Merge pull request #163 from shadps4-emu/video_core/surface_works

Basic detiling

CMakeLists.txt

@@ -530,6 +530,13 @@ endif()
 
 target_include_directories(shadps4 PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
 
+# Shaders sources
+set(HOST_SHADERS_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/src/video_core/host_shaders)
+
+add_subdirectory(${HOST_SHADERS_INCLUDE})
+add_dependencies(shadps4 host_shaders)
+target_include_directories(shadps4 PRIVATE ${HOST_SHADERS_INCLUDE})
+
 if (ENABLE_QT_GUI)
     set_target_properties(shadps4 PROPERTIES
         WIN32_EXECUTABLE ON

src/core/libraries/gnmdriver/gnmdriver.cpp

@@ -1462,7 +1462,7 @@ s32 PS4_SYSV_ABI sceGnmSubmitCommandBuffers(u32 count, const u32* dcb_gpu_addrs[
         if (Config::dumpPM4()) {
             static auto last_frame_num = frames_submitted;
             static u32 seq_num{};
-            if (last_frame_num == frames_submitted) {
+            if (last_frame_num && last_frame_num == frames_submitted) {
                 ++seq_num;
             } else {
                 last_frame_num = frames_submitted;

src/video_core/amdgpu/liverpool.cpp

@@ -199,7 +199,7 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
                 ASSERT(col_buf_id < NumColorBuffers);
 
                 const auto nop_offset = header->type3.count;
-                if (nop_offset == 0x0e) {
+                if (nop_offset == 0x0e || nop_offset == 0x0d) {
                     ASSERT_MSG(payload[nop_offset] == 0xc0001000,
                                "NOP hint is missing in CB setup sequence");
                     last_cb_extent[col_buf_id].raw = payload[nop_offset + 1];

src/video_core/amdgpu/liverpool.h

@@ -6,6 +6,7 @@
 #include "common/assert.h"
 #include "common/bit_field.h"
 #include "common/types.h"
+#include "resource.h"
 #include "video_core/amdgpu/pixel_format.h"
 
 #include <array>
@@ -622,7 +623,7 @@ struct Liverpool {
             BitField<19, 1, u32> cmask_is_linear;
         } info;
         union {
-            BitField<0, 5, u32> tile_mode_index;
+            BitField<0, 5, TilingMode> tile_mode_index;
             BitField<5, 5, u32> fmask_tile_mode_index;
             BitField<12, 3, u32> num_samples_log2;
             BitField<15, 3, u32> num_fragments_log2;
@@ -661,6 +662,22 @@ struct Liverpool {
             return u64(cmask_base_address) << 8;
         }
 
+        [[nodiscard]] size_t GetSizeAligned() const {
+            const auto num_bytes_per_element = NumBits(info.format) / 8u;
+            const auto slice_size = (slice.tile_max + 1) * 64u;
+            const auto total_size = slice_size * (view.slice_max + 1) * num_bytes_per_element;
+            ASSERT(total_size > 0);
+            return total_size;
+        }
+
+        [[nodiscard]] TilingMode GetTilingMode() const {
+            return attrib.tile_mode_index;
+        }
+
+        [[nodiscard]] bool IsTiled() const {
+            return !info.linear_general;
+        }
+
         NumberFormat NumFormat() const {
             // There is a small difference between T# and CB number types, account for it.
             return info.number_type == AmdGpu::NumberFormat::Uscaled ? AmdGpu::NumberFormat::Srgb
@@ -834,7 +851,9 @@ private:
             static constexpr std::suspend_always final_suspend() noexcept {
                 return {};
             }
-            void unhandled_exception() {}
+            void unhandled_exception() {
+                UNREACHABLE();
+            }
             void return_void() {}
             struct empty {};
             std::suspend_always yield_value(empty&&) {

src/video_core/amdgpu/pixel_format.cpp

@@ -40,17 +40,30 @@ std::string_view NameOf(NumberFormat fmt) {
     }
 }
 
-u32 NumComponents(DataFormat format) {
+int NumComponents(DataFormat format) {
-    constexpr std::array numComponentsPerElement = {
+    constexpr std::array num_components_per_element = {
         0,  1,  1,  2,  1,  2,  3,  3,  4,  4,  4,  2, 4, 3, 4, -1, 3, 4, 4, 4, 2,
         2,  2,  -1, -1, -1, -1, -1, -1, -1, -1, -1, 3, 3, 3, 4, 4,  4, 1, 2, 3, 4,
         -1, -1, 2,  2,  2,  2,  2,  2,  2,  2,  2,  2, 2, 2, 2, 2,  3, 1, 1};
 
     const u32 index = static_cast<u32>(format);
-    if (index >= numComponentsPerElement.size()) {
+    if (index >= num_components_per_element.size()) {
         return 0;
     }
-    return numComponentsPerElement[index];
+    return num_components_per_element[index];
+}
+
+int NumBits(DataFormat format) {
+    const std::array num_bits_per_element = {
+        0,  8,  16, 16, 32, 32, 32, 32, 32, 32, 32, 64, 64, 96, 128, -1, 16, 16, 16, 16, 32,
+        32, 64, -1, -1, -1, -1, -1, -1, -1, -1, -1, 16, 16, 32, 4,   8,  8,  4,  8,  8,  8,
+        -1, -1, 8,  8,  8,  8,  8,  8,  16, 16, 32, 32, 32, 64, 64,  8,  16, 1,  1};
+
+    const u32 index = static_cast<u32>(format);
+    if (index >= num_bits_per_element.size()) {
+        return 0;
+    }
+    return num_bits_per_element[index];
 }
 
 } // namespace AmdGpu

src/video_core/amdgpu/pixel_format.h

@@ -63,7 +63,8 @@ enum class NumberFormat : u32 {
 
 [[nodiscard]] std::string_view NameOf(NumberFormat fmt);
 
