Engines: Implement Accelerate DMA Texture.

src/video_core/buffer_cache/buffer_cache.h

@@ -55,6 +55,19 @@ constexpr u32 NUM_STORAGE_BUFFERS = 16;
 constexpr u32 NUM_TEXTURE_BUFFERS = 16;
 constexpr u32 NUM_STAGES = 5;
 
+enum class ObtainBufferSynchronize : u32 {
+    NoSynchronize = 0,
+    FullSynchronize = 1,
+    SynchronizeNoDirty = 2,
+};
+
+enum class ObtainBufferOperation : u32 {
+    DoNothing = 0,
+    MarkAsWritten = 1,
+    DiscardWrite = 2,
+    MarkQuery = 3,
+};
+
 using UniformBufferSizes = std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>;
 using ComputeUniformBufferSizes = std::array<u32, NUM_COMPUTE_UNIFORM_BUFFERS>;
 
@@ -191,6 +204,10 @@ public:
 
     bool DMAClear(GPUVAddr src_address, u64 amount, u32 value);
 
+    [[nodiscard]] std::pair<Buffer*, u32> ObtainBuffer(GPUVAddr gpu_addr, u32 size,
+                                                       ObtainBufferSynchronize sync_info,
+                                                       ObtainBufferOperation post_op);
+
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
@@ -641,6 +658,42 @@ bool BufferCache<P>::DMAClear(GPUVAddr dst_address, u64 amount, u32 value) {
     return true;
 }
 
+template <class P>
+std::pair<typename P::Buffer*, u32> BufferCache<P>::ObtainBuffer(GPUVAddr gpu_addr, u32 size,
+                                                                 ObtainBufferSynchronize sync_info,
+                                                                 ObtainBufferOperation post_op) {
+    const std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
+    if (!cpu_addr) {
+        return {&slot_buffers[NULL_BUFFER_ID], 0};
+    }
+    const BufferId buffer_id = FindBuffer(*cpu_addr, size);
+    Buffer& buffer = slot_buffers[buffer_id];
+
+    // synchronize op
+    switch (sync_info) {
+    case ObtainBufferSynchronize::FullSynchronize:
+        SynchronizeBuffer(buffer, *cpu_addr, size);
+        break;
+    default:
+        break;
+    }
+
+    switch (post_op) {
+    case ObtainBufferOperation::MarkAsWritten:
+        MarkWrittenBuffer(buffer_id, *cpu_addr, size);
+        break;
+    case ObtainBufferOperation::DiscardWrite: {
+        IntervalType interval{*cpu_addr, size};
+        ClearDownload(interval);
+        break;
+    }
+    default:
+        break;
+    }
+
+    return {&buffer, buffer.Offset(*cpu_addr)};
+}
+
 template <class P>
 void BufferCache<P>::BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
                                                u32 size) {

src/video_core/engines/maxwell_dma.cpp

@@ -14,7 +14,13 @@
 #include "video_core/textures/decoders.h"
 
 MICROPROFILE_DECLARE(GPU_DMAEngine);
+MICROPROFILE_DECLARE(GPU_DMAEngineBL);
+MICROPROFILE_DECLARE(GPU_DMAEngineLB);
+MICROPROFILE_DECLARE(GPU_DMAEngineBB);
 MICROPROFILE_DEFINE(GPU_DMAEngine, "GPU", "DMA Engine", MP_RGB(224, 224, 128));
+MICROPROFILE_DEFINE(GPU_DMAEngineBL, "GPU", "DMA Engine Block - Linear", MP_RGB(224, 224, 128));
+MICROPROFILE_DEFINE(GPU_DMAEngineLB, "GPU", "DMA Engine Linear - Block", MP_RGB(224, 224, 128));
+MICROPROFILE_DEFINE(GPU_DMAEngineBB, "GPU", "DMA Engine Block - Block", MP_RGB(224, 224, 128));
 
 namespace Tegra::Engines {
 
@@ -72,6 +78,7 @@ void MaxwellDMA::Launch() {
         memory_manager.FlushCaching();
         if (!is_src_pitch && !is_dst_pitch) {
             // If both the source and the destination are in block layout, assert.
+            MICROPROFILE_SCOPE(GPU_DMAEngineBB);
             CopyBlockLinearToBlockLinear();
             ReleaseSemaphore();
             return;
@@ -87,8 +94,10 @@ void MaxwellDMA::Launch() {
             }
         } else {
             if (!is_src_pitch && is_dst_pitch) {
+                MICROPROFILE_SCOPE(GPU_DMAEngineBL);
                 CopyBlockLinearToPitch();
             } else {
+                MICROPROFILE_SCOPE(GPU_DMAEngineLB);
                 CopyPitchToBlockLinear();
             }
         }
@@ -153,21 +162,35 @@ void MaxwellDMA::Launch() {
 }
 
