[vulkan] First intent on unswizzling via gpu shaders

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -25,6 +25,9 @@
 #include "video_core/host_shaders/vulkan_quad_indexed_comp_spv.h"
 #include "video_core/host_shaders/vulkan_uint8_comp_spv.h"
 #include "video_core/host_shaders/block_linear_unswizzle_3d_bcn_comp_spv.h"
+#include "video_core/host_shaders/block_linear_unswizzle_2d_comp_spv.h"
+#include "video_core/host_shaders/block_linear_unswizzle_3d_comp_spv.h"
+#include "video_core/host_shaders/pitch_unswizzle_comp_spv.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
@@ -232,6 +235,26 @@ struct QueriesPrefixScanPushConstants {
 struct ConditionalRenderingResolvePushConstants {
     u32 compare_to_zero;
 };
+
+struct BlockLinear3DImagePushConstants {
+    alignas(16) std::array<u32, 3> origin;
+    alignas(16) std::array<s32, 3> destination;
+    u32 bytes_per_block_log2;
+    u32 slice_size;
+    u32 block_size;
+    u32 x_shift;
+    u32 block_height;
+    u32 block_height_mask;
+    u32 block_depth;
+    u32 block_depth_mask;
+};
+
+struct PitchUnswizzlePushConstants {
+    std::array<u32, 2> origin;
+    std::array<s32, 2> destination;
+    u32 bytes_per_block;
+    u32 pitch;
+};
 } // Anonymous namespace
 
 ComputePass::ComputePass(const Device& device_, Scheduler& scheduler, DescriptorPool& descriptor_pool,
@@ -653,6 +676,306 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
     scheduler.Finish();
 }
 
