Add instance draw counts to vertex buffers

This reverts commit 2ea449f71a.

src/video_core/buffer_cache/buffer_base.h

@@ -20,8 +20,6 @@ namespace VideoCommon {
 
 enum class BufferFlagBits {
     Picked = 1 << 0,
-    CachedWrites = 1 << 1,
-    PreemtiveDownload = 1 << 2,
 };
 DECLARE_ENUM_FLAG_OPERATORS(BufferFlagBits)
 
@@ -58,15 +56,6 @@ public:
         flags |= BufferFlagBits::Picked;
     }
 
-    void MarkPreemtiveDownload() noexcept {
-        flags |= BufferFlagBits::PreemtiveDownload;
-    }
-
-    /// Unmark buffer as picked
-    void Unpick() noexcept {
-        flags &= ~BufferFlagBits::Picked;
-    }
-
     /// Increases the likeliness of this being a stream buffer
     void IncreaseStreamScore(int score) noexcept {
         stream_score += score;
@@ -87,15 +76,6 @@ public:
         return True(flags & BufferFlagBits::Picked);
     }
 
-    /// Returns true when the buffer has pending cached writes
-    [[nodiscard]] bool HasCachedWrites() const noexcept {
-        return True(flags & BufferFlagBits::CachedWrites);
-    }
-
-    bool IsPreemtiveDownload() const noexcept {
-        return True(flags & BufferFlagBits::PreemtiveDownload);
-    }
-
     /// Returns the base CPU address of the buffer
     [[nodiscard]] VAddr CpuAddr() const noexcept {
         return cpu_addr;

src/video_core/buffer_cache/buffer_cache.h

@@ -7,6 +7,7 @@
 #pragma once
 
 #include <algorithm>
+#include <limits>
 #include <memory>
 #include <numeric>
 
@@ -121,25 +122,6 @@ void BufferCache<P>::WriteMemory(DAddr device_addr, u64 size) {
     memory_tracker.MarkRegionAsCpuModified(device_addr, size);
 }
 
-template <class P>
-void BufferCache<P>::CachedWriteMemory(DAddr device_addr, u64 size) {
-    const bool is_dirty = IsRegionRegistered(device_addr, size);
-    if (!is_dirty) {
-        return;
-    }
-    DAddr aligned_start = Common::AlignDown(device_addr, DEVICE_PAGESIZE);
-    DAddr aligned_end = Common::AlignUp(device_addr + size, DEVICE_PAGESIZE);
-    if (!IsRegionGpuModified(aligned_start, aligned_end - aligned_start)) {
-        WriteMemory(device_addr, size);
-        return;
-    }
-
-    tmp_buffer.resize_destructive(size);
-    device_memory.ReadBlockUnsafe(device_addr, tmp_buffer.data(), size);
-
-    InlineMemoryImplementation(device_addr, size, tmp_buffer);
-}
-
 template <class P>
 bool BufferCache<P>::OnCPUWrite(DAddr device_addr, u64 size) {
     const bool is_dirty = IsRegionRegistered(device_addr, size);
@@ -422,7 +404,7 @@ void BufferCache<P>::UnbindGraphicsStorageBuffers(size_t stage) {
 }
 
 template <class P>
-bool BufferCache<P>::BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, u32 cbuf_index,
+void BufferCache<P>::BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, u32 cbuf_index,
                                                u32 cbuf_offset, bool is_written) {
     const bool already_enabled =
         ((channel_state->enabled_storage_buffers[stage] >> ssbo_index) & 1U) != 0;
@@ -433,7 +415,7 @@ bool BufferCache<P>::BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index,
                 LOG_WARNING(HW_GPU,
                             "Skipping graphics storage buffer {} due to driver limit {}",
                             ssbo_index, max_bindings);
-                return false;
+                return;
             }
         }
     }
@@ -449,7 +431,6 @@ bool BufferCache<P>::BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index,
     const GPUVAddr ssbo_addr = cbufs.const_buffers[cbuf_index].address + cbuf_offset;
     channel_state->storage_buffers[stage][ssbo_index] =
         StorageBufferBinding(ssbo_addr, cbuf_index, is_written);
-    return (channel_state->storage_buffers[stage][ssbo_index].buffer_id != NULL_BUFFER_ID);
 }
 
 template <class P>
@@ -762,16 +743,6 @@ void BufferCache<P>::BindHostIndexBuffer() {
     }
 }
 
-template <class P>
-void BufferCache<P>::BindHostVertexBuffer(u32 index, Buffer& buffer, u32 offset, u32 size,
-                                          u32 stride) {
-    if constexpr (IS_OPENGL) {
-        runtime.BindVertexBuffer(index, buffer, offset, size, stride);
-    } else {
-        runtime.BindVertexBuffer(index, buffer.Handle(), offset, size, stride);
-    }
-}
-
 template <class P>
 Binding& BufferCache<P>::VertexBufferSlot(u32 index) {
     ASSERT(index < NUM_VERTEX_BUFFERS);
@@ -1251,9 +1222,14 @@ void BufferCache<P>::UpdateIndexBuffer() {
     const GPUVAddr gpu_addr_begin = index_buffer_ref.StartAddress();
     const GPUVAddr gpu_addr_end = index_buffer_ref.EndAddress();
     const std::optional<DAddr> device_addr = gpu_memory->GpuToCpuAddress(gpu_addr_begin);
-    const u32 address_size = static_cast<u32>(gpu_addr_end - gpu_addr_begin);
-    const u32 draw_size = (index_buffer_ref.count + index_buffer_ref.first) * u32(index_buffer_ref.FormatSizeInBytes());
-    const u32 size = (std::min)(address_size, draw_size);
+    u64 address_size = 0;
+    if (gpu_addr_end > gpu_addr_begin) {
+        address_size = (std::min)(gpu_addr_end - gpu_addr_begin,
+                                  u64{(std::numeric_limits<u32>::max)()});
+    }
+    const u64 draw_size = (u64{index_buffer_ref.count} + u64{index_buffer_ref.first}) *
+                          u64{index_buffer_ref.FormatSizeInBytes()};
+    const u32 size = static_cast<u32>((std::min)(address_size, draw_size));
     if (size == 0 || !device_addr) {
         channel_state->index_buffer = NULL_BINDING;
         return;
@@ -1268,6 +1244,20 @@ void BufferCache<P>::UpdateIndexBuffer() {
 template <class P>
 void BufferCache<P>::UpdateVertexBuffers() {
     auto& flags = maxwell3d->dirty.flags;
+    const u32 base_instance = maxwell3d->draw_manager.draw_state.base_instance;
+    if (draw_instance_count != last_draw_instance_count ||
+        base_instance != last_draw_base_instance) {
+        last_draw_instance_count = draw_instance_count;
+        last_draw_base_instance = base_instance;
+        const auto& instances = maxwell3d->regs.vertex_stream_instances;
+        for (u32 index = 0; index < NUM_VERTEX_BUFFERS; ++index) {
+            if (!instances.IsInstancingEnabled(index)) {
+                continue;
+            }
+            flags[Dirty::VertexBuffer0 + index] = true;
+            flags[Dirty::VertexBuffers] = true;
+        }
+    }
     if (!maxwell3d->dirty.flags[Dirty::VertexBuffers]) {
         return;
     }
@@ -1288,15 +1278,37 @@ void BufferCache<P>::UpdateVertexBuffer(u32 index) {
     const GPUVAddr gpu_addr_begin = array.Address();
     const GPUVAddr gpu_addr_end = limit.Address() + 1;
     const std::optional<DAddr> device_addr = gpu_memory->GpuToCpuAddress(gpu_addr_begin);
-    const u32 address_size = static_cast<u32>(gpu_addr_end - gpu_addr_begin);
-    u32 size = address_size; // TODO: Analyze stride and number of vertices
-    if (array.enable == 0 || size == 0 || !device_addr) {
+    if (array.enable == 0 || !device_addr || gpu_addr_end <= gpu_addr_begin) {
         channel_state->vertex_buffers[index] = NULL_BINDING;
         UpdateVertexBufferSlot(index, NULL_BINDING);
         return;
     }
-    if (!gpu_memory->IsWithinGPUAddressRange(gpu_addr_end) || size >= 64_MiB) {
-        size = static_cast<u32>(gpu_memory->MaxContinuousRange(gpu_addr_begin, size));
+    // TODO: Analyze stride and number of vertices
+    constexpr u64 implausible_size = 64_MiB;
+    u64 address_size = gpu_addr_end - gpu_addr_begin;
+    if (address_size > u64{(std::numeric_limits<u32>::max)()}) {
+        address_size = implausible_size;
+    }
+    const bool is_instanced = maxwell3d->regs.vertex_stream_instances.IsInstancingEnabled(index);
+    const u64 stride = static_cast<u64>(array.stride);
+    if (is_instanced && stride != 0 && draw_instance_count != 0) {
+        const u64 base_instance = static_cast<u64>(maxwell3d->draw_manager.draw_state.base_instance);
+        u64 elements = base_instance + 1;
+        if (array.frequency != 0) {
+            elements = (base_instance + static_cast<u64>(draw_instance_count) - 1) /
+                           static_cast<u64>(array.frequency) +
+                       1;
+        }
+        address_size = (std::min)(address_size, (elements + 1) * stride);
+    }
+    if (!gpu_memory->IsWithinGPUAddressRange(gpu_addr_end) || address_size >= implausible_size) {
+        address_size = gpu_memory->MaxContinuousRange(gpu_addr_begin, address_size);
+    }
+    const u32 size = static_cast<u32>(address_size);
+    if (size == 0) {
+        channel_state->vertex_buffers[index] = NULL_BINDING;
+        UpdateVertexBufferSlot(index, NULL_BINDING);
+        return;
     }
     const BufferId buffer_id = FindBuffer(*device_addr, size);
     const Binding binding{
@@ -1578,9 +1590,10 @@ template <class P>
 BufferId BufferCache<P>::CreateBuffer(DAddr device_addr, u32 wanted_size) {
     DAddr device_addr_end = Common::AlignUp(device_addr + wanted_size, CACHING_PAGESIZE);
     device_addr = Common::AlignDown(device_addr, CACHING_PAGESIZE);
-    wanted_size = static_cast<u32>(device_addr_end - device_addr);
+    constexpr u64 max_buffer_size = u64{(std::numeric_limits<u32>::max)()};
+    wanted_size = static_cast<u32>((std::min)(device_addr_end - device_addr, max_buffer_size));
     const OverlapResult overlap = ResolveOverlaps(device_addr, wanted_size);
-    const u32 size = static_cast<u32>(overlap.end - overlap.begin);
+    const u32 size = static_cast<u32>((std::min)(overlap.end - overlap.begin, max_buffer_size));
     const BufferId new_buffer_id = slot_buffers.insert(runtime, overlap.begin, size);
     auto& new_buffer = slot_buffers[new_buffer_id];
     const size_t size_bytes = new_buffer.SizeBytes();