 void MaxwellDMA::CopyBlockLinearToPitch() {
-    UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0);
-    UNIMPLEMENTED_IF(regs.src_params.layer != 0);
-
-    const bool is_remapping = regs.launch_dma.remap_enable != 0;
-
-    // Optimized path for micro copies.
-    const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
-    if (!is_remapping && dst_size < GOB_SIZE && regs.pitch_out <= GOB_SIZE_X &&
-        regs.src_params.height > GOB_SIZE_Y) {
-        FastCopyBlockLinearToPitch();
+    UNIMPLEMENTED_IF(regs.launch_dma.remap_enable != 0);
+
+    u32 bytes_per_pixel = 1;
+    DMA::ImageOperand src_operand;
+    src_operand.bytes_per_pixel = bytes_per_pixel;
+    src_operand.params = regs.src_params;
+    src_operand.address = regs.offset_in;
+
+    DMA::BufferOperand dst_operand;
+    dst_operand.pitch = regs.pitch_out;
+    dst_operand.width = regs.line_length_in;
+    dst_operand.height = regs.line_count;
+    dst_operand.address = regs.offset_out;
+    DMA::ImageCopy copy_info{};
+    copy_info.length_x = regs.line_length_in;
+    copy_info.length_y = regs.line_count;
+    auto& accelerate = rasterizer->AccessAccelerateDMA();
+    if (accelerate.ImageToBuffer(copy_info, src_operand, dst_operand)) {
         return;
     }
 
+    UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0);
+    UNIMPLEMENTED_IF(regs.src_params.block_size.depth != 0);
+    UNIMPLEMENTED_IF(regs.src_params.block_size.depth == 0 && regs.src_params.depth != 1);
+
     // Deswizzle the input and copy it over.
-    const Parameters& src_params = regs.src_params;
+    const DMA::Parameters& src_params = regs.src_params;
+
+    const bool is_remapping = regs.launch_dma.remap_enable != 0;
 
     const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
     const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;
@@ -187,7 +210,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
         x_offset >>= bpp_shift;
     }
 
-    const u32 bytes_per_pixel = base_bpp << bpp_shift;
+    bytes_per_pixel = base_bpp << bpp_shift;
     const u32 height = src_params.height;
     const u32 depth = src_params.depth;
     const u32 block_height = src_params.block_size.height;
@@ -195,11 +218,12 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
     const size_t src_size =
         CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
 
+    const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
     read_buffer.resize_destructive(src_size);
     write_buffer.resize_destructive(dst_size);
 
-    memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
-    memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    memory_manager.ReadBlock(src_operand.address, read_buffer.data(), src_size);
+    memory_manager.ReadBlockUnsafe(dst_operand.address, write_buffer.data(), dst_size);
 
     UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
                      src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
@@ -216,6 +240,24 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
     const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
     const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;
 
+    u32 bytes_per_pixel = 1;
+    DMA::ImageOperand dst_operand;
+    dst_operand.bytes_per_pixel = bytes_per_pixel;
+    dst_operand.params = regs.dst_params;
+    dst_operand.address = regs.offset_out;
+    DMA::BufferOperand src_operand;
+    src_operand.pitch = regs.pitch_in;
+    src_operand.width = regs.line_length_in;
+    src_operand.height = regs.line_count;
+    src_operand.address = regs.offset_in;
+    DMA::ImageCopy copy_info{};
+    copy_info.length_x = regs.line_length_in;
+    copy_info.length_y = regs.line_count;
+    auto& accelerate = rasterizer->AccessAccelerateDMA();
+    if (accelerate.BufferToImage(copy_info, src_operand, dst_operand)) {
+        return;
+    }
+
     const auto& dst_params = regs.dst_params;
 
     const u32 base_bpp = !is_remapping ? 1U : num_remap_components * remap_components_size;
@@ -233,7 +275,7 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
         x_offset >>= bpp_shift;
     }
 
-    const u32 bytes_per_pixel = base_bpp << bpp_shift;
+    bytes_per_pixel = base_bpp << bpp_shift;
     const u32 height = dst_params.height;
     const u32 depth = dst_params.depth;
     const u32 block_height = dst_params.block_size.height;
@@ -260,45 +302,14 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
     memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
 }
 
-void MaxwellDMA::FastCopyBlockLinearToPitch() {
-    const u32 bytes_per_pixel = 1U;
-    const size_t src_size = GOB_SIZE;
-    const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
-    u32 pos_x = regs.src_params.origin.x;
-    u32 pos_y = regs.src_params.origin.y;
-    const u64 offset = GetGOBOffset(regs.src_params.width, regs.src_params.height, pos_x, pos_y,
-                                    regs.src_params.block_size.height, bytes_per_pixel);
-    const u32 x_in_gob = 64 / bytes_per_pixel;
-    pos_x = pos_x % x_in_gob;
-    pos_y = pos_y % 8;
-
-    read_buffer.resize_destructive(src_size);
-    write_buffer.resize_destructive(dst_size);
-
-    if (Settings::IsGPULevelExtreme()) {
-        memory_manager.ReadBlock(regs.offset_in + offset, read_buffer.data(), src_size);
-        memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
-    } else {
-        memory_manager.ReadBlockUnsafe(regs.offset_in + offset, read_buffer.data(), src_size);
-        memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
-    }
-
-    UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, regs.src_params.width,
-                     regs.src_params.height, 1, pos_x, pos_y, regs.line_length_in, regs.line_count,
-                     regs.src_params.block_size.height, regs.src_params.block_size.depth,
-                     regs.pitch_out);
-
-    memory_manager.WriteBlockCached(regs.offset_out, write_buffer.data(), dst_size);
-}
-
 void MaxwellDMA::CopyBlockLinearToBlockLinear() {
     UNIMPLEMENTED_IF(regs.src_params.block_size.width != 0);
 
     const bool is_remapping = regs.launch_dma.remap_enable != 0;
 
