaskNone, builder);
          builder.createNoResultOp(spv::OpKill);
          // OpKill terminates the block.
          stencil_kill_if.makeEndIf(false);
        }
      } break;
    }
  }

  // End the main function and make it the entry point.
  builder.leaveFunction();
  builder.addExecutionMode(main_function, spv::ExecutionModeOriginUpperLeft);
  if (output_fragment_depth != spv::NoResult) {
    builder.addExecutionMode(main_function, spv::ExecutionModeDepthReplacing);
  }
  if (output_fragment_stencil_ref != spv::NoResult) {
    builder.addExecutionMode(main_function,
                             spv::ExecutionModeStencilRefReplacingEXT);
  }
  spv::Instruction* entry_point =
      builder.addEntryPoint(spv::ExecutionModelFragment, main_function, "main");
  for (spv::Id interface_id : main_interface) {
    entry_point->addIdOperand(interface_id);
  }

  // Serialize the shader code.
  std::vector<unsigned int> shader_code;
  builder.dump(shader_code);

  // Create the shader module, and store the handle even if creation fails not
  // to try to create it again later.
  VkShaderModule shader_module = ui::vulkan::util::CreateShaderModule(
      vulkan_device, reinterpret_cast<const uint32_t*>(shader_code.data()),
      sizeof(uint32_t) * shader_code.size());
  if (shader_module == VK_NULL_HANDLE) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to create the render target ownership "
        "transfer shader 0x{:08X}",
        key.key);
  }
  transfer_shaders_.emplace(key, shader_module);
  return shader_module;
}

VkPipeline const* VulkanRenderTargetCache::GetTransferPipelines(
    TransferPipelineKey key) {
  auto pipeline_it = transfer_pipelines_.find(key);
  if (pipeline_it != transfer_pipelines_.end()) {
    return pipeline_it->second[0] != VK_NULL_HANDLE ? pipeline_it->second.data()
                                                    : nullptr;
  }

  VkRenderPass render_pass =
      GetHostRenderTargetsRenderPass(key.render_pass_key);
  VkShaderModule fragment_shader_module = GetTransferShader(key.shader_key);
  if (render_pass == VK_NULL_HANDLE ||
      fragment_shader_module == VK_NULL_HANDLE) {
    transfer_pipelines_.emplace(key, std::array<VkPipeline, 4>{});
    return nullptr;
  }

  const TransferModeInfo& mode = kTransferModes[size_t(key.shader_key.mode)];

  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  const ui::vulkan::VulkanDevice::Functions& dfn = vulkan_device->functions();
  const VkDevice device = vulkan_device->device();
  const ui::vulkan::VulkanDevice::Properties& device_properties =
      vulkan_device->properties();

  uint32_t dest_sample_count = uint32_t(1)
                               << uint32_t(key.shader_key.dest_msaa_samples);
  bool dest_is_masked_sample =
      dest_sample_count > 1 && !device_properties.sampleRateShading;

  VkPipelineShaderStageCreateInfo shader_stages[2];
  shader_stages[0].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
  shader_stages[0].pNext = nullptr;
  shader_stages[0].flags = 0;
  shader_stages[0].stage = VK_SHADER_STAGE_VERTEX_BIT;
  shader_stages[0].module = transfer_passthrough_vertex_shader_;
  shader_stages[0].pName = "main";
  shader_stages[0].pSpecializationInfo = nullptr;
  shader_stages[1].sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO;
  shader_stages[1].pNext = nullptr;
  shader_stages[1].flags = 0;
  shader_stages[1].stage = VK_SHADER_STAGE_FRAGMENT_BIT;
  shader_stages[1].module = fragment_shader_module;
  shader_stages[1].pName = "main";
  shader_stages[1].pSpecializationInfo = nullptr;
  VkSpecializationMapEntry sample_id_specialization_map_entry;
  uint32_t sample_id_specialization_constant;
  VkSpecializationInfo sample_id_specialization_info;
  if (dest_is_masked_sample) {
    sample_id_specialization_map_entry.constantID = 0;
    sample_id_specialization_map_entry.offset = 0;
    sample_id_specialization_map_entry.size = sizeof(uint32_t);
    sample_id_specialization_constant = 0;
    sample_id_specialization_info.mapEntryCount = 1;
    sample_id_specialization_info.pMapEntries =
        &sample_id_specialization_map_entry;
    sample_id_specialization_info.dataSize =
        sizeof(sample_id_specialization_constant);
    sample_id_specialization_info.pData = &sample_id_specialization_constant;
    shader_stages[1].pSpecializationInfo = &sample_id_specialization_info;
  }

  VkVertexInputBindingDescription vertex_input_binding;
  vertex_input_binding.binding = 0;
  vertex_input_binding.stride = sizeof(float) * 2;
  vertex_input_binding.inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
  VkVertexInputAttributeDescription vertex_input_attribute;
  vertex_input_attribute.location = 0;
  vertex_input_attribute.binding = 0;
  vertex_input_attribute.format = VK_FORMAT_R32G32_SFLOAT;
  vertex_input_attribute.offset = 0;
  VkPipelineVertexInputStateCreateInfo vertex_input_state;
  vertex_input_state.sType =
      VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO;
  vertex_input_state.pNext = nullptr;
  vertex_input_state.flags = 0;
  vertex_input_state.vertexBindingDescriptionCount = 1;
  vertex_input_state.pVertexBindingDescriptions = &vertex_input_binding;
  vertex_input_state.vertexAttributeDescriptionCount = 1;
  vertex_input_state.pVertexAttributeDescriptions = &vertex_input_attribute;

  VkPipelineInputAssemblyStateCreateInfo input_assembly_state;
  input_assembly_state.sType =
      VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO;
  input_assembly_state.pNext = nullptr;
  input_assembly_state.flags = 0;
  input_assembly_state.topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
  input_assembly_state.primitiveRestartEnable = VK_FALSE;

  // Dynamic, to stay within maxViewportDimensions while preferring a
  // power-of-two factor for converting from pixel coordinates to NDC for exact
  // precision.
  VkPipelineViewportStateCreateInfo viewport_state;
  viewport_state.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO;
  viewport_state.pNext = nullptr;
  viewport_state.flags = 0;
  viewport_state.viewportCount = 1;
  viewport_state.pViewports = nullptr;
  viewport_state.scissorCount = 1;
  viewport_state.pScissors = nullptr;

  VkPipelineRasterizationStateCreateInfo rasterization_state = {};
  rasterization_state.sType =
      VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO;
  rasterization_state.polygonMode = VK_POLYGON_MODE_FILL;
  rasterization_state.cullMode = VK_CULL_MODE_NONE;
  rasterization_state.frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE;
  rasterization_state.lineWidth = 1.0f;

  // For samples other than the first, will be changed for the pipelines for
  // other samples.
  VkSampleMask sample_mask = UINT32_MAX;
  VkPipelineMultisampleStateCreateInfo multisample_state = {};
  multisample_state.sType =
      VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO;
  multisample_state.rasterizationSamples =
      (dest_sample_count == 2 && !msaa_2x_attachments_supported_)
          ? VK_SAMPLE_COUNT_4_BIT
          : VkSampleCountFlagBits(dest_sample_count);
  if (dest_sample_count > 1) {
    if (device_properties.sampleRateShading) {
      multisample_state.sampleShadingEnable = VK_TRUE;
      multisample_state.minSampleShading = 1.0f;
      if (dest_sample_count == 2 && !msaa_2x_attachments_supported_) {
        // Emulating 2x MSAA as samples 0 and 3 of 4x MSAA when 2x is not
        // supported.
        sample_mask = 0b1001;
      }
    } else {
      sample_mask = 0b1;
    }
    if (sample_mask != UINT32_MAX) {
      multisample_state.pSampleMask = &sample_mask;
    }
  }

  // Whether the depth / stencil state is used depends on the presence of a
  // depth attachment in the render pass - but not making assumptions about
  // whether the render pass contains any specific attachments, so setting up
  // valid depth / stencil state unconditionally.
  VkPipelineDepthStencilStateCreateInfo depth_stencil_state = {};
  depth_stencil_state.sType =
      VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO;
  if (mode.output == TransferOutput::kDepth) {
    depth_stencil_state.depthTestEnable = VK_TRUE;
    depth_stencil_state.depthWriteEnable = VK_TRUE;
    depth_stencil_state.depthCompareOp = cvars::depth_transfer_not_equal_test
                                             ? VK_COMPARE_OP_NOT_EQUAL
                                             : VK_COMPARE_OP_ALWAYS;
  }
  if ((mode.output == TransferOutput::kDepth &&
       vulkan_device->extensions().ext_EXT_shader_stencil_export) ||
      mode.output == TransferOutput::kStencilBit) {
    depth_stencil_state.stencilTestEnable = VK_TRUE;
    depth_stencil_state.front.failOp = VK_STENCIL_OP_KEEP;
    depth_stencil_state.front.passOp = VK_STENCIL_OP_REPLACE;
    depth_stencil_state.front.depthFailOp = VK_STENCIL_OP_REPLACE;
    // Using ALWAYS, not NOT_EQUAL, so depth writing is unaffected by stencil
    // being different.
    depth_stencil_state.front.compareOp = VK_COMPARE_OP_ALWAYS;
    // Will be dynamic for stencil bit output.
    depth_stencil_state.front.writeMask = UINT8_MAX;
    depth_stencil_state.front.reference = UINT8_MAX;
    depth_stencil_state.back = depth_stencil_state.front;
  }

  // Whether the color blend state is used depends on the presence of color
  // attachments in the render pass - but not making assumptions about whether
  // the render pass contains any specific attachments, so setting up valid
  // color blend state unconditionally.
  VkPipelineColorBlendAttachmentState
      color_blend_attachments[xenos::kMaxColorRenderTargets] = {};
  VkPipelineColorBlendStateCreateInfo color_blend_state = {};
  color_blend_state.sType =
      VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO;
  color_blend_state.attachmentCount =
      32 - xe::lzcnt(key.render_pass_key.depth_and_color_used >> 1);
  color_blend_state.pAttachments = color_blend_attachments;
  if (mode.output == TransferOutput::kColor) {
    assert_true(device_properties.independentBlend);
    color_blend_attachments[key.shader_key.dest_color_rt_index].colorWriteMask =
        VK_COLOR_COMPONENT_R_BIT | VK_COLOR_COMPONENT_G_BIT |
        VK_COLOR_COMPONENT_B_BIT | VK_COLOR_COMPONENT_A_BIT;
  }

  std::array<VkDynamicState, 3> dynamic_states;
  VkPipelineDynamicStateCreateInfo dynamic_state;
  dynamic_state.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO;
  dynamic_state.pNext = nullptr;
  dynamic_state.flags = 0;
  dynamic_state.dynamicStateCount = 0;
  dynamic_state.pDynamicStates = dynamic_states.data();
  dynamic_states[dynamic_state.dynamicStateCount++] = VK_DYNAMIC_STATE_VIEWPORT;
  dynamic_states[dynamic_state.dynamicStateCount++] = VK_DYNAMIC_STATE_SCISSOR;
  if (mode.output == TransferOutput::kStencilBit) {
    dynamic_states[dynamic_state.dynamicStateCount++] =
        VK_DYNAMIC_STATE_STENCIL_WRITE_MASK;
  }

