Keep working on ASTC

2026-10-06 14:09:58 +00:00 · 2026-09-16 16:19:52 -04:00
parent 66a25a3d79
commit 9bedcf6202
8 changed files with 221 additions and 57 deletions
@@ -57,6 +57,7 @@ using VideoCore::Surface::SurfaceType;
 namespace {
 constexpr bool ENABLE_MSAA_TILER_RESOLVE = true;
 constexpr bool ENABLE_MSAA_RESOLVE_CONSUME = true;
+constexpr bool ENABLE_ACCELERATED_UNSWIZZLE = false;
 constexpr bool ENABLE_MSAA_COLOR_DISCARD = true;
 constexpr bool ENABLE_MSAA_DEPTH_STENCIL_DISCARD = true;

@@ -193,6 +194,9 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
 }

 [[nodiscard]] bool IsUnswizzleAcceleratedFormat(const Device& device, PixelFormat format) {
+    if (!ENABLE_ACCELERATED_UNSWIZZLE) {
+        return false;
+    }
    if (IsPixelFormatASTC(format) || VideoCore::Surface::IsPixelFormatBCn(format)) {
        return false;
    }
@@ -1018,7 +1022,7 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, Scheduler& sched
        astc_decoder_pass.emplace(device, scheduler, descriptor_pool, staging_buffer_pool,
                                  compute_pass_descriptor_queue, memory_allocator);
    }
-    if (device.IsKhrImageFormatListSupported()) {
+    if (ENABLE_ACCELERATED_UNSWIZZLE && device.IsKhrImageFormatListSupported()) {
        bl_unswizzle_2d_pass.emplace(device, scheduler, descriptor_pool, staging_buffer_pool,
                                     compute_pass_descriptor_queue);
        bl_unswizzle_3d_pass.emplace(device, scheduler, descriptor_pool, staging_buffer_pool,
@@ -1117,6 +1117,12 @@ void TextureCache<P>::UploadImageContents(Image& image, StagingBuffer& staging)
    Tegra::Memory::GpuGuestMemory<u8, Tegra::Memory::GuestMemoryFlags::UnsafeRead> swizzle_data(
        *gpu_memory, gpu_addr, image.guest_size_bytes, &swizzle_data_buffer);
    if (True(image.flags & ImageFlagBits::Converted)) {
+        if (CanConvertFromGuest(image.info)) {
+            const auto copies =
+                FixSmallVectorADL(ConvertImageFromGuest(swizzle_data, image.info, mapped_span));
+            image.UploadMemory(staging, copies);
+            return;
+        }
        unswizzle_data_buffer.resize_destructive(image.unswizzled_size_bytes);
        auto copies = FixSmallVectorADL(UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data, unswizzle_data_buffer));
        ConvertImage(unswizzle_data_buffer, image.info, mapped_span, copies);
@@ -1302,10 +1308,27 @@ void TextureCache<P>::QueueAsyncDecode(Image& image, ImageId image_id) {
    decode->image_id = image_id;
    async_decodes.push_back(std::move(decode));

+    const size_t out_size = MapSizeBytes(image);
+    if (CanConvertFromGuest(image.info)) {
+        decode_ptr->input_data.resize_destructive(image.guest_size_bytes);
+        gpu_memory->ReadBlockUnsafe(image.gpu_addr, decode_ptr->input_data.data(),
+                                    image.guest_size_bytes);
+
+        texture_decode_worker.QueueWork([out_size, info = image.info, async_decode = decode_ptr] {
+            async_decode->decoded_data.resize_destructive(out_size);
+            auto copies =
+                ConvertImageFromGuest(async_decode->input_data, info, async_decode->decoded_data);
+
+            std::unique_lock lock{async_decode->mutex};
+            async_decode->copies = std::move(copies);
+            async_decode->complete = true;
+        });
+        return;
+    }
+
    std::vector<u8> local_unswizzle_data_buffer(image.unswizzled_size_bytes, 0);
    Tegra::Memory::GpuGuestMemory<u8, Tegra::Memory::GuestMemoryFlags::UnsafeRead> swizzle_data(*gpu_memory, image.gpu_addr, image.guest_size_bytes, &swizzle_data_buffer);
    auto copies = UnswizzleImage(*gpu_memory, image.gpu_addr, image.info, swizzle_data, local_unswizzle_data_buffer);
-    const size_t out_size = MapSizeBytes(image);