-u32 NumComponents(DataFormat format);
+int NumComponents(DataFormat format);
+int NumBits(DataFormat format);
 
 } // namespace AmdGpu
 

src/video_core/amdgpu/resource.h

@@ -85,6 +85,12 @@ constexpr std::string_view NameOf(ImageType type) {
     }
 }
 
+enum class TilingMode : u32 {
+    Display_Linear = 0x8u,
+    Display_MacroTiled = 0xAu,
+    Texture_MicroTiled = 0xDu,
+};
+
 struct Image {
     union {
         BitField<0, 38, u64> base_address;
@@ -122,7 +128,7 @@ struct Image {
     }
 
     u32 Pitch() const {
-        return pitch;
+        return pitch + 1;
     }
 
     u32 NumLayers() const {
@@ -140,6 +146,19 @@ struct Image {
     NumberFormat GetNumberFmt() const noexcept {
         return static_cast<NumberFormat>(num_format.Value());
     }
+
+    [[nodiscard]] TilingMode GetTilingMode() const {
+        return static_cast<TilingMode>(tiling_index.Value());
+    }
+
+    [[nodiscard]] bool IsTiled() const {
+        return GetTilingMode() != TilingMode::Display_Linear;
+    }
+
+    [[nodiscard]] size_t GetSizeAligned() const {
+        // TODO: Derive this properly from tiling params
+        return (width + 1) * (height + 1) * NumComponents(GetDataFmt());
+    }
 };
 
 // 8.2.7. Image Sampler [RDNA 2 Instruction Set Architecture]

src/video_core/host_shaders/CMakeLists.txt

@@ -0,0 +1,42 @@
+# SPDX-FileCopyrightText: Copyright 2023 Citra Emulator Project
+# SPDX-License-Identifier: GPL-2.0-or-later
+
+set(SHADER_FILES
+    detile_m8x1.comp
+    detile_m8x4.comp
+)
+
+set(SHADER_INCLUDE ${CMAKE_CURRENT_BINARY_DIR}/include)
+set(SHADER_DIR ${SHADER_INCLUDE}/video_core/host_shaders)
+set(HOST_SHADERS_INCLUDE ${SHADER_INCLUDE} PARENT_SCOPE)
+
+set(INPUT_FILE ${CMAKE_CURRENT_SOURCE_DIR}/source_shader.h.in)
+set(HEADER_GENERATOR ${CMAKE_CURRENT_SOURCE_DIR}/StringShaderHeader.cmake)
+
+foreach(FILENAME IN ITEMS ${SHADER_FILES})
+    string(REPLACE "." "_" SHADER_NAME ${FILENAME})
+    set(SOURCE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/${FILENAME})
+    set(SOURCE_HEADER_FILE ${SHADER_DIR}/${SHADER_NAME}.h)
+    add_custom_command(
+        OUTPUT
+            ${SOURCE_HEADER_FILE}
+        COMMAND
+            ${CMAKE_COMMAND} -P ${HEADER_GENERATOR} ${SOURCE_FILE} ${SOURCE_HEADER_FILE} ${INPUT_FILE}
+        MAIN_DEPENDENCY
+            ${SOURCE_FILE}
+        DEPENDS
+            ${INPUT_FILE}
+            # HEADER_GENERATOR should be included here but msbuild seems to assume it's always modified
+    )
+    set(SHADER_HEADERS ${SHADER_HEADERS} ${SOURCE_HEADER_FILE})
+endforeach()
+
+set(SHADER_SOURCES ${SHADER_FILES})
+list(APPEND SHADER_SOURCES ${GLSL_INCLUDES})
+
+add_custom_target(host_shaders
+    DEPENDS
+        ${SHADER_HEADERS}
+    SOURCES
+        ${SHADER_SOURCES}
+)

src/video_core/host_shaders/StringShaderHeader.cmake

@@ -0,0 +1,36 @@
+# SPDX-FileCopyrightText: 2020 yuzu Emulator Project
+# SPDX-License-Identifier: GPL-2.0-or-later
+
+set(SOURCE_FILE ${CMAKE_ARGV3})
+set(HEADER_FILE ${CMAKE_ARGV4})
+set(INPUT_FILE ${CMAKE_ARGV5})
+
+get_filename_component(CONTENTS_NAME ${SOURCE_FILE} NAME)
+string(REPLACE "." "_" CONTENTS_NAME ${CONTENTS_NAME})
+string(TOUPPER ${CONTENTS_NAME} CONTENTS_NAME)
+
+FILE(READ ${SOURCE_FILE} line_contents)
+
+# Replace double quotes with single quotes,
+# as double quotes will be used to wrap the lines
+STRING(REGEX REPLACE "\"" "'" line_contents "${line_contents}")
+
+# CMake separates list elements with semicolons, but semicolons
+# are used extensively in the shader code.
+# Replace with a temporary marker, to be reverted later.
+STRING(REGEX REPLACE ";" "{{SEMICOLON}}" line_contents "${line_contents}")
+
+# Make every line an individual element in the CMake list.
+STRING(REGEX REPLACE "\n" ";" line_contents "${line_contents}")
+
+# Build the shader string, wrapping each line in double quotes.
+foreach(line IN LISTS line_contents)
+    string(CONCAT CONTENTS "${CONTENTS}" \"${line}\\n\"\n)
+endforeach()
+
+# Revert the original semicolons in the source.
+STRING(REGEX REPLACE "{{SEMICOLON}}" ";" CONTENTS "${CONTENTS}")
+
+get_filename_component(OUTPUT_DIR ${HEADER_FILE} DIRECTORY)
+make_directory(${OUTPUT_DIR})
+configure_file(${INPUT_FILE} ${HEADER_FILE} @ONLY)

