Merge pull request #4176 from ReinUsesLisp/compatible-formats

texture_cache: Check format compatibility before copying

src/video_core/CMakeLists.txt

@@ -3,6 +3,8 @@ add_library(video_core STATIC
     buffer_cache/buffer_cache.h
     buffer_cache/map_interval.cpp
     buffer_cache/map_interval.h
+    compatible_formats.cpp
+    compatible_formats.h
     dirty_flags.cpp
     dirty_flags.h
     dma_pusher.cpp

src/video_core/compatible_formats.cpp

@@ -0,0 +1,162 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <array>
+#include <bitset>
+#include <cstddef>
+
+#include "video_core/compatible_formats.h"
+#include "video_core/surface.h"
+
+namespace VideoCore::Surface {
+
+namespace {
+
+// Compatibility table taken from Table 3.X.2 in:
+// https://www.khronos.org/registry/OpenGL/extensions/ARB/ARB_texture_view.txt
+
+constexpr std::array VIEW_CLASS_128_BITS = {
+    PixelFormat::RGBA32F,
+    PixelFormat::RGBA32UI,
+};
+// Missing formats:
+// PixelFormat::RGBA32I
+
+constexpr std::array VIEW_CLASS_96_BITS = {
+    PixelFormat::RGB32F,
+};
+// Missing formats:
+// PixelFormat::RGB32UI,
+// PixelFormat::RGB32I,
+
+constexpr std::array VIEW_CLASS_64_BITS = {
+    PixelFormat::RGBA16F, PixelFormat::RG32F,   PixelFormat::RGBA16UI, PixelFormat::RG32UI,
+    PixelFormat::RGBA16U, PixelFormat::RGBA16F, PixelFormat::RGBA16S,
+};
+// Missing formats:
+// PixelFormat::RGBA16I
+// PixelFormat::RG32I
+
+// TODO: How should we handle 48 bits?
+
+constexpr std::array VIEW_CLASS_32_BITS = {
+    PixelFormat::RG16F,        PixelFormat::R11FG11FB10F, PixelFormat::R32F,
+    PixelFormat::A2B10G10R10U, PixelFormat::RG16UI,       PixelFormat::R32UI,
+    PixelFormat::RG16I,        PixelFormat::R32I,         PixelFormat::ABGR8U,
+    PixelFormat::RG16,         PixelFormat::ABGR8S,       PixelFormat::RG16S,
+    PixelFormat::RGBA8_SRGB,   PixelFormat::E5B9G9R9F,    PixelFormat::BGRA8,
+    PixelFormat::BGRA8_SRGB,
+};
+// Missing formats:
+// PixelFormat::RGBA8UI
+// PixelFormat::RGBA8I
+// PixelFormat::RGB10_A2_UI
+
+// TODO: How should we handle 24 bits?
+
+constexpr std::array VIEW_CLASS_16_BITS = {
+    PixelFormat::R16F, PixelFormat::RG8UI, PixelFormat::R16UI, PixelFormat::R16I,
+    PixelFormat::RG8U, PixelFormat::R16U,  PixelFormat::RG8S,  PixelFormat::R16S,
+};
+// Missing formats:
+// PixelFormat::RG8I
+
+constexpr std::array VIEW_CLASS_8_BITS = {
+    PixelFormat::R8UI,
+    PixelFormat::R8U,
+};
+// Missing formats:
+// PixelFormat::R8I
+// PixelFormat::R8S
+
+constexpr std::array VIEW_CLASS_RGTC1_RED = {
+    PixelFormat::DXN1,
+};
+// Missing formats:
+// COMPRESSED_SIGNED_RED_RGTC1
+
+constexpr std::array VIEW_CLASS_RGTC2_RG = {
+    PixelFormat::DXN2UNORM,