src/video_core/buffer_cache/buffer_cache_base.h

@@ -217,8 +217,6 @@ public:
 
     void WriteMemory(DAddr device_addr, u64 size);
 
-    void CachedWriteMemory(DAddr device_addr, u64 size);
-
     bool OnCPUWrite(DAddr device_addr, u64 size);
 
     void DownloadMemory(DAddr device_addr, u64 size);
@@ -248,7 +246,7 @@ public:
 
     void UnbindGraphicsStorageBuffers(size_t stage);
 
-    bool BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, u32 cbuf_index, u32 cbuf_offset,
+    void BindGraphicsStorageBuffer(size_t stage, size_t ssbo_index, u32 cbuf_index, u32 cbuf_offset,
                                    bool is_written);
 
     void UnbindGraphicsTextureBuffers(size_t stage);
@@ -309,6 +307,10 @@ public:
         current_draw_indirect = current_draw_indirect_;
     }
 
+    void SetDrawInstanceCount(u32 draw_instance_count_) {
+        draw_instance_count = draw_instance_count_;
+    }
+
     [[nodiscard]] std::pair<Buffer*, u32> GetDrawIndirectCount();
 
     [[nodiscard]] std::pair<Buffer*, u32> GetDrawIndirectBuffer();
@@ -376,8 +378,6 @@ private:
 
     void BindHostTransformFeedbackBuffers();
 
-    void BindHostVertexBuffer(u32 index, Buffer& buffer, u32 offset, u32 size, u32 stride);
-
     void BindHostComputeUniformBuffers();
 
     void BindHostComputeStorageBuffers();
@@ -484,6 +484,10 @@ private:
 
     const Tegra::Engines::Maxwell3D::DrawManager::IndirectParams* current_draw_indirect{};
 
+    u32 draw_instance_count = 0;
+    u32 last_draw_instance_count = 0;
+    u32 last_draw_base_instance = 0;
+
     u32 last_index_count = 0;
 
     u32 enabled_vertex_buffers_mask = 0;

src/video_core/host_shaders/CMakeLists.txt

@@ -206,6 +206,40 @@ foreach(VARIANT IN ITEMS ${SHADER_TYPE_VARIANTS})
     set(SHADER_HEADERS ${SHADER_HEADERS} ${VARIANT_HEADER_FILE})
 endforeach()
 
+set(SHADER_DEFINE_VARIANTS
+    "block_linear_unswizzle_2d.comp|nonarrow|HAS_EXTENDED_TYPES=0"
+    "pitch_unswizzle.comp|nonarrow|HAS_EXTENDED_TYPES=0"
+    "block_linear_unswizzle_3d.comp|nonarrow|HAS_EXTENDED_TYPES=0"
+)
+
+foreach(VARIANT IN ITEMS ${SHADER_DEFINE_VARIANTS})
+    string(REPLACE "|" ";" VARIANT_PARTS ${VARIANT})
+    list(GET VARIANT_PARTS 0 VARIANT_FILENAME)
+    list(GET VARIANT_PARTS 1 VARIANT_SUFFIX)
+    list(GET VARIANT_PARTS 2 VARIANT_DEFINE)
+
+    set(VARIANT_SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/${VARIANT_FILENAME})
+    get_filename_component(VARIANT_STEM ${VARIANT_FILENAME} NAME_WE)
+    get_filename_component(VARIANT_EXT ${VARIANT_FILENAME} EXT)
+    string(REPLACE "." "" VARIANT_EXT ${VARIANT_EXT})
+    set(VARIANT_NAME ${VARIANT_STEM}_${VARIANT_SUFFIX}_${VARIANT_EXT})
+
+    string(TOUPPER ${VARIANT_NAME}_SPV VARIANT_VARIABLE_NAME)
+    set(VARIANT_HEADER_FILE ${SHADER_DIR}/${VARIANT_NAME}_spv.h)
+    add_custom_command(
+        OUTPUT
+            ${VARIANT_HEADER_FILE}
+        COMMAND
+            ${GLSLANGVALIDATOR} -V ${QUIET_FLAG} -I"${FIDELITYFX_INCLUDE_DIR}" ${GLSL_FLAGS}
+            -D${VARIANT_DEFINE}
+            --variable-name ${VARIANT_VARIABLE_NAME} -o ${VARIANT_HEADER_FILE} ${VARIANT_SOURCE}
+            --target-env ${SPIR_V_VERSION}
+        MAIN_DEPENDENCY
+            ${VARIANT_SOURCE}
+    )
+    set(SHADER_HEADERS ${SHADER_HEADERS} ${VARIANT_HEADER_FILE})
+endforeach()
+
 foreach(FILEPATH IN ITEMS ${FIDELITYFX_FILES})
     get_filename_component(FILENAME ${FILEPATH} NAME)
     string(REPLACE "." "_" HEADER_NAME ${FILENAME})

src/video_core/host_shaders/block_linear_unswizzle_2d.comp

@@ -5,9 +5,13 @@
 
 #ifdef VULKAN
 
+#ifndef HAS_EXTENDED_TYPES
+#define HAS_EXTENDED_TYPES 1
+#endif
+#if HAS_EXTENDED_TYPES
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
-#define HAS_EXTENDED_TYPES 1
+#endif
 #define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants {
 #define END_PUSH_CONSTANTS };
 #define UNIFORM(n)

src/video_core/host_shaders/block_linear_unswizzle_3d.comp

@@ -5,9 +5,13 @@
 
 #ifdef VULKAN
 
+#ifndef HAS_EXTENDED_TYPES
+#define HAS_EXTENDED_TYPES 1
+#endif
+#if HAS_EXTENDED_TYPES
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
-#define HAS_EXTENDED_TYPES 1
+#endif
 #define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants {
 #define END_PUSH_CONSTANTS };
 #define UNIFORM(n)

src/video_core/host_shaders/pitch_unswizzle.comp

@@ -5,9 +5,13 @@
 
 #ifdef VULKAN
 
+#ifndef HAS_EXTENDED_TYPES
+#define HAS_EXTENDED_TYPES 1
+#endif
+#if HAS_EXTENDED_TYPES
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
-#define HAS_EXTENDED_TYPES 1
+#endif
 #define BEGIN_PUSH_CONSTANTS layout(push_constant) uniform PushConstants {
 #define END_PUSH_CONSTANTS };
 #define UNIFORM(n)