src/video_core/host_shaders/detile_m8x1.comp

@@ -0,0 +1,48 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#version 450
+#extension GL_KHR_shader_subgroup_shuffle : require
+
+// NOTE: Current subgroup utilization is subotimal on most GPUs, so
+// it will be nice to process two tiles at once here.
+layout (local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
+
+layout(std430, binding = 0) buffer input_buf {
+    uint in_data[];
+};
+layout(r8ui, binding = 1) uniform writeonly uimage2D output_img;
+
+layout(push_constant) uniform image_info {
+    uint pitch;
+} info;
+
+#define MICRO_TILE_DIM      8
+#define TEXELS_PER_ELEMENT  4
+
+void main() {
+    uint p0 = in_data[gl_GlobalInvocationID.x];
+    uint p1 = subgroupShuffleXor(p0, 1);
+    uint hword = gl_LocalInvocationID.x & 1;
+    uint dst_tx = (hword == 1)
+        ? (p0 & 0xffff0000) | (p1 >> 16)
+        : (p0 & 0x0000ffff) | (p1 << 16);
+
+    uint col = (gl_LocalInvocationID.x >> 2) & 1;
+    uint row = (gl_LocalInvocationID.x % TEXELS_PER_ELEMENT)
+                + TEXELS_PER_ELEMENT * (gl_LocalInvocationID.x >> 3);
+
+    uint tiles_per_pitch = info.pitch >> 3; // log2(MICRO_TILE_DIM)
+    uint target_tile_x = gl_WorkGroupID.x % tiles_per_pitch;
+    uint target_tile_y = gl_WorkGroupID.x / tiles_per_pitch;
+    uint dw_ofs_x = target_tile_x * MICRO_TILE_DIM + TEXELS_PER_ELEMENT * col;
+    uint dw_ofs_y = target_tile_y * MICRO_TILE_DIM + row;
+
+    ivec2 img_pos = ivec2(dw_ofs_x, dw_ofs_y);
+
+    #pragma unroll
+    for (int ofs = 0; ofs < TEXELS_PER_ELEMENT; ++ofs) {
+        imageStore(output_img, img_pos + ivec2(ofs, 0), uvec4(dst_tx & 0xff));
+        dst_tx >>= 8;
+    }
+}

src/video_core/host_shaders/detile_m8x4.comp

@@ -0,0 +1,57 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#version 450
+
+layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout(std430, binding = 0) buffer input_buf {
+    uint in_data[];
+};
+layout(rgba8ui, binding = 1) uniform writeonly uimage2D output_img;
+
+layout(push_constant) uniform image_info {
+    uint pitch;
+} info;
+
+// Inverse morton LUT, small enough to fit into K$
+uint lut_8x4[16] = {
+    0x11011000, 0x31213020,
+    0x13031202, 0x33233222,
+    0x51415040, 0x71617060,
+    0x53435242, 0x73637262,
+
+    0x15051404, 0x35253424,
+    0x17071606, 0x37273626,
+    0x55455444, 0x75657464,
+    0x57475646, 0x77677666,
+};
+
+#define MICRO_TILE_DIM      8
+#define TEXELS_PER_ELEMENT  1
+
+void main() {
+    uint src_tx = in_data[gl_GlobalInvocationID.x];
+
+    uint bit_ofs = 8 * (gl_LocalInvocationID.x % 4);
+    uint packed_pos = lut_8x4[gl_LocalInvocationID.x >> 2] >> bit_ofs;
+    uint col = bitfieldExtract(packed_pos, 4, 4);
+    uint row = bitfieldExtract(packed_pos, 0, 4);
+
+    uint tiles_per_pitch = info.pitch >> 3; // log2(MICRO_TILE_DIM)
+    uint target_tile_x = gl_WorkGroupID.x % tiles_per_pitch;
+    uint target_tile_y = gl_WorkGroupID.x / tiles_per_pitch;
+
+    uint dw_ofs_x = target_tile_x * MICRO_TILE_DIM + TEXELS_PER_ELEMENT * col;
+    uint dw_ofs_y = target_tile_y * MICRO_TILE_DIM + row;
+
+    ivec2 img_pos = ivec2(dw_ofs_x, dw_ofs_y);
+
+    uvec4 dst_tx = uvec4(
+        bitfieldExtract(src_tx,  0, 8),
+        bitfieldExtract(src_tx,  8, 8),
+        bitfieldExtract(src_tx, 16, 8),
+        bitfieldExtract(src_tx, 24, 8)
+    );
+    imageStore(output_img, img_pos, dst_tx);
+}

src/video_core/host_shaders/source_shader.h.in

@@ -0,0 +1,14 @@
+// SPDX-FileCopyrightText: Copyright 2022 Citra Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include <string_view>
+
+namespace HostShaders {
+
+constexpr std::string_view @CONTENTS_NAME@ = {
+@CONTENTS@
+};
+
+} // namespace HostShaders

src/video_core/renderer_vulkan/vk_instance.cpp

@@ -150,7 +150,6 @@ bool Instance::CreateDevice() {
     external_memory_host = add_extension(VK_EXT_EXTERNAL_MEMORY_HOST_EXTENSION_NAME);
     tooling_info = add_extension(VK_EXT_TOOLING_INFO_EXTENSION_NAME);
     custom_border_color = add_extension(VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME);
-    index_type_uint8 = add_extension(VK_KHR_INDEX_TYPE_UINT8_EXTENSION_NAME);
     add_extension(VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME);
     add_extension(VK_KHR_MAINTENANCE_4_EXTENSION_NAME);
     // The next two extensions are required to be available together in order to support write masks
@@ -219,9 +218,6 @@ bool Instance::CreateDevice() {
             .customBorderColors = true,
             .customBorderColorWithoutFormat = true,
         },
-        vk::PhysicalDeviceIndexTypeUint8FeaturesEXT{
-            .indexTypeUint8 = true,
-        },
         vk::PhysicalDeviceColorWriteEnableFeaturesEXT{
             .colorWriteEnable = true,
         },
@@ -230,10 +226,6 @@ bool Instance::CreateDevice() {
         },
     };
 
