Engines: Implement Accelerate DMA Texture.

7 files changed, 218 insertions, 12 deletions

diff --git a/src/video_core/texture_cache/image_info.cpp b/src/video_core/texture_cache/image_info.cpp
index e9100091e..a1296b574 100644
--- a/src/video_core/texture_cache/image_info.cpp
+++ b/src/video_core/texture_cache/image_info.cpp
@@ -216,10 +216,51 @@ ImageInfo::ImageInfo(const Tegra::Engines::Fermi2D::Surface& config) noexcept {
             .height = config.height,
             .depth = 1,
         };
-        rescaleable = block.depth == 0;
-        rescaleable &= size.height > 256;
+        rescaleable = block.depth == 0 && size.height > 256;
         downscaleable = size.height > 512;
     }
 }
 
+static PixelFormat ByteSizeToFormat(u32 bytes_per_pixel) {
+    switch (bytes_per_pixel) {
+    case 1:
+        return PixelFormat::R8_UINT;
+    case 2:
+        return PixelFormat::R8G8_UINT;
+    case 4:
+        return PixelFormat::A8B8G8R8_UINT;
+    case 8:
+        return PixelFormat::R16G16B16A16_UINT;
+    case 16:
+        return PixelFormat::R32G32B32A32_UINT;
+    default:
+        UNIMPLEMENTED();
+        return PixelFormat::Invalid;
+    }
+}
+
+ImageInfo::ImageInfo(const Tegra::DMA::ImageOperand& config) noexcept {
+    const u32 bytes_per_pixel = config.bytes_per_pixel;
+    format = ByteSizeToFormat(bytes_per_pixel);
+    type = config.params.block_size.depth > 0 ? ImageType::e3D : ImageType::e2D;
+    num_samples = 1;
+    block = Extent3D{
+        .width = config.params.block_size.width,
+        .height = config.params.block_size.height,
+        .depth = config.params.block_size.depth,
+    };
+    size = Extent3D{
+        .width = config.params.width,
+        .height = config.params.height,
+        .depth = config.params.depth,
+    };
+    tile_width_spacing = 0;
+    resources.levels = 1;
+    resources.layers = 1;
+    layer_stride = CalculateLayerStride(*this);
+    maybe_unaligned_layer_stride = CalculateLayerSize(*this);
+    rescaleable = block.depth == 0 && size.height > 256;
+    downscaleable = size.height > 512;
+}
+
 } // namespace VideoCommon
diff --git a/src/video_core/texture_cache/image_info.h b/src/video_core/texture_cache/image_info.h
index 93755e15e..a12f5b44f 100644
--- a/src/video_core/texture_cache/image_info.h
+++ b/src/video_core/texture_cache/image_info.h
@@ -5,6 +5,7 @@
 
 #include "video_core/engines/fermi_2d.h"
 #include "video_core/engines/maxwell_3d.h"
+#include "video_core/engines/maxwell_dma.h"
 #include "video_core/surface.h"
 #include "video_core/texture_cache/types.h"
 
@@ -19,6 +20,7 @@ struct ImageInfo {
     explicit ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs, size_t index) noexcept;
     explicit ImageInfo(const Tegra::Engines::Maxwell3D::Regs& regs) noexcept;
     explicit ImageInfo(const Tegra::Engines::Fermi2D::Surface& config) noexcept;
+    explicit ImageInfo(const Tegra::DMA::ImageOperand& config) noexcept;
 
     PixelFormat format = PixelFormat::Invalid;
     ImageType type = ImageType::e1D;
diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h
index 9dd152fbe..335338434 100644
--- a/src/video_core/texture_cache/texture_cache.h
+++ b/src/video_core/texture_cache/texture_cache.h
@@ -1359,6 +1359,75 @@ std::optional<typename TextureCache<P>::BlitImages> TextureCache<P>::GetBlitImag
 }
 