+    PixelFormat::DXN2SNORM,
+};
+
+constexpr std::array VIEW_CLASS_BPTC_UNORM = {
+    PixelFormat::BC7U,
+    PixelFormat::BC7U_SRGB,
+};
+
+constexpr std::array VIEW_CLASS_BPTC_FLOAT = {
+    PixelFormat::BC6H_SF16,
+    PixelFormat::BC6H_UF16,
+};
+
+// Compatibility table taken from Table 4.X.1 in:
+// https://www.khronos.org/registry/OpenGL/extensions/ARB/ARB_copy_image.txt
+
+constexpr std::array COPY_CLASS_128_BITS = {
+    PixelFormat::RGBA32UI,   PixelFormat::RGBA32F,   PixelFormat::DXT23,
+    PixelFormat::DXT23_SRGB, PixelFormat::DXT45,     PixelFormat::DXT45_SRGB,
+    PixelFormat::DXN2SNORM,  PixelFormat::BC7U,      PixelFormat::BC7U_SRGB,
+    PixelFormat::BC6H_SF16,  PixelFormat::BC6H_UF16,
+};
+// Missing formats:
+// PixelFormat::RGBA32I
+// COMPRESSED_RG_RGTC2
+
+constexpr std::array COPY_CLASS_64_BITS = {
+    PixelFormat::RGBA16F, PixelFormat::RG32F,   PixelFormat::RGBA16UI,  PixelFormat::RG32UI,
+    PixelFormat::RGBA16U, PixelFormat::RGBA16S, PixelFormat::DXT1_SRGB, PixelFormat::DXT1,
+
+};
+// Missing formats:
+// PixelFormat::RGBA16I
+// PixelFormat::RG32I,
+// COMPRESSED_RGB_S3TC_DXT1_EXT
+// COMPRESSED_SRGB_S3TC_DXT1_EXT
+// COMPRESSED_RGBA_S3TC_DXT1_EXT
+// COMPRESSED_SIGNED_RED_RGTC1
+
+void Enable(FormatCompatibility::Table& compatiblity, size_t format_a, size_t format_b) {
+    compatiblity[format_a][format_b] = true;
+    compatiblity[format_b][format_a] = true;
+}
+
+void Enable(FormatCompatibility::Table& compatibility, PixelFormat format_a, PixelFormat format_b) {
+    Enable(compatibility, static_cast<size_t>(format_a), static_cast<size_t>(format_b));
+}
+
+template <typename Range>
+void EnableRange(FormatCompatibility::Table& compatibility, const Range& range) {
+    for (auto it_a = range.begin(); it_a != range.end(); ++it_a) {
+        for (auto it_b = it_a; it_b != range.end(); ++it_b) {
+            Enable(compatibility, *it_a, *it_b);
+        }
+    }
+}
+
+} // Anonymous namespace
+
+FormatCompatibility::FormatCompatibility() {
+    for (size_t i = 0; i < MaxPixelFormat; ++i) {
+        // Identity is allowed
+        Enable(view, i, i);
+    }
+
+    EnableRange(view, VIEW_CLASS_128_BITS);
+    EnableRange(view, VIEW_CLASS_96_BITS);
+    EnableRange(view, VIEW_CLASS_64_BITS);
+    EnableRange(view, VIEW_CLASS_32_BITS);
+    EnableRange(view, VIEW_CLASS_16_BITS);
+    EnableRange(view, VIEW_CLASS_8_BITS);
+    EnableRange(view, VIEW_CLASS_RGTC1_RED);
+    EnableRange(view, VIEW_CLASS_RGTC2_RG);
+    EnableRange(view, VIEW_CLASS_BPTC_UNORM);
+    EnableRange(view, VIEW_CLASS_BPTC_FLOAT);
+
+    copy = view;
+    EnableRange(copy, COPY_CLASS_128_BITS);
+    EnableRange(copy, COPY_CLASS_64_BITS);
+}
+
+} // namespace VideoCore::Surface