     // Deswizzle the input and copy it over.
-    const Parameters& src = regs.src_params;
-    const Parameters& dst = regs.dst_params;
+    const DMA::Parameters& src = regs.src_params;
+    const DMA::Parameters& dst = regs.dst_params;
 
     const u32 num_remap_components = regs.remap_const.num_dst_components_minus_one + 1;
     const u32 remap_components_size = regs.remap_const.component_size_minus_one + 1;

src/video_core/engines/maxwell_dma.h

@@ -24,6 +24,54 @@ namespace VideoCore {
 class RasterizerInterface;
 }
 
+namespace Tegra {
+namespace DMA {
+
+union Origin {
+    BitField<0, 16, u32> x;
+    BitField<16, 16, u32> y;
+};
+static_assert(sizeof(Origin) == 4);
+
+struct ImageCopy {
+    u32 length_x{};
+    u32 length_y{};
+};
+
+union BlockSize {
+    BitField<0, 4, u32> width;
+    BitField<4, 4, u32> height;
+    BitField<8, 4, u32> depth;
+    BitField<12, 4, u32> gob_height;
+};
+static_assert(sizeof(BlockSize) == 4);
+
+struct Parameters {
+    BlockSize block_size;
+    u32 width;
+    u32 height;
+    u32 depth;
+    u32 layer;
+    Origin origin;
+};
+static_assert(sizeof(Parameters) == 24);
+
+struct ImageOperand {
+    u32 bytes_per_pixel;
+    Parameters params;
+    GPUVAddr address;
+};
+
+struct BufferOperand {
+    u32 pitch;
+    u32 width;
+    u32 height;
+    GPUVAddr address;
+};
+
+} // namespace DMA
+} // namespace Tegra
+
 namespace Tegra::Engines {
 
 class AccelerateDMAInterface {
@@ -32,6 +80,12 @@ public:
     virtual bool BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64 amount) = 0;
 
     virtual bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) = 0;
+
+    virtual bool ImageToBuffer(const DMA::ImageCopy& copy_info, const DMA::ImageOperand& src,
+                               const DMA::BufferOperand& dst) = 0;
+
+    virtual bool BufferToImage(const DMA::ImageCopy& copy_info, const DMA::BufferOperand& src,
+                               const DMA::ImageOperand& dst) = 0;
 };
 
 /**
@@ -51,30 +105,6 @@ public:
         }
     };
 
-    union BlockSize {
-        BitField<0, 4, u32> width;
-        BitField<4, 4, u32> height;
-        BitField<8, 4, u32> depth;
-        BitField<12, 4, u32> gob_height;
-    };
-    static_assert(sizeof(BlockSize) == 4);
-
-    union Origin {
-        BitField<0, 16, u32> x;
-        BitField<16, 16, u32> y;
-    };
-    static_assert(sizeof(Origin) == 4);
-
-    struct Parameters {
-        BlockSize block_size;
-        u32 width;
-        u32 height;
-        u32 depth;
-        u32 layer;
-        Origin origin;
-    };
-    static_assert(sizeof(Parameters) == 24);
-
     struct Semaphore {
         PackedGPUVAddr address;
         u32 payload;
@@ -227,8 +257,6 @@ private:
 
     void CopyBlockLinearToBlockLinear();
 
-    void FastCopyBlockLinearToPitch();
-
     void ReleaseSemaphore();
 
     void ConsumeSinkImpl() override;
@@ -261,17 +289,17 @@ private:
                 u32 reserved05[0x3f];
                 PackedGPUVAddr offset_in;
                 PackedGPUVAddr offset_out;
-                u32 pitch_in;
-                u32 pitch_out;
+                s32 pitch_in;
+                s32 pitch_out;
                 u32 line_length_in;
                 u32 line_count;
                 u32 reserved06[0xb6];
                 u32 remap_consta_value;
                 u32 remap_constb_value;
                 RemapConst remap_const;
-                Parameters dst_params;
+                DMA::Parameters dst_params;
                 u32 reserved07[0x1];
-                Parameters src_params;
+                DMA::Parameters src_params;
                 u32 reserved08[0x275];
                 u32 pm_trigger_end;
                 u32 reserved09[0x3ba];

src/video_core/renderer_null/null_rasterizer.h

@@ -22,6 +22,14 @@ public:
     explicit AccelerateDMA();
     bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override;
     bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
+    bool ImageToBuffer(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::ImageOperand& src,
+                       const Tegra::DMA::BufferOperand& dst) override {
+        return false;
+    }
+    bool BufferToImage(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::BufferOperand& src,
+                       const Tegra::DMA::ImageOperand& dst) override {
+        return false;
+    }
 };
 
 class RasterizerNull final : public VideoCore::RasterizerAccelerated,

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -56,6 +56,16 @@ public:
 
     bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
 
+    bool ImageToBuffer(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::ImageOperand& src,
+                       const Tegra::DMA::BufferOperand& dst) override {
+        return false;
+    }
+
+    bool BufferToImage(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::BufferOperand& src,
+                       const Tegra::DMA::ImageOperand& dst) override {
+        return false;
+    }
+
 private:
     BufferCache& buffer_cache;
 };