+BlockLinearUnswizzleImage2DPass::BlockLinearUnswizzleImage2DPass(
+    const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+    StagingBufferPool& staging_buffer_pool_,
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(device_, scheduler_, descriptor_pool_, ASTC_DESCRIPTOR_SET_BINDINGS,
+                  ASTC_PASS_DESCRIPTOR_UPDATE_TEMPLATE_ENTRY, ASTC_BANK_INFO,
+                  COMPUTE_PUSH_CONSTANT_RANGE<sizeof(BlockLinearSwizzle2DParams)>,
+                  BLOCK_LINEAR_UNSWIZZLE_2D_COMP_SPV),
+      scheduler{scheduler_}, staging_buffer_pool{staging_buffer_pool_},
+      compute_pass_descriptor_queue{compute_pass_descriptor_queue_} {}
+
+BlockLinearUnswizzleImage2DPass::~BlockLinearUnswizzleImage2DPass() = default;
+
+void BlockLinearUnswizzleImage2DPass::Unswizzle(
+    Image& image, const StagingBufferRef& map,
+    std::span<const VideoCommon::SwizzleParameters> swizzles) {
+    using namespace VideoCommon::Accelerated;
+    scheduler.RequestOutsideRenderPassOperationContext();
+    const VkPipeline vk_pipeline = *pipeline;
+    const VkImageAspectFlags aspect_mask = image.AspectMask();
+    const VkImage vk_image = image.Handle();
+    const bool is_initialized = image.ExchangeInitialization();
+    scheduler.Record([vk_pipeline, vk_image, aspect_mask,
+                      is_initialized](vk::CommandBuffer cmdbuf) {
+        const VkImageMemoryBarrier image_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = static_cast<VkAccessFlags>(is_initialized ? VK_ACCESS_SHADER_WRITE_BIT
+                                                                       : VK_ACCESS_NONE),
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            .oldLayout = is_initialized ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_UNDEFINED,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = vk_image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        cmdbuf.PipelineBarrier(is_initialized ? vk::PIPELINE_STAGE_GRAPHICS_COMPUTE_TRANSFER
+                              : VkPipelineStageFlags(VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT),
+                       VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, image_barrier);
+        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, vk_pipeline);
+    });
+    for (const VideoCommon::SwizzleParameters& swizzle : swizzles) {
+        const size_t input_offset = swizzle.buffer_offset + map.offset;
+        const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
+        const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
+        const u32 num_dispatches_z = image.info.resources.layers;
+
+        compute_pass_descriptor_queue.Acquire(scheduler, 2);
+        compute_pass_descriptor_queue.AddBuffer(map.buffer, input_offset,
+                                                image.guest_size_bytes - swizzle.buffer_offset);
+        compute_pass_descriptor_queue.AddImage(image.StorageImageView(swizzle.level));
+        const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
+
+        const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
+        scheduler.Record([this, num_dispatches_x, num_dispatches_y, num_dispatches_z, params,
+                          descriptor_data](vk::CommandBuffer cmdbuf) {
+            const VkDescriptorSet set = descriptor_allocator.Commit();
+            device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
+            cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *layout, 0, set, {});
+            cmdbuf.PushConstants(*layout, VK_SHADER_STAGE_COMPUTE_BIT, params);
+            cmdbuf.Dispatch(num_dispatches_x, num_dispatches_y, num_dispatches_z);
+        });
+    }
+    scheduler.Record([vk_image, aspect_mask](vk::CommandBuffer cmdbuf) {
+        const VkImageMemoryBarrier image_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = vk_image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+                       vk::PIPELINE_STAGE_GRAPHICS_COMPUTE, 0, image_barrier);
+    });
+    scheduler.Finish();
+}
+
+BlockLinearUnswizzleImage3DPass::BlockLinearUnswizzleImage3DPass(
+    const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+    StagingBufferPool& staging_buffer_pool_,
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(device_, scheduler_, descriptor_pool_, ASTC_DESCRIPTOR_SET_BINDINGS,
+                  ASTC_PASS_DESCRIPTOR_UPDATE_TEMPLATE_ENTRY, ASTC_BANK_INFO,
+                  COMPUTE_PUSH_CONSTANT_RANGE<sizeof(BlockLinear3DImagePushConstants)>,
+                  BLOCK_LINEAR_UNSWIZZLE_3D_COMP_SPV),
+      scheduler{scheduler_}, staging_buffer_pool{staging_buffer_pool_},
+      compute_pass_descriptor_queue{compute_pass_descriptor_queue_} {}
+
+BlockLinearUnswizzleImage3DPass::~BlockLinearUnswizzleImage3DPass() = default;
+
+void BlockLinearUnswizzleImage3DPass::Unswizzle(
+    Image& image, const StagingBufferRef& map,
+    std::span<const VideoCommon::SwizzleParameters> swizzles) {
+    using namespace VideoCommon::Accelerated;
+    scheduler.RequestOutsideRenderPassOperationContext();
+    const VkPipeline vk_pipeline = *pipeline;
+    const VkImageAspectFlags aspect_mask = image.AspectMask();
+    const VkImage vk_image = image.Handle();
+    const bool is_initialized = image.ExchangeInitialization();
+    scheduler.Record([vk_pipeline, vk_image, aspect_mask,
+                      is_initialized](vk::CommandBuffer cmdbuf) {
+        const VkImageMemoryBarrier image_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = static_cast<VkAccessFlags>(is_initialized ? VK_ACCESS_SHADER_WRITE_BIT
+                                                                       : VK_ACCESS_NONE),
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            .oldLayout = is_initialized ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_UNDEFINED,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = vk_image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        cmdbuf.PipelineBarrier(is_initialized ? vk::PIPELINE_STAGE_GRAPHICS_COMPUTE_TRANSFER
+                              : VkPipelineStageFlags(VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT),
+                       VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, image_barrier);
+        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, vk_pipeline);
+    });
+    for (const VideoCommon::SwizzleParameters& swizzle : swizzles) {
+        const size_t input_offset = swizzle.buffer_offset + map.offset;