-    if (!index_type_uint8) {
-        device_chain.unlink<vk::PhysicalDeviceIndexTypeUint8FeaturesEXT>();
-    }
-
     if (!color_write_en) {
         device_chain.unlink<vk::PhysicalDeviceColorWriteEnableFeaturesEXT>();
         device_chain.unlink<vk::PhysicalDeviceExtendedDynamicState3FeaturesEXT>();

src/video_core/renderer_vulkan/vk_instance.h

@@ -81,11 +81,6 @@ public:
         return custom_border_color;
     }
 
-    /// Returns true when VK_EXT_index_type_uint8 is supported
-    bool IsIndexTypeUint8Supported() const {
-        return index_type_uint8;
-    }
-
     /// Returns true when VK_EXT_fragment_shader_interlock is supported
     bool IsFragmentShaderInterlockSupported() const {
         return fragment_shader_interlock;
@@ -216,7 +211,6 @@ private:
     bool image_view_reinterpretation{true};
     bool timeline_semaphores{};
     bool custom_border_color{};
-    bool index_type_uint8{};
     bool fragment_shader_interlock{};
     bool image_format_list{};
     bool pipeline_creation_cache_control{};

src/video_core/texture_cache/image.cpp

@@ -7,6 +7,7 @@
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/texture_cache/image.h"
+#include "video_core/texture_cache/tile_manager.h"
 
 #include <vk_mem_alloc.h>
 
@@ -16,7 +17,7 @@ using namespace Vulkan;
 using VideoOutFormat = Libraries::VideoOut::PixelFormat;
 using Libraries::VideoOut::TilingMode;
 
-[[nodiscard]] vk::Format ConvertPixelFormat(const VideoOutFormat format) {
+static vk::Format ConvertPixelFormat(const VideoOutFormat format) {
     switch (format) {
     case VideoOutFormat::A8R8G8B8Srgb:
         return vk::Format::eB8G8R8A8Srgb;
@@ -32,7 +33,7 @@ using Libraries::VideoOut::TilingMode;
     return {};
 }
 
-[[nodiscard]] vk::ImageUsageFlags ImageUsageFlags(const vk::Format format) {
+static vk::ImageUsageFlags ImageUsageFlags(const vk::Format format) {
     vk::ImageUsageFlags usage = vk::ImageUsageFlagBits::eTransferSrc |
                                 vk::ImageUsageFlagBits::eTransferDst |
                                 vk::ImageUsageFlagBits::eSampled;
@@ -46,7 +47,7 @@ using Libraries::VideoOut::TilingMode;
     return usage;
 }
 
-[[nodiscard]] vk::ImageType ConvertImageType(AmdGpu::ImageType type) noexcept {
+static vk::ImageType ConvertImageType(AmdGpu::ImageType type) noexcept {
     switch (type) {
     case AmdGpu::ImageType::Color1D:
         return vk::ImageType::e1D;
@@ -86,18 +87,19 @@ ImageInfo::ImageInfo(const Libraries::VideoOut::BufferAttributeGroup& group) noe
 
 ImageInfo::ImageInfo(const AmdGpu::Liverpool::ColorBuffer& buffer,
                      const AmdGpu::Liverpool::CbDbExtent& hint /*= {}*/) noexcept {
-    is_tiled = true;
+    is_tiled = buffer.IsTiled();
     pixel_format = LiverpoolToVK::SurfaceFormat(buffer.info.format, buffer.NumFormat());
     type = vk::ImageType::e2D;
     size.width = hint.Valid() ? hint.width : buffer.Pitch();
     size.height = hint.Valid() ? hint.height : buffer.Height();
     size.depth = 1;
     pitch = size.width;
-    guest_size_bytes = buffer.slice.tile_max * (buffer.view.slice_max + 1);
+    guest_size_bytes = buffer.GetSizeAligned();
 }
 
 ImageInfo::ImageInfo(const AmdGpu::Image& image) noexcept {
-    is_tiled = false;
+    is_tiled = image.IsTiled();
+    tiling_mode = image.GetTilingMode();
     pixel_format = LiverpoolToVK::SurfaceFormat(image.GetDataFmt(), image.GetNumberFmt());
     type = ConvertImageType(image.type);
     size.width = image.width + 1;
@@ -106,8 +108,7 @@ ImageInfo::ImageInfo(const AmdGpu::Image& image) noexcept {
     pitch = image.Pitch();
     resources.levels = image.NumLevels();
     resources.layers = image.NumLayers();
-    // TODO: Derive this properly from tiling params
+    guest_size_bytes = image.GetSizeAligned();
-    guest_size_bytes = size.width * size.height * 4;
 }
 
 UniqueImage::UniqueImage(vk::Device device_, VmaAllocator allocator_)
@@ -151,6 +152,18 @@ Image::Image(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_,
     if (info.type == vk::ImageType::e3D) {
         flags |= vk::ImageCreateFlagBits::e2DArrayCompatible;
     }
+    if (info.is_tiled) {
+        flags |= vk::ImageCreateFlagBits::eExtendedUsage;
+        if (false) { // IsBlockCodedFormat()
+            flags |= vk::ImageCreateFlagBits::eBlockTexelViewCompatible;
+        }
+    }
+
+    info.usage = ImageUsageFlags(info.pixel_format);
+    if (info.is_tiled || info.is_storage) {
+        info.usage |= vk::ImageUsageFlagBits::eStorage;
+    }
+
     const vk::ImageCreateInfo image_ci = {
         .flags = flags,
         .imageType = info.type,
@@ -163,12 +176,20 @@ Image::Image(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_,
         .mipLevels = static_cast<u32>(info.resources.levels),
         .arrayLayers = static_cast<u32>(info.resources.layers),
         .tiling = vk::ImageTiling::eOptimal,
-        .usage = ImageUsageFlags(info.pixel_format),
+        .usage = info.usage,
         .initialLayout = vk::ImageLayout::eUndefined,
     };
 
     image.Create(image_ci);
 