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -172,7 +172,7 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
       buffer_cache(*this, cpu_memory_, buffer_cache_runtime),
       pipeline_cache(*this, device, scheduler, descriptor_pool, update_descriptor_queue,
                      render_pass_cache, buffer_cache, texture_cache, gpu.ShaderNotify()),
-      query_cache{*this, device, scheduler}, accelerate_dma{buffer_cache},
+      query_cache{*this, device, scheduler}, accelerate_dma(buffer_cache, texture_cache, scheduler),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()) {
     scheduler.SetQueryCache(query_cache);
@@ -756,7 +756,9 @@ void RasterizerVulkan::FlushWork() {
     draw_counter = 0;
 }
 
-AccelerateDMA::AccelerateDMA(BufferCache& buffer_cache_) : buffer_cache{buffer_cache_} {}
+AccelerateDMA::AccelerateDMA(BufferCache& buffer_cache_, TextureCache& texture_cache_,
+                             Scheduler& scheduler_)
+    : buffer_cache{buffer_cache_}, texture_cache{texture_cache_}, scheduler{scheduler_} {}
 
 bool AccelerateDMA::BufferClear(GPUVAddr src_address, u64 amount, u32 value) {
     std::scoped_lock lock{buffer_cache.mutex};
@@ -768,6 +770,234 @@ bool AccelerateDMA::BufferCopy(GPUVAddr src_address, GPUVAddr dest_address, u64
     return buffer_cache.DMACopy(src_address, dest_address, amount);
 }
 
+bool AccelerateDMA::ImageToBuffer(const Tegra::DMA::ImageCopy& copy_info,
+                                  const Tegra::DMA::ImageOperand& src,
+                                  const Tegra::DMA::BufferOperand& dst) {
+    std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
+    auto query_image = texture_cache.ObtainImage(src, false);
+    if (!query_image) {
+        return false;
+    }
+    auto* image = query_image->first;
+    auto [level, base] = query_image->second;
+    const u32 buffer_size = static_cast<u32>(dst.pitch * dst.height);
+    const auto [buffer, offset] = buffer_cache.ObtainBuffer(
+        dst.address, buffer_size, VideoCommon::ObtainBufferSynchronize::FullSynchronize,
+        VideoCommon::ObtainBufferOperation::MarkAsWritten);
+
+    const bool is_rescaled = image->IsRescaled();
+    if (is_rescaled) {
+        image->ScaleDown();
+    }
+    VkImageSubresourceLayers subresources{
+        .aspectMask = image->AspectMask(),
+        .mipLevel = level,
+        .baseArrayLayer = base,
+        .layerCount = 1,
+    };
+    const u32 bpp = VideoCore::Surface::BytesPerBlock(image->info.format);
+    const auto convert = [old_bpp = src.bytes_per_pixel, bpp](u32 value) {
+        return (old_bpp * value) / bpp;
+    };
+    const u32 base_x = convert(src.params.origin.x.Value());
+    const u32 base_y = src.params.origin.y.Value();
+    const u32 length_x = convert(copy_info.length_x);
+    const u32 length_y = copy_info.length_y;
+    VkOffset3D image_offset{
+        .x = static_cast<s32>(base_x),
+        .y = static_cast<s32>(base_y),
+        .z = 0,
+    };
+    VkExtent3D image_extent{
+        .width = length_x,
+        .height = length_y,
+        .depth = 1,
+    };
+    auto buff_info(dst);
+    buff_info.pitch = convert(dst.pitch);
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([src_image = image->Handle(), dst_buffer = buffer->Handle(),
+                      buffer_offset = offset, subresources, image_offset, image_extent,
+                      buff_info](vk::CommandBuffer cmdbuf) {
+        const std::array buffer_copy_info{
+            VkBufferImageCopy{
+                .bufferOffset = buffer_offset,
+                .bufferRowLength = buff_info.pitch,
+                .bufferImageHeight = buff_info.height,
+                .imageSubresource = subresources,
+                .imageOffset = image_offset,
+                .imageExtent = image_extent,
+            },
+        };
+        const VkImageSubresourceRange range{
+            .aspectMask = subresources.aspectMask,
+            .baseMipLevel = subresources.mipLevel,
+            .levelCount = 1,
+            .baseArrayLayer = subresources.baseArrayLayer,
+            .layerCount = 1,
+        };
+        static constexpr VkMemoryBarrier WRITE_BARRIER{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_MEMORY_READ_BIT | VK_ACCESS_MEMORY_WRITE_BIT,
+        };
+        const std::array pre_barriers{
+            VkImageMemoryBarrier{
+                .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
+                                 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
+                                 VK_ACCESS_TRANSFER_WRITE_BIT,
+                .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
+                .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+                .newLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .image = src_image,
+                .subresourceRange = range,
+            },
+        };
+        const std::array post_barriers{
+            VkImageMemoryBarrier{
+                .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = 0,
+                .dstAccessMask = 0,
+                .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL,
+                .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .image = src_image,
+                .subresourceRange = range,
+            },
+        };
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                               0, {}, {}, pre_barriers);
+        cmdbuf.CopyImageToBuffer(src_image, VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL, dst_buffer,
+                                 buffer_copy_info);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                               0, WRITE_BARRIER, nullptr, post_barriers);
+    });
+    if (is_rescaled) {
+        image->ScaleUp(true);
+    }
+    return true;
+}
+
+bool AccelerateDMA::BufferToImage(const Tegra::DMA::ImageCopy& copy_info,
+                                  const Tegra::DMA::BufferOperand& src,
+                                  const Tegra::DMA::ImageOperand& dst) {
+    std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
+    auto query_image = texture_cache.ObtainImage(dst, true);
+    if (!query_image) {
+        return false;
+    }
+    auto* image = query_image->first;
+    auto [level, base] = query_image->second;
+    const u32 buffer_size = static_cast<u32>(src.pitch * src.height);
+    const auto [buffer, offset] = buffer_cache.ObtainBuffer(
+        src.address, buffer_size, VideoCommon::ObtainBufferSynchronize::FullSynchronize,
+        VideoCommon::ObtainBufferOperation::DoNothing);
+    const bool is_rescaled = image->IsRescaled();
+    if (is_rescaled) {
+        image->ScaleDown(true);
+    }
+    VkImageSubresourceLayers subresources{
+        .aspectMask = image->AspectMask(),
+        .mipLevel = level,
+        .baseArrayLayer = base,
+        .layerCount = 1,
+    };
+    const u32 bpp = VideoCore::Surface::BytesPerBlock(image->info.format);
+    const auto convert = [old_bpp = dst.bytes_per_pixel, bpp](u32 value) {
+        return (old_bpp * value) / bpp;
+    };
+    const u32 base_x = convert(dst.params.origin.x.Value());
+    const u32 base_y = dst.params.origin.y.Value();
+    const u32 length_x = convert(copy_info.length_x);
+    const u32 length_y = copy_info.length_y;
+    VkOffset3D image_offset{
+        .x = static_cast<s32>(base_x),
+        .y = static_cast<s32>(base_y),
+        .z = 0,
+    };
+    VkExtent3D image_extent{
+        .width = length_x,
+        .height = length_y,
+        .depth = 1,
+    };
+    auto buff_info(src);
+    buff_info.pitch = convert(src.pitch);
+    scheduler.RequestOutsideRenderPassOperationContext();
+    scheduler.Record([dst_image = image->Handle(), src_buffer = buffer->Handle(),
+                      buffer_offset = offset, subresources, image_offset, image_extent,
+                      buff_info](vk::CommandBuffer cmdbuf) {
+        const std::array buffer_copy_info{
+            VkBufferImageCopy{
+                .bufferOffset = buffer_offset,
+                .bufferRowLength = buff_info.pitch,
+                .bufferImageHeight = buff_info.height,
+                .imageSubresource = subresources,
+                .imageOffset = image_offset,
+                .imageExtent = image_extent,
+            },
+        };
+        const VkImageSubresourceRange range{
+            .aspectMask = subresources.aspectMask,
+            .baseMipLevel = subresources.mipLevel,
+            .levelCount = 1,
+            .baseArrayLayer = subresources.baseArrayLayer,
+            .layerCount = 1,
+        };
+        static constexpr VkMemoryBarrier READ_BARRIER{
+            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_MEMORY_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT | VK_ACCESS_TRANSFER_WRITE_BIT,
+        };
+        const std::array pre_barriers{
+            VkImageMemoryBarrier{
+                .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
+                                 VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
+                                 VK_ACCESS_TRANSFER_WRITE_BIT,
+                .dstAccessMask = VK_ACCESS_TRANSFER_READ_BIT,
+                .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+                .newLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .image = dst_image,
+                .subresourceRange = range,
+            },
+        };
+        const std::array post_barriers{
+            VkImageMemoryBarrier{
+                .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+                .pNext = nullptr,
+                .srcAccessMask = 0,
+                .dstAccessMask = 0,
+                .oldLayout = VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL,
+                .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+                .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+                .image = dst_image,
+                .subresourceRange = range,
+            },
+        };
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_ALL_COMMANDS_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT,
+                               0, READ_BARRIER, {}, pre_barriers);
+        cmdbuf.CopyBufferToImage(src_buffer, dst_image, VK_IMAGE_LAYOUT_GENERAL, buffer_copy_info);
+        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
+                               0, nullptr, nullptr, post_barriers);
+    });
+    if (is_rescaled) {
+        image->ScaleUp();
+    }
+    return true;
+}
+
 void RasterizerVulkan::UpdateDynamicStates() {
     auto& regs = maxwell3d->regs;
     UpdateViewportsState(regs);

