video_core: detiler: display micro 64bpp (#2137)

src/video_core/amdgpu/resource.h

@@ -119,6 +119,7 @@ constexpr std::string_view NameOf(ImageType type) {
 enum class TilingMode : u32 {
     Depth_MacroTiled = 0u,
     Display_Linear = 0x8u,
+    Display_MicroTiled = 0x9u,
     Display_MacroTiled = 0xAu,
     Texture_MicroTiled = 0xDu,
     Texture_MacroTiled = 0xEu,
@@ -131,6 +132,8 @@ constexpr std::string_view NameOf(TilingMode type) {
         return "Depth_MacroTiled";
     case TilingMode::Display_Linear:
         return "Display_Linear";
+    case TilingMode::Display_MicroTiled:
+        return "Display_MicroTiled";
     case TilingMode::Display_MacroTiled:
         return "Display_MacroTiled";
     case TilingMode::Texture_MicroTiled:

src/video_core/host_shaders/CMakeLists.txt

@@ -2,6 +2,7 @@
 # SPDX-License-Identifier: GPL-2.0-or-later
 
 set(SHADER_FILES
+    detilers/display_micro_64bpp.comp
     detilers/macro_32bpp.comp
     detilers/macro_64bpp.comp
     detilers/macro_8bpp.comp

src/video_core/host_shaders/detilers/display_micro_64bpp.comp

@@ -0,0 +1,60 @@
+// 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(std430, binding = 1) buffer output_buf {
+    uint out_data[];
+};
+
+layout(push_constant) uniform image_info {
+    uint num_levels;
+    uint pitch;
+    uint height;
+    uint c0;
+    uint c1;
+} info;
+
+const uint lut_64bpp[16] = {
+    0x05040100, 0x0d0c0908,
+    0x07060302, 0x0f0e0b0a,
+    0x15141110, 0x1d1c1918,
+    0x17161312, 0x1f1e1b1a,
+    0x25242120, 0x2d2c2928,
+    0x27262322, 0x2f2e2b2a,
+    0x35343130, 0x3d3c3938,
+    0x37363332, 0x3f3e3b3a,
+};
+
+#define MICRO_TILE_DIM      (8)
+#define MICRO_TILE_SZ       (512)
+#define TEXELS_PER_ELEMENT  (1)
+#define BPP                 (64)
+
+void main() {
+    uint x =  gl_GlobalInvocationID.x % info.pitch;
+    uint y = (gl_GlobalInvocationID.x / info.pitch) % info.height;
+    uint z =  gl_GlobalInvocationID.x / (info.pitch * info.height);
+
+    uint col = bitfieldExtract(x, 0, 3);
+    uint row = bitfieldExtract(y, 0, 3);
+    uint idx_dw = lut_64bpp[(col + row * MICRO_TILE_DIM) >> 2u];
+    uint byte_ofs = gl_LocalInvocationID.x & 3u;
+    uint idx = bitfieldExtract(idx_dw >> (8 * byte_ofs), 0, 8);
+
+    uint slice_offs = z * info.c1 * MICRO_TILE_SZ;
+    uint tile_row = y / MICRO_TILE_DIM;
+    uint tile_column = x / MICRO_TILE_DIM;
+    uint tile_offs = ((tile_row * info.c0) + tile_column) * MICRO_TILE_SZ;
+    uint offs = slice_offs + tile_offs + ((idx * BPP) / 8u);
+
+    uint p0 = in_data[(offs >> 2) + 0];
+    uint p1 = in_data[(offs >> 2) + 1];
+    out_data[2 * gl_GlobalInvocationID.x + 0] = p0;
+    out_data[2 * gl_GlobalInvocationID.x + 1] = p1;
+}

src/video_core/texture_cache/image_info.cpp

@@ -182,6 +182,7 @@ void ImageInfo::UpdateSize() {
         case AmdGpu::TilingMode::Texture_Volume:
             mip_d += (-mip_d) & 3u;
             [[fallthrough]];
+        case AmdGpu::TilingMode::Display_MicroTiled:
         case AmdGpu::TilingMode::Texture_MicroTiled: {
             std::tie(mip_info.pitch, mip_info.size) =
                 ImageSizeMicroTiled(mip_w, mip_h, bpp, num_samples);

src/video_core/texture_cache/texture_cache.cpp

@@ -469,9 +469,6 @@ ImageView& TextureCache::FindDepthTarget(BaseDesc& desc) {
 }
 
 void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_scheduler /*= nullptr*/) {
-    RENDERER_TRACE;
-    TRACE_HINT(fmt::format("{:x}:{:x}", image.info.guest_address, image.info.guest_size));
-
     if (False(image.flags & ImageFlagBits::Dirty)) {
         return;
     }
@@ -480,6 +477,9 @@ void TextureCache::RefreshImage(Image& image, Vulkan::Scheduler* custom_schedule
         return;
     }
 