 template <class P>
+ImageId TextureCache<P>::FindDMAImage(const ImageInfo& info, GPUVAddr gpu_addr) {
+    std::optional<VAddr> cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr);
+    if (!cpu_addr) {
+        cpu_addr = gpu_memory->GpuToCpuAddress(gpu_addr, CalculateGuestSizeInBytes(info));
+        if (!cpu_addr) {
+            return ImageId{};
+        }
+    }
+    ImageId image_id{};
+    boost::container::small_vector<ImageId, 1> image_ids;
+    const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) {
+        if (True(existing_image.flags & ImageFlagBits::Remapped)) {
+            return false;
+        }
+        if (info.type == ImageType::Linear || existing_image.info.type == ImageType::Linear)
+            [[unlikely]] {
+            const bool strict_size = True(existing_image.flags & ImageFlagBits::Strong);
+            const ImageInfo& existing = existing_image.info;
+            if (existing_image.gpu_addr == gpu_addr && existing.type == info.type &&
+                existing.pitch == info.pitch &&
+                IsPitchLinearSameSize(existing, info, strict_size) &&
+                IsViewCompatible(existing.format, info.format, false, true)) {
+                image_id = existing_image_id;
+                image_ids.push_back(existing_image_id);
+                return true;
+            }
+        } else if (IsSubCopy(info, existing_image, gpu_addr)) {
+            image_id = existing_image_id;
+            image_ids.push_back(existing_image_id);
+            return true;
+        }
+        return false;
+    };
+    ForEachImageInRegion(*cpu_addr, CalculateGuestSizeInBytes(info), lambda);
+    if (image_ids.size() <= 1) [[likely]] {
+        return image_id;
+    }
+    auto image_ids_compare = [this](ImageId a, ImageId b) {
+        auto& image_a = slot_images[a];
+        auto& image_b = slot_images[b];
+        return image_a.modification_tick < image_b.modification_tick;
+    };
+    return *std::ranges::max_element(image_ids, image_ids_compare);
+}
+
+template <class P>
+std::optional<std::pair<typename TextureCache<P>::Image*, std::pair<u32, u32>>>
+TextureCache<P>::ObtainImage(const Tegra::DMA::ImageOperand& operand, bool mark_as_modified) {
+    ImageInfo dst_info(operand);
+    ImageId dst_id = FindDMAImage(dst_info, operand.address);
+    if (!dst_id) {
+        return std::nullopt;
+    }
+    auto& image = slot_images[dst_id];
+    auto base = image.TryFindBase(operand.address);
+    if (!base) {
+        return std::nullopt;
+    }
+    if (False(image.flags & ImageFlagBits::GpuModified)) {
+        // No need to waste time on an image that's synced with guest
+        return std::nullopt;
+    }
+    PrepareImage(dst_id, mark_as_modified, false);
+    auto& new_image = slot_images[dst_id];
+    lru_cache.Touch(new_image.lru_index, frame_tick);
+    return std::make_pair(&new_image, std::make_pair(base->level, base->layer));
+}
+
+template <class P>
 SamplerId TextureCache<P>::FindSampler(const TSCEntry& config) {
     if (std::ranges::all_of(config.raw, [](u64 value) { return value == 0; })) {
         return NULL_SAMPLER_ID;
diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h
index 013836933..848a5d9ea 100644
--- a/src/video_core/texture_cache/texture_cache_base.h
+++ b/src/video_core/texture_cache/texture_cache_base.h
@@ -209,6 +209,9 @@ public:
     /// Pop asynchronous downloads
     void PopAsyncFlushes();
 
+    [[nodiscard]] std::optional<std::pair<Image*, std::pair<u32, u32>>> ObtainImage(
+        const Tegra::DMA::ImageOperand& operand, bool mark_as_modified);
+
     /// Return true when a CPU region is modified from the GPU
     [[nodiscard]] bool IsRegionGpuModified(VAddr addr, size_t size);
 
@@ -300,6 +303,8 @@ private:
     /// Remove joined images from the cache
     [[nodiscard]] ImageId JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VAddr cpu_addr);
 