+        const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 16U);
+        const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 8U);
+        const u32 num_dispatches_z = Common::DivCeil(swizzle.num_tiles.depth, 8U);
+
+        compute_pass_descriptor_queue.Acquire(scheduler, 2);
+        compute_pass_descriptor_queue.AddBuffer(map.buffer, input_offset,
+                                                image.guest_size_bytes - swizzle.buffer_offset);
+        compute_pass_descriptor_queue.AddImage(image.StorageImageView(swizzle.level));
+        const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
+
+        const auto p = MakeBlockLinearSwizzle3DParams(swizzle, image.info);
+        const BlockLinear3DImagePushConstants params{
+            .origin = p.origin,
+            .destination = p.destination,
+            .bytes_per_block_log2 = p.bytes_per_block_log2,
+            .slice_size = p.slice_size,
+            .block_size = p.block_size,
+            .x_shift = p.x_shift,
+            .block_height = p.block_height,
+            .block_height_mask = p.block_height_mask,
+            .block_depth = p.block_depth,
+            .block_depth_mask = p.block_depth_mask,
+        };
+        scheduler.Record([this, num_dispatches_x, num_dispatches_y, num_dispatches_z, params,
+                          descriptor_data](vk::CommandBuffer cmdbuf) {
+            const VkDescriptorSet set = descriptor_allocator.Commit();
+            device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
+            cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *layout, 0, set, {});
+            cmdbuf.PushConstants(*layout, VK_SHADER_STAGE_COMPUTE_BIT, params);
+            cmdbuf.Dispatch(num_dispatches_x, num_dispatches_y, num_dispatches_z);
+        });
+    }
+    scheduler.Record([vk_image, aspect_mask](vk::CommandBuffer cmdbuf) {
+        const VkImageMemoryBarrier image_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = vk_image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+                       vk::PIPELINE_STAGE_GRAPHICS_COMPUTE, 0, image_barrier);
+    });
+    scheduler.Finish();
+}
+
+PitchUnswizzlePass::PitchUnswizzlePass(const Device& device_, Scheduler& scheduler_,
+                                       DescriptorPool& descriptor_pool_,
+                                       StagingBufferPool& staging_buffer_pool_,
+                                       ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(device_, scheduler_, descriptor_pool_, ASTC_DESCRIPTOR_SET_BINDINGS,
+                  ASTC_PASS_DESCRIPTOR_UPDATE_TEMPLATE_ENTRY, ASTC_BANK_INFO,
+                  COMPUTE_PUSH_CONSTANT_RANGE<sizeof(PitchUnswizzlePushConstants)>,
+                  PITCH_UNSWIZZLE_COMP_SPV),
+      scheduler{scheduler_}, staging_buffer_pool{staging_buffer_pool_},
+      compute_pass_descriptor_queue{compute_pass_descriptor_queue_} {}
+
+PitchUnswizzlePass::~PitchUnswizzlePass() = default;
+
+void PitchUnswizzlePass::Unswizzle(Image& image, const StagingBufferRef& map,
+                                   std::span<const VideoCommon::SwizzleParameters> swizzles) {
+    scheduler.RequestOutsideRenderPassOperationContext();
+    const VkPipeline vk_pipeline = *pipeline;
+    const VkImageAspectFlags aspect_mask = image.AspectMask();
+    const VkImage vk_image = image.Handle();
+    const bool is_initialized = image.ExchangeInitialization();
+    scheduler.Record([vk_pipeline, vk_image, aspect_mask,
+                      is_initialized](vk::CommandBuffer cmdbuf) {
+        const VkImageMemoryBarrier image_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = static_cast<VkAccessFlags>(is_initialized ? VK_ACCESS_SHADER_WRITE_BIT
+                                                                       : VK_ACCESS_NONE),
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            .oldLayout = is_initialized ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_UNDEFINED,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = vk_image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        cmdbuf.PipelineBarrier(is_initialized ? vk::PIPELINE_STAGE_GRAPHICS_COMPUTE_TRANSFER
+                              : VkPipelineStageFlags(VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT),
+                       VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, image_barrier);
+        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, vk_pipeline);
+    });
+    const u32 bytes_per_block = VideoCore::Surface::BytesPerBlock(image.info.format);
+    for (const VideoCommon::SwizzleParameters& swizzle : swizzles) {
+        const size_t input_offset = swizzle.buffer_offset + map.offset;
+        const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
+        const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
+
+        compute_pass_descriptor_queue.Acquire(scheduler, 2);
+        compute_pass_descriptor_queue.AddBuffer(map.buffer, input_offset,
+                                                image.guest_size_bytes - swizzle.buffer_offset);
+        compute_pass_descriptor_queue.AddImage(image.StorageImageView(swizzle.level));
+        const void* const descriptor_data{compute_pass_descriptor_queue.UpdateData()};
+
+        const PitchUnswizzlePushConstants params{
+            .origin = {0, 0},
+            .destination = {0, 0},
+            .bytes_per_block = bytes_per_block,
+            .pitch = image.info.pitch,
+        };
+        scheduler.Record([this, num_dispatches_x, num_dispatches_y, params,
+                          descriptor_data](vk::CommandBuffer cmdbuf) {
+            const VkDescriptorSet set = descriptor_allocator.Commit();
+            device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);
+            cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, *layout, 0, set, {});
+            cmdbuf.PushConstants(*layout, VK_SHADER_STAGE_COMPUTE_BIT, params);
+            cmdbuf.Dispatch(num_dispatches_x, num_dispatches_y, 1);
+        });
+    }
+    scheduler.Record([vk_image, aspect_mask](vk::CommandBuffer cmdbuf) {
+        const VkImageMemoryBarrier image_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = vk_image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+                       vk::PIPELINE_STAGE_GRAPHICS_COMPUTE, 0, image_barrier);
+    });
+    scheduler.Finish();
+}
+
 constexpr u32 BL3D_BINDING_INPUT_BUFFER  = 0;
 constexpr u32 BL3D_BINDING_OUTPUT_BUFFER = 1;
 