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -45,14 +45,23 @@ class StateTracker;
 
 class AccelerateDMA : public Tegra::Engines::AccelerateDMAInterface {
 public:
-    explicit AccelerateDMA(BufferCache& buffer_cache);
+    explicit AccelerateDMA(BufferCache& buffer_cache, TextureCache& texture_cache,
+                           Scheduler& scheduler);
 
     bool BufferCopy(GPUVAddr start_address, GPUVAddr end_address, u64 amount) override;
 
     bool BufferClear(GPUVAddr src_address, u64 amount, u32 value) override;
 
+    bool ImageToBuffer(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::ImageOperand& src,
+                       const Tegra::DMA::BufferOperand& dst) override;
+
+    bool BufferToImage(const Tegra::DMA::ImageCopy& copy_info, const Tegra::DMA::BufferOperand& src,
+                       const Tegra::DMA::ImageOperand& dst) override;
+
 private:
     BufferCache& buffer_cache;
+    TextureCache& texture_cache;
+    Scheduler& scheduler;
 };
 
 class RasterizerVulkan final : public VideoCore::RasterizerAccelerated,

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -864,13 +864,19 @@ void TextureCacheRuntime::ReinterpretImage(Image& dst, Image& src,
     const VkImageAspectFlags src_aspect_mask = src.AspectMask();
     const VkImageAspectFlags dst_aspect_mask = dst.AspectMask();
 