+    // Create a special view for detiler
+    if (info.is_tiled) {
+        ImageViewInfo view_info;
+        view_info.format = DemoteImageFormatForDetiling(info.pixel_format);
+        view_info.used_for_detiling = true;
+        view_for_detiler.emplace(*instance, view_info, image);
+    }
+
     Transit(vk::ImageLayout::eGeneral, vk::AccessFlagBits::eNone);
 }
 

src/video_core/texture_cache/image.h

@@ -12,6 +12,8 @@
 #include "video_core/texture_cache/image_view.h"
 #include "video_core/texture_cache/types.h"
 
+#include <optional>
+
 namespace Vulkan {
 class Instance;
 class Scheduler;
@@ -39,12 +41,15 @@ struct ImageInfo {
     explicit ImageInfo(const AmdGpu::Image& image) noexcept;
 
     bool is_tiled = false;
+    bool is_storage = false;
     vk::Format pixel_format = vk::Format::eUndefined;
     vk::ImageType type = vk::ImageType::e1D;
+    vk::ImageUsageFlags usage;
     SubresourceExtent resources;
     Extent3D size{1, 1, 1};
     u32 pitch = 0;
     u32 guest_size_bytes = 0;
+    AmdGpu::TilingMode tiling_mode{AmdGpu::TilingMode::Display_Linear};
 };
 
 struct UniqueImage {
@@ -114,6 +119,7 @@ struct Image {
     VAddr cpu_addr_end = 0;
     std::vector<ImageViewInfo> image_view_infos;
     std::vector<ImageViewId> image_view_ids;
+    std::optional<ImageView> view_for_detiler;
 
     // Resource state tracking
     vk::Flags<vk::PipelineStageFlagBits> pl_stage = vk::PipelineStageFlagBits::eAllCommands;

src/video_core/texture_cache/image_view.cpp

@@ -58,10 +58,16 @@ ImageViewInfo::ImageViewInfo(const AmdGpu::Image& image) noexcept {
     mapping.a = ConvertComponentSwizzle(image.dst_sel_w);
 }
 
-ImageView::ImageView(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
+ImageView::ImageView(const Vulkan::Instance& instance, const ImageViewInfo& info_, vk::Image image,
-                     const ImageViewInfo& info_, vk::Image image)
+                     std::optional<vk::ImageUsageFlags> usage_override /*= {}*/)
     : info{info_} {
+    vk::ImageViewUsageCreateInfo usage_ci{};
+    if (usage_override) {
+        usage_ci.usage = usage_override.value();
+    }
+
     const vk::ImageViewCreateInfo image_view_ci = {
+        .pNext = usage_override.has_value() ? &usage_ci : nullptr,
         .image = image,
         .viewType = info.type,
         .format = info.format,

src/video_core/texture_cache/image_view.h

@@ -7,6 +7,8 @@
 #include "video_core/renderer_vulkan/vk_common.h"
 #include "video_core/texture_cache/types.h"
 
+#include <optional>
+
 namespace Vulkan {
 class Instance;
 class Scheduler;
@@ -22,13 +24,14 @@ struct ImageViewInfo {
     vk::Format format = vk::Format::eR8G8B8A8Unorm;
     SubresourceRange range;
     vk::ComponentMapping mapping{};
+    bool used_for_detiling = false;
 
     auto operator<=>(const ImageViewInfo&) const = default;
 };
 
 struct ImageView {
-    explicit ImageView(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler,
+    explicit ImageView(const Vulkan::Instance& instance, const ImageViewInfo& info, vk::Image image,
-                       const ImageViewInfo& info, vk::Image image);
+                       std::optional<vk::ImageUsageFlags> usage_override = {});
     ~ImageView();
 
     ImageView(const ImageView&) = delete;

src/video_core/texture_cache/texture_cache.cpp

@@ -5,6 +5,7 @@
 #include "common/assert.h"
 #include "common/config.h"
 #include "core/virtual_memory.h"
+#include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/texture_cache/texture_cache.h"
 #include "video_core/texture_cache/tile_manager.h"
@@ -64,7 +65,8 @@ static constexpr u64 PageShift = 12;
 TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler& scheduler_)
     : instance{instance_}, scheduler{scheduler_},
       staging{instance, scheduler, vk::BufferUsageFlagBits::eTransferSrc, StreamBufferSize,
-              Vulkan::BufferType::Upload} {
+              Vulkan::BufferType::Upload},
+      tile_manager{instance, scheduler} {
 
 #ifndef _WIN64
     sigset_t signal_mask;
@@ -91,7 +93,7 @@ TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler&
     ASSERT(null_id.index == 0);
 
     ImageViewInfo view_info;
-    void(slot_image_views.insert(instance, scheduler, view_info, slot_images[null_id].image));
+    void(slot_image_views.insert(instance, view_info, slot_images[null_id].image));
 }
 
 TextureCache::~TextureCache() {
@@ -138,21 +140,33 @@ Image& TextureCache::FindImage(const ImageInfo& info, VAddr cpu_address) {
     return image;
 }
 
-ImageView& TextureCache::FindImageView(const AmdGpu::Image& desc) {
+ImageView& TextureCache::RegisterImageView(Image& image, const ImageViewInfo& view_info) {
-    Image& image = FindImage(ImageInfo{desc}, desc.Address());
-
-    const ImageViewInfo view_info{desc};
     if (const ImageViewId view_id = image.FindView(view_info); view_id) {
         return slot_image_views[view_id];
     }
 