  std::array<VkPipeline, 4> pipelines{};
  VkGraphicsPipelineCreateInfo pipeline_create_info;
  pipeline_create_info.sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO;
  pipeline_create_info.pNext = nullptr;
  pipeline_create_info.flags = 0;
  if (dest_is_masked_sample) {
    pipeline_create_info.flags |= VK_PIPELINE_CREATE_ALLOW_DERIVATIVES_BIT;
  }
  pipeline_create_info.stageCount = uint32_t(xe::countof(shader_stages));
  pipeline_create_info.pStages = shader_stages;
  pipeline_create_info.pVertexInputState = &vertex_input_state;
  pipeline_create_info.pInputAssemblyState = &input_assembly_state;
  pipeline_create_info.pTessellationState = nullptr;
  pipeline_create_info.pViewportState = &viewport_state;
  pipeline_create_info.pRasterizationState = &rasterization_state;
  pipeline_create_info.pMultisampleState = &multisample_state;
  pipeline_create_info.pDepthStencilState = &depth_stencil_state;
  pipeline_create_info.pColorBlendState = &color_blend_state;
  pipeline_create_info.pDynamicState = &dynamic_state;
  pipeline_create_info.layout =
      transfer_pipeline_layouts_[size_t(mode.pipeline_layout)];
  pipeline_create_info.renderPass = render_pass;
  pipeline_create_info.subpass = 0;
  pipeline_create_info.basePipelineHandle = VK_NULL_HANDLE;
  pipeline_create_info.basePipelineIndex = -1;
  if (dfn.vkCreateGraphicsPipelines(device, VK_NULL_HANDLE, 1,
                                    &pipeline_create_info, nullptr,
                                    &pipelines[0]) != VK_SUCCESS) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to create the render target ownership "
        "transfer pipeline for render pass 0x{:08X}, shader 0x{:08X}",
        key.render_pass_key.key, key.shader_key.key);
    transfer_pipelines_.emplace(key, std::array<VkPipeline, 4>{});
    return nullptr;
  }
  if (dest_is_masked_sample) {
    assert_true(multisample_state.pSampleMask == &sample_mask);
    pipeline_create_info.flags = (pipeline_create_info.flags &
                                  ~VK_PIPELINE_CREATE_ALLOW_DERIVATIVES_BIT) |
                                 VK_PIPELINE_CREATE_DERIVATIVE_BIT;
    pipeline_create_info.basePipelineHandle = pipelines[0];
    for (uint32_t i = 1; i < dest_sample_count; ++i) {
      // Emulating 2x MSAA as samples 0 and 3 of 4x MSAA when 2x is not
      // supported.
      uint32_t host_sample_index =
          (dest_sample_count == 2 && !msaa_2x_attachments_supported_ && i == 1)
              ? 3
              : i;
      sample_id_specialization_constant = host_sample_index;
      sample_mask = uint32_t(1) << host_sample_index;
      if (dfn.vkCreateGraphicsPipelines(device, VK_NULL_HANDLE, 1,
                                        &pipeline_create_info, nullptr,
                                        &pipelines[i]) != VK_SUCCESS) {
        XELOGE(
            "VulkanRenderTargetCache: Failed to create the render target "
            "ownership transfer pipeline for render pass 0x{:08X}, shader "
            "0x{:08X}, sample {}",
            key.render_pass_key.key, key.shader_key.key, i);
        for (uint32_t j = 0; j < i; ++j) {
          dfn.vkDestroyPipeline(device, pipelines[j], nullptr);
        }
        transfer_pipelines_.emplace(key, std::array<VkPipeline, 4>{});
        return nullptr;
      }
    }
  }
  return transfer_pipelines_.emplace(key, pipelines).first->second.data();
}

void VulkanRenderTargetCache::PerformTransfersAndResolveClears(
    uint32_t render_target_count, RenderTarget* const* render_targets,
    const std::vector<Transfer>* render_target_transfers,
    const uint64_t* render_target_resolve_clear_values,
    const Transfer::Rectangle* resolve_clear_rectangle) {
  assert_true(GetPath() == Path::kHostRenderTargets);

  const ui::vulkan::VulkanDevice* const vulkan_device =
      command_processor_.GetVulkanDevice();
  uint64_t current_submission = command_processor_.GetCurrentSubmission();
  DeferredCommandBuffer& command_buffer =
      command_processor_.deferred_command_buffer();

  bool resolve_clear_needed =
      render_target_resolve_clear_values && resolve_clear_rectangle;
  VkClearRect resolve_clear_rect;
  if (resolve_clear_needed) {
    // Assuming the rectangle is already clamped by the setup function from the
    // common render target cache.
    resolve_clear_rect.rect.offset.x =
        int32_t(resolve_clear_rectangle->x_pixels * draw_resolution_scale_x());
    resolve_clear_rect.rect.offset.y =
        int32_t(resolve_clear_rectangle->y_pixels * draw_resolution_scale_y());
    resolve_clear_rect.rect.extent.width =
        resolve_clear_rectangle->width_pixels * draw_resolution_scale_x();
    resolve_clear_rect.rect.extent.height =
        resolve_clear_rectangle->height_pixels * draw_resolution_scale_y();
    resolve_clear_rect.baseArrayLayer = 0;
    resolve_clear_rect.layerCount = 1;
  }

  // Do host depth storing for the depth destination (assuming there can be only
  // one depth destination) where depth destination == host depth source.
  bool host_depth_store_set_up = false;
  for (uint32_t i = 0; i < render_target_count; ++i) {
    RenderTarget* dest_rt = render_targets[i];
    if (!dest_rt) {
      continue;
    }
    auto& dest_vulkan_rt = *static_cast<VulkanRenderTarget*>(dest_rt);
    RenderTargetKey dest_rt_key = dest_vulkan_rt.key();
    if (!dest_rt_key.is_depth) {
      continue;
    }
    const std::vector<Transfer>& depth_transfers = render_target_transfers[i];
    for (const Transfer& transfer : depth_transfers) {
      if (transfer.host_depth_source != dest_rt) {
        continue;
      }
      if (!host_depth_store_set_up) {
        // Pipeline.
        command_processor_.BindExternalComputePipeline(
            host_depth_store_pipelines_[size_t(dest_rt_key.msaa_samples)]);
        // Descriptor set bindings.
        VkDescriptorSet host_depth_store_descriptor_sets[] = {
            edram_storage_buffer_descriptor_set_,
            dest_vulkan_rt.GetDescriptorSetTransferSource(),
        };
        command_buffer.CmdVkBindDescriptorSets(
            VK_PIPELINE_BIND_POINT_COMPUTE, host_depth_store_pipeline_layout_,
            0, uint32_t(xe::countof(host_depth_store_descriptor_sets)),
            host_depth_store_descriptor_sets, 0, nullptr);
        // Render target constant.
        HostDepthStoreRenderTargetConstant
            host_depth_store_render_target_constant =
                GetHostDepthStoreRenderTargetConstant(
                    dest_rt_key.pitch_tiles_at_32bpp,
                    msaa_2x_attachments_supported_);
        command_buffer.CmdVkPushConstants(
            host_depth_store_pipeline_layout_, VK_SHADER_STAGE_COMPUTE_BIT,
            uint32_t(offsetof(HostDepthStoreConstants, render_target)),
            sizeof(host_depth_store_render_target_constant),
            &host_depth_store_render_target_constant);
        // Barriers - don't need to try to combine them with the rest of
        // render target transfer barriers now - if this happens, after host
        // depth storing, SHADER_READ -> DEPTH_STENCIL_ATTACHMENT_WRITE will be
        // done anyway even in the best case, so it's not possible to have all
        // the barriers in one place here.
        UseEdramBuffer(EdramBufferUsage::kComputeWrite);
        // Always transitioning both depth and stencil, not storing separate
        // usage flags for depth and stencil.
        command_processor_.PushImageMemoryBarrier(
            dest_vulkan_rt.image(),
            ui::vulkan::util::InitializeSubresourceRange(
                VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT),
            dest_vulkan_rt.current_stage_mask(),
            VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
            dest_vulkan_rt.current_access_mask(), VK_ACCESS_SHADER_READ_BIT,
            dest_vulkan_rt.current_layout(),
            VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
        dest_vulkan_rt.SetUsage(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
                                VK_ACCESS_SHADER_READ_BIT,
                                VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
        host_depth_store_set_up = true;
      }
      Transfer::Rectangle
          transfer_rectangles[Transfer::kMaxRectanglesWithCutout];
      uint32_t transfer_rectangle_count = transfer.GetRectangles(
          dest_rt_key.base_tiles, dest_rt_key.pitch_tiles_at_32bpp,
          dest_rt_key.msaa_samples, false, transfer_rectangles,
          resolve_clear_rectangle);
      assert_not_zero(transfer_rectangle_count);
      HostDepthStoreRectangleConstant host_depth_store_rectangle_constant;
      for (uint32_t j = 0; j < transfer_rectangle_count; ++j) {
        uint32_t group_count_x, group_count_y;
        GetHostDepthStoreRectangleInfo(
            transfer_rectangles[j], dest_rt_key.msaa_samples,
            host_depth_store_rectangle_constant, group_count_x, group_count_y);
        command_buffer.CmdVkPushConstants(
            host_depth_store_pipeline_layout_, VK_SHADER_STAGE_COMPUTE_BIT,
            uint32_t(offsetof(HostDepthStoreConstants, rectangle)),
            sizeof(host_depth_store_rectangle_constant),
            &host_depth_store_rectangle_constant);
        command_processor_.SubmitBarriers(true);
        command_buffer.CmdVkDispatch(group_count_x, group_count_y, 1);
        MarkEdramBufferModified();
      }
    }
    break;
  }

  constexpr VkPipelineStageFlags kSourceStageMask =
      VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT;
  constexpr VkAccessFlags kSourceAccessMask = VK_ACCESS_SHADER_READ_BIT;
  constexpr VkImageLayout kSourceLayout =
      VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;