src/video_core/compatible_formats.h

@@ -0,0 +1,32 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <array>
+#include <bitset>
+#include <cstddef>
+
+#include "video_core/surface.h"
+
+namespace VideoCore::Surface {
+
+class FormatCompatibility {
+public:
+    using Table = std::array<std::bitset<MaxPixelFormat>, MaxPixelFormat>;
+
+    explicit FormatCompatibility();
+
+    bool TestView(PixelFormat format_a, PixelFormat format_b) const noexcept {
+        return view[static_cast<size_t>(format_a)][static_cast<size_t>(format_b)];
+    }
+
+    bool TestCopy(PixelFormat format_a, PixelFormat format_b) const noexcept {
+        return copy[static_cast<size_t>(format_a)][static_cast<size_t>(format_b)];
+    }
+
+private:
+    Table view;
+    Table copy;
+};
+
+} // namespace VideoCore::Surface

src/video_core/texture_cache/texture_cache.h

@@ -24,6 +24,7 @@
 #include "core/core.h"
 #include "core/memory.h"
 #include "core/settings.h"
+#include "video_core/compatible_formats.h"
 #include "video_core/dirty_flags.h"
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/engines/maxwell_3d.h"
@@ -47,8 +48,8 @@ class RasterizerInterface;
 
 namespace VideoCommon {
 
+using VideoCore::Surface::FormatCompatibility;
 using VideoCore::Surface::PixelFormat;
-
 using VideoCore::Surface::SurfaceTarget;
 using RenderTargetConfig = Tegra::Engines::Maxwell3D::Regs::RenderTargetConfig;
 
@@ -595,7 +596,7 @@ private:
         } else {
             new_surface = GetUncachedSurface(gpu_addr, params);
         }
-        const auto& final_params = new_surface->GetSurfaceParams();
+        const SurfaceParams& final_params = new_surface->GetSurfaceParams();
         if (cr_params.type != final_params.type) {
             if (Settings::IsGPULevelExtreme()) {
                 BufferCopy(current_surface, new_surface);
@@ -603,7 +604,7 @@ private:
         } else {
             std::vector<CopyParams> bricks = current_surface->BreakDown(final_params);
             for (auto& brick : bricks) {
-                ImageCopy(current_surface, new_surface, brick);
+                TryCopyImage(current_surface, new_surface, brick);
             }
         }
         Unregister(current_surface);
@@ -694,7 +695,7 @@ private:
                 }
                 const CopyParams copy_params(0, 0, 0, 0, 0, base_layer, 0, mipmap, width, height,
                                              src_params.depth);
-                ImageCopy(surface, new_surface, copy_params);
+                TryCopyImage(surface, new_surface, copy_params);
             }
         }
         if (passed_tests == 0) {
@@ -791,7 +792,7 @@ private:
             const u32 width = params.width;
             const u32 height = params.height;
             const CopyParams copy_params(0, 0, 0, 0, 0, slice, 0, 0, width, height, 1);
-            ImageCopy(surface, new_surface, copy_params);
+            TryCopyImage(surface, new_surface, copy_params);
         }
         for (const auto& surface : overlaps) {
             Unregister(surface);
@@ -1192,6 +1193,19 @@ private:
         return {};
     }
 
+    /// Try to do an image copy logging when formats are incompatible.
+    void TryCopyImage(TSurface& src, TSurface& dst, const CopyParams& copy) {
+        const SurfaceParams& src_params = src->GetSurfaceParams();
+        const SurfaceParams& dst_params = dst->GetSurfaceParams();
+        if (!format_compatibility.TestCopy(src_params.pixel_format, dst_params.pixel_format)) {
+            LOG_ERROR(HW_GPU, "Illegal copy between formats={{{}, {}}}",
+                      static_cast<int>(dst_params.pixel_format),
+                      static_cast<int>(src_params.pixel_format));
+            return;
+        }
+        ImageCopy(src, dst, copy);
+    }
+
     constexpr PixelFormat GetSiblingFormat(PixelFormat format) const {
         return siblings_table[static_cast<std::size_t>(format)];
     }
@@ -1241,6 +1255,7 @@ private:
     VideoCore::RasterizerInterface& rasterizer;
 
     FormatLookupTable format_lookup_table;
+    FormatCompatibility format_compatibility;
 
     u64 ticks{};