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -226,22 +226,6 @@ void BufferCacheRuntime::BindIndexBuffer(Buffer& buffer, u32 offset, u32 size) {
     }
 }
 
-void BufferCacheRuntime::BindVertexBuffer(u32 index, Buffer& buffer, u32 offset, u32 size,
-                                          u32 stride) {
-    if (index >= max_attributes) {
-        return;
-    }
-    if (has_unified_vertex_buffers) {
-        buffer.MakeResident(GL_READ_ONLY);
-        glBindVertexBuffer(index, 0, 0, static_cast<GLsizei>(stride));
-        glBufferAddressRangeNV(GL_VERTEX_ATTRIB_ARRAY_ADDRESS_NV, index,
-                               buffer.HostGpuAddr() + offset, static_cast<GLsizeiptr>(size));
-    } else {
-        glBindVertexBuffer(index, buffer.Handle(), static_cast<GLintptr>(offset),
-                           static_cast<GLsizei>(stride));
-    }
-}
-
 void BufferCacheRuntime::BindVertexBuffers(VideoCommon::HostBindings<Buffer>& bindings) {
     // TODO: Should HostBindings provide the correct runtime types to avoid these transforms?
     std::array<GLuint, 32> buffer_handles;

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -99,8 +99,6 @@ public:
 
     void BindIndexBuffer(Buffer& buffer, u32 offset, u32 size);
 
-    void BindVertexBuffer(u32 index, Buffer& buffer, u32 offset, u32 size, u32 stride);
-
     void BindVertexBuffers(VideoCommon::HostBindings<Buffer>& bindings);
 
     void BindUniformBuffer(size_t stage, u32 binding_index, Buffer& buffer, u32 offset, u32 size);

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -259,6 +259,7 @@ void RasterizerOpenGL::PrepareDraw(bool is_indexed, Func&& draw_func) {
 }
 
 void RasterizerOpenGL::Draw(bool is_indexed, u32 instance_count) {
+    buffer_cache.SetDrawInstanceCount(instance_count);
     PrepareDraw(is_indexed, [this, is_indexed, instance_count](GLenum primitive_mode) {
         const auto& draw_state = maxwell3d->draw_manager.draw_state;
         const GLuint base_instance = GLuint(draw_state.base_instance);
@@ -304,6 +305,7 @@ void RasterizerOpenGL::Draw(bool is_indexed, u32 instance_count) {
 void RasterizerOpenGL::DrawIndirect() {
     const auto& params = maxwell3d->draw_manager.indirect_state;
     buffer_cache.SetDrawIndirect(&params);
+    buffer_cache.SetDrawInstanceCount(0);
     PrepareDraw(params.is_indexed, [this, &params](GLenum primitive_mode) {
         if (params.is_byte_count) {
             const GPUVAddr tfb_object_base_addr = params.indirect_start_address - 4U;

src/video_core/renderer_vulkan/vk_buffer_cache.cpp

@@ -585,29 +585,6 @@ void BufferCacheRuntime::BindQuadIndexBuffer(PrimitiveTopology topology, u32 fir
     }
 }
 
-void BufferCacheRuntime::BindVertexBuffer(u32 index, VkBuffer buffer, u32 offset, u32 size, u32 stride) {
-    if (index >= device.GetMaxVertexInputBindings()) {
-        return;
-    }
-    if (device.IsExtExtendedDynamicStateSupported()) {
-        scheduler.Record([index, buffer, offset, size, stride](vk::CommandBuffer cmdbuf) {
-            const VkDeviceSize vk_offset = buffer != VK_NULL_HANDLE ? offset : 0;
-            const VkDeviceSize vk_size = buffer != VK_NULL_HANDLE ? size : VK_WHOLE_SIZE;
-            const VkDeviceSize vk_stride = stride;
-            cmdbuf.BindVertexBuffers2EXT(index, 1, &buffer, &vk_offset, &vk_size, &vk_stride);
-        });
-    } else {
-        if (!device.HasNullDescriptor() && buffer == VK_NULL_HANDLE) {
-            ReserveNullBuffer();
-            buffer = *null_buffer;
-            offset = 0;
-        }
-        scheduler.Record([index, buffer, offset](vk::CommandBuffer cmdbuf) {
-            cmdbuf.BindVertexBuffer(index, buffer, offset);
-        });
-    }
-}
-
 void BufferCacheRuntime::BindVertexBuffers(VideoCommon::HostBindings<Buffer>& bindings) {
     boost::container::static_vector<VkBuffer, VideoCommon::NUM_VERTEX_BUFFERS> buffer_handles(bindings.buffers.size());
     for (u32 i = 0; i < bindings.buffers.size(); ++i) {

src/video_core/renderer_vulkan/vk_buffer_cache.h

@@ -138,8 +138,6 @@ public:
 
     void BindQuadIndexBuffer(PrimitiveTopology topology, u32 first, u32 count);
 
-    void BindVertexBuffer(u32 index, VkBuffer buffer, u32 offset, u32 size, u32 stride);
-
     void BindVertexBuffers(VideoCommon::HostBindings<Buffer>& bindings);
 
     void BindTransformFeedbackBuffer(u32 index, VkBuffer buffer, u32 offset, u32 size);

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -22,7 +22,13 @@
 #include "video_core/host_shaders/resolve_conditional_render_comp_spv.h"
 #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_2d_comp_spv.h"
+#include "video_core/host_shaders/block_linear_unswizzle_2d_nonarrow_comp_spv.h"
 #include "video_core/host_shaders/block_linear_unswizzle_3d_bcn_comp_spv.h"
+#include "video_core/host_shaders/block_linear_unswizzle_3d_comp_spv.h"
+#include "video_core/host_shaders/block_linear_unswizzle_3d_nonarrow_comp_spv.h"
+#include "video_core/host_shaders/pitch_unswizzle_comp_spv.h"
+#include "video_core/host_shaders/pitch_unswizzle_nonarrow_comp_spv.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
 #include "video_core/surface.h"
 #include "video_core/renderer_vulkan/vk_descriptor_pool.h"
@@ -872,4 +878,291 @@ void BlockLinearUnswizzle3DPass::UnswizzleChunk(
     });
 }
 