  // Try to insert as many barriers as possible in one place, hoping that in the
  // best case (no cross-copying between current render targets), barriers will
  // need to be only inserted here, not between transfers. In case of
  // cross-copying, if the destination use is going to happen before the source
  // use, choose the destination state, otherwise the source state - to match
  // the order in which transfers will actually happen (otherwise there will be
  // just a useless switch back and forth).
  for (uint32_t i = 0; i < render_target_count; ++i) {
    RenderTarget* dest_rt = render_targets[i];
    if (!dest_rt) {
      continue;
    }
    const std::vector<Transfer>& dest_transfers = render_target_transfers[i];
    if (!resolve_clear_needed && dest_transfers.empty()) {
      continue;
    }
    // Transition the destination, only if not going to be used as a source
    // earlier.
    bool dest_used_previously_as_source = false;
    for (uint32_t j = 0; j < i; ++j) {
      for (const Transfer& previous_transfer : render_target_transfers[j]) {
        if (previous_transfer.source == dest_rt ||
            previous_transfer.host_depth_source == dest_rt) {
          dest_used_previously_as_source = true;
          break;
        }
      }
    }
    if (!dest_used_previously_as_source) {
      auto& dest_vulkan_rt = *static_cast<VulkanRenderTarget*>(dest_rt);
      VkPipelineStageFlags dest_dst_stage_mask;
      VkAccessFlags dest_dst_access_mask;
      VkImageLayout dest_new_layout;
      dest_vulkan_rt.GetDrawUsage(&dest_dst_stage_mask, &dest_dst_access_mask,
                                  &dest_new_layout);
      command_processor_.PushImageMemoryBarrier(
          dest_vulkan_rt.image(),
          ui::vulkan::util::InitializeSubresourceRange(
              dest_vulkan_rt.key().is_depth
                  ? (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)
                  : VK_IMAGE_ASPECT_COLOR_BIT),
          dest_vulkan_rt.current_stage_mask(), dest_dst_stage_mask,
          dest_vulkan_rt.current_access_mask(), dest_dst_access_mask,
          dest_vulkan_rt.current_layout(), dest_new_layout);
      dest_vulkan_rt.SetUsage(dest_dst_stage_mask, dest_dst_access_mask,
                              dest_new_layout);
    }
    // Transition the sources, only if not going to be used as destinations
    // earlier.
    for (const Transfer& transfer : dest_transfers) {
      bool source_previously_used_as_dest = false;
      bool host_depth_source_previously_used_as_dest = false;
      for (uint32_t j = 0; j < i; ++j) {
        if (render_target_transfers[j].empty()) {
          continue;
        }
        const RenderTarget* previous_rt = render_targets[j];
        if (transfer.source == previous_rt) {
          source_previously_used_as_dest = true;
        }
        if (transfer.host_depth_source == previous_rt) {
          host_depth_source_previously_used_as_dest = true;
        }
      }
      if (!source_previously_used_as_dest) {
        auto& source_vulkan_rt =
            *static_cast<VulkanRenderTarget*>(transfer.source);
        command_processor_.PushImageMemoryBarrier(
            source_vulkan_rt.image(),
            ui::vulkan::util::InitializeSubresourceRange(
                source_vulkan_rt.key().is_depth
                    ? (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)
                    : VK_IMAGE_ASPECT_COLOR_BIT),
            source_vulkan_rt.current_stage_mask(), kSourceStageMask,
            source_vulkan_rt.current_access_mask(), kSourceAccessMask,
            source_vulkan_rt.current_layout(), kSourceLayout);
        source_vulkan_rt.SetUsage(kSourceStageMask, kSourceAccessMask,
                                  kSourceLayout);
      }
      // transfer.host_depth_source == dest_rt means the EDRAM buffer will be
      // used instead, no need to transition.
      if (transfer.host_depth_source && transfer.host_depth_source != dest_rt &&
          !host_depth_source_previously_used_as_dest) {
        auto& host_depth_source_vulkan_rt =
            *static_cast<VulkanRenderTarget*>(transfer.host_depth_source);
        command_processor_.PushImageMemoryBarrier(
            host_depth_source_vulkan_rt.image(),
            ui::vulkan::util::InitializeSubresourceRange(
                VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT),
            host_depth_source_vulkan_rt.current_stage_mask(), kSourceStageMask,
            host_depth_source_vulkan_rt.current_access_mask(),
            kSourceAccessMask, host_depth_source_vulkan_rt.current_layout(),
            kSourceLayout);
        host_depth_source_vulkan_rt.SetUsage(kSourceStageMask,
                                             kSourceAccessMask, kSourceLayout);
      }
    }
  }
  if (host_depth_store_set_up) {
    // Will be reading copied host depth from the EDRAM buffer.
    UseEdramBuffer(EdramBufferUsage::kFragmentRead);
  }

  // Perform the transfers and clears.

  TransferPipelineLayoutIndex last_transfer_pipeline_layout_index =
      TransferPipelineLayoutIndex::kCount;
  uint32_t transfer_descriptor_sets_bound = 0;
  uint32_t transfer_push_constants_set = 0;
  VkDescriptorSet last_descriptor_set_host_depth_stencil_textures =
      VK_NULL_HANDLE;
  VkDescriptorSet last_descriptor_set_depth_stencil_textures = VK_NULL_HANDLE;
  VkDescriptorSet last_descriptor_set_color_texture = VK_NULL_HANDLE;
  TransferAddressConstant last_host_depth_address_constant;
  TransferAddressConstant last_address_constant;

  for (uint32_t i = 0; i < render_target_count; ++i) {
    RenderTarget* dest_rt = render_targets[i];
    if (!dest_rt) {
      continue;
    }

    const std::vector<Transfer>& current_transfers = render_target_transfers[i];
    if (current_transfers.empty() && !resolve_clear_needed) {
      continue;
    }

    auto& dest_vulkan_rt = *static_cast<VulkanRenderTarget*>(dest_rt);
    RenderTargetKey dest_rt_key = dest_vulkan_rt.key();

    // Late barriers in case there was cross-copying that prevented merging of
    // barriers.
    {
      VkPipelineStageFlags dest_dst_stage_mask;
      VkAccessFlags dest_dst_access_mask;
      VkImageLayout dest_new_layout;
      dest_vulkan_rt.GetDrawUsage(&dest_dst_stage_mask, &dest_dst_access_mask,
                                  &dest_new_layout);
      command_processor_.PushImageMemoryBarrier(
          dest_vulkan_rt.image(),
          ui::vulkan::util::InitializeSubresourceRange(
              dest_rt_key.is_depth
                  ? (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)
                  : VK_IMAGE_ASPECT_COLOR_BIT),
          dest_vulkan_rt.current_stage_mask(), dest_dst_stage_mask,
          dest_vulkan_rt.current_access_mask(), dest_dst_access_mask,
          dest_vulkan_rt.current_layout(), dest_new_layout);
      dest_vulkan_rt.SetUsage(dest_dst_stage_mask, dest_dst_access_mask,
                              dest_new_layout);
    }

    // Get the objects needed for transfers to the destination.
    // TODO(Triang3l): Reuse the guest render pass for transfers where possible
    // (if the Vulkan format used for drawing is also usable for transfers - for
    // instance, R8G8B8A8_UNORM can be used for both, so the guest pass can be
    // reused, but R16G16B16A16_SFLOAT render targets use R16G16B16A16_UINT for
    // transfers, so the transfer pass has to be separate) to avoid stores and
    // loads on tile-based devices to make this actually applicable. Also
    // overall perform all non-cross-copying transfers for the current
    // framebuffer configuration in a single pass, to load / store only once.
    RenderPassKey transfer_render_pass_key;
    transfer_render_pass_key.msaa_samples = dest_rt_key.msaa_samples;
    if (dest_rt_key.is_depth) {
      transfer_render_pass_key.depth_and_color_used = 0b1;
      transfer_render_pass_key.depth_format = dest_rt_key.GetDepthFormat();
    } else {
      transfer_render_pass_key.depth_and_color_used = 0b1 << 1;
      transfer_render_pass_key.color_0_view_format =
          dest_rt_key.GetColorFormat();
      transfer_render_pass_key.color_rts_use_transfer_formats = 1;
    }
    VkRenderPass transfer_render_pass =
        GetHostRenderTargetsRenderPass(transfer_render_pass_key);
    if (transfer_render_pass == VK_NULL_HANDLE) {
      continue;
    }
    const RenderTarget*
        transfer_framebuffer_render_targets[1 + xenos::kMaxColorRenderTargets] =
            {};
    transfer_framebuffer_render_targets[dest_rt_key.is_depth ? 0 : 1] = dest_rt;
    const Framebuffer* transfer_framebuffer = GetHostRenderTargetsFramebuffer(
        transfer_render_pass_key, dest_rt_key.pitch_tiles_at_32bpp,
        transfer_framebuffer_render_targets);
    if (!transfer_framebuffer) {
      continue;
    }
    // Don't enter the render pass immediately - may still insert source
    // barriers later.

    if (!current_transfers.empty()) {
      uint32_t dest_pitch_tiles = dest_rt_key.GetPitchTiles();
      bool dest_is_64bpp = dest_rt_key.Is64bpp();

      // Gather shader keys and sort to reduce pipeline state and binding
      // switches. Also gather stencil rectangles to clear if needed.
      bool need_stencil_bit_draws =
          dest_rt_key.is_depth &&
          !vulkan_device->extensions().ext_EXT_shader_stencil_export;
      current_transfer_invocations_.clear();
      current_transfer_invocations_.reserve(
          current_transfers.size() << uint32_t(need_stencil_bit_draws));
      uint32_t rt_sort_index = 0;
      TransferShaderKey new_transfer_shader_key;
      new_transfer_shader_key.dest_msaa_samples = dest_rt_key.msaa_samples;
      new_transfer_shader_key.dest_resource_format =
          dest_rt_key.resource_format;
      uint32_t stencil_clear_rectangle_count = 0;
      for (uint32_t j = 0; j <= uint32_t(need_stencil_bit_draws); ++j) {
        // j == 0 - color or depth.
        // j == 1 - stencil bits.
        // Stencil bit writing always requires a different root signature,
        // handle these separately. Stencil never has a host depth source.
        // Clear previously set sort indices.
        for (const Transfer& transfer : current_transfers) {
          auto host_depth_source_vulkan_rt =
              static_cast<VulkanRenderTarget*>(transfer.host_depth_source);
          if (host_depth_source_vulkan_rt) {
            host_depth_source_vulkan_rt->SetTemporarySortIndex(UINT32_MAX);
          }
          assert_not_null(transfer.source);
          auto& source_vulkan_rt =
              *static_cast<VulkanRenderTarget*>(transfer.source);
          source_vulkan_rt.SetTemporarySortIndex(UINT32_MAX);
        }
        for (const Transfer& transfer : current_transfers) {
          assert_not_null(transfer.source);
          auto& source_vulkan_rt =
              *static_cast<VulkanRenderTarget*>(transfer.source);
          VulkanRenderTarget* host_depth_source_vulkan_rt =
              j ? nullptr
                : static_cast<VulkanRenderTarget*>(transfer.host_depth_source);
          if (host_depth_source_vulkan_rt &&
              host_depth_source_vulkan_rt->temporary_sort_index() ==
                  UINT32_MAX) {
            host_depth_source_vulkan_rt->SetTemporarySortIndex(rt_sort_index++);
          }
          if (source_vulkan_rt.temporary_sort_index() == UINT32_MAX) {
            source_vulkan_rt.SetTemporarySortIndex(rt_sort_index++);
          }
          RenderTargetKey source_rt_key = source_vulkan_rt.key();
          new_transfer_shader_key.source_msaa_samples =
              source_rt_key.msaa_samples;
          new_transfer_shader_key.source_resource_format =
              source_rt_key.resource_format;
          bool host_depth_source_is_copy =
              host_depth_source_vulkan_rt == &dest_vulkan_rt;
          // The host depth copy buffer has only raw samples.
          new_transfer_shader_key.host_depth_source_msaa_samples =
              (host_depth_source_vulkan_rt && !host_depth_source_is_copy)
                  ? host_depth_source_vulkan_rt->key().msaa_samples
                  : xenos::MsaaSamples::k1X;
          if (j) {
            new_transfer_shader_key.mode =
                source_rt_key.is_depth ? TransferMode::kDepthToStencilBit
                                       : TransferMode::kColorToStencilBit;
            stencil_clear_rectangle_count +=
                transfer.GetRectangles(dest_rt_key.base_tiles, dest_pitch_tiles,
                                       dest_rt_key.msaa_samples, dest_is_64bpp,
                                       nullptr, resolve_clear_rectangle);
          } else {
            if (dest_rt_key.is_depth) {
              if (host_depth_source_vulkan_rt) {
                if (host_depth_source_is_copy) {
                  new_transfer_shader_key.mode =
                      source_rt_key.is_depth
                          ? TransferMode::kDepthAndHostDepthCopyToDepth
                          : TransferMode::kColorAndHostDepthCopyToDepth;
                } else {
                  new_transfer_shader_key.mode =
                      source_rt_key.is_depth
                          ? TransferMode::kDepthAndHostDepthToDepth
                          : TransferMode::kColorAndHostDepthToDepth;
                }
              } else {
                new_transfer_shader_key.mode =
                    source_rt_key.is_depth ? TransferMode::kDepthToDepth
                                           : TransferMode::kColorToDepth;
              }
            } else {
              new_transfer_shader_key.mode = source_rt_key.is_depth
                                                 ? TransferMode::kDepthToColor
                                                 : TransferMode::kColorToColor;
            }
          }
          current_transfer_invocations_.emplace_back(transfer,
                                                     new_transfer_shader_key);
          if (j) {
            current_transfer_invocations_.back().transfer.host_depth_source =
                nullptr;
          }
        }
      }
      std::sort(current_transfer_invocations_.begin(),
                current_transfer_invocations_.end());

