Texture Cache: Improve documentation

This commit adds a series of HLE methods for handling 3D textures in
general. This helps games that generate 3D textures on every frame and
may reduce loading times for certain games.

src/video_core/renderer_vulkan/shaders/blit.frag

@@ -0,0 +1,24 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+/*
+ * Build instructions:
+ * $ glslangValidator -V $THIS_FILE -o output.spv
+ * $ spirv-opt -O --strip-debug output.spv -o optimized.spv
+ * $ xxd -i optimized.spv
+ *
+ * Then copy that bytecode to the C++ file
+ */
+
+#version 460 core
+
+layout (location = 0) in vec2 frag_tex_coord;
+
+layout (location = 0) out vec4 color;
+
+layout (binding = 1) uniform sampler2D color_texture;
+
+void main() {
+    color = texture(color_texture, frag_tex_coord);
+}

src/video_core/renderer_vulkan/shaders/blit.vert

@@ -0,0 +1,28 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+/*
+ * Build instructions:
+ * $ glslangValidator -V $THIS_FILE -o output.spv
+ * $ spirv-opt -O --strip-debug output.spv -o optimized.spv
+ * $ xxd -i optimized.spv
+ *
+ * Then copy that bytecode to the C++ file
+ */
+
+#version 460 core
+
+layout (location = 0) in vec2 vert_position;
+layout (location = 1) in vec2 vert_tex_coord;
+
+layout (location = 0) out vec2 frag_tex_coord;
+
+layout (set = 0, binding = 0) uniform MatrixBlock {
+    mat4 modelview_matrix;
+};
+
+void main() {
+    gl_Position = modelview_matrix * vec4(vert_position, 0.0, 1.0);
+    frag_tex_coord = vert_tex_coord;
+}

src/video_core/renderer_vulkan/shaders/quad_array.comp

@@ -0,0 +1,37 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+/*
+ * Build instructions:
+ * $ glslangValidator -V $THIS_FILE -o output.spv
+ * $ spirv-opt -O --strip-debug output.spv -o optimized.spv
+ * $ xxd -i optimized.spv
+ *
+ * Then copy that bytecode to the C++ file
+ */
+
+#version 460 core
+
+layout (local_size_x = 1024) in;
+
+layout (std430, set = 0, binding = 0) buffer OutputBuffer {
+    uint output_indexes[];
+};
+
+layout (push_constant) uniform PushConstants {
+    uint first;
+};
+
+void main() {
+    uint primitive = gl_GlobalInvocationID.x;
+    if (primitive * 6 >= output_indexes.length()) {
+        return;
+    }
+
+    const uint quad_map[6] = uint[](0, 1, 2, 0, 2, 3);
+    for (uint vertex = 0; vertex < 6; ++vertex) {
+        uint index = first + primitive * 4 + quad_map[vertex];
+        output_indexes[primitive * 6 + vertex] = index;
+    }
+}

src/video_core/renderer_vulkan/shaders/uint8.comp

@@ -0,0 +1,33 @@
+// Copyright 2019 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+/*
+ * Build instructions:
+ * $ glslangValidator -V $THIS_FILE -o output.spv
+ * $ spirv-opt -O --strip-debug output.spv -o optimized.spv
+ * $ xxd -i optimized.spv
+ *
+ * Then copy that bytecode to the C++ file
+ */
+
+#version 460 core
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_8bit_storage : require
+
+layout (local_size_x = 1024) in;
+
+layout (std430, set = 0, binding = 0) readonly buffer InputBuffer {
+    uint8_t input_indexes[];
+};
+
+layout (std430, set = 0, binding = 1) writeonly buffer OutputBuffer {
+    uint16_t output_indexes[];
+};
+
+void main() {
+    uint id = gl_GlobalInvocationID.x;
+    if (id < input_indexes.length()) {
+        output_indexes[id] = uint16_t(input_indexes[id]);
+    }
+}

src/video_core/shader/decode/memory.cpp

@@ -22,7 +22,23 @@ using Tegra::Shader::Register;
 
 namespace {
 
-u32 GetUniformTypeElementsCount(Tegra::Shader::UniformType uniform_type) {
+u32 GetLdgMemorySize(Tegra::Shader::UniformType uniform_type) {
+    switch (uniform_type) {
+    case Tegra::Shader::UniformType::UnsignedByte:
+    case Tegra::Shader::UniformType::Single:
+        return 1;
+    case Tegra::Shader::UniformType::Double:
+        return 2;
+    case Tegra::Shader::UniformType::Quad:
+    case Tegra::Shader::UniformType::UnsignedQuad:
+        return 4;
+    default:
+        UNIMPLEMENTED_MSG("Unimplemented size={}!", static_cast<u32>(uniform_type));
+        return 1;
+    }
+}
+
+u32 GetStgMemorySize(Tegra::Shader::UniformType uniform_type) {
     switch (uniform_type) {
     case Tegra::Shader::UniformType::Single:
         return 1;
@@ -170,7 +186,7 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
         const auto [real_address_base, base_address, descriptor] =
             TrackGlobalMemory(bb, instr, false);
 