+    // All tiled images are created with storage usage flag. This makes set of formats (e.g. sRGB)
+    // impossible to use. However, during view creation, if an image isn't used as storage and not a
+    // target for the detiler, we can temporary remove its storage bit.
+    std::optional<vk::ImageUsageFlags> usage_override;
+    if (!image.info.is_storage && !view_info.used_for_detiling) {
+        usage_override = image.info.usage & ~vk::ImageUsageFlagBits::eStorage;
+    }
+
     const ImageViewId view_id =
-        slot_image_views.insert(instance, scheduler, view_info, image.image);
+        slot_image_views.insert(instance, view_info, image.image, usage_override);
     image.image_view_infos.emplace_back(view_info);
     image.image_view_ids.emplace_back(view_id);
     return slot_image_views[view_id];
 }
 
+ImageView& TextureCache::FindImageView(const AmdGpu::Image& desc) {
+    Image& image = FindImage(ImageInfo{desc}, desc.Address());
+
+    const ImageViewInfo view_info{desc};
+    return RegisterImageView(image, view_info);
+}
+
 ImageView& TextureCache::RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
                                       const AmdGpu::Liverpool::CbDbExtent& hint) {
     const ImageInfo info{buffer, hint};
@@ -160,15 +174,7 @@ ImageView& TextureCache::RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buff
 
     ImageViewInfo view_info;
     view_info.format = info.pixel_format;
-    if (const ImageViewId view_id = image.FindView(view_info); view_id) {
+    return RegisterImageView(image, view_info);
-        return slot_image_views[view_id];
-    }
-
-    const ImageViewId view_id =
-        slot_image_views.insert(instance, scheduler, view_info, image.image);
-    image.image_view_infos.emplace_back(view_info);
-    image.image_view_ids.emplace_back(view_id);
-    return slot_image_views[view_id];
 }
 
 void TextureCache::RefreshImage(Image& image) {
@@ -176,52 +182,48 @@ void TextureCache::RefreshImage(Image& image) {
     image.flags &= ~ImageFlagBits::CpuModified;
 
     {
-
+        if (!tile_manager.TryDetile(image)) {
-        // Upload data to the staging buffer.
+            // Upload data to the staging buffer.
-        const auto [data, offset, _] = staging.Map(image.info.guest_size_bytes, 4);
+            const auto& [data, offset, _] = staging.Map(image.info.guest_size_bytes, 4);
-        const u8* image_data = reinterpret_cast<const u8*>(image.cpu_addr);
+            const u8* image_data = reinterpret_cast<const u8*>(image.cpu_addr);
-        if (image.info.is_tiled) {
-            ConvertTileToLinear(data, image_data, image.info.size.width, image.info.size.height,
-                                Config::isNeoMode());
-        } else {
             std::memcpy(data, image_data, image.info.guest_size_bytes);
+            staging.Commit(image.info.guest_size_bytes);
+
+            const auto cmdbuf = scheduler.CommandBuffer();
+            image.Transit(vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits::eTransferWrite);
+
+            // Copy to the image.
+            const vk::BufferImageCopy image_copy = {
+                .bufferOffset = offset,
+                .bufferRowLength = 0,
+                .bufferImageHeight = 0,
+                .imageSubresource{
+                    .aspectMask = vk::ImageAspectFlagBits::eColor,
+                    .mipLevel = 0,
+                    .baseArrayLayer = 0,
+                    .layerCount = 1,
+                },
+                .imageOffset = {0, 0, 0},
+                .imageExtent = {image.info.size.width, image.info.size.height, 1},
+            };
+
+            cmdbuf.copyBufferToImage(staging.Handle(), image.image,
+                                     vk::ImageLayout::eTransferDstOptimal, image_copy);
         }
-        staging.Commit(image.info.guest_size_bytes);
-
-        // Copy to the image.
-        const vk::BufferImageCopy image_copy = {
-            .bufferOffset = offset,
-            .bufferRowLength = 0,
-            .bufferImageHeight = 0,
-            .imageSubresource{
-                .aspectMask = vk::ImageAspectFlagBits::eColor,
-                .mipLevel = 0,
-                .baseArrayLayer = 0,
-                .layerCount = 1,
-            },
-            .imageOffset = {0, 0, 0},
-            .imageExtent = {image.info.size.width, image.info.size.height, 1},
-        };
-
-        const auto cmdbuf = scheduler.CommandBuffer();
-        const vk::ImageSubresourceRange range = {
-            .aspectMask = vk::ImageAspectFlagBits::eColor,
-            .baseMipLevel = 0,
-            .levelCount = 1,
-            .baseArrayLayer = 0,
-            .layerCount = VK_REMAINING_ARRAY_LAYERS,
-        };
-
-        image.Transit(vk::ImageLayout::eTransferDstOptimal, vk::AccessFlagBits::eTransferWrite);
-
-        cmdbuf.copyBufferToImage(staging.Handle(), image.image,
-                                 vk::ImageLayout::eTransferDstOptimal, image_copy);
 
         image.Transit(vk::ImageLayout::eGeneral,
                       vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eTransferRead);
         return;
     }
 
+    const vk::ImageSubresourceRange range = {
+        .aspectMask = vk::ImageAspectFlagBits::eColor,
+        .baseMipLevel = 0,
+        .levelCount = 1,
+        .baseArrayLayer = 0,
+        .layerCount = VK_REMAINING_ARRAY_LAYERS,
+    };
+
     const u8* image_data = reinterpret_cast<const u8*>(image.cpu_addr);
     for (u32 l = 0; l < image.info.resources.layers; l++) {
         // Upload data to the staging buffer.

src/video_core/texture_cache/texture_cache.h

@@ -13,6 +13,7 @@
 #include "video_core/texture_cache/image_view.h"
 #include "video_core/texture_cache/sampler.h"
 #include "video_core/texture_cache/slot_vector.h"
+#include "video_core/texture_cache/tile_manager.h"
 
 namespace Core::Libraries::VideoOut {
 struct BufferAttributeGroup;
@@ -36,22 +37,24 @@ public:
     void OnCpuWrite(VAddr address);
 