+    RENDERER_TRACE;
+    TRACE_HINT(fmt::format("{:x}:{:x}", image.info.guest_address, image.info.guest_size));
+
     if (True(image.flags & ImageFlagBits::MaybeCpuDirty) &&
         False(image.flags & ImageFlagBits::CpuDirty)) {
         // The image size should be less than page size to be considered MaybeCpuDirty

src/video_core/texture_cache/tile_manager.cpp

@@ -8,6 +8,7 @@
 #include "video_core/texture_cache/image_view.h"
 #include "video_core/texture_cache/tile_manager.h"
 
+#include "video_core/host_shaders/detilers/display_micro_64bpp_comp.h"
 #include "video_core/host_shaders/detilers/macro_32bpp_comp.h"
 #include "video_core/host_shaders/detilers/macro_64bpp_comp.h"
 #include "video_core/host_shaders/detilers/macro_8bpp_comp.h"
@@ -53,6 +54,14 @@ const DetilerContext* TileManager::GetDetiler(const ImageInfo& info) const {
             return nullptr;
         }
         break;
+    case AmdGpu::TilingMode::Display_MicroTiled:
+        switch (bpp) {
+        case 64:
+            return &detilers[DetilerType::Display_Micro64];
+        default:
+            return nullptr;
+        }
+        break;
     default:
         return nullptr;
     }
@@ -68,10 +77,11 @@ struct DetilerParams {
 TileManager::TileManager(const Vulkan::Instance& instance, Vulkan::Scheduler& scheduler)
     : instance{instance}, scheduler{scheduler} {
     static const std::array detiler_shaders{
-        HostShaders::MICRO_8BPP_COMP,   HostShaders::MICRO_16BPP_COMP,
-        HostShaders::MICRO_32BPP_COMP,  HostShaders::MICRO_64BPP_COMP,
-        HostShaders::MICRO_128BPP_COMP, HostShaders::MACRO_8BPP_COMP,
-        HostShaders::MACRO_32BPP_COMP,  HostShaders::MACRO_64BPP_COMP,
+        HostShaders::MICRO_8BPP_COMP,          HostShaders::MICRO_16BPP_COMP,
+        HostShaders::MICRO_32BPP_COMP,         HostShaders::MICRO_64BPP_COMP,
+        HostShaders::MICRO_128BPP_COMP,        HostShaders::MACRO_8BPP_COMP,
+        HostShaders::MACRO_32BPP_COMP,         HostShaders::MACRO_64BPP_COMP,
+        HostShaders::DISPLAY_MICRO_64BPP_COMP,
     };
 
     boost::container::static_vector<vk::DescriptorSetLayoutBinding, 2> bindings{
@@ -258,7 +268,8 @@ std::pair<vk::Buffer, u32> TileManager::TryDetile(vk::Buffer in_buffer, u32 in_o
     params.num_levels = info.resources.levels;
     params.pitch0 = info.pitch >> (info.props.is_block ? 2u : 0u);
     params.height = info.size.height;
-    if (info.tiling_mode == AmdGpu::TilingMode::Texture_Volume) {
+    if (info.tiling_mode == AmdGpu::TilingMode::Texture_Volume ||
+        info.tiling_mode == AmdGpu::TilingMode::Display_MicroTiled) {
         ASSERT(info.resources.levels == 1);
         const auto tiles_per_row = info.pitch / 8u;
         const auto tiles_per_slice = tiles_per_row * ((info.size.height + 7u) / 8u);

src/video_core/texture_cache/tile_manager.h

@@ -22,6 +22,8 @@ enum DetilerType : u32 {
     Macro32,
     Macro64,
 
+    Display_Micro64,
+
     Max
 };