-        const u32 count = GetUniformTypeElementsCount(type);
+        const u32 count = GetLdgMemorySize(type);
         if (!real_address_base || !base_address) {
             // Tracking failed, load zeroes.
             for (u32 i = 0; i < count; ++i) {
@@ -181,12 +197,22 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
 
         for (u32 i = 0; i < count; ++i) {
             const Node it_offset = Immediate(i * 4);
-            const Node real_address =
-                Operation(OperationCode::UAdd, NO_PRECISE, real_address_base, it_offset);
-            const Node gmem = MakeNode<GmemNode>(real_address, base_address, descriptor);
+            const Node real_address = Operation(OperationCode::UAdd, real_address_base, it_offset);
+            Node gmem = MakeNode<GmemNode>(real_address, base_address, descriptor);
+
+            if (type == Tegra::Shader::UniformType::UnsignedByte) {
+                // To handle unaligned loads get the byte used to dereferenced global memory
+                // and extract that byte from the loaded uint32.
+                Node byte = Operation(OperationCode::UBitwiseAnd, real_address, Immediate(3));
+                byte = Operation(OperationCode::ULogicalShiftLeft, std::move(byte), Immediate(3));
+
+                gmem = Operation(OperationCode::UBitfieldExtract, std::move(gmem), std::move(byte),
+                                 Immediate(8));
+            }
 
             SetTemporary(bb, i, gmem);
         }
+
         for (u32 i = 0; i < count; ++i) {
             SetRegister(bb, instr.gpr0.Value() + i, GetTemporary(i));
         }
@@ -276,7 +302,7 @@ u32 ShaderIR::DecodeMemory(NodeBlock& bb, u32 pc) {
             break;
         }
 
-        const u32 count = GetUniformTypeElementsCount(type);
+        const u32 count = GetStgMemorySize(type);
         for (u32 i = 0; i < count; ++i) {
             const Node it_offset = Immediate(i * 4);
             const Node real_address = Operation(OperationCode::UAdd, real_address_base, it_offset);

src/video_core/texture_cache/surface_params.cpp

@@ -392,4 +392,42 @@ std::string SurfaceParams::TargetName() const {
     }
 }
 
+u32 SurfaceParams::GetBlockSize() const {
+    const u32 x = 64U << block_width;
+    const u32 y = 8U << block_height;
+    const u32 z = 1U << block_depth;
+    return x * y * z;
+}
+
+std::pair<u32, u32> SurfaceParams::GetBlockXY() const {
+    const u32 x_pixels = 64U / GetBytesPerPixel();
+    const u32 x = x_pixels << block_width;
+    const u32 y = 8U << block_height;
+    return {x, y};
+}
+
+std::tuple<u32, u32, u32> SurfaceParams::GetBlockOffsetXYZ(u32 offset) const {
+    const auto div_ceil = [](const u32 x, const u32 y) { return ((x + y - 1) / y); };
+    const u32 block_size = GetBlockSize();
+    const u32 block_index = offset / block_size;
+    const u32 gob_offset = offset % block_size;
+    const u32 gob_index = gob_offset / static_cast<u32>(Tegra::Texture::GetGOBSize());
+    const u32 x_gob_pixels = 64U / GetBytesPerPixel();
+    const u32 x_block_pixels = x_gob_pixels << block_width;
+    const u32 y_block_pixels = 8U << block_height;
+    const u32 z_block_pixels = 1U << block_depth;
+    const u32 x_blocks = div_ceil(width, x_block_pixels);
+    const u32 y_blocks = div_ceil(height, y_block_pixels);
+    const u32 z_blocks = div_ceil(depth, z_block_pixels);
+    const u32 base_x = block_index % x_blocks;
+    const u32 base_y = (block_index / x_blocks) % y_blocks;
+    const u32 base_z = (block_index / (x_blocks * y_blocks)) % z_blocks;
+    u32 x = base_x * x_block_pixels;
+    u32 y = base_y * y_block_pixels;
+    u32 z = base_z * z_block_pixels;
+    z += gob_index >> block_height;
+    y += (gob_index * 8U) % y_block_pixels;
+    return {x, y, z};
+}
+
 } // namespace VideoCommon

src/video_core/texture_cache/surface_params.h

@@ -4,6 +4,8 @@
 
 #pragma once
 
+#include <utility>
+
 #include "common/alignment.h"
 #include "common/bit_util.h"
 #include "common/cityhash.h"
@@ -136,6 +138,15 @@ public:
 
     std::size_t GetConvertedMipmapSize(u32 level) const;
 
+    /// Get this texture Tegra Block size in guest memory layout
+    u32 GetBlockSize() const;
+
+    /// Get X, Y coordinates max sizes of a single block.
+    std::pair<u32, u32> GetBlockXY() const;
+
+    /// Get the offset in x, y, z coordinates from a memory offset
+    std::tuple<u32, u32, u32> GetBlockOffsetXYZ(u32 offset) const;
+
     /// Returns the size of a layer in bytes in guest memory.
     std::size_t GetGuestLayerSize() const {
         return GetLayerSize(false, false);
@@ -269,7 +280,8 @@ private:
 