     /// Retrieves the image handle of the image with the provided attributes and address.
-    Image& FindImage(const ImageInfo& info, VAddr cpu_address);
+    [[nodiscard]] Image& FindImage(const ImageInfo& info, VAddr cpu_address);
 
     /// Retrieves an image view with the properties of the specified image descriptor.
-    ImageView& FindImageView(const AmdGpu::Image& image);
+    [[nodiscard]] ImageView& FindImageView(const AmdGpu::Image& image);
 
     /// Retrieves the render target with specified properties
-    ImageView& RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
+    [[nodiscard]] ImageView& RenderTarget(const AmdGpu::Liverpool::ColorBuffer& buffer,
-                            const AmdGpu::Liverpool::CbDbExtent& hint);
+                                          const AmdGpu::Liverpool::CbDbExtent& hint);
 
     /// Reuploads image contents.
     void RefreshImage(Image& image);
 
     /// Retrieves the sampler that matches the provided S# descriptor.
-    vk::Sampler GetSampler(const AmdGpu::Sampler& sampler);
+    [[nodiscard]] vk::Sampler GetSampler(const AmdGpu::Sampler& sampler);
 
 private:
+    ImageView& RegisterImageView(Image& image, const ImageViewInfo& view_info);
+
     /// Iterate over all page indices in a range
     template <typename Func>
     static void ForEachPage(PAddr addr, size_t size, Func&& func) {
@@ -128,6 +131,7 @@ private:
     const Vulkan::Instance& instance;
     Vulkan::Scheduler& scheduler;
     Vulkan::StreamBuffer staging;
+    TileManager tile_manager;
     SlotVector<Image> slot_images;
     SlotVector<ImageView> slot_image_views;
     tsl::robin_map<u64, Sampler> samplers;

src/video_core/texture_cache/tile_manager.cpp

@@ -1,10 +1,20 @@
 // SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
-#include <cstring>
+#include "video_core/renderer_vulkan/vk_instance.h"
-#include "common/assert.h"
+#include "video_core/renderer_vulkan/vk_scheduler.h"
+#include "video_core/renderer_vulkan/vk_shader_util.h"
+#include "video_core/texture_cache/image_view.h"
+#include "video_core/texture_cache/texture_cache.h"
 #include "video_core/texture_cache/tile_manager.h"
 
+#include "video_core/host_shaders/detile_m8x1_comp.h"
+#include "video_core/host_shaders/detile_m8x4_comp.h"
+
+#include <boost/container/static_vector.hpp>
+#include <magic_enum.hpp>
+#include <vulkan/vulkan_to_string.hpp>
+
 namespace VideoCore {
 
 static u32 IntLog2(u32 i) {
@@ -162,4 +172,188 @@ void ConvertTileToLinear(u8* dst, const u8* src, u32 width, u32 height, bool is_
     }
 }
 