+namespace {
+
+constexpr u32 UNSWIZZLE_BINDING_INPUT_BUFFER = 0;
+constexpr u32 UNSWIZZLE_BINDING_OUTPUT_IMAGE = 1;
+constexpr size_t UNSWIZZLE_NUM_BINDINGS = 2;
+
+constexpr std::array<VkDescriptorSetLayoutBinding, UNSWIZZLE_NUM_BINDINGS>
+    UNSWIZZLE_DESCRIPTOR_SET_BINDINGS{{
+        {
+            .binding = UNSWIZZLE_BINDING_INPUT_BUFFER,
+            .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+            .descriptorCount = 1,
+            .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+            .pImmutableSamplers = nullptr,
+        },
+        {
+            .binding = UNSWIZZLE_BINDING_OUTPUT_IMAGE,
+            .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+            .descriptorCount = 1,
+            .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
+            .pImmutableSamplers = nullptr,
+        },
+    }};
+
+constexpr std::array<VkDescriptorUpdateTemplateEntry, UNSWIZZLE_NUM_BINDINGS>
+    UNSWIZZLE_DESCRIPTOR_UPDATE_TEMPLATE{{
+        {
+            .dstBinding = UNSWIZZLE_BINDING_INPUT_BUFFER,
+            .dstArrayElement = 0,
+            .descriptorCount = 1,
+            .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
+            .offset = UNSWIZZLE_BINDING_INPUT_BUFFER * sizeof(DescriptorUpdateEntry),
+            .stride = sizeof(DescriptorUpdateEntry),
+        },
+        {
+            .dstBinding = UNSWIZZLE_BINDING_OUTPUT_IMAGE,
+            .dstArrayElement = 0,
+            .descriptorCount = 1,
+            .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
+            .offset = UNSWIZZLE_BINDING_OUTPUT_IMAGE * sizeof(DescriptorUpdateEntry),
+            .stride = sizeof(DescriptorUpdateEntry),
+        },
+    }};
+
+constexpr DescriptorBankInfo UNSWIZZLE_BANK_INFO{
+    .uniform_buffers = 0,
+    .storage_buffers = 1,
+    .texture_buffers = 0,
+    .image_buffers = 0,
+    .textures = 0,
+    .images = 1,
+    .score = 2,
+};
+
+[[nodiscard]] std::span<const u32> UnswizzleSpv(const Device& device,
+                                                std::span<const u32> extended,
+                                                std::span<const u32> narrow) {
+    if (device.IsStorageBuffer8BitAccessSupported() &&
+        device.IsStorageBuffer16BitAccessSupported()) {
+        return extended;
+    }
+    return narrow;
+}
+
+struct PitchUnswizzlePushConstants {
+    alignas(8) std::array<u32, 2> origin;
+    alignas(8) std::array<s32, 2> destination;
+    u32 bytes_per_block;
+    u32 pitch;
+};
+
+void RecordUnswizzleEntryBarrier(Scheduler& scheduler, VkPipeline vk_pipeline, VkImage vk_image,
+                                 VkImageAspectFlags aspect_mask, bool is_initialized) {
+    scheduler.Record([vk_pipeline, vk_image, aspect_mask,
+                      is_initialized](vk::CommandBuffer cmdbuf) {
+        VkAccessFlags src_access = VK_ACCESS_NONE;
+        VkImageLayout old_layout = VK_IMAGE_LAYOUT_UNDEFINED;
+        if (is_initialized) {
+            src_access = VK_ACCESS_SHADER_WRITE_BIT | VK_ACCESS_TRANSFER_WRITE_BIT |
+                         VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT;
+            old_layout = VK_IMAGE_LAYOUT_GENERAL;
+        }
+        const VkImageMemoryBarrier image_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = src_access,
+            .dstAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+            .oldLayout = old_layout,
+            .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,
+            },
+        };
+        VkPipelineStageFlags src_stage = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
+        if (is_initialized) {
+            src_stage = vk::PIPELINE_STAGE_GRAPHICS_COMPUTE_TRANSFER;
+        }
+        cmdbuf.PipelineBarrier(src_stage, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, image_barrier);
+        cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, vk_pipeline);
+    });
+}
+
+void RecordUnswizzleExitBarrier(Scheduler& scheduler, VkImage vk_image,
+                                VkImageAspectFlags aspect_mask) {
+    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_TRANSFER_READ_BIT |
+                             VK_ACCESS_COLOR_ATTACHMENT_READ_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_TRANSFER, 0, image_barrier);
+    });
+}
+
+} // Anonymous namespace
+
+BlockLinearUnswizzle2DPass::BlockLinearUnswizzle2DPass(
+    const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(device_, scheduler_, descriptor_pool_, UNSWIZZLE_DESCRIPTOR_SET_BINDINGS,
+                  UNSWIZZLE_DESCRIPTOR_UPDATE_TEMPLATE, UNSWIZZLE_BANK_INFO,
+                  COMPUTE_PUSH_CONSTANT_RANGE<sizeof(
+                      VideoCommon::Accelerated::BlockLinearSwizzle2DParams)>,
+                  UnswizzleSpv(device_, BLOCK_LINEAR_UNSWIZZLE_2D_COMP_SPV,
+                               BLOCK_LINEAR_UNSWIZZLE_2D_NONARROW_COMP_SPV)),
+      scheduler{scheduler_}, compute_pass_descriptor_queue{compute_pass_descriptor_queue_} {}
+
+BlockLinearUnswizzle2DPass::~BlockLinearUnswizzle2DPass() = default;
+
+void BlockLinearUnswizzle2DPass::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();
+    RecordUnswizzleEntryBarrier(scheduler, vk_pipeline, vk_image, aspect_mask, is_initialized);
+
+    const u32 num_layers = static_cast<u32>(image.info.resources.layers);
+    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 auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
+        scheduler.Record([this, num_dispatches_x, num_dispatches_y, num_layers, 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_layers);
+        });
+    }
+    RecordUnswizzleExitBarrier(scheduler, vk_image, aspect_mask);
+}
+
+BlockLinearUnswizzleImage3DPass::BlockLinearUnswizzleImage3DPass(
+    const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(device_, scheduler_, descriptor_pool_, UNSWIZZLE_DESCRIPTOR_SET_BINDINGS,
+                  UNSWIZZLE_DESCRIPTOR_UPDATE_TEMPLATE, UNSWIZZLE_BANK_INFO,
+                  COMPUTE_PUSH_CONSTANT_RANGE<sizeof(BlockLinearSwizzle3DParams)>,
+                  UnswizzleSpv(device_, BLOCK_LINEAR_UNSWIZZLE_3D_COMP_SPV,
+                               BLOCK_LINEAR_UNSWIZZLE_3D_NONARROW_COMP_SPV)),
+      scheduler{scheduler_}, 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();
+    RecordUnswizzleEntryBarrier(scheduler, vk_pipeline, vk_image, aspect_mask, is_initialized);
+
+    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 params = MakeBlockLinearSwizzle3DParams(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);
+        });
+    }
+    RecordUnswizzleExitBarrier(scheduler, vk_image, aspect_mask);
+}
+
+PitchUnswizzlePass::PitchUnswizzlePass(
+    const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue_)
+    : ComputePass(device_, scheduler_, descriptor_pool_, UNSWIZZLE_DESCRIPTOR_SET_BINDINGS,
+                  UNSWIZZLE_DESCRIPTOR_UPDATE_TEMPLATE, UNSWIZZLE_BANK_INFO,
+                  COMPUTE_PUSH_CONSTANT_RANGE<sizeof(PitchUnswizzlePushConstants)>,
+                  UnswizzleSpv(device_, PITCH_UNSWIZZLE_COMP_SPV,
+                               PITCH_UNSWIZZLE_NONARROW_COMP_SPV)),
+      scheduler{scheduler_}, 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();
+    RecordUnswizzleEntryBarrier(scheduler, vk_pipeline, vk_image, aspect_mask, is_initialized);
+
+    const u32 bytes_per_block = VideoCore::Surface::BytesPerBlock(image.info.format);
+    const u32 pitch = image.info.pitch;
+    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 = 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);
+        });
+    }
+    RecordUnswizzleExitBarrier(scheduler, vk_image, aspect_mask);
+}
+
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_compute_pass.h

@@ -164,4 +164,49 @@ private:
     ComputePassDescriptorQueue& compute_pass_descriptor_queue;
 };
 
+class BlockLinearUnswizzle2DPass final : public ComputePass {
+public:
+    explicit BlockLinearUnswizzle2DPass(
+        const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+        ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+    ~BlockLinearUnswizzle2DPass();
+
+    void Unswizzle(Image& image, const StagingBufferRef& map,
+                   std::span<const VideoCommon::SwizzleParameters> swizzles);
+
+private:
+    Scheduler& scheduler;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
+
+class BlockLinearUnswizzleImage3DPass final : public ComputePass {
+public:
+    explicit BlockLinearUnswizzleImage3DPass(
+        const Device& device_, Scheduler& scheduler_, DescriptorPool& descriptor_pool_,
+        ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+    ~BlockLinearUnswizzleImage3DPass();
+
+    void Unswizzle(Image& image, const StagingBufferRef& map,
+                   std::span<const VideoCommon::SwizzleParameters> swizzles);
+
+private:
+    Scheduler& scheduler;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
+
+class PitchUnswizzlePass final : public ComputePass {
+public:
+    explicit PitchUnswizzlePass(const Device& device_, Scheduler& scheduler_,
+                                DescriptorPool& descriptor_pool_,
+                                ComputePassDescriptorQueue& compute_pass_descriptor_queue_);
+    ~PitchUnswizzlePass();
+
+    void Unswizzle(Image& image, const StagingBufferRef& map,
+                   std::span<const VideoCommon::SwizzleParameters> swizzles);
+
+private:
+    Scheduler& scheduler;
+    ComputePassDescriptorQueue& compute_pass_descriptor_queue;
+};
+
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -694,9 +694,11 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
         const size_t num_vertex_arrays = (std::min)(
             Maxwell::NumVertexArrays, static_cast<size_t>(device.GetMaxVertexInputBindings()));
         for (size_t index = 0; index < num_vertex_arrays; ++index) {
-            const bool instanced = key.state.binding_divisors[index] != 0;
-            const auto rate =
-                instanced ? VK_VERTEX_INPUT_RATE_INSTANCE : VK_VERTEX_INPUT_RATE_VERTEX;
+            const bool instanced = ((key.state.enabled_divisors >> index) & 1) != 0;
+            auto rate = VK_VERTEX_INPUT_RATE_VERTEX;
+            if (instanced) {
+                rate = VK_VERTEX_INPUT_RATE_INSTANCE;
+            }
             vertex_bindings.push_back({
                 .binding = static_cast<u32>(index),
                 .stride = key.state.vertex_strides[index],
@@ -705,7 +707,7 @@ void GraphicsPipeline::MakePipeline(VkRenderPass render_pass) {
             if (instanced) {
                 vertex_binding_divisors.push_back({
                     .binding = static_cast<u32>(index),
-                    .divisor = key.state.binding_divisors[index],
+                    .divisor = device.GetVertexAttribDivisor(key.state.binding_divisors[index]),
                 });
             }
         }