    auto func = [out_size, copies, info = image.info,
                 input = std::move(local_unswizzle_data_buffer),
@@ -63,6 +63,7 @@ struct ImageViewInOut {

 struct AsyncDecodeContext {
    ImageId image_id;
+    Common::ScratchBuffer<u8> input_data;
    Common::ScratchBuffer<u8> decoded_data;
    boost::container::small_vector<BufferImageCopy, 16> copies;
    std::mutex mutex;
@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later

 // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
@@ -994,6 +994,78 @@ void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8
    }
 }

+bool CanConvertFromGuest(const ImageInfo& info) {
+    if (!IsPixelFormatASTC(info.format) || info.type == ImageType::Linear) {
+        return false;
+    }
+    return Settings::values.astc_recompression.GetValue() ==
+           Settings::AstcRecompression::Uncompressed;
+}
+
+boost::container::small_vector<BufferImageCopy, 16> ConvertImageFromGuest(
+    std::span<const u8> input, const ImageInfo& info, std::span<u8> output) {
+    const u32 bpp_log2 = BytesPerBlockLog2(info.format);
+    const Extent2D tile_size = DefaultBlockSize(info.format);
+    const Extent3D size = info.size;
+    const LevelInfo level_info = MakeLevelInfo(info);
+    const s32 num_layers = info.resources.layers;
+    const s32 num_levels = info.resources.levels;
+    const std::array level_sizes = CalculateLevelSizes(level_info, num_levels);
+    const Extent2D gob = GobSize(bpp_log2, info.block.height, info.tile_width_spacing);
+    const u32 layer_size = CalculateLevelBytes(level_sizes, num_levels);
+    const u32 layer_stride = AlignLayerSize(layer_size, size, level_info.block, tile_size.height,
+                                            info.tile_width_spacing);
+    const u32 out_bytes_per_texel = BytesPerBlock(PixelFormat::A8B8G8R8_UNORM);
+    size_t guest_offset = 0;
+    u32 output_offset = 0;
+    boost::container::small_vector<BufferImageCopy, 16> copies(num_levels);
+
+    for (s32 level = 0; level < num_levels; ++level) {
+        const Extent3D level_size = AdjustMipSize(size, level);
+        const Extent3D num_tiles = AdjustTileSize(level_size, tile_size);
+        const Extent3D block =
+            AdjustMipBlockSize(num_tiles, level_info.block, level, level_info.num_levels);
+        const u32 stride_alignment = StrideAlignment(num_tiles, info.block, gob, bpp_log2);
+        const u32 stride = Common::AlignUpLog2(num_tiles.width, stride_alignment) << bpp_log2;
+        const u32 gobs_in_x = Common::DivCeilLog2(stride, GOB_SIZE_X_SHIFT);
+        const u32 gob_block_size = gobs_in_x << (GOB_SIZE_SHIFT + block.height + block.depth);
+        const u32 level_bytes = level_size.width * level_size.height * level_size.depth *
+                                num_layers * out_bytes_per_texel;
+        copies[level] = BufferImageCopy{
+            .buffer_offset = output_offset,
+            .buffer_size = level_bytes,
+            .buffer_row_length = level_size.width,
+            .buffer_image_height = level_size.height,
+            .image_subresource =
+                {
+                    .base_level = level,
+                    .base_layer = 0,
+                    .num_layers = num_layers,
+                },
+            .image_offset = {0, 0, 0},
+            .image_extent = level_size,
+        };
+        const Tegra::Texture::ASTC::BlockLinearLayout layout{
+            .layer_stride = layer_stride,
+            .slice_size =
+                Common::DivCeilLog2(num_tiles.height, block.height + GOB_SIZE_Y_SHIFT) *
+                gob_block_size,
+            .block_size = gob_block_size,
+            .x_shift = GOB_SIZE_SHIFT + block.height + block.depth,
+            .gob_height = block.height,
+            .gob_height_mask = (1U << block.height) - 1,
+            .gob_depth = block.depth,
+            .gob_depth_mask = (1U << block.depth) - 1,
+        };
+        Tegra::Texture::ASTC::DecompressBlockLinear(
+            input.subspan(guest_offset), level_size.width, level_size.height, level_size.depth,
+            num_layers, tile_size.width, tile_size.height, layout, output.subspan(output_offset));
+        output_offset += level_bytes;
+        guest_offset += level_sizes[level];
+    }
+    return copies;
+}
+
 boost::container::small_vector<BufferImageCopy, 16> FullDownloadCopies(const ImageInfo& info) {
    const Extent3D size = info.size;
    const u32 bytes_per_block = BytesPerBlock(info.format);
@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Eden Emulator Project
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later

 // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
@@ -72,6 +72,11 @@ struct OverlapResult {
 void ConvertImage(std::span<const u8> input, const ImageInfo& info, std::span<u8> output,
                  std::span<BufferImageCopy> copies);

+[[nodiscard]] bool CanConvertFromGuest(const ImageInfo& info);
+
+[[nodiscard]] boost::container::small_vector<BufferImageCopy, 16> ConvertImageFromGuest(
+    std::span<const u8> input, const ImageInfo& info, std::span<u8> output);
+
 [[nodiscard]] boost::container::small_vector<BufferImageCopy, 16> FullDownloadCopies(
    const ImageInfo& info);

@@ -20,6 +20,7 @@
 #include "common/common_types.h"
 #include <ranges>
 #include "video_core/textures/astc.h"
+#include "video_core/textures/decoders.h"
 #include "video_core/textures/workers.h"

 class InputBitStream {
@@ -1671,7 +1672,7 @@ static void ComputeEndpoints(Pixel& ep1, Pixel& ep2, const u32*& colorValues,
 }

 static void FillVoidExtentLDR(InputBitStream& strm, std::span<u32> outBuf, u32 blockWidth,
-                              u32 blockHeight) {
+                              u32 blockHeight, u32 stride) {
    // Don't actually care about the void extent, just read the bits...
    for (s32 i = 0; i < 4; ++i) {
        strm.ReadBits<13>();
@@ -1688,7 +1689,7 @@ static void FillVoidExtentLDR(InputBitStream& strm, std::span<u32> outBuf, u32 b

    for (u32 j = 0; j < blockHeight; j++) {
        for (u32 i = 0; i < blockWidth; i++) {
-            outBuf[j * blockWidth + i] = rgba;
+            outBuf[j * stride + i] = rgba;
        }
    }
 }
@@ -1726,52 +1727,52 @@ static u16 HalfToClampedByte(u16 half_bits) {
    return static_cast<u16>(clamped * 255.0f + 0.5f);
 }

-static void FillError(std::span<u32> outBuf, u32 blockWidth, u32 blockHeight) {
+static void FillError(std::span<u32> outBuf, u32 blockWidth, u32 blockHeight, u32 stride) {
    for (u32 j = 0; j < blockHeight; j++) {
        for (u32 i = 0; i < blockWidth; i++) {
-            outBuf[j * blockWidth + i] = 0x00000000;
+            outBuf[j * stride + i] = 0x00000000;
        }
    }
 }

 static void DecompressBlock(std::span<const u8, 16> inBuf, const u32 blockWidth,
-                            const u32 blockHeight, std::span<u32, 12 * 12> outBuf) {
+                            const u32 blockHeight, std::span<u32> outBuf, u32 stride) {
    InputBitStream strm(inBuf);
    TexelWeightParams weightParams = DecodeBlockInfo(strm);