+vk::Format DemoteImageFormatForDetiling(vk::Format format) {
+    switch (format) {
+    case vk::Format::eB8G8R8A8Srgb:
+    case vk::Format::eR8G8B8A8Unorm:
+        return vk::Format::eR8G8B8A8Uint;
+    case vk::Format::eR8Unorm:
+        return vk::Format::eR8Uint;
+    default:
+        LOG_ERROR(Render_Vulkan, "Unexpected format for demotion {}", vk::to_string(format));
+        break;
+    }
+    return format;
+}
+
+const DetilerContext* TileManager::GetDetiler(const Image& image) const {
+    const auto format = DemoteImageFormatForDetiling(image.info.pixel_format);
+
+    if (image.info.tiling_mode == AmdGpu::TilingMode::Texture_MicroTiled) {
+        switch (format) {
+        case vk::Format::eR8Uint:
+            return &detilers[DetilerType::Micro8x1];
+        case vk::Format::eR8G8B8A8Uint:
+            return &detilers[DetilerType::Micro8x4];
+        default:
+            return nullptr;
+        }
+    }
+    return nullptr;
+}
+
+static constexpr vk::BufferUsageFlags StagingFlags = vk::BufferUsageFlagBits::eTransferDst |
+                                                     vk::BufferUsageFlagBits::eUniformBuffer |
+                                                     vk::BufferUsageFlagBits::eStorageBuffer;
+
+TileManager::TileManager(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler)
+    : instance{instance}, scheduler{scheduler}, staging{instance, scheduler, StagingFlags, 64_MB} {
+
+    static const std::array detiler_shaders{
+        HostShaders::DETILE_M8X1_COMP,
+        HostShaders::DETILE_M8X4_COMP,
+    };
+
+    for (int pl_id = 0; pl_id < DetilerType::Max; ++pl_id) {
+        auto& ctx = detilers[pl_id];
+
+        const auto& module = Vulkan::Compile(
+            detiler_shaders[pl_id], vk::ShaderStageFlagBits::eCompute, instance.GetDevice());
+
+        // Set module debug name
+        auto module_name = magic_enum::enum_name(static_cast<DetilerType>(pl_id));
+        const vk::DebugUtilsObjectNameInfoEXT name_info = {
+            .objectType = vk::ObjectType::eShaderModule,
+            .objectHandle = std::bit_cast<u64>(module),
+            .pObjectName = module_name.data(),
+        };
+        instance.GetDevice().setDebugUtilsObjectNameEXT(name_info);
+
+        const vk::PipelineShaderStageCreateInfo shader_ci = {
+            .stage = vk::ShaderStageFlagBits::eCompute,
+            .module = module,
+            .pName = "main",
+        };
+
+        boost::container::static_vector<vk::DescriptorSetLayoutBinding, 2> bindings{
+            {
+                .binding = 0,
+                .descriptorType = vk::DescriptorType::eStorageBuffer,
+                .descriptorCount = 1,
+                .stageFlags = vk::ShaderStageFlagBits::eCompute,
+            },
+            {
+                .binding = 1,
+                .descriptorType = vk::DescriptorType::eStorageImage,
+                .descriptorCount = 1,
+                .stageFlags = vk::ShaderStageFlagBits::eCompute,
+            },
+        };
+
+        const vk::DescriptorSetLayoutCreateInfo desc_layout_ci = {
+            .flags = vk::DescriptorSetLayoutCreateFlagBits::ePushDescriptorKHR,
+            .bindingCount = static_cast<u32>(bindings.size()),
+            .pBindings = bindings.data(),
+        };
+        static auto desc_layout =
+            instance.GetDevice().createDescriptorSetLayoutUnique(desc_layout_ci);
+
+        const vk::PushConstantRange push_constants = {
+            .stageFlags = vk::ShaderStageFlagBits::eCompute,
+            .offset = 0,
+            .size = sizeof(u32),
+        };
+
+        const vk::DescriptorSetLayout set_layout = *desc_layout;
+        const vk::PipelineLayoutCreateInfo layout_info = {
+            .setLayoutCount = 1U,
+            .pSetLayouts = &set_layout,
+            .pushConstantRangeCount = 1,
+            .pPushConstantRanges = &push_constants,
+        };
+        ctx.pl_layout = instance.GetDevice().createPipelineLayoutUnique(layout_info);
+
+        const vk::ComputePipelineCreateInfo compute_pipeline_ci = {
+            .stage = shader_ci,
+            .layout = *ctx.pl_layout,
+        };
+        auto result = instance.GetDevice().createComputePipelineUnique(
+            /*pipeline_cache*/ {}, compute_pipeline_ci);
+        if (result.result == vk::Result::eSuccess) {
+            ctx.pl = std::move(result.value);
+        } else {
+            UNREACHABLE_MSG("Detiler pipeline creation failed!");
+        }
+
+        // Once pipeline is compiled, we don't need the shader module anymore
+        instance.GetDevice().destroyShaderModule(module);
+    }
+}
+
+TileManager::~TileManager() = default;
+
+bool TileManager::TryDetile(Image& image) {
+    if (!image.info.is_tiled) {
+        return false;
+    }
+
+    const auto* detiler = GetDetiler(image);
+    if (!detiler) {
+        LOG_ERROR(Render_Vulkan, "Unsupported tiled image: {} {}",
+                  vk::to_string(image.info.pixel_format), static_cast<u32>(image.info.tiling_mode));
+        return false;
+    }
+
+    const auto& [data, offset, _] = staging.Map(image.info.guest_size_bytes, 4);
+    const u8* image_data = reinterpret_cast<const u8*>(image.cpu_addr);
+    std::memcpy(data, image_data, image.info.guest_size_bytes);
+    staging.Commit(image.info.guest_size_bytes);
+
+    auto cmdbuf = scheduler.CommandBuffer();
+    cmdbuf.bindPipeline(vk::PipelineBindPoint::eCompute, *detiler->pl);
+
+    image.Transit(vk::ImageLayout::eGeneral, vk::AccessFlagBits::eShaderWrite);
+
+    const vk::DescriptorBufferInfo input_buffer_info{
+        .buffer = staging.Handle(),
+        .offset = offset,
+        .range = image.info.guest_size_bytes,
+    };
+
+    ASSERT(image.view_for_detiler.has_value());
+    const vk::DescriptorImageInfo output_image_info{
+        .imageView = *image.view_for_detiler->image_view,
+        .imageLayout = image.layout,
+    };
+
+    std::vector<vk::WriteDescriptorSet> set_writes{
+        {
+            .dstSet = VK_NULL_HANDLE,
+            .dstBinding = 0,
+            .dstArrayElement = 0,
+            .descriptorCount = 1,
+            .descriptorType = vk::DescriptorType::eStorageBuffer,
+            .pBufferInfo = &input_buffer_info,
+        },
+        {
+            .dstSet = VK_NULL_HANDLE,
+            .dstBinding = 1,
+            .dstArrayElement = 0,
+            .descriptorCount = 1,
+            .descriptorType = vk::DescriptorType::eStorageImage,
+            .pImageInfo = &output_image_info,
+        },
+    };
+    cmdbuf.pushDescriptorSetKHR(vk::PipelineBindPoint::eCompute, *detiler->pl_layout, 0,
+                                set_writes);
+
+    cmdbuf.pushConstants(*detiler->pl_layout, vk::ShaderStageFlagBits::eCompute, 0u,
+                         sizeof(image.info.pitch), &image.info.pitch);
+
+    cmdbuf.dispatch((image.info.size.width * image.info.size.height) / 64, 1,
+                    1); // round to 64
+
+    return true;
+}
+
 } // namespace VideoCore

src/video_core/texture_cache/tile_manager.h

@@ -4,10 +4,46 @@
 #pragma once
 
 #include "common/types.h"
+#include "video_core/renderer_vulkan/vk_stream_buffer.h"
+#include "video_core/texture_cache/image.h"
 
 namespace VideoCore {
 
+class TextureCache;
+
 /// Converts tiled texture data to linear format.
 void ConvertTileToLinear(u8* dst, const u8* src, u32 width, u32 height, bool neo);
 
+/// Converts image format to the one used internally by detiler.
+vk::Format DemoteImageFormatForDetiling(vk::Format format);
+
+enum DetilerType : u32 {
+    Micro8x1,
+    Micro8x4,
+
+    Max
+};
+
+struct DetilerContext {
+    vk::UniquePipeline pl;
+    vk::UniquePipelineLayout pl_layout;
+};
+
+class TileManager {
+public:
+    TileManager(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler);
+    ~TileManager();
+
+    bool TryDetile(Image& image);
+
+private:
+    const DetilerContext* GetDetiler(const Image& image) const;
+
+private:
+    const Vulkan::Instance& instance;
+    Vulkan::Scheduler& scheduler;
+    Vulkan::StreamBuffer staging;
+    std::array<DetilerContext, DetilerType::Max> detilers;
+};
+
 } // namespace VideoCore