src/video_core/renderer_vulkan/vk_compute_pass.h

@@ -25,6 +25,7 @@ struct SwizzleParameters;
 
 namespace Vulkan {
 
+using VideoCommon::Accelerated::BlockLinearSwizzle2DParams;
 using VideoCommon::Accelerated::BlockLinearSwizzle3DParams;
 
 class Device;
@@ -164,6 +165,56 @@ private:
     ComputePassDescriptorQueue& compute_pass_descriptor_queue;
 };
 
+class BlockLinearUnswizzleImage2DPass final : public ComputePass {
+public:
+    explicit BlockLinearUnswizzleImage2DPass(const Device& device_, Scheduler& scheduler_,
+                                             DescriptorPool& descriptor_pool_,
+                                             StagingBufferPool& staging_buffer_pool_,
+                                             ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+    ~BlockLinearUnswizzleImage2DPass();
+
+    void Unswizzle(Image& image, const StagingBufferRef& map,
+                   std::span<const VideoCommon::SwizzleParameters> swizzles);
+
+private:
+    Scheduler& scheduler;
+    StagingBufferPool& staging_buffer_pool;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
+
+class BlockLinearUnswizzleImage3DPass final : public ComputePass {
+public:
+    explicit BlockLinearUnswizzleImage3DPass(const Device& device_, Scheduler& scheduler_,
+                                             DescriptorPool& descriptor_pool_,
+                                             StagingBufferPool& staging_buffer_pool_,
+                                             ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+    ~BlockLinearUnswizzleImage3DPass();
+
+    void Unswizzle(Image& image, const StagingBufferRef& map,
+                   std::span<const VideoCommon::SwizzleParameters> swizzles);
+
+private:
+    Scheduler& scheduler;
+    StagingBufferPool& staging_buffer_pool;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
+
+class PitchUnswizzlePass final : public ComputePass {
+public:
+    explicit PitchUnswizzlePass(const Device& device_, Scheduler& scheduler_,
+                                DescriptorPool& descriptor_pool_,
+                                StagingBufferPool& staging_buffer_pool_,
+                                ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+    ~PitchUnswizzlePass();
+
+    void Unswizzle(Image& image, const StagingBufferRef& map,
+                   std::span<const VideoCommon::SwizzleParameters> swizzles);
+
+private:
+    Scheduler& scheduler;
+    StagingBufferPool& staging_buffer_pool;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
 
 class MSAACopyPass final : public ComputePass {
 public:

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -195,6 +195,29 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     return allocator.CreateImage(image_ci);
 }
 
+[[nodiscard]] VkFormat UnswizzleStorageFormat(u32 bytes_per_block) {
+    switch (bytes_per_block) {
+    case 1:
+        return VK_FORMAT_R8_UINT;
+    case 2:
+        return VK_FORMAT_R16_UINT;
+    case 4:
+        return VK_FORMAT_R32_UINT;
+    case 8:
+        return VK_FORMAT_R32G32_UINT;
+    case 16:
+        return VK_FORMAT_R32G32B32A32_UINT;
+    default:
+        ASSERT_MSG(false, "Invalid bytes_per_block={} for accelerated unswizzle", bytes_per_block);
+        return VK_FORMAT_R32_UINT;
+    }
+}
+
+[[nodiscard]] bool IsUnswizzleStorageFormatSupported(u32 bytes_per_block) {
+    return bytes_per_block == 1 || bytes_per_block == 2 || bytes_per_block == 4 ||
+           bytes_per_block == 8 || bytes_per_block == 16;
+}
+
 [[nodiscard]] vk::ImageView MakeStorageView(const vk::Device& device, u32 level, VkImage image,
                                             VkFormat format) {
     static constexpr VkImageViewUsageCreateInfo storage_image_view_usage_create_info{
@@ -887,6 +910,12 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, Scheduler& sched
     if (device.IsStorageImageMultisampleSupported()) {
         msaa_copy_pass.emplace(device, scheduler, descriptor_pool, staging_buffer_pool, compute_pass_descriptor_queue);
     }
+    bl_unswizzle_2d_pass.emplace(device, scheduler, descriptor_pool, staging_buffer_pool,
+                                 compute_pass_descriptor_queue);
+    bl_unswizzle_3d_pass.emplace(device, scheduler, descriptor_pool, staging_buffer_pool,
+                                 compute_pass_descriptor_queue);
+    pitch_unswizzle_pass.emplace(device, scheduler, descriptor_pool, staging_buffer_pool,
+                                 compute_pass_descriptor_queue);
     if (!device.IsKhrImageFormatListSupported()) {
         return;
     }
@@ -894,6 +923,14 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, Scheduler& sched
         const auto image_format = static_cast<PixelFormat>(index_a);
         if (IsPixelFormatASTC(image_format) && !device.IsOptimalAstcSupported()) {
             view_formats[index_a].push_back(VK_FORMAT_A8B8G8R8_UNORM_PACK32);
+        } else if (!IsPixelFormatASTC(image_format) && !IsPixelFormatBCn(image_format)) {
+            // Generic block-linear/pitch accelerated unswizzle needs an alternate UINT
+            // storage view registered up-front so MakeImage() enables mutable format +
+            // storage usage for these images (see Image::Image / storage_image_views below).
+            const u32 bpp = VideoCore::Surface::BytesPerBlock(image_format);
+            if (IsUnswizzleStorageFormatSupported(bpp)) {
+                view_formats[index_a].push_back(UnswizzleStorageFormat(bpp));
+            }
         }
         for (size_t index_b = 0; index_b < VideoCore::Surface::MaxPixelFormat; index_b++) {
             const auto view_format = static_cast<PixelFormat>(index_b);
@@ -1580,6 +1617,15 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
         flags |= VideoCommon::ImageFlagBits::Converted;
         flags |= VideoCommon::ImageFlagBits::CostlyLoad;
     }
+    if (!IsPixelFormatASTC(info.format) && !IsPixelFormatBCn(info.format)) {
+        if (info.type == ImageType::e2D || info.type == ImageType::e3D) {
+            flags |= VideoCommon::ImageFlagBits::AcceleratedUpload;
+        } else if (info.type == ImageType::Linear) {
+            if (IsUnswizzleStorageFormatSupported(VideoCore::Surface::BytesPerBlock(info.format))) {
+                flags |= VideoCommon::ImageFlagBits::AcceleratedUpload;
+            }
+        }
+    }
     if (runtime->device.HasDebuggingToolAttached()) {
         original_image.SetObjectNameEXT(VideoCommon::Name(*this).c_str());
     }
@@ -1593,6 +1639,13 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
             storage_image_views[level] =
                 MakeStorageView(device, level, *original_image, VK_FORMAT_A8B8G8R8_UNORM_PACK32);
         }
+    } else if (True(flags & VideoCommon::ImageFlagBits::AcceleratedUpload)) {
+        const auto& device = runtime->device.GetLogical();
+        const VkFormat storage_format =
+            UnswizzleStorageFormat(VideoCore::Surface::BytesPerBlock(info.format));
+        for (s32 level = 0; level < info.resources.levels; ++level) {
+            storage_image_views[level] = MakeStorageView(device, level, *original_image, storage_format);
+        }
     }
 }
 