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -260,6 +260,7 @@ void RasterizerVulkan::PrepareDraw(bool is_indexed, Func&& draw_func) {
 }
 
 void RasterizerVulkan::Draw(bool is_indexed, u32 instance_count) {
+    buffer_cache.SetDrawInstanceCount(instance_count);
     PrepareDraw(is_indexed, [this, is_indexed, instance_count] {
         const auto& draw_state = maxwell3d->draw_manager.draw_state;
         const u32 num_instances{instance_count};
@@ -295,6 +296,7 @@ void RasterizerVulkan::Draw(bool is_indexed, u32 instance_count) {
 void RasterizerVulkan::DrawIndirect() {
     const auto& params = maxwell3d->draw_manager.indirect_state;
     buffer_cache.SetDrawIndirect(&params);
+    buffer_cache.SetDrawInstanceCount(0);
     PrepareDraw(params.is_indexed, [this, &params] {
         const auto indirect_buffer = buffer_cache.GetDrawIndirectBuffer();
         const auto& buffer = indirect_buffer.first;
@@ -1917,13 +1919,19 @@ void RasterizerVulkan::UpdateVertexInput(Tegra::Engines::Maxwell3D::Regs& regs)
     for (u32 binding = 0; binding < max_bindings; ++binding) {
         const auto& input_binding{regs.vertex_streams[binding]};
         const bool is_instanced{regs.vertex_stream_instances.IsInstancingEnabled(binding)};
+        auto input_rate = VK_VERTEX_INPUT_RATE_VERTEX;
+        u32 divisor = 1;
+        if (is_instanced) {
+            input_rate = VK_VERTEX_INPUT_RATE_INSTANCE;
+            divisor = device.GetVertexAttribDivisor(input_binding.frequency);
+        }
         bindings.push_back({
             .sType = VK_STRUCTURE_TYPE_VERTEX_INPUT_BINDING_DESCRIPTION_2_EXT,
             .pNext = nullptr,
             .binding = binding,
             .stride = input_binding.stride,
-            .inputRate = is_instanced ? VK_VERTEX_INPUT_RATE_INSTANCE : VK_VERTEX_INPUT_RATE_VERTEX,
-            .divisor = is_instanced ? input_binding.frequency : 1,
+            .inputRate = input_rate,
+            .divisor = divisor,
         });
     }
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -449,7 +449,9 @@ void Scheduler::EndRenderPass()
                                          | VK_ACCESS_COLOR_ATTACHMENT_READ_BIT
                                          | VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT
                                          | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT
-                                         | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,
+                                         | VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT
+                                         | VK_ACCESS_TRANSFER_READ_BIT
+                                         | VK_ACCESS_TRANSFER_WRITE_BIT,
                         .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
                         .newLayout = VK_IMAGE_LAYOUT_GENERAL,
                         .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
@@ -460,7 +462,7 @@ void Scheduler::EndRenderPass()
             }
             cmdbuf.EndRenderPass();
             cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT | VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
-                                   VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, vk::PIPELINE_STAGE_GRAPHICS_COMPUTE,
+                                   VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT, vk::PIPELINE_STAGE_GRAPHICS_COMPUTE_TRANSFER,
                                    0, nullptr, nullptr, vk::Span(barriers.data(), num_images));
             if (has_transform_feedback) {
                 static constexpr VkMemoryBarrier XFB_OUTPUT_BARRIER{

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -160,6 +160,55 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
           info.size.depth == 1;
 }
 
+[[nodiscard]] PixelFormat UnswizzleViewFormat(u32 bytes_per_block) {
+    switch (bytes_per_block) {
+    case 1:
+        return PixelFormat::R8_UINT;
+    case 2:
+        return PixelFormat::R16_UINT;
+    case 4:
+        return PixelFormat::R32_UINT;
+    case 8:
+        return PixelFormat::R32G32_UINT;
+    case 16:
+        return PixelFormat::R32G32B32A32_UINT;
+    default:
+        return PixelFormat::Invalid;
+    }
+}
+
+constexpr u32 UNSWIZZLE_WORKGROUP_INVOCATIONS = 32 * 32;
+
+[[nodiscard]] bool SupportsAcceleratedUnswizzleDevice(const Device& device) {
+    return device.IsKhrImageFormatListSupported() &&
+           device.GetMaxComputeWorkGroupInvocations() >= UNSWIZZLE_WORKGROUP_INVOCATIONS;
+}
+
+[[nodiscard]] bool SupportsAcceleratedUnswizzle(const Device& device, const ImageInfo& info) {
+    if (!SupportsAcceleratedUnswizzleDevice(device)) {
+        return false;
+    }
+    if (info.num_samples > 1) {
+        return false;
+    }
+    if (info.type != ImageType::e2D && info.type != ImageType::e3D &&
+        info.type != ImageType::Linear) {
+        return false;
+    }
+    const PixelFormat view_format =
+        UnswizzleViewFormat(VideoCore::Surface::BytesPerBlock(info.format));
+    if (view_format == PixelFormat::Invalid) {
+        return false;
+    }
+    if (!VideoCore::Surface::IsViewCompatible(info.format, view_format, false, true)) {
+        return false;
+    }
+    const auto host_format =
+        MaxwellToVK::SurfaceFormat(device, FormatType::Optimal, false, view_format);
+    return device.IsFormatSupported(host_format.format, VK_FORMAT_FEATURE_STORAGE_IMAGE_BIT,
+                                    FormatType::Optimal);
+}
+
 [[nodiscard]] VkImageCreateInfo MakeImageCreateInfo(const Device& device, const ImageInfo& info,
                                                     std::optional<VkFormat> format_override = {}) {
     auto format_info =
@@ -248,8 +297,18 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     return allocator.CreateImage(image_ci);
 }
 
+[[nodiscard]] VkImageViewType StorageViewType(ImageType type) {
+    if (type == ImageType::e3D) {
+        return VK_IMAGE_VIEW_TYPE_3D;
+    }
+    if (type == ImageType::Linear) {
+        return VK_IMAGE_VIEW_TYPE_2D;
+    }
+    return VK_IMAGE_VIEW_TYPE_2D_ARRAY;
+}
+
 [[nodiscard]] vk::ImageView MakeStorageView(const vk::Device& device, u32 level, VkImage image,
-                                            VkFormat format) {
+                                            VkFormat format, VkImageViewType view_type) {
     static constexpr VkImageViewUsageCreateInfo storage_image_view_usage_create_info{
         .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_USAGE_CREATE_INFO,
         .pNext = nullptr,
@@ -260,7 +319,7 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
         .pNext = &storage_image_view_usage_create_info,
         .flags = 0,
         .image = image,
-        .viewType = VK_IMAGE_VIEW_TYPE_2D_ARRAY,
+        .viewType = view_type,
         .format = format,
         .components{
             .r = VK_COMPONENT_SWIZZLE_IDENTITY,
@@ -658,18 +717,11 @@ void CopyBufferToImage(vk::CommandBuffer cmdbuf, VkBuffer src_buffer, VkImage im
             .subresourceRange = subresource_range,
     };
 