-    std::ranges::transform(copies, vk_in_copies.begin(), [src_aspect_mask](const auto& copy) {
-        return MakeBufferImageCopy(copy, true, src_aspect_mask);
+    const auto bpp_in = BytesPerBlock(src.info.format) / DefaultBlockWidth(src.info.format);
+    const auto bpp_out = BytesPerBlock(dst.info.format) / DefaultBlockWidth(dst.info.format);
+    std::ranges::transform(copies, vk_in_copies.begin(),
+                           [src_aspect_mask, bpp_in, bpp_out](const auto& copy) {
+                               auto copy2 = copy;
+                               copy2.src_offset.x = (bpp_out * copy.src_offset.x) / bpp_in;
+                               copy2.extent.width = (bpp_out * copy.extent.width) / bpp_in;
+                               return MakeBufferImageCopy(copy2, true, src_aspect_mask);
                            });
     std::ranges::transform(copies, vk_out_copies.begin(), [dst_aspect_mask](const auto& copy) {
         return MakeBufferImageCopy(copy, false, dst_aspect_mask);
     });
-    const u32 img_bpp = BytesPerBlock(src.info.format);
+    const u32 img_bpp = BytesPerBlock(dst.info.format);
     size_t total_size = 0;
     for (const auto& copy : copies) {
         total_size += copy.extent.width * copy.extent.height * copy.extent.depth * img_bpp;

src/video_core/texture_cache/image_info.cpp

@@ -216,10 +216,51 @@ ImageInfo::ImageInfo(const Tegra::Engines::Fermi2D::Surface& config) noexcept {
             .height = config.height,
             .depth = 1,
         };
-        rescaleable = block.depth == 0;
-        rescaleable &= size.height > 256;
+        rescaleable = block.depth == 0 && size.height > 256;
         downscaleable = size.height > 512;
     }
 }
 
+static PixelFormat ByteSizeToFormat(u32 bytes_per_pixel) {
+    switch (bytes_per_pixel) {
+    case 1:
+        return PixelFormat::R8_UINT;
+    case 2:
+        return PixelFormat::R8G8_UINT;
+    case 4:
+        return PixelFormat::A8B8G8R8_UINT;
+    case 8:
+        return PixelFormat::R16G16B16A16_UINT;
+    case 16:
+        return PixelFormat::R32G32B32A32_UINT;
+    default:
+        UNIMPLEMENTED();
+        return PixelFormat::Invalid;
+    }
+}
+
+ImageInfo::ImageInfo(const Tegra::DMA::ImageOperand& config) noexcept {
+    const u32 bytes_per_pixel = config.bytes_per_pixel;
+    format = ByteSizeToFormat(bytes_per_pixel);
+    type = config.params.block_size.depth > 0 ? ImageType::e3D : ImageType::e2D;
+    num_samples = 1;
+    block = Extent3D{
+        .width = config.params.block_size.width,
+        .height = config.params.block_size.height,
+        .depth = config.params.block_size.depth,
+    };
+    size = Extent3D{
+        .width = config.params.width,
+        .height = config.params.height,
+        .depth = config.params.depth,
+    };
+    tile_width_spacing = 0;
+    resources.levels = 1;
+    resources.layers = 1;
+    layer_stride = CalculateLayerStride(*this);
+    maybe_unaligned_layer_stride = CalculateLayerSize(*this);
+    rescaleable = block.depth == 0 && size.height > 256;
+    downscaleable = size.height > 512;
+}
+
 } // namespace VideoCommon

src/video_core/texture_cache/image_info.h

@@ -5,6 +5,7 @@
 
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/engines/maxwell_3d.h"
+#include "video_core/engines/maxwell_dma.h"
 #include "video_core/surface.h"
 #include "video_core/texture_cache/types.h"
 
@@ -19,6 +20,7 @@ struct ImageInfo {
     explicit ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs, size_t index) noexcept;
     explicit ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs) noexcept;
     explicit ImageInfo(const Tegra::Engines::Fermi2D::Surface& config) noexcept;
+    explicit ImageInfo(const Tegra::DMA::ImageOperand& config) noexcept;
 
     PixelFormat format = PixelFormat::Invalid;
     ImageType type = ImageType::e1D;

src/video_core/texture_cache/texture_cache.h

@@ -1358,6 +1358,75 @@ std::optional<typename TextureCache<P>::BlitImages> TextureCache<P>::GetBlitImag
     }};
 }
 