    // Was there an error?
    if (weightParams.m_bError) {
        assert(false && "Invalid block mode");
-        FillError(outBuf, blockWidth, blockHeight);
+        FillError(outBuf, blockWidth, blockHeight, stride);
        return;
    }

    if (weightParams.m_bVoidExtentLDR) {
-        FillVoidExtentLDR(strm, outBuf, blockWidth, blockHeight);
+        FillVoidExtentLDR(strm, outBuf, blockWidth, blockHeight, stride);
        return;
    }

    if (weightParams.m_bVoidExtentHDR) {
        assert(false && "HDR void extent blocks are unsupported!");
-        FillError(outBuf, blockWidth, blockHeight);
+        FillError(outBuf, blockWidth, blockHeight, stride);
        return;
    }

    if (weightParams.m_Width > blockWidth) {
        assert(false && "Texel weight grid width should be smaller than block width");
-        FillError(outBuf, blockWidth, blockHeight);
+        FillError(outBuf, blockWidth, blockHeight, stride);
        return;
    }

    if (weightParams.m_Height > blockHeight) {
        assert(false && "Texel weight grid height should be smaller than block height");
-        FillError(outBuf, blockWidth, blockHeight);
+        FillError(outBuf, blockWidth, blockHeight, stride);
        return;
    }

    if (weightParams.GetNumWeightValues() > 64) {
        assert(false && "Too many weights in the weight grid");
-        FillError(outBuf, blockWidth, blockHeight);
+        FillError(outBuf, blockWidth, blockHeight, stride);
        return;
    }

@@ -1781,7 +1782,7 @@ static void DecompressBlock(std::span<const u8, 16> inBuf, const u32 blockWidth,

    if (nPartitions == 4 && weightParams.m_bDualPlane) {
        assert(false && "Dual plane mode is incompatible with four partition blocks");
-        FillError(outBuf, blockWidth, blockHeight);
+        FillError(outBuf, blockWidth, blockHeight, stride);
        return;
    }

@@ -1813,7 +1814,7 @@ static void DecompressBlock(std::span<const u8, 16> inBuf, const u32 blockWidth,
    u32 nWeightBits = weightParams.GetPackedBitSize();
    if (nWeightBits < 24 || nWeightBits > 96) {
        assert(false && "Invalid weight bit count");
-        FillError(outBuf, blockWidth, blockHeight);
+        FillError(outBuf, blockWidth, blockHeight, stride);
        return;
    }
    s32 remainingBits = 128 - nWeightBits - static_cast<int>(strm.GetBitsRead());
@@ -1995,41 +1996,54 @@ static void DecompressBlock(std::span<const u8, 16> inBuf, const u32 blockWidth,
                }
            }

-            outBuf[j * blockWidth + i] = p.Pack();
+            outBuf[j * stride + i] = p.Pack();
        }
 }

+static void DecodeAndStoreBlock(std::span<const uint8_t> data, size_t src_offset, u32 x, u32 y,
+                                u32 width, u32 height, u32 block_width, u32 block_height,
+                                std::span<u32> out_slice) {
+    const u32 decomp_width = (std::min)(block_width, width - x);
+    const u32 decomp_height = (std::min)(block_height, height - y);
+    u32* const dst = out_slice.data() + size_t{y} * width + x;
+    if (src_offset + 16 > data.size()) {
+        for (u32 h = 0; h < decomp_height; ++h) {
+            std::memset(dst + size_t{h} * width, 0, decomp_width * 4);
+        }
+        return;
+    }
+    const std::span<const u8, 16> block_ptr{data.subspan(src_offset, 16)};
+    if (decomp_width == block_width && decomp_height == block_height) {
+        const size_t touched = size_t{block_height - 1} * width + block_width;
+        DecompressBlock(block_ptr, block_width, block_height, std::span<u32>{dst, touched}, width);
+        return;
+    }
+    std::array<u32, 12 * 12> staging;
+    DecompressBlock(block_ptr, block_width, block_height, staging, block_width);
+    for (u32 h = 0; h < decomp_height; ++h) {
+        std::memcpy(dst + size_t{h} * width, staging.data() + h * block_width, decomp_width * 4);
+    }
+}
+
 void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
                uint32_t block_width, uint32_t block_height, std::span<uint8_t> output) {
    const u32 rows = Common::DivideUp(height, block_height);
    const u32 cols = Common::DivideUp(width, block_width);