     /// Returns the size of all mipmap levels and aligns as needed.
     std::size_t GetInnerMemorySize(bool as_host_size, bool layer_only, bool uncompressed) const {
-        return GetLayerSize(as_host_size, uncompressed) * (layer_only ? 1U : depth);
+        return GetLayerSize(as_host_size, uncompressed) *
+               (layer_only ? 1U : (is_layered ? depth : 1U));
     }
 
     /// Returns the size of a layer

src/video_core/texture_cache/texture_cache.h

@@ -615,6 +615,86 @@ private:
         return {{new_surface, new_surface->GetMainView()}};
     }
 
+    /**
+     * Takes care of managing 3D textures and its slices. Does HLE methods for reconstructing the 3D
+     * textures within the GPU if possible. Falls back to LLE when it isn't possible to use any of
+     * the HLE methods.
+     *
+     * @param overlaps          The overlapping surfaces registered in the cache.
+     * @param params            The parameters on the new surface.
+     * @param gpu_addr          The starting address of the new surface.
+     * @param cache_addr        The starting address of the new surface on physical memory.
+     * @param preserve_contents Indicates that the new surface should be loaded from memory or
+     *                          left blank.
+     */
+    std::optional<std::pair<TSurface, TView>> Manage3DSurfaces(std::vector<TSurface>& overlaps,
+                                                               const SurfaceParams& params,
+                                                               const GPUVAddr gpu_addr,
+                                                               const CacheAddr cache_addr,
+                                                               bool preserve_contents) {
+        if (params.target == SurfaceTarget::Texture3D) {
+            bool failed = false;
+            if (params.num_levels > 1) {
+                // We can't handle mipmaps in 3D textures yet, better fallback to LLE approach
+                return std::nullopt;
+            }
+            TSurface new_surface = GetUncachedSurface(gpu_addr, params);
+            bool modified = false;
+            for (auto& surface : overlaps) {
+                const SurfaceParams& src_params = surface->GetSurfaceParams();
+                if (src_params.target != SurfaceTarget::Texture2D) {
+                    failed = true;
+                    break;
+                }
+                if (src_params.height != params.height) {
+                    failed = true;
+                    break;
+                }
+                if (src_params.block_depth != params.block_depth ||
+                    src_params.block_height != params.block_height) {
+                    failed = true;
+                    break;
+                }
+                const u32 offset = static_cast<u32>(surface->GetCacheAddr() - cache_addr);
+                const auto [x, y, z] = params.GetBlockOffsetXYZ(offset);
+                modified |= surface->IsModified();
+                const CopyParams copy_params(0, 0, 0, 0, 0, z, 0, 0, params.width, params.height,
+                                             1);
+                ImageCopy(surface, new_surface, copy_params);
+            }
+            if (failed) {
+                return std::nullopt;
+            }
+            for (const auto& surface : overlaps) {
+                Unregister(surface);
+            }
+            new_surface->MarkAsModified(modified, Tick());
+            Register(new_surface);
+            auto view = new_surface->GetMainView();
+            return {{std::move(new_surface), view}};
+        } else {
+            for (const auto& surface : overlaps) {
+                if (!surface->MatchTarget(params.target)) {
+                    if (overlaps.size() == 1 && surface->GetCacheAddr() == cache_addr) {
+                        if (Settings::values.use_accurate_gpu_emulation) {
+                            return std::nullopt;
+                        }
+                        Unregister(surface);
+                        return InitializeSurface(gpu_addr, params, preserve_contents);
+                    }
+                    return std::nullopt;
+                }
+                if (surface->GetCacheAddr() != cache_addr) {
+                    continue;
+                }
+                if (surface->MatchesStructure(params) == MatchStructureResult::FullMatch) {
+                    return {{surface, surface->GetMainView()}};
+                }
+            }
+            return InitializeSurface(gpu_addr, params, preserve_contents);
+        }
+    }
+
     /**
      * Gets the starting address and parameters of a candidate surface and tries
      * to find a matching surface within the cache. This is done in 3 big steps:
@@ -687,6 +767,15 @@ private:
             }
         }
 
+        // Check if it's a 3D texture
+        if (params.block_depth > 0) {
+            auto surface =
+                Manage3DSurfaces(overlaps, params, gpu_addr, cache_addr, preserve_contents);
+            if (surface) {
+                return *surface;
+            }
+        }
+
         // Split cases between 1 overlap or many.
         if (overlaps.size() == 1) {
             TSurface current_surface = overlaps[0];

src/video_core/textures/decoders.h

@@ -12,6 +12,10 @@ namespace Tegra::Texture {
 
 // GOBSize constant. Calculated by 64 bytes in x multiplied by 8 y coords, represents
 // an small rect of (64/bytes_per_pixel)X8.
+inline std::size_t GetGOBSize() {
+    return 512;
+}
+
 inline std::size_t GetGOBSizeShift() {
     return 9;
 }