+template <class P>
+ImageId TextureCache<P>::FindDMAImage(const ImageInfo& info, GPUVAddr gpu_addr) {
+    std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
+    if (!cpu_addr) {
+        cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr, CalculateGuestSizeInBytes(info));
+        if (!cpu_addr) {
+            return ImageId{};
+        }
+    }
+    ImageId image_id{};
+    boost::container::small_vector<ImageId, 1> image_ids;
+    const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
+        if (True(existing_image.flags & ImageFlagBits::Remapped)) {
+            return false;
+        }
+        if (info.type == ImageType::Linear || existing_image.info.type == ImageType::Linear)
+            [[unlikely]] {
+            const bool strict_size = True(existing_image.flags & ImageFlagBits::Strong);
+            const ImageInfo& existing = existing_image.info;
+            if (existing_image.gpu_addr == gpu_addr && existing.type == info.type &&
+                existing.pitch == info.pitch &&
+                IsPitchLinearSameSize(existing, info, strict_size) &&
+                IsViewCompatible(existing.format, info.format, false, true)) {
+                image_id = existing_image_id;
+                image_ids.push_back(existing_image_id);
+                return true;
+            }
+        } else if (IsSubCopy(info, existing_image, gpu_addr)) {
+            image_id = existing_image_id;
+            image_ids.push_back(existing_image_id);
+            return true;
+        }
+        return false;
+    };
+    ForEachImageInRegion(*cpu_addr, CalculateGuestSizeInBytes(info), lambda);
+    if (image_ids.size() <= 1) [[likely]] {
+        return image_id;
+    }
+    auto image_ids_compare = [this](ImageId a, ImageId b) {
+        auto& image_a = slot_images[a];
+        auto& image_b = slot_images[b];
+        return image_a.modification_tick < image_b.modification_tick;
+    };
+    return *std::ranges::max_element(image_ids, image_ids_compare);
+}
+
+template <class P>
+std::optional<std::pair<typename TextureCache<P>::Image*, std::pair<u32, u32>>>
+TextureCache<P>::ObtainImage(const Tegra::DMA::ImageOperand& operand, bool mark_as_modified) {
+    ImageInfo dst_info(operand);
+    ImageId dst_id = FindDMAImage(dst_info, operand.address);
+    if (!dst_id) {
+        return std::nullopt;
+    }
+    auto& image = slot_images[dst_id];
+    auto base = image.TryFindBase(operand.address);
+    if (!base) {
+        return std::nullopt;
+    }
+    if (False(image.flags & ImageFlagBits::GpuModified)) {
+        // No need to waste time on an image that's synced with guest
+        return std::nullopt;
+    }
+    PrepareImage(dst_id, mark_as_modified, false);
+    auto& new_image = slot_images[dst_id];
+    lru_cache.Touch(new_image.lru_index, frame_tick);
+    return std::make_pair(&new_image, std::make_pair(base->level, base->layer));
+}
+
 template <class P>
 SamplerId TextureCache<P>::FindSampler(const TSCEntry& config) {
     if (std::ranges::all_of(config.raw, [](u64 value) { return value == 0; })) {

src/video_core/texture_cache/texture_cache_base.h

@@ -209,6 +209,9 @@ public:
     /// Pop asynchronous downloads
     void PopAsyncFlushes();
 
+    [[nodiscard]] std::optional<std::pair<Image*, std::pair<u32, u32>>> ObtainImage(
+        const Tegra::DMA::ImageOperand& operand, bool mark_as_modified);
+
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
@@ -300,6 +303,8 @@ private:
     /// Remove joined images from the cache
     [[nodiscard]] ImageId JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VAddr cpu_addr);
 
+    [[nodiscard]] ImageId FindDMAImage(const ImageInfo& info, GPUVAddr gpu_addr);
+
     /// Return a blit image pair from the given guest blit parameters
     [[nodiscard]] std::optional<BlitImages> GetBlitImages(
         const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src,

src/video_core/texture_cache/types.h

@@ -54,6 +54,7 @@ enum class RelaxedOptions : u32 {
     Format = 1 << 1,
     Samples = 1 << 2,
     ForceBrokenViews = 1 << 3,
+    FormatBpp = 1 << 4,
 };
 DECLARE_ENUM_FLAG_OPERATORS(RelaxedOptions)
 

src/video_core/texture_cache/util.cpp

@@ -743,6 +743,44 @@ std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageIn
     return copies;
 }
 
+std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src, u32 up_scale,
+                                                  u32 down_shift) {
+    std::vector<ImageCopy> copies;
+    copies.reserve(src.resources.levels);
+    const bool is_3d = src.type == ImageType::e3D;
+    for (s32 level = 0; level < src.resources.levels; ++level) {
+        ImageCopy& copy = copies.emplace_back();
+        copy.src_subresource = SubresourceLayers{
+            .base_level = level,
+            .base_layer = 0,
+            .num_layers = src.resources.layers,
+        };
+        copy.dst_subresource = SubresourceLayers{
+            .base_level = level,
+            .base_layer = 0,
+            .num_layers = src.resources.layers,
+        };
+        copy.src_offset = Offset3D{
+            .x = 0,
+            .y = 0,
+            .z = 0,
+        };
+        copy.dst_offset = Offset3D{
+            .x = 0,
+            .y = 0,
+            .z = 0,
+        };
+        const Extent3D mip_size = AdjustMipSize(src.size, level);
+        copy.extent = AdjustSamplesSize(mip_size, src.num_samples);
+        if (is_3d) {
+            copy.extent.depth = src.size.depth;
+        }
+        copy.extent.width = std::max<u32>((copy.extent.width * up_scale) >> down_shift, 1);
+        copy.extent.height = std::max<u32>((copy.extent.height * up_scale) >> down_shift, 1);
+    }
+    return copies;
+}
+
 bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config) {
     const GPUVAddr address = config.Address();
     if (address == 0) {
@@ -999,6 +1037,20 @@ bool IsBlockLinearSizeCompatible(const ImageInfo& lhs, const ImageInfo& rhs, u32
     }
 }
 