@@ -2448,16 +2501,24 @@ void TextureCacheRuntime::AccelerateImageUpload(
         return astc_decoder_pass->Assemble(image, map, swizzles);
     }
 
-    if (!Settings::values.gpu_unswizzle_enabled.GetValue() || !bl3d_unswizzle_pass) {
-        if (IsPixelFormatBCn(image.info.format) && image.info.type == ImageType::e3D) {
-            ASSERT(false && "GPU unswizzle is disabled for BCn 3D texture");
+    if (IsPixelFormatBCn(image.info.format)) {
+        if (Settings::values.gpu_unswizzle_enabled.GetValue() && bl3d_unswizzle_pass &&
+            image.info.type == ImageType::e3D && image.info.resources.levels == 1 &&
+            image.info.resources.layers == 1) {
+            return bl3d_unswizzle_pass->Unswizzle(image, map, swizzles, z_start, z_count);
         }
-        ASSERT(false);
+        ASSERT(false && "GPU unswizzle is disabled for BCn 3D texture");
         return;
     }
 
-    if (bl3d_unswizzle_pass && IsPixelFormatBCn(image.info.format) && image.info.type == ImageType::e3D && image.info.resources.levels == 1 && image.info.resources.layers == 1) {
-        return bl3d_unswizzle_pass->Unswizzle(image, map, swizzles, z_start, z_count);
+    if (image.info.type == ImageType::e2D) {
+        return bl_unswizzle_2d_pass->Unswizzle(image, map, swizzles);
+    }
+    if (image.info.type == ImageType::e3D) {
+        return bl_unswizzle_3d_pass->Unswizzle(image, map, swizzles);
+    }
+    if (image.info.type == ImageType::Linear) {
+        return pitch_unswizzle_pass->Unswizzle(image, map, swizzles);
     }
 
     ASSERT(false);

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -130,6 +130,9 @@ public:
     std::optional<ASTCDecoderPass> astc_decoder_pass;
 
     std::optional<BlockLinearUnswizzle3DPass> bl3d_unswizzle_pass;
+    std::optional<BlockLinearUnswizzleImage2DPass> bl_unswizzle_2d_pass;
+    std::optional<BlockLinearUnswizzleImage3DPass> bl_unswizzle_3d_pass;
+    std::optional<PitchUnswizzlePass> pitch_unswizzle_pass;
     std::optional<MSAACopyPass> msaa_copy_pass;
     const Settings::ResolutionScalingInfo& resolution;
     std::array<std::vector<VkFormat>, VideoCore::Surface::MaxPixelFormat> view_formats;