      for (auto it = current_transfer_invocations_.cbegin();
           it != current_transfer_invocations_.cend(); ++it) {
        assert_not_null(it->transfer.source);
        auto& source_vulkan_rt =
            *static_cast<VulkanRenderTarget*>(it->transfer.source);
        command_processor_.PushImageMemoryBarrier(
            source_vulkan_rt.image(),
            ui::vulkan::util::InitializeSubresourceRange(
                source_vulkan_rt.key().is_depth
                    ? (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)
                    : VK_IMAGE_ASPECT_COLOR_BIT),
            source_vulkan_rt.current_stage_mask(), kSourceStageMask,
            source_vulkan_rt.current_access_mask(), kSourceAccessMask,
            source_vulkan_rt.current_layout(), kSourceLayout);
        source_vulkan_rt.SetUsage(kSourceStageMask, kSourceAccessMask,
                                  kSourceLayout);
        auto host_depth_source_vulkan_rt =
            static_cast<VulkanRenderTarget*>(it->transfer.host_depth_source);
        if (host_depth_source_vulkan_rt) {
          TransferShaderKey transfer_shader_key = it->shader_key;
          if (transfer_shader_key.mode ==
                  TransferMode::kDepthAndHostDepthCopyToDepth ||
              transfer_shader_key.mode ==
                  TransferMode::kColorAndHostDepthCopyToDepth) {
            // Reading copied host depth from the EDRAM buffer.
            UseEdramBuffer(EdramBufferUsage::kFragmentRead);
          } else {
            // Reading host depth from the texture.
            command_processor_.PushImageMemoryBarrier(
                host_depth_source_vulkan_rt->image(),
                ui::vulkan::util::InitializeSubresourceRange(
                    VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT),
                host_depth_source_vulkan_rt->current_stage_mask(),
                kSourceStageMask,
                host_depth_source_vulkan_rt->current_access_mask(),
                kSourceAccessMask,
                host_depth_source_vulkan_rt->current_layout(), kSourceLayout);
            host_depth_source_vulkan_rt->SetUsage(
                kSourceStageMask, kSourceAccessMask, kSourceLayout);
          }
        }
      }

      // Perform the transfers for the render target.

      command_processor_.SubmitBarriersAndEnterRenderTargetCacheRenderPass(
          transfer_render_pass, transfer_framebuffer);

      if (stencil_clear_rectangle_count) {
        VkClearAttachment* stencil_clear_attachment;
        VkClearRect* stencil_clear_rect_write_ptr;
        command_buffer.CmdClearAttachmentsEmplace(1, stencil_clear_attachment,
                                                  stencil_clear_rectangle_count,
                                                  stencil_clear_rect_write_ptr);
        stencil_clear_attachment->aspectMask = VK_IMAGE_ASPECT_STENCIL_BIT;
        stencil_clear_attachment->colorAttachment = 0;
        stencil_clear_attachment->clearValue.depthStencil.depth = 0.0f;
        stencil_clear_attachment->clearValue.depthStencil.stencil = 0;
        for (const Transfer& transfer : current_transfers) {
          Transfer::Rectangle transfer_stencil_clear_rectangles
              [Transfer::kMaxRectanglesWithCutout];
          uint32_t transfer_stencil_clear_rectangle_count =
              transfer.GetRectangles(dest_rt_key.base_tiles, dest_pitch_tiles,
                                     dest_rt_key.msaa_samples, dest_is_64bpp,
                                     transfer_stencil_clear_rectangles,
                                     resolve_clear_rectangle);
          for (uint32_t j = 0; j < transfer_stencil_clear_rectangle_count;
               ++j) {
            const Transfer::Rectangle& stencil_clear_rectangle =
                transfer_stencil_clear_rectangles[j];
            stencil_clear_rect_write_ptr->rect.offset.x = int32_t(
                stencil_clear_rectangle.x_pixels * draw_resolution_scale_x());
            stencil_clear_rect_write_ptr->rect.offset.y = int32_t(
                stencil_clear_rectangle.y_pixels * draw_resolution_scale_y());
            stencil_clear_rect_write_ptr->rect.extent.width =
                stencil_clear_rectangle.width_pixels *
                draw_resolution_scale_x();
            stencil_clear_rect_write_ptr->rect.extent.height =
                stencil_clear_rectangle.height_pixels *
                draw_resolution_scale_y();
            stencil_clear_rect_write_ptr->baseArrayLayer = 0;
            stencil_clear_rect_write_ptr->layerCount = 1;
            ++stencil_clear_rect_write_ptr;
          }
        }
      }

      // Prefer power of two viewports for exact division by simply biasing the
      // exponent.
      VkViewport transfer_viewport;
      transfer_viewport.x = 0.0f;
      transfer_viewport.y = 0.0f;
      transfer_viewport.width =
          float(std::min(xe::next_pow2(transfer_framebuffer->host_extent.width),
                         vulkan_device->properties().maxViewportDimensions[0]));
      transfer_viewport.height = float(
          std::min(xe::next_pow2(transfer_framebuffer->host_extent.height),
                   vulkan_device->properties().maxViewportDimensions[1]));
      transfer_viewport.minDepth = 0.0f;
      transfer_viewport.maxDepth = 1.0f;
      command_processor_.SetViewport(transfer_viewport);
      // GetRectangles returns coordinates in guest pixels, so scale
      // pixels_to_ndc to convert guest pixels to NDC correctly.
      float pixels_to_ndc_x =
          2.0f / transfer_viewport.width * draw_resolution_scale_x();
      float pixels_to_ndc_y =
          2.0f / transfer_viewport.height * draw_resolution_scale_y();
      VkRect2D transfer_scissor;
      transfer_scissor.offset.x = 0;
      transfer_scissor.offset.y = 0;
      transfer_scissor.extent = transfer_framebuffer->host_extent;
      command_processor_.SetScissor(transfer_scissor);

      for (auto it = current_transfer_invocations_.cbegin();
           it != current_transfer_invocations_.cend(); ++it) {
        const TransferInvocation& transfer_invocation_first = *it;
        // Will be merging transfers from the same source into one mesh.
        auto it_merged_first = it, it_merged_last = it;
        uint32_t transfer_rectangle_count =
            transfer_invocation_first.transfer.GetRectangles(
                dest_rt_key.base_tiles, dest_pitch_tiles,
                dest_rt_key.msaa_samples, dest_is_64bpp, nullptr,
                resolve_clear_rectangle);
        for (auto it_merge = std::next(it_merged_first);
             it_merge != current_transfer_invocations_.cend(); ++it_merge) {
          if (!transfer_invocation_first.CanBeMergedIntoOneDraw(*it_merge)) {
            break;
          }
          transfer_rectangle_count += it_merge->transfer.GetRectangles(
              dest_rt_key.base_tiles, dest_pitch_tiles,
              dest_rt_key.msaa_samples, dest_is_64bpp, nullptr,
              resolve_clear_rectangle);
          it_merged_last = it_merge;
        }
        assert_not_zero(transfer_rectangle_count);
        // Skip the merged transfers in the subsequent iterations.
        it = it_merged_last;

        assert_not_null(it->transfer.source);
        auto& source_vulkan_rt =
            *static_cast<VulkanRenderTarget*>(it->transfer.source);
        auto host_depth_source_vulkan_rt =
            static_cast<VulkanRenderTarget*>(it->transfer.host_depth_source);
        TransferShaderKey transfer_shader_key = it->shader_key;
        const TransferModeInfo& transfer_mode_info =
            kTransferModes[size_t(transfer_shader_key.mode)];
        TransferPipelineLayoutIndex transfer_pipeline_layout_index =
            transfer_mode_info.pipeline_layout;
        const TransferPipelineLayoutInfo& transfer_pipeline_layout_info =
            kTransferPipelineLayoutInfos[size_t(
                transfer_pipeline_layout_index)];
        uint32_t transfer_sample_pipeline_count =
            vulkan_device->properties().sampleRateShading
                ? 1
                : uint32_t(1) << uint32_t(dest_rt_key.msaa_samples);
        bool transfer_is_stencil_bit =
            (transfer_pipeline_layout_info.used_push_constant_dwords &
             kTransferUsedPushConstantDwordStencilMaskBit) != 0;

        uint32_t transfer_vertex_count = 6 * transfer_rectangle_count;
        VkBuffer transfer_vertex_buffer;
        VkDeviceSize transfer_vertex_buffer_offset;
        float* transfer_rectangle_write_ptr =
            reinterpret_cast<float*>(transfer_vertex_buffer_pool_->Request(
                current_submission, sizeof(float) * 2 * transfer_vertex_count,
                sizeof(float), transfer_vertex_buffer,
                transfer_vertex_buffer_offset));
        if (!transfer_rectangle_write_ptr) {
          continue;
        }
        for (auto it_merged = it_merged_first; it_merged <= it_merged_last;
             ++it_merged) {
          Transfer::Rectangle transfer_invocation_rectangles
              [Transfer::kMaxRectanglesWithCutout];
          uint32_t transfer_invocation_rectangle_count =
              it_merged->transfer.GetRectangles(
                  dest_rt_key.base_tiles, dest_pitch_tiles,
                  dest_rt_key.msaa_samples, dest_is_64bpp,
                  transfer_invocation_rectangles, resolve_clear_rectangle);
          assert_not_zero(transfer_invocation_rectangle_count);
          for (uint32_t j = 0; j < transfer_invocation_rectangle_count; ++j) {
            const Transfer::Rectangle& transfer_rectangle =
                transfer_invocation_rectangles[j];
            float transfer_rectangle_x0 =
                -1.0f + transfer_rectangle.x_pixels * pixels_to_ndc_x;
            float transfer_rectangle_y0 =
                -1.0f + transfer_rectangle.y_pixels * pixels_to_ndc_y;
            float transfer_rectangle_x1 =
                transfer_rectangle_x0 +
                transfer_rectangle.width_pixels * pixels_to_ndc_x;
            float transfer_rectangle_y1 =
                transfer_rectangle_y0 +
                transfer_rectangle.height_pixels * pixels_to_ndc_y;
            // O-*
            // |/
            // *
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_x0;
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_y0;
            // *-*
            // |/
            // O
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_x0;
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_y1;
            // *-O
            // |/
            // *
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_x1;
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_y0;
            //   O
            //  /|
            // *-*
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_x1;
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_y0;
            //   *
            //  /|
            // O-*
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_x0;
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_y1;
            //   *
            //  /|
            // *-O
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_x1;
            *(transfer_rectangle_write_ptr++) = transfer_rectangle_y1;
          }
        }
        command_buffer.CmdVkBindVertexBuffers(0, 1, &transfer_vertex_buffer,
                                              &transfer_vertex_buffer_offset);

        const VkPipeline* transfer_pipelines = GetTransferPipelines(
            TransferPipelineKey(transfer_render_pass_key, transfer_shader_key));
        if (!transfer_pipelines) {
          continue;
        }
        command_processor_.BindExternalGraphicsPipeline(transfer_pipelines[0]);
        if (last_transfer_pipeline_layout_index !=
            transfer_pipeline_layout_index) {
          last_transfer_pipeline_layout_index = transfer_pipeline_layout_index;
          transfer_descriptor_sets_bound = 0;
          transfer_push_constants_set = 0;
        }