-    cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
-                           VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT |
-                           VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
+    cmdbuf.PipelineBarrier(vk::PIPELINE_STAGE_GRAPHICS_COMPUTE, VK_PIPELINE_STAGE_TRANSFER_BIT, 0,
                            read_barrier);
     cmdbuf.CopyBufferToImage(src_buffer, image, VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL, copies);
-    // TODO: Move this to another API
-    cmdbuf.PipelineBarrier(
-            VK_PIPELINE_STAGE_TRANSFER_BIT,
-            VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
-            VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
-            VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT,
-            0, nullptr, nullptr, write_barrier);
+    cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, vk::PIPELINE_STAGE_GRAPHICS_COMPUTE, 0,
+                           nullptr, nullptr, write_barrier);
 }
 
 [[nodiscard]] VkImageBlit MakeImageBlit(const Region2D& dst_region, const Region2D& src_region,
@@ -968,6 +1020,14 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, Scheduler& sched
         bl3d_unswizzle_pass.emplace(device, scheduler, descriptor_pool,
                                    staging_buffer_pool, compute_pass_descriptor_queue);
     }
+    if (SupportsAcceleratedUnswizzleDevice(device)) {
+        bl_unswizzle_2d_pass.emplace(device, scheduler, descriptor_pool,
+                                     compute_pass_descriptor_queue);
+        bl_unswizzle_image_3d_pass.emplace(device, scheduler, descriptor_pool,
+                                           compute_pass_descriptor_queue);
+        pitch_unswizzle_pass.emplace(device, scheduler, descriptor_pool,
+                                     compute_pass_descriptor_queue);
+    }
 }
 
 void TextureCacheRuntime::Finish() {
@@ -1902,16 +1962,12 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
     if (runtime->device.HasDebuggingToolAttached()) {
         original_image.SetObjectNameEXT(VideoCommon::Name(*this).c_str());
     }
+    if (False(flags & VideoCommon::ImageFlagBits::Converted) &&
+        SupportsAcceleratedUnswizzle(runtime->device, info)) {
+        flags |= VideoCommon::ImageFlagBits::AcceleratedUpload;
+    }
     current_image = &Image::original_image;
     storage_image_views.resize(info.resources.levels);
-    if (WillUseAcceleratedAstcDecode(runtime->device, info)) {
-        const auto& device = runtime->device.GetLogical();
-        const VkFormat storage_format = VK_FORMAT_A8B8G8R8_UNORM_PACK32;
-        for (s32 level = 0; level < info.resources.levels; ++level) {
-            storage_image_views[level] =
-                MakeStorageView(device, level, *original_image, storage_format);
-        }
-    }
 }
 
 Image::Image(const VideoCommon::NullImageParams& params) : VideoCommon::ImageBase{params} {}
@@ -2035,7 +2091,7 @@ void Image::UploadMemory(VkBuffer buffer, VkDeviceSize offset,
                                                 temp_vk_image, info.format, info.num_samples,
                                                 {image_copies.data(), image_copies.size()}, false);
         }
-        initialized = true;
+        InitializationFor(current_image) = true;
         runtime->ReleaseMsaaScratchImage(temp_vk_image);
 
         if (is_rescaled) {
@@ -2056,7 +2112,7 @@ void Image::UploadMemory(VkBuffer buffer, VkDeviceSize offset,
     const VkBuffer src_buffer = buffer;
     const VkImage vk_image = *original_image;
     const VkImageAspectFlags vk_aspect_mask = aspect_mask;
-    const bool was_initialized = std::exchange(initialized, true);
+    const bool was_initialized = std::exchange(InitializationFor(&Image::original_image), true);
 
     scheduler->Record([src_buffer, vk_image, vk_aspect_mask, was_initialized,
                        vk_copies](vk::CommandBuffer cmdbuf) {
@@ -2321,16 +2377,46 @@ void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferIm
     DownloadMemory(buffers, offsets, copies);
 }
 
+std::vector<vk::ImageView>& Image::StorageViewsFor(vk::Image Image::*image) {
+    if (image == &Image::scaled_image) {
+        if (scaled_storage_image_views.empty()) {
+            scaled_storage_image_views.resize(info.resources.levels);
+        }
+        return scaled_storage_image_views;
+    }
+    return storage_image_views;
+}
+
+bool& Image::InitializationFor(vk::Image Image::*image) noexcept {
+    if (image == &Image::scaled_image) {
+        return scaled_initialized;
+    }
+    return original_initialized;
+}
+
 VkImageView Image::StorageImageView(s32 level) noexcept {
-    auto& view = storage_image_views[level];
+    const bool astc_decode = WillUseAcceleratedAstcDecode(runtime->device, info);
+    const bool unswizzle_upload =
+        !astc_decode && True(flags & ImageFlagBits::AcceleratedUpload);
+    vk::Image Image::*target = current_image;
+    if (astc_decode || unswizzle_upload) {
+        target = &Image::original_image;
+    }
+    auto& view = StorageViewsFor(target)[level];
     if (!view) {
         auto format_info =
             MaxwellToVK::SurfaceFormat(runtime->device, FormatType::Optimal, true, info.format);
-        if (WillUseAcceleratedAstcDecode(runtime->device, info)) {
+        if (astc_decode) {
             format_info.format = VK_FORMAT_A8B8G8R8_UNORM_PACK32;
         }
-        view = MakeStorageView(runtime->device.GetLogical(), level, *(this->*current_image),
-                               format_info.format);
+        if (unswizzle_upload) {
+            const PixelFormat view_format =
+                UnswizzleViewFormat(VideoCore::Surface::BytesPerBlock(info.format));
+            format_info = MaxwellToVK::SurfaceFormat(runtime->device, FormatType::Optimal, false,
+                                                     view_format);
+        }
+        view = MakeStorageView(runtime->device.GetLogical(), level, *(this->*target),
+                               format_info.format, StorageViewType(info.type));
     }
     return *view;
 }
@@ -2370,6 +2456,7 @@ bool Image::ScaleUp(bool ignore) {
     }
     if (NeedsScaleHelper()) {
         if (!BlitScaleHelper(true)) {
+            flags &= ~ImageFlagBits::Rescaled;
             current_image = &Image::original_image;
             return false;
         }
@@ -2559,6 +2646,10 @@ ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewI
     if (device->IsExtAstcDecodeModeSupported() && IsLdrAstcFormat(format_info.format)) {
         view_next = &astc_decode_mode;
     }
+    auto subresource_range = MakeSubresourceRange(aspect_mask, info.range);
+    if (True(flags & VideoCommon::ImageViewFlagBits::Slice)) {
+        subresource_range.levelCount = 1;
+    }
     const VkImageViewCreateInfo create_info{
         .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
         .pNext = view_next,
@@ -2567,7 +2658,7 @@ ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewI
         .viewType = VkImageViewType{},
         .format = format_info.format,
         .components = swizzle_mapping,
-        .subresourceRange = MakeSubresourceRange(aspect_mask, info.range),
+        .subresourceRange = subresource_range,
     };
     const auto create = [&](TextureType tex_type, std::optional<u32> num_layers) {
         VkImageViewCreateInfo ci{create_info};
@@ -3141,6 +3232,19 @@ void TextureCacheRuntime::AccelerateImageUpload(
         return astc_decoder_pass->Assemble(image, map, swizzles);
     }
 
+    if (bl_unswizzle_2d_pass && image.info.type == ImageType::e2D) {
+        return bl_unswizzle_2d_pass->Unswizzle(image, map, swizzles);
+    }
+
+    if (bl_unswizzle_image_3d_pass && image.info.type == ImageType::e3D &&
+        !IsPixelFormatBCn(image.info.format)) {
+        return bl_unswizzle_image_3d_pass->Unswizzle(image, map, swizzles);
+    }
+
+    if (pitch_unswizzle_pass && image.info.type == ImageType::Linear) {
+        return pitch_unswizzle_pass->Unswizzle(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");

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -159,6 +159,9 @@ public:
     std::optional<ASTCDecoderPass> astc_decoder_pass;
 
     std::optional<BlockLinearUnswizzle3DPass> bl3d_unswizzle_pass;
+    std::optional<BlockLinearUnswizzle2DPass> bl_unswizzle_2d_pass;
+    std::optional<BlockLinearUnswizzleImage3DPass> bl_unswizzle_image_3d_pass;
+    std::optional<PitchUnswizzlePass> pitch_unswizzle_pass;
     const Settings::ResolutionScalingInfo& resolution;
     std::array<std::vector<VkFormat>, VideoCore::Surface::MaxPixelFormat> view_formats;
 
@@ -350,7 +353,7 @@ public:
 