+    [[nodiscard]] ImageId FindDMAImage(const ImageInfo& info, GPUVAddr gpu_addr);
+
     /// Return a blit image pair from the given guest blit parameters
     [[nodiscard]] std::optional<BlitImages> GetBlitImages(
         const Tegra::Engines::Fermi2D::Surface& dst, const Tegra::Engines::Fermi2D::Surface& src,
diff --git a/src/video_core/texture_cache/types.h b/src/video_core/texture_cache/types.h
index 0453456b4..a0e10643f 100644
--- a/src/video_core/texture_cache/types.h
+++ b/src/video_core/texture_cache/types.h
@@ -54,6 +54,7 @@ enum class RelaxedOptions : u32 {
     Format = 1 << 1,
     Samples = 1 << 2,
     ForceBrokenViews = 1 << 3,
+    FormatBpp = 1 << 4,
 };
 DECLARE_ENUM_FLAG_OPERATORS(RelaxedOptions)
 
diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp
index 697f86641..de37db684 100644
--- a/src/video_core/texture_cache/util.cpp
+++ b/src/video_core/texture_cache/util.cpp
@@ -743,6 +743,44 @@ std::vector<ImageCopy> MakeShrinkImageCopies(const ImageInfo& dst, const ImageIn
     return copies;
 }
 
+std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src, u32 up_scale,
+                                                  u32 down_shift) {
+    std::vector<ImageCopy> copies;
+    copies.reserve(src.resources.levels);
+    const bool is_3d = src.type == ImageType::e3D;
+    for (s32 level = 0; level < src.resources.levels; ++level) {
+        ImageCopy& copy = copies.emplace_back();
+        copy.src_subresource = SubresourceLayers{
+            .base_level = level,
+            .base_layer = 0,
+            .num_layers = src.resources.layers,
+        };
+        copy.dst_subresource = SubresourceLayers{
+            .base_level = level,
+            .base_layer = 0,
+            .num_layers = src.resources.layers,
+        };
+        copy.src_offset = Offset3D{
+            .x = 0,
+            .y = 0,
+            .z = 0,
+        };
+        copy.dst_offset = Offset3D{
+            .x = 0,
+            .y = 0,
+            .z = 0,
+        };
+        const Extent3D mip_size = AdjustMipSize(src.size, level);
+        copy.extent = AdjustSamplesSize(mip_size, src.num_samples);
+        if (is_3d) {
+            copy.extent.depth = src.size.depth;
+        }
+        copy.extent.width = std::max<u32>((copy.extent.width * up_scale) >> down_shift, 1);
+        copy.extent.height = std::max<u32>((copy.extent.height * up_scale) >> down_shift, 1);
+    }
+    return copies;
+}
+
 bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config) {
     const GPUVAddr address = config.Address();
     if (address == 0) {
@@ -999,6 +1037,20 @@ bool IsBlockLinearSizeCompatible(const ImageInfo& lhs, const ImageInfo& rhs, u32
     }
 }
 
+bool IsBlockLinearSizeCompatibleBPPRelaxed(const ImageInfo& lhs, const ImageInfo& rhs,
+                                           u32 lhs_level, u32 rhs_level) noexcept {
+    ASSERT(lhs.type != ImageType::Linear);
+    ASSERT(rhs.type != ImageType::Linear);
+    const auto lhs_bpp = BytesPerBlock(lhs.format);
+    const auto rhs_bpp = BytesPerBlock(rhs.format);
+    const Extent3D lhs_size = AdjustMipSize(lhs.size, lhs_level);
+    const Extent3D rhs_size = AdjustMipSize(rhs.size, rhs_level);
+    return Common::AlignUpLog2(lhs_size.width * lhs_bpp, GOB_SIZE_X_SHIFT) ==
+               Common::AlignUpLog2(rhs_size.width * rhs_bpp, GOB_SIZE_X_SHIFT) &&
+           Common::AlignUpLog2(lhs_size.height, GOB_SIZE_Y_SHIFT) ==
+               Common::AlignUpLog2(rhs_size.height, GOB_SIZE_Y_SHIFT);
+}
+
 bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs, bool strict_size) noexcept {
     ASSERT(lhs.type == ImageType::Linear);
     ASSERT(rhs.type == ImageType::Linear);
@@ -1073,7 +1125,8 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
         // Format checking is relaxed, but we still have to check for matching bytes per block.
         // This avoids creating a view for blits on UE4 titles where formats with different bytes
         // per block are aliased.
-        if (BytesPerBlock(existing.format) != BytesPerBlock(candidate.format)) {
+        if (BytesPerBlock(existing.format) != BytesPerBlock(candidate.format) &&
+            False(options & RelaxedOptions::FormatBpp)) {
             return std::nullopt;
         }
     } else {
@@ -1088,10 +1141,8 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
     if (existing.type != candidate.type) {
         return std::nullopt;
     }
-    if (False(options & RelaxedOptions::Samples)) {
-        if (existing.num_samples != candidate.num_samples) {
-            return std::nullopt;
-        }
+    if (False(options & RelaxedOptions::Samples) && existing.num_samples != candidate.num_samples) {
+        return std::nullopt;
     }
     if (existing.resources.levels < candidate.resources.levels + base->level) {
         return std::nullopt;
@@ -1101,14 +1152,16 @@ std::optional<SubresourceBase> FindSubresource(const ImageInfo& candidate, const
         if (mip_depth < candidate.size.depth + base->layer) {
             return std::nullopt;
         }
-    } else {
-        if (existing.resources.layers < candidate.resources.layers + base->layer) {
-            return std::nullopt;
-        }
+    } else if (existing.resources.layers < candidate.resources.layers + base->layer) {
+        return std::nullopt;
     }
     const bool strict_size = False(options & RelaxedOptions::Size);
     if (!IsBlockLinearSizeCompatible(existing, candidate, base->level, 0, strict_size)) {
-        return std::nullopt;
+        if (False(options & RelaxedOptions::FormatBpp)) {
+            return std::nullopt;
+        } else if (!IsBlockLinearSizeCompatibleBPPRelaxed(existing, candidate, base->level, 0)) {
+            return std::nullopt;
+        }
     }
     // TODO: compare block sizes
     return base;
@@ -1120,6 +1173,31 @@ bool IsSubresource(const ImageInfo& candidate, const ImageBase& image, GPUVAddr
         .has_value();
 }
 