        // Invalidate outdated bindings.
        if (transfer_pipeline_layout_info.used_descriptor_sets &
            kTransferUsedDescriptorSetHostDepthStencilTexturesBit) {
          assert_not_null(host_depth_source_vulkan_rt);
          VkDescriptorSet descriptor_set_host_depth_stencil_textures =
              host_depth_source_vulkan_rt->GetDescriptorSetTransferSource();
          if (last_descriptor_set_host_depth_stencil_textures !=
              descriptor_set_host_depth_stencil_textures) {
            last_descriptor_set_host_depth_stencil_textures =
                descriptor_set_host_depth_stencil_textures;
            transfer_descriptor_sets_bound &=
                ~kTransferUsedDescriptorSetHostDepthStencilTexturesBit;
          }
        }
        if (transfer_pipeline_layout_info.used_descriptor_sets &
            kTransferUsedDescriptorSetDepthStencilTexturesBit) {
          VkDescriptorSet descriptor_set_depth_stencil_textures =
              source_vulkan_rt.GetDescriptorSetTransferSource();
          if (last_descriptor_set_depth_stencil_textures !=
              descriptor_set_depth_stencil_textures) {
            last_descriptor_set_depth_stencil_textures =
                descriptor_set_depth_stencil_textures;
            transfer_descriptor_sets_bound &=
                ~kTransferUsedDescriptorSetDepthStencilTexturesBit;
          }
        }
        if (transfer_pipeline_layout_info.used_descriptor_sets &
            kTransferUsedDescriptorSetColorTextureBit) {
          VkDescriptorSet descriptor_set_color_texture =
              source_vulkan_rt.GetDescriptorSetTransferSource();
          if (last_descriptor_set_color_texture !=
              descriptor_set_color_texture) {
            last_descriptor_set_color_texture = descriptor_set_color_texture;
            transfer_descriptor_sets_bound &=
                ~kTransferUsedDescriptorSetColorTextureBit;
          }
        }
        if (transfer_pipeline_layout_info.used_push_constant_dwords &
            kTransferUsedPushConstantDwordHostDepthAddressBit) {
          assert_not_null(host_depth_source_vulkan_rt);
          RenderTargetKey host_depth_source_rt_key =
              host_depth_source_vulkan_rt->key();
          TransferAddressConstant host_depth_address_constant;
          host_depth_address_constant.dest_pitch = dest_pitch_tiles;
          host_depth_address_constant.source_pitch =
              host_depth_source_rt_key.GetPitchTiles();
          host_depth_address_constant.source_to_dest =
              int32_t(dest_rt_key.base_tiles) -
              int32_t(host_depth_source_rt_key.base_tiles);
          if (last_host_depth_address_constant != host_depth_address_constant) {
            last_host_depth_address_constant = host_depth_address_constant;
            transfer_push_constants_set &=
                ~kTransferUsedPushConstantDwordHostDepthAddressBit;
          }
        }
        if (transfer_pipeline_layout_info.used_push_constant_dwords &
            kTransferUsedPushConstantDwordAddressBit) {
          RenderTargetKey source_rt_key = source_vulkan_rt.key();
          TransferAddressConstant address_constant;
          address_constant.dest_pitch = dest_pitch_tiles;
          address_constant.source_pitch = source_rt_key.GetPitchTiles();
          address_constant.source_to_dest = int32_t(dest_rt_key.base_tiles) -
                                            int32_t(source_rt_key.base_tiles);
          if (last_address_constant != address_constant) {
            last_address_constant = address_constant;
            transfer_push_constants_set &=
                ~kTransferUsedPushConstantDwordAddressBit;
          }
        }

        // Apply the new bindings.
        // TODO(Triang3l): Merge binding updates into spans.
        VkPipelineLayout transfer_pipeline_layout =
            transfer_pipeline_layouts_[size_t(transfer_pipeline_layout_index)];
        uint32_t transfer_descriptor_sets_unbound =
            transfer_pipeline_layout_info.used_descriptor_sets &
            ~transfer_descriptor_sets_bound;
        if (transfer_descriptor_sets_unbound &
            kTransferUsedDescriptorSetHostDepthBufferBit) {
          command_buffer.CmdVkBindDescriptorSets(
              VK_PIPELINE_BIND_POINT_GRAPHICS, transfer_pipeline_layout,
              xe::bit_count(transfer_pipeline_layout_info.used_descriptor_sets &
                            (kTransferUsedDescriptorSetHostDepthBufferBit - 1)),
              1, &edram_storage_buffer_descriptor_set_, 0, nullptr);
          transfer_descriptor_sets_bound |=
              kTransferUsedDescriptorSetHostDepthBufferBit;
        }
        if (transfer_descriptor_sets_unbound &
            kTransferUsedDescriptorSetHostDepthStencilTexturesBit) {
          command_buffer.CmdVkBindDescriptorSets(
              VK_PIPELINE_BIND_POINT_GRAPHICS, transfer_pipeline_layout,
              xe::bit_count(
                  transfer_pipeline_layout_info.used_descriptor_sets &
                  (kTransferUsedDescriptorSetHostDepthStencilTexturesBit - 1)),
              1, &last_descriptor_set_host_depth_stencil_textures, 0, nullptr);
          transfer_descriptor_sets_bound |=
              kTransferUsedDescriptorSetHostDepthStencilTexturesBit;
        }
        if (transfer_descriptor_sets_unbound &
            kTransferUsedDescriptorSetDepthStencilTexturesBit) {
          command_buffer.CmdVkBindDescriptorSets(
              VK_PIPELINE_BIND_POINT_GRAPHICS, transfer_pipeline_layout,
              xe::bit_count(
                  transfer_pipeline_layout_info.used_descriptor_sets &
                  (kTransferUsedDescriptorSetDepthStencilTexturesBit - 1)),
              1, &last_descriptor_set_depth_stencil_textures, 0, nullptr);
          transfer_descriptor_sets_bound |=
              kTransferUsedDescriptorSetDepthStencilTexturesBit;
        }
        if (transfer_descriptor_sets_unbound &
            kTransferUsedDescriptorSetColorTextureBit) {
          command_buffer.CmdVkBindDescriptorSets(
              VK_PIPELINE_BIND_POINT_GRAPHICS, transfer_pipeline_layout,
              xe::bit_count(transfer_pipeline_layout_info.used_descriptor_sets &
                            (kTransferUsedDescriptorSetColorTextureBit - 1)),
              1, &last_descriptor_set_color_texture, 0, nullptr);
          transfer_descriptor_sets_bound |=
              kTransferUsedDescriptorSetColorTextureBit;
        }
        uint32_t transfer_push_constants_unset =
            transfer_pipeline_layout_info.used_push_constant_dwords &
            ~transfer_push_constants_set;
        if (transfer_push_constants_unset &
            kTransferUsedPushConstantDwordHostDepthAddressBit) {
          command_buffer.CmdVkPushConstants(
              transfer_pipeline_layout, VK_SHADER_STAGE_FRAGMENT_BIT,
              sizeof(uint32_t) *
                  xe::bit_count(
                      transfer_pipeline_layout_info.used_push_constant_dwords &
                      (kTransferUsedPushConstantDwordHostDepthAddressBit - 1)),
              sizeof(uint32_t), &last_host_depth_address_constant);
          transfer_push_constants_set |=
              kTransferUsedPushConstantDwordHostDepthAddressBit;
        }
        if (transfer_push_constants_unset &
            kTransferUsedPushConstantDwordAddressBit) {
          command_buffer.CmdVkPushConstants(
              transfer_pipeline_layout, VK_SHADER_STAGE_FRAGMENT_BIT,
              sizeof(uint32_t) *
                  xe::bit_count(
                      transfer_pipeline_layout_info.used_push_constant_dwords &
                      (kTransferUsedPushConstantDwordAddressBit - 1)),
              sizeof(uint32_t), &last_address_constant);
          transfer_push_constants_set |=
              kTransferUsedPushConstantDwordAddressBit;
        }

        for (uint32_t j = 0; j < transfer_sample_pipeline_count; ++j) {
          if (j) {
            command_processor_.BindExternalGraphicsPipeline(
                transfer_pipelines[j]);
          }
          for (uint32_t k = 0; k < uint32_t(transfer_is_stencil_bit ? 8 : 1);
               ++k) {
            if (transfer_is_stencil_bit) {
              uint32_t transfer_stencil_bit = uint32_t(1) << k;
              command_buffer.CmdVkPushConstants(
                  transfer_pipeline_layout, VK_SHADER_STAGE_FRAGMENT_BIT,
                  sizeof(uint32_t) *
                      xe::bit_count(
                          transfer_pipeline_layout_info
                              .used_push_constant_dwords &
                          (kTransferUsedPushConstantDwordStencilMaskBit - 1)),
                  sizeof(uint32_t), &transfer_stencil_bit);
              command_buffer.CmdVkSetStencilWriteMask(
                  VK_STENCIL_FACE_FRONT_AND_BACK, transfer_stencil_bit);
            }
            command_buffer.CmdVkDraw(transfer_vertex_count, 1, 0, 0);
          }
        }
      }
    }

    // Perform the clear.
    if (resolve_clear_needed) {
      command_processor_.SubmitBarriersAndEnterRenderTargetCacheRenderPass(
          transfer_render_pass, transfer_framebuffer);
      VkClearAttachment resolve_clear_attachment;
      resolve_clear_attachment.colorAttachment = 0;
      std::memset(&resolve_clear_attachment.clearValue, 0,
                  sizeof(resolve_clear_attachment.clearValue));
      uint64_t clear_value = render_target_resolve_clear_values[i];
      if (dest_rt_key.is_depth) {
        resolve_clear_attachment.aspectMask =
            VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT;
        uint32_t depth_guest_clear_value =
            (uint32_t(clear_value) >> 8) & 0xFFFFFF;
        switch (dest_rt_key.GetDepthFormat()) {
          case xenos::DepthRenderTargetFormat::kD24S8:
            resolve_clear_attachment.clearValue.depthStencil.depth =
                xenos::UNorm24To32(depth_guest_clear_value);
            break;
          case xenos::DepthRenderTargetFormat::kD24FS8:
            // Taking [0, 2) -> [0, 1) remapping into account.
            resolve_clear_attachment.clearValue.depthStencil.depth =
                xenos::Float20e4To32(depth_guest_clear_value) * 0.5f;
            break;
        }
        resolve_clear_attachment.clearValue.depthStencil.stencil =
            uint32_t(clear_value) & 0xFF;
      } else {
        resolve_clear_attachment.aspectMask = VK_IMAGE_ASPECT_COLOR_BIT;
        switch (dest_rt_key.GetColorFormat()) {
          case xenos::ColorRenderTargetFormat::k_8_8_8_8:
          case xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA: {
            for (uint32_t j = 0; j < 4; ++j) {
              resolve_clear_attachment.clearValue.color.float32[j] =
                  ((clear_value >> (j * 8)) & 0xFF) * (1.0f / 0xFF);
            }
          } break;
          case xenos::ColorRenderTargetFormat::k_2_10_10_10:
          case xenos::ColorRenderTargetFormat::k_2_10_10_10_AS_10_10_10_10: {
            for (uint32_t j = 0; j < 3; ++j) {
              resolve_clear_attachment.clearValue.color.float32[j] =
                  ((clear_value >> (j * 10)) & 0x3FF) * (1.0f / 0x3FF);
            }
            resolve_clear_attachment.clearValue.color.float32[3] =
                ((clear_value >> 30) & 0x3) * (1.0f / 0x3);
          } break;
          case xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT:
          case xenos::ColorRenderTargetFormat::
              k_2_10_10_10_FLOAT_AS_16_16_16_16: {
            for (uint32_t j = 0; j < 3; ++j) {
              resolve_clear_attachment.clearValue.color.float32[j] =
                  xenos::Float7e3To32((clear_value >> (j * 10)) & 0x3FF);
            }
            resolve_clear_attachment.clearValue.color.float32[3] =
                ((clear_value >> 30) & 0x3) * (1.0f / 0x3);
          } break;
          case xenos::ColorRenderTargetFormat::k_16_16:
          case xenos::ColorRenderTargetFormat::k_16_16_FLOAT: {
            // Using uint for transfers and clears of both. Disregarding the
            // current -32...32 vs. -1...1 settings for consistency with color
            // clear via depth aliasing.
            // TODO(Triang3l): Handle cases of unsupported multisampled 16_UINT
            // and completely unsupported 16_UNORM.
            for (uint32_t j = 0; j < 2; ++j) {
              resolve_clear_attachment.clearValue.color.uint32[j] =
                  uint32_t(clear_value >> (j * 16)) & 0xFFFF;
            }
          } break;
          case xenos::ColorRenderTargetFormat::k_16_16_16_16:
          case xenos::ColorRenderTargetFormat::k_16_16_16_16_FLOAT: {
            // Using uint for transfers and clears of both. Disregarding the
            // current -32...32 vs. -1...1 settings for consistency with color
            // clear via depth aliasing.
            // TODO(Triang3l): Handle cases of unsupported multisampled 16_UINT
            // and completely unsupported 16_UNORM.
            for (uint32_t j = 0; j < 4; ++j) {
              resolve_clear_attachment.clearValue.color.uint32[j] =
                  uint32_t(clear_value >> (j * 16)) & 0xFFFF;
            }
          } break;
          case xenos::ColorRenderTargetFormat::k_32_FLOAT: {
            // Using uint for proper denormal and NaN handling.
            resolve_clear_attachment.clearValue.color.uint32[0] =
                uint32_t(clear_value);
          } break;
          case xenos::ColorRenderTargetFormat::k_32_32_FLOAT: {
            // Using uint for proper denormal and NaN handling.
            resolve_clear_attachment.clearValue.color.uint32[0] =
                uint32_t(clear_value);
            resolve_clear_attachment.clearValue.color.uint32[1] =
                uint32_t(clear_value >> 32);
          } break;
        }
      }
      command_buffer.CmdVkClearAttachments(1, &resolve_clear_attachment, 1,
                                           &resolve_clear_rect);
    }
  }
}