     /// Returns true when the image is already initialized and mark it as initialized
     [[nodiscard]] bool ExchangeInitialization() noexcept {
-        return std::exchange(initialized, true);
+        return std::exchange(InitializationFor(current_image), true);
     }
 
     VkImageView StorageImageView(s32 level) noexcept;
@@ -370,6 +373,10 @@ private:
 
     bool NeedsScaleHelper() const;
 
+    std::vector<vk::ImageView>& StorageViewsFor(vk::Image Image::*image);
+
+    bool& InitializationFor(vk::Image Image::*image) noexcept;
+
     Scheduler* scheduler{};
     TextureCacheRuntime* runtime{};
 
@@ -387,8 +394,10 @@ private:
     vk::Image Image::*current_image{};
 
     std::vector<vk::ImageView> storage_image_views;
+    std::vector<vk::ImageView> scaled_storage_image_views;
     VkImageAspectFlags aspect_mask = 0;
-    bool initialized = false;
+    bool original_initialized = false;
+    bool scaled_initialized = false;
 
     std::optional<Framebuffer> scale_framebuffer;
     std::optional<Framebuffer> normal_framebuffer;

src/video_core/texture_cache/accelerated_swizzle.h

@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -23,8 +26,8 @@ struct BlockLinearSwizzle2DParams {
 };
 
 struct BlockLinearSwizzle3DParams {
-    std::array<u32, 3> origin;
-    std::array<s32, 3> destination;
+    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;

src/video_core/texture_cache/texture_cache.h

@@ -20,6 +20,7 @@
 #include "video_core/engines/kepler_compute.h"
 #include "video_core/guest_memory.h"
 #include "video_core/host1x/gpu_device_memory_manager.h"
+#include "video_core/texture_cache/accelerated_swizzle.h"
 #include "video_core/texture_cache/image_view_base.h"
 #include "video_core/texture_cache/samples_helper.h"
 #include "video_core/texture_cache/texture_cache_base.h"
@@ -279,11 +280,11 @@ void TextureCache<P>::CheckFeedbackLoop(std::span<const ImageViewInOut> views) {
 
             const ImageId view_image_id = slot_image_views[view.id].image_id;
             {
-                bool is_continue = false;
+                bool is_feedback = false;
                 for (size_t i = 0; i < 8; ++i)
-                    is_continue |= (rt_active_mask & (1u << i)) && view_image_id == rt_image_id[i];
-                if (is_continue)
-                    continue;
+                    is_feedback |= (rt_active_mask & (1u << i)) && view_image_id == rt_image_id[i];
+                if (is_feedback)
+                    return true;
             }
             if (depth_active && view_image_id == rt_depth_image_id) {
                 return true;
@@ -626,14 +627,25 @@ void TextureCache<P>::DownloadMemory(DAddr cpu_addr, size_t size) {
     std::ranges::sort(images, [this](ImageId lhs, ImageId rhs) {
         return slot_images[lhs].modification_tick < slot_images[rhs].modification_tick;
     });
+    size_t total_size_bytes = 0;
+    for (const ImageId image_id : images) {
+        total_size_bytes += slot_images[image_id].unswizzled_size_bytes;
+    }
+    auto download_map = runtime.DownloadStagingBuffer(total_size_bytes);
     for (const ImageId image_id : images) {
         Image& image = slot_images[image_id];
-        auto map = runtime.DownloadStagingBuffer(image.unswizzled_size_bytes);
         const auto copies = FixSmallVectorADL(FullDownloadCopies(image.info));
-        image.DownloadMemory(map, copies);
-        runtime.Finish();
-        SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span,
+        image.DownloadMemory(download_map, copies);
+        download_map.offset += image.unswizzled_size_bytes;
+    }
+    runtime.Finish();
+    std::span<u8> download_span = download_map.mapped_span;
+    for (const ImageId image_id : images) {
+        const ImageBase& image = slot_images[image_id];
+        const auto copies = FixSmallVectorADL(FullDownloadCopies(image.info));
+        SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span,
                      swizzle_data_buffer);
+        download_span = download_span.subspan(image.unswizzled_size_bytes);
     }
 }
 
@@ -1122,6 +1134,12 @@ void TextureCache<P>::RefreshContents(Image& image, ImageId image_id) {
 
     TrackImage(image, image_id);
 
+    if (image.info.rescaleable &&
+        IsRegionGpuModified(image.cpu_addr, image.guest_size_bytes)) {
+        runtime.TransitionImageLayout(image);
+        return;
+    }
+
     if (image.info.num_samples > 1 && !runtime.CanUploadMSAA()) {
         LOG_WARNING(HW_GPU, "MSAA image uploads are not implemented");
         runtime.TransitionImageLayout(image);
@@ -1157,8 +1175,7 @@ void TextureCache<P>::UploadImageContents(Image& image, StagingBuffer& staging)
     const GPUVAddr gpu_addr = image.gpu_addr;
 
     if (True(image.flags & ImageFlagBits::AcceleratedUpload)) {
-        gpu_memory->ReadBlock(gpu_addr, mapped_span.data(), mapped_span.size_bytes(),
-                              VideoCommon::CacheType::NoTextureCache);
+        gpu_memory->ReadBlockUnsafe(gpu_addr, mapped_span.data(), image.guest_size_bytes);
         const auto uploads = FullUploadSwizzles(image.info);
         runtime.AccelerateImageUpload(image, staging, FixSmallVectorADL(uploads), 0, 0);
         return;
@@ -1265,6 +1282,9 @@ ImageId TextureCache<P>::FindImage(const ImageInfo& info, GPUVAddr gpu_addr,
 
 template <class P>
 bool TextureCache<P>::ImageCanRescale(ImageBase& image) {
+    if (!Settings::values.resolution_info.active) {
+        return false;
+    }
     if (!image.info.rescaleable) {
         return false;
     }
@@ -1352,12 +1372,12 @@ void TextureCache<P>::QueueAsyncDecode(Image& image, ImageId image_id) {
     decode->image_id = image_id;
     async_decodes.push_back(std::move(decode));
 
-    std::vector<u8> local_unswizzle_data_buffer(image.unswizzled_size_bytes, 0);
+    Common::ScratchBuffer<u8> local_unswizzle_data_buffer(image.unswizzled_size_bytes);
     Tegra::Memory::GpuGuestMemory<u8, Tegra::Memory::GuestMemoryFlags::UnsafeRead> swizzle_data(*gpu_memory, image.gpu_addr, image.guest_size_bytes, &swizzle_data_buffer);
     auto copies = UnswizzleImage(*gpu_memory, image.gpu_addr, image.info, swizzle_data, local_unswizzle_data_buffer);
     const size_t out_size = MapSizeBytes(image);
 
-    auto func = [out_size, copies, info = image.info,
+    auto func = [out_size, copies = std::move(copies), info = image.info,
                  input = std::move(local_unswizzle_data_buffer),
                  async_decode = decode_ptr]() mutable {
         async_decode->decoded_data.resize_destructive(out_size);
@@ -1426,18 +1446,16 @@ void TextureCache<P>::TickAsyncUnswizzle() {
     Image& image = slot_images[task.image_id];
 
     if (!task.initialized) {
-        task.total_size = MapSizeBytes(image);
+        task.total_size = image.guest_size_bytes;
         task.staging_buffer = runtime.UploadStagingBuffer(task.total_size, true);
 
-        const auto& info = image.info;
-        const u32 bytes_per_block = BytesPerBlock(info.format);
-        const u32 width_blocks = Common::DivCeil(info.size.width, 4u);
-        const u32 height_blocks = Common::DivCeil(info.size.height, 4u);
-
-        const u32 stride = width_blocks * bytes_per_block;
-        const u32 aligned_height = height_blocks;
-        task.bytes_per_slice = static_cast<size_t>(stride) * aligned_height;
+        const auto layout = FullUploadSwizzles(task.info);
+        const auto params =
+            VideoCommon::Accelerated::MakeBlockLinearSwizzle3DParams(layout.front(), task.info);
+        task.bytes_per_slice = params.slice_size;
+        task.chunked = task.info.block.depth == 0;
         task.last_submitted_offset = 0;
+        task.slices_submitted = 0;
         task.initialized = true;
     }
 
@@ -1452,31 +1470,39 @@ void TextureCache<P>::TickAsyncUnswizzle() {
             if (copy_amount == 0) copy_amount = task.bytes_per_slice;
         }
 