+bool IsSubCopy(const ImageInfo& candidate, const ImageBase& image, GPUVAddr candidate_addr) {
+    const std::optional<SubresourceBase> base = image.TryFindBase(candidate_addr);
+    if (!base) {
+        return false;
+    }
+    const ImageInfo& existing = image.info;
+    if (existing.resources.levels < candidate.resources.levels + base->level) {
+        return false;
+    }
+    if (existing.type == ImageType::e3D) {
+        const u32 mip_depth = std::max(1U, existing.size.depth << base->level);
+        if (mip_depth < candidate.size.depth + base->layer) {
+            return false;
+        }
+    } else {
+        if (existing.resources.layers < candidate.resources.layers + base->layer) {
+            return false;
+        }
+    }
+    if (!IsBlockLinearSizeCompatibleBPPRelaxed(existing, candidate, base->level, 0)) {
+        return false;
+    }
+    return true;
+}
+
 void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
                       const ImageBase* src) {
     const auto original_dst_format = dst_info.format;
diff --git a/src/video_core/texture_cache/util.h b/src/video_core/texture_cache/util.h
index d103db8ae..84aa6880d 100644
--- a/src/video_core/texture_cache/util.h
+++ b/src/video_core/texture_cache/util.h
@@ -56,6 +56,10 @@ struct OverlapResult {
                                                            SubresourceBase base, u32 up_scale = 1,
                                                            u32 down_shift = 0);
 
+[[nodiscard]] std::vector<ImageCopy> MakeReinterpretImageCopies(const ImageInfo& src,
+                                                                u32 up_scale = 1,
+                                                                u32 down_shift = 0);
+
 [[nodiscard]] bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config);
 
 [[nodiscard]] std::vector<BufferImageCopy> UnswizzleImage(Tegra::MemoryManager& gpu_memory,
@@ -88,6 +92,9 @@ void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const Ima
 [[nodiscard]] bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs,
                                          bool strict_size) noexcept;
 
+[[nodiscard]] bool IsBlockLinearSizeCompatibleBPPRelaxed(const ImageInfo& lhs, const ImageInfo& rhs,
+                                                         u32 lhs_level, u32 rhs_level) noexcept;
+
 [[nodiscard]] std::optional<OverlapResult> ResolveOverlap(const ImageInfo& new_info,
                                                           GPUVAddr gpu_addr, VAddr cpu_addr,
                                                           const ImageBase& overlap,
@@ -106,6 +113,9 @@ void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const Ima
                                  GPUVAddr candidate_addr, RelaxedOptions options, bool broken_views,
                                  bool native_bgr);
 
+[[nodiscard]] bool IsSubCopy(const ImageInfo& candidate, const ImageBase& image,
+                             GPUVAddr candidate_addr);
+
 void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst,
                       const ImageBase* src);