VkPipeline VulkanRenderTargetCache::GetDumpPipeline(DumpPipelineKey key) {
  auto pipeline_it = dump_pipelines_.find(key);
  if (pipeline_it != dump_pipelines_.end()) {
    return pipeline_it->second;
  }

  std::vector<spv::Id> id_vector_temp;

  SpirvBuilder builder(spv::Spv_1_0,
                       (SpirvShaderTranslator::kSpirvMagicToolId << 16) | 1,
                       nullptr);
  spv::Id ext_inst_glsl_std_450 = builder.import("GLSL.std.450");
  builder.addCapability(spv::CapabilityShader);
  builder.setMemoryModel(spv::AddressingModelLogical, spv::MemoryModelGLSL450);
  builder.setSource(spv::SourceLanguageUnknown, 0);

  spv::Id type_void = builder.makeVoidType();
  spv::Id type_int = builder.makeIntType(32);
  spv::Id type_int2 = builder.makeVectorType(type_int, 2);
  spv::Id type_uint = builder.makeUintType(32);
  spv::Id type_uint2 = builder.makeVectorType(type_uint, 2);
  spv::Id type_uint3 = builder.makeVectorType(type_uint, 3);
  spv::Id type_float = builder.makeFloatType(32);

  // Bindings.
  // EDRAM buffer.
  bool format_is_64bpp = !key.is_depth && xenos::IsColorRenderTargetFormat64bpp(
                                              key.GetColorFormat());
  id_vector_temp.clear();
  id_vector_temp.push_back(
      builder.makeRuntimeArray(format_is_64bpp ? type_uint2 : type_uint));
  // Storage buffers have std430 packing, no padding to 4-component vectors.
  builder.addDecoration(id_vector_temp.back(), spv::DecorationArrayStride,
                        sizeof(uint32_t) << uint32_t(format_is_64bpp));
  spv::Id type_edram = builder.makeStructType(id_vector_temp, "XeEdram");
  builder.addMemberName(type_edram, 0, "edram");
  builder.addMemberDecoration(type_edram, 0, spv::DecorationNonReadable);
  builder.addMemberDecoration(type_edram, 0, spv::DecorationOffset, 0);
  // Block since SPIR-V 1.3, but since SPIR-V 1.0 is generated, it's
  // BufferBlock.
  builder.addDecoration(type_edram, spv::DecorationBufferBlock);
  // StorageBuffer since SPIR-V 1.3, but since SPIR-V 1.0 is generated, it's
  // Uniform.
  spv::Id edram_buffer = builder.createVariable(
      spv::NoPrecision, spv::StorageClassUniform, type_edram, "xe_edram");
  builder.addDecoration(edram_buffer, spv::DecorationDescriptorSet,
                        kDumpDescriptorSetEdram);
  builder.addDecoration(edram_buffer, spv::DecorationBinding, 0);
  // Color or depth source.
  bool source_is_multisampled = key.msaa_samples != xenos::MsaaSamples::k1X;
  bool source_is_uint;
  if (key.is_depth) {
    source_is_uint = false;
  } else {
    GetColorOwnershipTransferVulkanFormat(key.GetColorFormat(),
                                          &source_is_uint);
  }
  spv::Id source_component_type = source_is_uint ? type_uint : type_float;
  spv::Id source_texture = builder.createVariable(
      spv::NoPrecision, spv::StorageClassUniformConstant,
      builder.makeImageType(source_component_type, spv::Dim2D, false, false,
                            source_is_multisampled, 1, spv::ImageFormatUnknown),
      "xe_edram_dump_source");
  builder.addDecoration(source_texture, spv::DecorationDescriptorSet,
                        kDumpDescriptorSetSource);
  builder.addDecoration(source_texture, spv::DecorationBinding, 0);
  // Stencil source.
  spv::Id source_stencil_texture = spv::NoResult;
  if (key.is_depth) {
    source_stencil_texture = builder.createVariable(
        spv::NoPrecision, spv::StorageClassUniformConstant,
        builder.makeImageType(type_uint, spv::Dim2D, false, false,
                              source_is_multisampled, 1,
                              spv::ImageFormatUnknown),
        "xe_edram_dump_stencil");
    builder.addDecoration(source_stencil_texture, spv::DecorationDescriptorSet,
                          kDumpDescriptorSetSource);
    builder.addDecoration(source_stencil_texture, spv::DecorationBinding, 1);
  }
  // Push constants.
  id_vector_temp.clear();
  id_vector_temp.reserve(kDumpPushConstantCount);
  for (uint32_t i = 0; i < kDumpPushConstantCount; ++i) {
    id_vector_temp.push_back(type_uint);
  }
  spv::Id type_push_constants =
      builder.makeStructType(id_vector_temp, "XeEdramDumpPushConstants");
  builder.addMemberName(type_push_constants, kDumpPushConstantPitches,
                        "pitches");
  builder.addMemberDecoration(type_push_constants, kDumpPushConstantPitches,
                              spv::DecorationOffset,
                              int(sizeof(uint32_t) * kDumpPushConstantPitches));
  builder.addMemberName(type_push_constants, kDumpPushConstantOffsets,
                        "offsets");
  builder.addMemberDecoration(type_push_constants, kDumpPushConstantOffsets,
                              spv::DecorationOffset,
                              int(sizeof(uint32_t) * kDumpPushConstantOffsets));
  builder.addDecoration(type_push_constants, spv::DecorationBlock);
  spv::Id push_constants = builder.createVariable(
      spv::NoPrecision, spv::StorageClassPushConstant, type_push_constants,
      "xe_edram_dump_push_constants");

  // gl_GlobalInvocationID input.
  spv::Id input_global_invocation_id =
      builder.createVariable(spv::NoPrecision, spv::StorageClassInput,
                             type_uint3, "gl_GlobalInvocationID");
  builder.addDecoration(input_global_invocation_id, spv::DecorationBuiltIn,
                        spv::BuiltInGlobalInvocationId);

  // Begin the main function.
  std::vector<spv::Id> main_param_types;
  std::vector<std::vector<spv::Decoration>> main_precisions;
  spv::Block* main_entry;
  spv::Function* main_function =
      builder.makeFunctionEntry(spv::NoPrecision, type_void, "main",
                                main_param_types, main_precisions, &main_entry);

  // For now, as the exact addressing in 64bpp render targets relatively to
  // 32bpp is unknown, treating 64bpp tiles as storing 40x16 samples rather than
  // 80x16 for simplicity of addressing into the texture.

  // Split the destination sample index into the 32bpp tile and the
  // 32bpp-tile-relative sample index.
  // Note that division by non-power-of-two constants will include a 4-cycle
  // 32*32 multiplication on AMD, even though so many bits are not needed for
  // the sample position - however, if an OpUnreachable path is inserted for the
  // case when the position has upper bits set, for some reason, the code for it
  // is not eliminated when compiling the shader for AMD via RenderDoc on
  // Windows, as of June 2022.
  spv::Id global_invocation_id =
      builder.createLoad(input_global_invocation_id, spv::NoPrecision);
  spv::Id rectangle_sample_x =
      builder.createCompositeExtract(global_invocation_id, type_uint, 0);
  uint32_t tile_width =
      (xenos::kEdramTileWidthSamples >> uint32_t(format_is_64bpp)) *
      draw_resolution_scale_x();
  spv::Id const_tile_width = builder.makeUintConstant(tile_width);
  spv::Id rectangle_tile_index_x = builder.createBinOp(
      spv::OpUDiv, type_uint, rectangle_sample_x, const_tile_width);
  spv::Id tile_sample_x = builder.createBinOp(
      spv::OpUMod, type_uint, rectangle_sample_x, const_tile_width);
  spv::Id rectangle_sample_y =
      builder.createCompositeExtract(global_invocation_id, type_uint, 1);
  uint32_t tile_height =
      xenos::kEdramTileHeightSamples * draw_resolution_scale_y();
  spv::Id const_tile_height = builder.makeUintConstant(tile_height);
  spv::Id rectangle_tile_index_y = builder.createBinOp(
      spv::OpUDiv, type_uint, rectangle_sample_y, const_tile_height);
  spv::Id tile_sample_y = builder.createBinOp(
      spv::OpUMod, type_uint, rectangle_sample_y, const_tile_height);

  // Get the tile index in the EDRAM relative to the dump rectangle base tile.
  id_vector_temp.clear();
  id_vector_temp.push_back(builder.makeIntConstant(kDumpPushConstantPitches));
  spv::Id pitches_constant = builder.createLoad(
      builder.createAccessChain(spv::StorageClassPushConstant, push_constants,
                                id_vector_temp),
      spv::NoPrecision);
  spv::Id const_uint_0 = builder.makeUintConstant(0);
  spv::Id const_edram_pitch_tiles_bits =
      builder.makeUintConstant(xenos::kEdramPitchTilesBits);
  spv::Id rectangle_tile_index = builder.createBinOp(
      spv::OpIAdd, type_uint,
      builder.createBinOp(
          spv::OpIMul, type_uint,
          builder.createTriOp(spv::OpBitFieldUExtract, type_uint,
                              pitches_constant, const_uint_0,
                              const_edram_pitch_tiles_bits),
          rectangle_tile_index_y),
      rectangle_tile_index_x);
  // Add the base tile in the dispatch to the dispatch-local tile index, not
  // wrapping yet so in case of a wraparound, the address relative to the base
  // in the image after subtraction of the base won't be negative.
  id_vector_temp.clear();
  id_vector_temp.push_back(builder.makeIntConstant(kDumpPushConstantOffsets));
  spv::Id offsets_constant = builder.createLoad(
      builder.createAccessChain(spv::StorageClassPushConstant, push_constants,
                                id_vector_temp),
      spv::NoPrecision);
  spv::Id const_edram_base_tiles_bits_plus_1 =
      builder.makeUintConstant(xenos::kEdramBaseTilesBits + 1);
  spv::Id edram_tile_index_non_wrapped = builder.createBinOp(
      spv::OpIAdd, type_uint,
      builder.createTriOp(spv::OpBitFieldUExtract, type_uint, offsets_constant,
                          const_uint_0, const_edram_base_tiles_bits_plus_1),
      rectangle_tile_index);