+bool IsBlockLinearSizeCompatibleBPPRelaxed(const ImageInfo& lhs, const ImageInfo& rhs,
+                                           u32 lhs_level, u32 rhs_level) noexcept {
+    ASSERT(lhs.type != ImageType::Linear);
+    ASSERT(rhs.type != ImageType::Linear);
+    const auto lhs_bpp = BytesPerBlock(lhs.format);
+    const auto rhs_bpp = BytesPerBlock(rhs.format);
+    const Extent3D lhs_size = AdjustMipSize(lhs.size, lhs_level);
+    const Extent3D rhs_size = AdjustMipSize(rhs.size, rhs_level);
+    return Common::AlignUpLog2(lhs_size.width * lhs_bpp, GOB_SIZE_X_SHIFT) ==
+               Common::AlignUpLog2(rhs_size.width * rhs_bpp, GOB_SIZE_X_SHIFT) &&
+           Common::AlignUpLog2(lhs_size.height, GOB_SIZE_Y_SHIFT) ==
+               Common::AlignUpLog2(rhs_size.height, GOB_SIZE_Y_SHIFT);
+}
+
 bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs, bool strict_size) noexcept {
     ASSERT(lhs.type == ImageType::Linear);
     ASSERT(rhs.type == ImageType::Linear);
@@ -1073,7 +1125,8 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
         // Format checking is relaxed, but we still have to check for matching bytes per block.
         // This avoids creating a view for blits on UE4 titles where formats with different bytes
         // per block are aliased.
-        if (BytesPerBlock(existing.format) != BytesPerBlock(candidate.format)) {
+        if (BytesPerBlock(existing.format) != BytesPerBlock(candidate.format) &&
+            False(options & RelaxedOptions::FormatBpp)) {
             return std::nullopt;
         }
     } else {
@@ -1088,11 +1141,9 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
     if (existing.type != candidate.type) {
         return std::nullopt;
     }
-    if (False(options & RelaxedOptions::Samples)) {
-        if (existing.num_samples != candidate.num_samples) {
+    if (False(options & RelaxedOptions::Samples) && existing.num_samples != candidate.num_samples) {
         return std::nullopt;
     }
-    }
     if (existing.resources.levels < candidate.resources.levels + base->level) {
         return std::nullopt;
     }
@@ -1101,15 +1152,17 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
         if (mip_depth < candidate.size.depth + base->layer) {
             return std::nullopt;
         }
-    } else {
-        if (existing.resources.layers < candidate.resources.layers + base->layer) {
+    } else if (existing.resources.layers < candidate.resources.layers + base->layer) {
         return std::nullopt;
     }
-    }
     const bool strict_size = False(options & RelaxedOptions::Size);
     if (!IsBlockLinearSizeCompatible(existing, candidate, base->level, 0, strict_size)) {
+        if (False(options & RelaxedOptions::FormatBpp)) {
+            return std::nullopt;
+        } else if (!IsBlockLinearSizeCompatibleBPPRelaxed(existing, candidate, base->level, 0)) {
             return std::nullopt;
         }
+    }
     // TODO: compare block sizes
     return base;
 }
@@ -1120,6 +1173,31 @@ bool IsSubresource(const ImageInfo& candidate, const ImageBase& image, GPUVAddr
         .has_value();
 }
 
+bool IsSubCopy(const ImageInfo& candidate, const ImageBase& image, GPUVAddr candidate_addr) {
+    const std::optional<SubresourceBase> base = image.TryFindBase(candidate_addr);
+    if (!base) {
+        return false;
+    }
+    const ImageInfo& existing = image.info;
+    if (existing.resources.levels < candidate.resources.levels + base->level) {
+        return false;
+    }
+    if (existing.type == ImageType::e3D) {
+        const u32 mip_depth = std::max(1U, existing.size.depth << base->level);
+        if (mip_depth < candidate.size.depth + base->layer) {
+            return false;
+        }
+    } else {
+        if (existing.resources.layers < candidate.resources.layers + base->layer) {
+            return false;
+        }
+    }
+    if (!IsBlockLinearSizeCompatibleBPPRelaxed(existing, candidate, base->level, 0)) {
+        return false;
+    }
+    return true;
+}
+
 void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
                       const ImageBase* src) {
     const auto original_dst_format = dst_info.format;

src/video_core/texture_cache/util.h

@@ -56,6 +56,10 @@ struct OverlapResult {
                                                            SubresourceBase base, u32 up_scale = 1,
                                                            u32 down_shift = 0);
 
+[[nodiscard]] std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src,
+                                                                u32 up_scale = 1,
+                                                                u32 down_shift = 0);
+
 [[nodiscard]] bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config);
 
 [[nodiscard]] std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory,
@@ -88,6 +92,9 @@ void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const Ima
 [[nodiscard]] bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs,
                                          bool strict_size) noexcept;
 
+[[nodiscard]] bool IsBlockLinearSizeCompatibleBPPRelaxed(const ImageInfo& lhs, const ImageInfo& rhs,
+                                                         u32 lhs_level, u32 rhs_level) noexcept;
+
 [[nodiscard]] std::optional<OverlapResult> ResolveOverlap(const ImageInfo& new_info,
                                                           GPUVAddr gpu_addr, VAddr cpu_addr,
                                                           const ImageBase& overlap,
@@ -106,6 +113,9 @@ void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const Ima
                                  GPUVAddr candidate_addr, RelaxedOptions options, bool broken_views,
                                  bool native_bgr);
 
+[[nodiscard]] bool IsSubCopy(const ImageInfo& candidate, const ImageBase& image,
+                             GPUVAddr candidate_addr);
+
 void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
                       const ImageBase* src);