texture_cache: 8bpp macro detiler

src/video_core/host_shaders/CMakeLists.txt

@@ -7,6 +7,7 @@ set(SHADER_FILES
     detile_m32x1.comp
     detile_m32x2.comp
     detile_m32x4.comp
+    detile_macro8x1.comp
     detile_macro32x1.comp
     detile_macro32x2.comp
     fs_tri.vert

src/video_core/host_shaders/detile_macro8x1.comp

@@ -0,0 +1,101 @@
+// 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_8bpp[][16] = {
+    {
+        0x05040100, 0x45444140,
+        0x07060302, 0x47464342,
+        0x0d0c0908, 0x4d4c4948,
+        0x0f0e0b0a, 0x4f4e4b4a,
+        0x85848180, 0xc5c4c1c0,
+        0x87868382, 0xc7c6c3c2,
+        0x8d8c8988, 0xcdccc9c8,
+        0x8f8e8b8a, 0xcfcecbca,
+    },
+    {
+        0x15141110, 0x55545150,
+        0x17161312, 0x57565352,
+        0x1d1c1918, 0x5d5c5958,
+        0x1f1e1b1a, 0x5f5e5b5a,
+        0x95949190, 0xd5d4d1d0,
+        0x97969392, 0xd7d6d3d2,
+        0x9d9c9998, 0xdddcd9d8,
+        0x9f9e9b9a, 0xdfdedbda,
+    },
+    {
+        0x25242120, 0x65646160,
+        0x27262322, 0x67666362,
+        0x2d2c2928, 0x6d6c6968,
+        0x2f2e2b2a, 0x6f6e6b6a,
+        0xa5a4a1a0, 0xe5e4e1e0,
+        0xa7a6a3a2, 0xe7e6e3e2,
+        0xadaca9a8, 0xedece9e8,
+        0xafaeabaa, 0xefeeebea,
+    },
+    {
+        0x35343130, 0x75747170,
+        0x37363332, 0x77767372,
+        0x3d3c3938, 0x7d7c7978,
+        0x3f3e3b3a, 0x7f7e7b7a,
+        0xb5b4b1b0, 0xf5f4f1f0,
+        0xb7b6b3b2, 0xf7f6f3f2,
+        0xbdbcb9b8, 0xfdfcf9f8,
+        0xbfbebbba, 0xfffefbfa,
+    },
+};
+
+#define MICRO_TILE_DIM      (8)
+#define MICRO_TILE_SZ       (256)
+#define TEXELS_PER_ELEMENT  (1)
+#define BPP                 (8)
+
+shared uint scratch[16];
+
+void main() {
+    uint slot = gl_LocalInvocationID.x >> 2u;
+    atomicAnd(scratch[slot], 0);
+
+    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 lut = bitfieldExtract(z, 0, 2);
+    uint idx_dw = lut_8bpp[lut][(col + row * MICRO_TILE_DIM) >> 2u];
+    uint byte_ofs = (gl_LocalInvocationID.x & 3u) * 8;
+    uint idx = bitfieldExtract(idx_dw >> byte_ofs, 0, 8);
+
+    uint slice_offs = (z >> 2u) * 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 / 8);
+
+    uint p0 = in_data[offs >> 2u];
+    uint byte = bitfieldExtract(p0 >> (offs * 8), 0, 8);
+    atomicOr(scratch[slot], byte << byte_ofs);
+
+    if (byte_ofs == 0) {
+        out_data[gl_GlobalInvocationID.x >> 2u] = scratch[slot];
+    }
+}

src/video_core/texture_cache/tile_manager.cpp

@@ -15,6 +15,7 @@
 #include "video_core/host_shaders/detile_m8x2_comp.h"
 #include "video_core/host_shaders/detile_macro32x1_comp.h"
 #include "video_core/host_shaders/detile_macro32x2_comp.h"
+#include "video_core/host_shaders/detile_macro8x1_comp.h"
 
 #include <boost/container/static_vector.hpp>
 #include <magic_enum/magic_enum.hpp>
@@ -108,6 +109,8 @@ const DetilerContext* TileManager::GetDetiler(const ImageInfo& info) const {
         }
     case AmdGpu::TilingMode::Texture_Volume:
         switch (format) {
+        case vk::Format::eR8Uint:
+            return &detilers[DetilerType::Macro8x1];
         case vk::Format::eR32Uint:
             return &detilers[DetilerType::Macro32x1];
         case vk::Format::eR32G32Uint:
@@ -133,8 +136,8 @@ TileManager::TileManager(const Vulkan::Instance& instance, Vulkan::Scheduler& sc
     static const std::array detiler_shaders{
         HostShaders::DETILE_M8X1_COMP,      HostShaders::DETILE_M8X2_COMP,
         HostShaders::DETILE_M32X1_COMP,     HostShaders::DETILE_M32X2_COMP,
-        HostShaders::DETILE_M32X4_COMP,     HostShaders::DETILE_MACRO32X1_COMP,
+        HostShaders::DETILE_M32X4_COMP,     HostShaders::DETILE_MACRO8X1_COMP,
-        HostShaders::DETILE_MACRO32X2_COMP,
+        HostShaders::DETILE_MACRO32X1_COMP, HostShaders::DETILE_MACRO32X2_COMP,
     };
 
     boost::container::static_vector<vk::DescriptorSetLayoutBinding, 2> bindings{
@@ -323,7 +326,6 @@ std::pair<vk::Buffer, u32> TileManager::TryDetile(vk::Buffer in_buffer, u32 in_o
     params.height = info.size.height;
     if (info.tiling_mode == AmdGpu::TilingMode::Texture_Volume) {
         ASSERT(info.resources.levels == 1);
-        ASSERT(info.num_bits >= 32);
         const auto tiles_per_row = info.pitch / 8u;
         const auto tiles_per_slice = tiles_per_row * ((info.size.height + 7u) / 8u);
         params.sizes[0] = tiles_per_row;

src/video_core/texture_cache/tile_manager.h

@@ -18,6 +18,7 @@ enum DetilerType : u32 {
     Micro32x2,
     Micro32x4,
 
+    Macro8x1,
     Macro32x1,
     Macro32x2,