  // Combine the tile sample index and the tile index, wrapping the tile
  // addressing, into the EDRAM sample index.
  spv::Id edram_sample_address = builder.createBinOp(
      spv::OpIAdd, type_uint,
      builder.createBinOp(
          spv::OpIMul, type_uint,
          builder.makeUintConstant(tile_width * tile_height),
          builder.createBinOp(
              spv::OpBitwiseAnd, type_uint, edram_tile_index_non_wrapped,
              builder.makeUintConstant(xenos::kEdramTileCount - 1))),
      builder.createBinOp(spv::OpIAdd, type_uint,
                          builder.createBinOp(spv::OpIMul, type_uint,
                                              const_tile_width, tile_sample_y),
                          tile_sample_x));
  if (key.is_depth) {
    // Swap 40-sample columns in the depth buffer in the destination address to
    // get the final address of the sample in the EDRAM.
    uint32_t tile_width_half = tile_width >> 1;
    edram_sample_address = builder.createUnaryOp(
        spv::OpBitcast, type_uint,
        builder.createBinOp(
            spv::OpIAdd, type_int,
            builder.createUnaryOp(spv::OpBitcast, type_int,
                                  edram_sample_address),
            builder.createTriOp(
                spv::OpSelect, type_int,
                builder.createBinOp(spv::OpULessThan, builder.makeBoolType(),
                                    tile_sample_x,
                                    builder.makeUintConstant(tile_width_half)),
                builder.makeIntConstant(int32_t(tile_width_half)),
                builder.makeIntConstant(-int32_t(tile_width_half)))));
  }

  // Get the linear tile index within the source texture.
  spv::Id source_tile_index = builder.createBinOp(
      spv::OpISub, type_uint, edram_tile_index_non_wrapped,
      builder.createTriOp(
          spv::OpBitFieldUExtract, type_uint, offsets_constant,
          const_edram_base_tiles_bits_plus_1,
          builder.makeUintConstant(xenos::kEdramBaseTilesBits)));
  // Split the linear tile index in the source texture into X and Y in tiles.
  spv::Id source_pitch_tiles = builder.createTriOp(
      spv::OpBitFieldUExtract, type_uint, pitches_constant,
      const_edram_pitch_tiles_bits, const_edram_pitch_tiles_bits);
  spv::Id source_tile_index_y = builder.createBinOp(
      spv::OpUDiv, type_uint, source_tile_index, source_pitch_tiles);
  spv::Id source_tile_index_x = builder.createBinOp(
      spv::OpUMod, type_uint, source_tile_index, source_pitch_tiles);
  // Combine the source tile offset and the sample index within the tile.
  spv::Id source_sample_x = builder.createBinOp(
      spv::OpIAdd, type_uint,
      builder.createBinOp(spv::OpIMul, type_uint, const_tile_width,
                          source_tile_index_x),
      tile_sample_x);
  spv::Id source_sample_y = builder.createBinOp(
      spv::OpIAdd, type_uint,
      builder.createBinOp(spv::OpIMul, type_uint, const_tile_height,
                          source_tile_index_y),
      tile_sample_y);
  // Get the source pixel coordinate and the sample index within the pixel.
  spv::Id source_pixel_x = source_sample_x, source_pixel_y = source_sample_y;
  spv::Id source_sample_id = spv::NoResult;
  if (source_is_multisampled) {
    spv::Id const_uint_1 = builder.makeUintConstant(1);
    source_pixel_y = builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                         source_sample_y, const_uint_1);
    if (key.msaa_samples >= xenos::MsaaSamples::k4X) {
      source_pixel_x = builder.createBinOp(spv::OpShiftRightLogical, type_uint,
                                           source_sample_x, const_uint_1);
      // 4x MSAA source texture sample index - bit 0 for horizontal, bit 1 for
      // vertical.
      source_sample_id = builder.createQuadOp(
          spv::OpBitFieldInsert, type_uint,
          builder.createBinOp(spv::OpBitwiseAnd, type_uint, source_sample_x,
                              const_uint_1),
          source_sample_y, const_uint_1, const_uint_1);
    } else {
      // 2x MSAA source texture sample index - convert from the guest to
      // the Vulkan standard sample locations.
      source_sample_id = builder.createTriOp(
          spv::OpSelect, type_uint,
          builder.createBinOp(
              spv::OpINotEqual, builder.makeBoolType(),
              builder.createBinOp(spv::OpBitwiseAnd, type_uint, source_sample_y,
                                  const_uint_1),
              const_uint_0),
          builder.makeUintConstant(draw_util::GetD3D10SampleIndexForGuest2xMSAA(
              1, msaa_2x_attachments_supported_)),
          builder.makeUintConstant(draw_util::GetD3D10SampleIndexForGuest2xMSAA(
              0, msaa_2x_attachments_supported_)));
    }
  }

  // Load the source, and pack the value into one or two 32-bit integers.
  spv::Id packed[2] = {};
  spv::Builder::TextureParameters source_texture_parameters = {};
  source_texture_parameters.sampler =
      builder.createLoad(source_texture, spv::NoPrecision);
  id_vector_temp.clear();
  id_vector_temp.push_back(
      builder.createUnaryOp(spv::OpBitcast, type_int, source_pixel_x));
  id_vector_temp.push_back(
      builder.createUnaryOp(spv::OpBitcast, type_int, source_pixel_y));
  source_texture_parameters.coords =
      builder.createCompositeConstruct(type_int2, id_vector_temp);
  if (source_is_multisampled) {
    source_texture_parameters.sample =
        builder.createUnaryOp(spv::OpBitcast, type_int, source_sample_id);
  } else {
    source_texture_parameters.lod = builder.makeIntConstant(0);
  }
  spv::Id source_vec4 = builder.createTextureCall(
      spv::NoPrecision, builder.makeVectorType(source_component_type, 4), false,
      true, false, false, false, source_texture_parameters,
      spv::ImageOperandsMaskNone);
  if (key.is_depth) {
    source_texture_parameters.sampler =
        builder.createLoad(source_stencil_texture, spv::NoPrecision);
    spv::Id source_stencil = builder.createCompositeExtract(
        builder.createTextureCall(
            spv::NoPrecision, builder.makeVectorType(type_uint, 4), false, true,
            false, false, false, source_texture_parameters,
            spv::ImageOperandsMaskNone),
        type_uint, 0);
    spv::Id source_depth32 =
        builder.createCompositeExtract(source_vec4, type_float, 0);
    switch (key.GetDepthFormat()) {
      case xenos::DepthRenderTargetFormat::kD24S8: {
        // Round to the nearest even integer. This seems to be the correct
        // conversion, adding +0.5 and rounding towards zero results in red
        // instead of black in the 4D5307E6 clear shader.
        packed[0] = builder.createUnaryOp(
            spv::OpConvertFToU, type_uint,
            builder.createUnaryBuiltinCall(
                type_float, ext_inst_glsl_std_450, GLSLstd450RoundEven,
                builder.createBinOp(
                    spv::OpFMul, type_float, source_depth32,
                    builder.makeFloatConstant(float(0xFFFFFF)))));
      } break;
      case xenos::DepthRenderTargetFormat::kD24FS8: {
        packed[0] = SpirvShaderTranslator::PreClampedDepthTo20e4(
            builder, source_depth32, depth_float24_round(), true,
            ext_inst_glsl_std_450);
      } break;
    }
    packed[0] = builder.createQuadOp(
        spv::OpBitFieldInsert, type_uint, source_stencil, packed[0],
        builder.makeUintConstant(8), builder.makeUintConstant(24));
  } else {
    switch (key.GetColorFormat()) {
      case xenos::ColorRenderTargetFormat::k_8_8_8_8:
      case xenos::ColorRenderTargetFormat::k_8_8_8_8_GAMMA: {
        spv::Id unorm_round_offset = builder.makeFloatConstant(0.5f);
        spv::Id unorm_scale = builder.makeFloatConstant(255.0f);
        packed[0] = builder.createUnaryOp(
            spv::OpConvertFToU, type_uint,
            builder.createBinOp(
                spv::OpFAdd, type_float,
                builder.createBinOp(
                    spv::OpFMul, type_float,
                    builder.createCompositeExtract(source_vec4, type_float, 0),
                    unorm_scale),
                unorm_round_offset));
        spv::Id component_width = builder.makeUintConstant(8);
        for (uint32_t i = 1; i < 4; ++i) {
          packed[0] = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint, packed[0],
              builder.createUnaryOp(
                  spv::OpConvertFToU, type_uint,
                  builder.createBinOp(
                      spv::OpFAdd, type_float,
                      builder.createBinOp(spv::OpFMul, type_float,
                                          builder.createCompositeExtract(
                                              source_vec4, type_float, i),
                                          unorm_scale),
                      unorm_round_offset)),
              builder.makeUintConstant(8 * i), component_width);
        }
      } break;
      case xenos::ColorRenderTargetFormat::k_2_10_10_10:
      case xenos::ColorRenderTargetFormat::k_2_10_10_10_AS_10_10_10_10: {
        spv::Id unorm_round_offset = builder.makeFloatConstant(0.5f);
        spv::Id unorm_scale_rgb = builder.makeFloatConstant(1023.0f);
        packed[0] = builder.createUnaryOp(
            spv::OpConvertFToU, type_uint,
            builder.createBinOp(
                spv::OpFAdd, type_float,
                builder.createBinOp(
                    spv::OpFMul, type_float,
                    builder.createCompositeExtract(source_vec4, type_float, 0),
                    unorm_scale_rgb),
                unorm_round_offset));
        spv::Id width_rgb = builder.makeUintConstant(10);
        spv::Id unorm_scale_a = builder.makeFloatConstant(3.0f);
        spv::Id width_a = builder.makeUintConstant(2);
        for (uint32_t i = 1; i < 4; ++i) {
          packed[0] = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint, packed[0],
              builder.createUnaryOp(
                  spv::OpConvertFToU, type_uint,
                  builder.createBinOp(
                      spv::OpFAdd, type_float,
                      builder.createBinOp(
                          spv::OpFMul, type_float,
                          builder.createCompositeExtract(source_vec4,
                                                         type_float, i),
                          i == 3 ? unorm_scale_a : unorm_scale_rgb),
                      unorm_round_offset)),
              builder.makeUintConstant(10 * i), i == 3 ? width_a : width_rgb);
        }
      } break;
      case xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT:
      case xenos::ColorRenderTargetFormat::k_2_10_10_10_FLOAT_AS_16_16_16_16: {
        // Float16 has a wider range for both color and alpha, also NaNs - clamp
        // and convert.
        packed[0] = SpirvShaderTranslator::UnclampedFloat32To7e3(
            builder, builder.createCompositeExtract(source_vec4, type_float, 0),
            ext_inst_glsl_std_450);
        spv::Id width_rgb = builder.makeUintConstant(10);
        for (uint32_t i = 1; i < 3; ++i) {
          packed[0] = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint, packed[0],
              SpirvShaderTranslator::UnclampedFloat32To7e3(
                  builder,
                  builder.createCompositeExtract(source_vec4, type_float, i),
                  ext_inst_glsl_std_450),
              builder.makeUintConstant(10 * i), width_rgb);
        }
        // Saturate and convert the alpha.
        spv::Id alpha_saturated = builder.createTriBuiltinCall(
            type_float, ext_inst_glsl_std_450, GLSLstd450NClamp,
            builder.createCompositeExtract(source_vec4, type_float, 3),
            builder.makeFloatConstant(0.0f), builder.makeFloatConstant(1.0f));
        packed[0] = builder.createQuadOp(
            spv::OpBitFieldInsert, type_uint, packed[0],
            builder.createUnaryOp(
                spv::OpConvertFToU, type_uint,
                builder.createBinOp(
                    spv::OpFAdd, type_float,
                    builder.createBinOp(spv::OpFMul, type_float,
                                        alpha_saturated,
                                        builder.makeFloatConstant(3.0f)),
                    builder.makeFloatConstant(0.5f))),
            builder.makeUintConstant(30), builder.makeUintConstant(2));
      } break;
      case xenos::ColorRenderTargetFormat::k_16_16:
      case xenos::ColorRenderTargetFormat::k_16_16_16_16:
      case xenos::ColorRenderTargetFormat::k_16_16_FLOAT:
      case xenos::ColorRenderTargetFormat::k_16_16_16_16_FLOAT: {
        // All 64bpp formats, and all 16 bits per component formats, are
        // represented as integers in ownership transfer for safe handling of
        // NaN encodings and -32768 / -32767.
        // TODO(Triang3l): Handle the case when that's not true (no multisampled
        // sampled images, no 16-bit UNORM, no cross-packing 32bpp aliasing on a
        // portability subset device or a 64bpp format where that wouldn't help
        // anyway).
        spv::Id component_offset_width = builder.makeUintConstant(16);
        for (uint32_t i = 0; i <= uint32_t(format_is_64bpp); ++i) {
          packed[i] = builder.createQuadOp(
              spv::OpBitFieldInsert, type_uint,
              builder.createCompositeExtract(source_vec4, type_uint, 2 * i),
              builder.createCompositeExtract(source_vec4, type_uint, 2 * i + 1),
              component_offset_width, component_offset_width);
        }
      } break;
      // Float32 is transferred as uint32 to preserve NaN encodings. However,
      // multisampled sampled image support is optional in Vulkan.
      case xenos::ColorRenderTargetFormat::k_32_FLOAT:
      case xenos::ColorRenderTargetFormat::k_32_32_FLOAT: {
        for (uint32_t i = 0; i <= uint32_t(format_is_64bpp); ++i) {
          spv::Id& packed_ref = packed[i];
          packed_ref = builder.createCompositeExtract(source_vec4,
                                                      source_component_type, i);
          if (!source_is_uint) {
            packed_ref =
                builder.createUnaryOp(spv::OpBitcast, type_uint, packed_ref);
          }
        }
      } break;
    }
  }