-        gpu_memory->ReadBlock(image.gpu_addr + task.current_offset,
-                              task.staging_buffer.mapped_span.data() + task.current_offset,
-                              copy_amount);
+        gpu_memory->ReadBlockUnsafe(image.gpu_addr + task.current_offset,
+                                    task.staging_buffer.mapped_span.data() + task.current_offset,
+                                    copy_amount);
         task.current_offset += copy_amount;
     }
 
     const bool is_final_batch = task.current_offset >= task.total_size;
-    const size_t bytes_ready = task.current_offset - task.last_submitted_offset;
-    const u32 complete_slices = static_cast<u32>(bytes_ready / task.bytes_per_slice);
 
-    if (complete_slices >= swizzle_slices_per_batch || (is_final_batch && complete_slices > 0)) {
-        const u32 z_start = static_cast<u32>(task.last_submitted_offset / task.bytes_per_slice);
-        const u32 slices_to_process = (std::min)(complete_slices, swizzle_slices_per_batch);
-        const u32 z_count = (std::min)(slices_to_process, image.info.size.depth - z_start);
+    if (task.chunked) {
+        const size_t bytes_ready = task.current_offset - task.last_submitted_offset;
+        const u32 complete_slices = static_cast<u32>(bytes_ready / task.bytes_per_slice);
 
-        if (z_count > 0) {
-            const auto uploads = FullUploadSwizzles(task.info);
-            runtime.AccelerateImageUpload(image, task.staging_buffer, FixSmallVectorADL(uploads), z_start, z_count);
-            task.last_submitted_offset += (static_cast<size_t>(z_count) * task.bytes_per_slice);
+        if (complete_slices >= swizzle_slices_per_batch || (is_final_batch && complete_slices > 0)) {
+            const u32 z_start = task.slices_submitted;
+            const u32 slices_to_process = (std::min)(complete_slices, swizzle_slices_per_batch);
+            const u32 z_count = (std::min)(slices_to_process, image.info.size.depth - z_start);
+
+            if (z_count > 0) {
+                const auto uploads = FullUploadSwizzles(task.info);
+                runtime.AccelerateImageUpload(image, task.staging_buffer,
+                                              FixSmallVectorADL(uploads), z_start, z_count);
+                task.last_submitted_offset += static_cast<size_t>(z_count) * task.bytes_per_slice;
+                task.slices_submitted += z_count;
+            }
         }
+    } else if (is_final_batch && task.slices_submitted == 0) {
+        const auto uploads = FullUploadSwizzles(task.info);
+        runtime.AccelerateImageUpload(image, task.staging_buffer, FixSmallVectorADL(uploads), 0,
+                                      image.info.size.depth);
+        task.slices_submitted = image.info.size.depth;
     }
 
-    // Check if complete
-    const u32 slices_submitted = static_cast<u32>(task.last_submitted_offset / task.bytes_per_slice);
-    const bool all_slices_submitted = slices_submitted >= image.info.size.depth;
+    const bool all_slices_submitted = task.slices_submitted >= image.info.size.depth;
 
     if (is_final_batch && all_slices_submitted) {
         runtime.FreeDeferredStagingBuffer(task.staging_buffer);

src/video_core/texture_cache/texture_cache_base.h

@@ -138,6 +138,8 @@ class TextureCache : public VideoCommon::ChannelSetupCaches<TextureCacheChannelI
         AsyncBuffer staging_buffer;
         size_t last_submitted_offset = 0;
         size_t bytes_per_slice;
+        u32 slices_submitted = 0;
+        bool chunked = false;
         bool initialized = false;
     };
 

src/video_core/vulkan_common/vulkan_device.cpp

@@ -1118,6 +1118,11 @@ bool Device::GetSuitability(bool requires_swapchain) {
             VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_TRANSFORM_FEEDBACK_PROPERTIES_EXT;
         SetNext(next, properties.transform_feedback);
     }
+    if (extensions.vertex_attribute_divisor) {
+        properties.vertex_attribute_divisor.sType =
+            VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VERTEX_ATTRIBUTE_DIVISOR_PROPERTIES_EXT;
+        SetNext(next, properties.vertex_attribute_divisor);
+    }
     if (extensions.maintenance5) {
         properties.maintenance5.sType =
             VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MAINTENANCE_5_PROPERTIES_KHR;

src/video_core/vulkan_common/vulkan_device.h

@@ -70,6 +70,7 @@ VK_DEFINE_HANDLE(VmaAllocator)
     FEATURE(EXT, ProvokingVertex, PROVOKING_VERTEX, provoking_vertex)                              \
     FEATURE(EXT, Robustness2, ROBUSTNESS_2, robustness2)                                           \
     FEATURE(EXT, TransformFeedback, TRANSFORM_FEEDBACK, transform_feedback)                        \
+    FEATURE(EXT, VertexAttributeDivisor, VERTEX_ATTRIBUTE_DIVISOR, vertex_attribute_divisor)       \
     FEATURE(EXT, VertexInputDynamicState, VERTEX_INPUT_DYNAMIC_STATE, vertex_input_dynamic_state)  \
     FEATURE(KHR, Maintenance5, MAINTENANCE_5, maintenance5)                                        \
     FEATURE(KHR, Maintenance6, MAINTENANCE_6, maintenance6)                                        \
@@ -91,7 +92,6 @@ VK_DEFINE_HANDLE(VmaAllocator)
     EXTENSION(EXT, SHADER_STENCIL_EXPORT, shader_stencil_export)                                   \
     EXTENSION(EXT, SHADER_VIEWPORT_INDEX_LAYER, shader_viewport_index_layer)                       \
     EXTENSION(EXT, TOOLING_INFO, tooling_info)                                                     \
-    EXTENSION(EXT, VERTEX_ATTRIBUTE_DIVISOR, vertex_attribute_divisor)                             \
     EXTENSION(KHR, CREATE_RENDERPASS_2, create_renderpass2)                                        \
     EXTENSION(KHR, DEPTH_STENCIL_RESOLVE, depth_stencil_resolve)                                   \
     EXTENSION(KHR, DRAW_INDIRECT_COUNT, draw_indirect_count)                                       \
@@ -354,6 +354,7 @@ public:
 
 #define FN_MAX_LIMIT_LIST \
     FN_MAX_LIMIT_ELEM(ComputeSharedMemorySize) \
+    FN_MAX_LIMIT_ELEM(ComputeWorkGroupInvocations) \
     FN_MAX_LIMIT_ELEM(PerStageDescriptorSampledImages) \
     FN_MAX_LIMIT_ELEM(PerStageResources) \
     FN_MAX_LIMIT_ELEM(DescriptorSetSamplers) \
@@ -679,6 +680,32 @@ FN_MAX_LIMIT_LIST
         return features.host_query_reset.hostQueryReset != VK_FALSE;
     }
 
+    u32 GetMaxVertexAttribDivisor() const {
+        const u32 reported = properties.vertex_attribute_divisor.maxVertexAttribDivisor;
+        if (reported == 0) {
+            return 1;
+        }
+        return reported;
+    }
+
+    bool IsVertexAttributeInstanceRateZeroDivisorSupported() const {
+        return features.vertex_attribute_divisor.vertexAttributeInstanceRateZeroDivisor == VK_TRUE;
+    }
+
+    u32 GetVertexAttribDivisor(u32 frequency) const {
+        const u32 max_divisor = GetMaxVertexAttribDivisor();
+        if (frequency == 0) {
+            if (IsVertexAttributeInstanceRateZeroDivisorSupported()) {
+                return 0;
+            }
+            return max_divisor;
+        }
+        if (frequency > max_divisor) {
+            return max_divisor;
+        }
+        return frequency;
+    }
+
     /// Returns true if the device supports VK_EXT_transform_feedback.
     bool IsExtTransformFeedbackSupported() const {
         return extensions.transform_feedback;
@@ -1174,6 +1201,7 @@ private:
         VkPhysicalDeviceDescriptorBufferPropertiesEXT descriptor_buffer{};
         VkPhysicalDeviceSubgroupSizeControlProperties subgroup_size_control{};
         VkPhysicalDeviceTransformFeedbackPropertiesEXT transform_feedback{};
+        VkPhysicalDeviceVertexAttributeDivisorPropertiesEXT vertex_attribute_divisor{};
         VkPhysicalDeviceMaintenance5PropertiesKHR maintenance5{};
         VkPhysicalDeviceDepthStencilResolveProperties depth_stencil_resolve{};
         VkPhysicalDeviceCustomBorderColorPropertiesEXT custom_border_color{};