  // Write the packed value to the EDRAM buffer.
  spv::Id store_value = packed[0];
  if (format_is_64bpp) {
    id_vector_temp.clear();
    id_vector_temp.push_back(packed[0]);
    id_vector_temp.push_back(packed[1]);
    store_value = builder.createCompositeConstruct(type_uint2, id_vector_temp);
  }
  id_vector_temp.clear();
  // The only SSBO structure member.
  id_vector_temp.push_back(builder.makeIntConstant(0));
  id_vector_temp.push_back(
      builder.createUnaryOp(spv::OpBitcast, type_int, edram_sample_address));
  // StorageBuffer since SPIR-V 1.3, but since SPIR-V 1.0 is generated, it's
  // Uniform.
  builder.createStore(store_value,
                      builder.createAccessChain(spv::StorageClassUniform,
                                                edram_buffer, id_vector_temp));

  // End the main function and make it the entry point.
  builder.leaveFunction();
  builder.addExecutionMode(main_function, spv::ExecutionModeLocalSize,
                           kDumpSamplesPerGroupX, kDumpSamplesPerGroupY, 1);
  spv::Instruction* entry_point = builder.addEntryPoint(
      spv::ExecutionModelGLCompute, main_function, "main");
  // Bindings only need to be added to the entry point's interface starting with
  // SPIR-V 1.4 - emitting 1.0 here, so only inputs / outputs.
  entry_point->addIdOperand(input_global_invocation_id);

  // Serialize the shader code.
  std::vector<unsigned int> shader_code;
  builder.dump(shader_code);

  // Create the pipeline, and store the handle even if creation fails not to try
  // to create it again later.
  VkPipeline pipeline = ui::vulkan::util::CreateComputePipeline(
      command_processor_.GetVulkanDevice(),
      key.is_depth ? dump_pipeline_layout_depth_ : dump_pipeline_layout_color_,
      reinterpret_cast<const uint32_t*>(shader_code.data()),
      sizeof(uint32_t) * shader_code.size());
  if (pipeline == VK_NULL_HANDLE) {
    XELOGE(
        "VulkanRenderTargetCache: Failed to create a render target dumping "
        "pipeline for {}-sample render targets with format {}",
        UINT32_C(1) << uint32_t(key.msaa_samples),
        key.is_depth
            ? xenos::GetDepthRenderTargetFormatName(key.GetDepthFormat())
            : xenos::GetColorRenderTargetFormatName(key.GetColorFormat()));
  }
  dump_pipelines_.emplace(key, pipeline);
  return pipeline;
}

void VulkanRenderTargetCache::DumpRenderTargets(uint32_t dump_base,
                                                uint32_t dump_row_length_used,
                                                uint32_t dump_rows,
                                                uint32_t dump_pitch) {
  assert_true(GetPath() == Path::kHostRenderTargets);

  GetResolveCopyRectanglesToDump(dump_base, dump_row_length_used, dump_rows,
                                 dump_pitch, dump_rectangles_);
  if (dump_rectangles_.empty()) {
    return;
  }

  // Clear previously set temporary indices.
  for (const ResolveCopyDumpRectangle& rectangle : dump_rectangles_) {
    static_cast<VulkanRenderTarget*>(rectangle.render_target)
        ->SetTemporarySortIndex(UINT32_MAX);
  }
  // Gather all needed barriers and info needed to sort the invocations.
  UseEdramBuffer(EdramBufferUsage::kComputeWrite);
  dump_invocations_.clear();
  dump_invocations_.reserve(dump_rectangles_.size());
  constexpr VkPipelineStageFlags kRenderTargetDstStageMask =
      VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT;
  constexpr VkAccessFlags kRenderTargetDstAccessMask =
      VK_ACCESS_SHADER_READ_BIT;
  constexpr VkImageLayout kRenderTargetNewLayout =
      VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL;
  uint32_t rt_sort_index = 0;
  for (const ResolveCopyDumpRectangle& rectangle : dump_rectangles_) {
    auto& vulkan_rt =
        *static_cast<VulkanRenderTarget*>(rectangle.render_target);
    RenderTargetKey rt_key = vulkan_rt.key();
    command_processor_.PushImageMemoryBarrier(
        vulkan_rt.image(),
        ui::vulkan::util::InitializeSubresourceRange(
            rt_key.is_depth
                ? (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)
                : VK_IMAGE_ASPECT_COLOR_BIT),
        vulkan_rt.current_stage_mask(), VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
        vulkan_rt.current_access_mask(), VK_ACCESS_SHADER_READ_BIT,
        vulkan_rt.current_layout(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
    vulkan_rt.SetUsage(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
                       VK_ACCESS_SHADER_READ_BIT,
                       VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
    if (vulkan_rt.temporary_sort_index() == UINT32_MAX) {
      vulkan_rt.SetTemporarySortIndex(rt_sort_index++);
    }
    DumpPipelineKey pipeline_key;
    pipeline_key.msaa_samples = rt_key.msaa_samples;
    pipeline_key.resource_format = rt_key.resource_format;
    pipeline_key.is_depth = rt_key.is_depth;
    dump_invocations_.emplace_back(rectangle, pipeline_key);
  }

  // Sort the invocations to reduce context and binding switches.
  std::sort(dump_invocations_.begin(), dump_invocations_.end());

  // Dump the render targets.
  DeferredCommandBuffer& command_buffer =
      command_processor_.deferred_command_buffer();
  bool edram_buffer_bound = false;
  VkDescriptorSet last_source_descriptor_set = VK_NULL_HANDLE;
  DumpPitches last_pitches;
  DumpOffsets last_offsets;
  bool pitches_bound = false, offsets_bound = false;
  for (const DumpInvocation& invocation : dump_invocations_) {
    const ResolveCopyDumpRectangle& rectangle = invocation.rectangle;
    auto& vulkan_rt =
        *static_cast<VulkanRenderTarget*>(rectangle.render_target);
    RenderTargetKey rt_key = vulkan_rt.key();
    DumpPipelineKey pipeline_key = invocation.pipeline_key;
    VkPipeline pipeline = GetDumpPipeline(pipeline_key);
    if (!pipeline) {
      continue;
    }
    command_processor_.BindExternalComputePipeline(pipeline);

    VkPipelineLayout pipeline_layout = rt_key.is_depth
                                           ? dump_pipeline_layout_depth_
                                           : dump_pipeline_layout_color_;

    // Only need to bind the EDRAM buffer once (relying on pipeline layout
    // compatibility).
    if (!edram_buffer_bound) {
      edram_buffer_bound = true;
      command_buffer.CmdVkBindDescriptorSets(
          VK_PIPELINE_BIND_POINT_COMPUTE, pipeline_layout,
          kDumpDescriptorSetEdram, 1, &edram_storage_buffer_descriptor_set_, 0,
          nullptr);
    }

    VkDescriptorSet source_descriptor_set =
        vulkan_rt.GetDescriptorSetTransferSource();
    if (last_source_descriptor_set != source_descriptor_set) {
      last_source_descriptor_set = source_descriptor_set;
      command_buffer.CmdVkBindDescriptorSets(
          VK_PIPELINE_BIND_POINT_COMPUTE, pipeline_layout,
          kDumpDescriptorSetSource, 1, &source_descriptor_set, 0, nullptr);
    }

    DumpPitches pitches;
    pitches.dest_pitch = dump_pitch;
    pitches.source_pitch = rt_key.GetPitchTiles();
    if (last_pitches != pitches) {
      last_pitches = pitches;
      pitches_bound = false;
    }
    if (!pitches_bound) {
      pitches_bound = true;
      command_buffer.CmdVkPushConstants(
          pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT,
          sizeof(uint32_t) * kDumpPushConstantPitches, sizeof(last_pitches),
          &last_pitches);
    }

    DumpOffsets offsets;
    offsets.source_base_tiles = rt_key.base_tiles;
    ResolveCopyDumpRectangle::Dispatch
        dispatches[ResolveCopyDumpRectangle::kMaxDispatches];
    uint32_t dispatch_count =
        rectangle.GetDispatches(dump_pitch, dump_row_length_used, dispatches);
    for (uint32_t i = 0; i < dispatch_count; ++i) {
      const ResolveCopyDumpRectangle::Dispatch& dispatch = dispatches[i];
      offsets.dispatch_first_tile = dump_base + dispatch.offset;
      if (last_offsets != offsets) {
        last_offsets = offsets;
        offsets_bound = false;
      }
      if (!offsets_bound) {
        offsets_bound = true;
        command_buffer.CmdVkPushConstants(
            pipeline_layout, VK_SHADER_STAGE_COMPUTE_BIT,
            sizeof(uint32_t) * kDumpPushConstantOffsets, sizeof(last_offsets),
            &last_offsets);
      }
      command_processor_.SubmitBarriers(true);
      command_buffer.CmdVkDispatch(
          (draw_resolution_scale_x() *
               (xenos::kEdramTileWidthSamples >> uint32_t(rt_key.Is64bpp())) *
               dispatch.width_tiles +
           (kDumpSamplesPerGroupX - 1)) /
              kDumpSamplesPerGroupX,
          (draw_resolution_scale_y() * xenos::kEdramTileHeightSamples *
               dispatch.height_tiles +
           (kDumpSamplesPerGroupY - 1)) /
              kDumpSamplesPerGroupY,
          1);
    }
    MarkEdramBufferModified();
  }
}

}  // namespace vulkan
}  // namespace gpu
}  // namespace xe