    Common::ThreadWorker& workers{GetThreadWorkers()};
+    const std::span<u32> out{reinterpret_cast<u32*>(output.data()), output.size() / 4};
+    const size_t slice_texels = size_t{height} * width;

    for (u32 z = 0; z < depth; ++z) {
-        const u32 depth_offset = z * height * width * 4;
+        const std::span<u32> out_slice = out.subspan(z * slice_texels);
        for (u32 y_index = 0; y_index < rows; ++y_index) {
-            auto decompress_stride = [data, width, height, block_width, block_height, output, rows,
-                                      cols, z, depth_offset, y_index] {
+            auto decompress_stride = [data, width, height, block_width, block_height, out_slice,
+                                      rows, cols, z, y_index] {
                const u32 y = y_index * block_height;
                for (u32 x_index = 0; x_index < cols; ++x_index) {
                    const u32 block_index = (z * rows * cols) + (y_index * cols) + x_index;
-                    const u32 x = x_index * block_width;
-
-                    const std::span<const u8, 16> blockPtr{data.subspan(block_index * 16, 16)};
-
-                    // Blocks can be at most 12x12
-                    std::array<u32, 12 * 12> uncompData;
-                    DecompressBlock(blockPtr, block_width, block_height, uncompData);
-
-                    u32 decompWidth = (std::min)(block_width, width - x);
-                    u32 decompHeight = (std::min)(block_height, height - y);
-
-                    const std::span<u8> outRow = output.subspan(depth_offset + (y * width + x) * 4);
-                    for (u32 h = 0; h < decompHeight; ++h) {
-                        std::memcpy(outRow.data() + h * width * 4,
-                                    uncompData.data() + h * block_width, decompWidth * 4);
-                    }
+                    DecodeAndStoreBlock(data, size_t{block_index} * 16, x_index * block_width, y,
+                                        width, height, block_width, block_height, out_slice);
                }
            };
            workers.QueueWork(std::move(decompress_stride));
@@ -2038,4 +2052,50 @@ void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height,
    }
 }

+void DecompressBlockLinear(std::span<const uint8_t> data, uint32_t width, uint32_t height,
+                           uint32_t depth, uint32_t layers, uint32_t block_width,
+                           uint32_t block_height, const BlockLinearLayout& layout,
+                           std::span<uint8_t> output) {
+    const u32 rows = Common::DivideUp(height, block_height);
+    const u32 cols = Common::DivideUp(width, block_width);
+    const size_t slice_texels = size_t{height} * width;
+
+    Common::ThreadWorker& workers{GetThreadWorkers()};
+    const std::span<u32> out{reinterpret_cast<u32*>(output.data()), output.size() / 4};
+
+    for (u32 layer = 0; layer < layers; ++layer) {
+        const size_t layer_offset = size_t{layer} * layout.layer_stride;
+        for (u32 z = 0; z < depth; ++z) {
+            const size_t slice_offset =
+                layer_offset + (z >> layout.gob_depth) * size_t{layout.slice_size} +
+                ((z & layout.gob_depth_mask) << (GOB_SIZE_SHIFT + layout.gob_height));
+            const std::span<u32> out_slice =
+                out.subspan((size_t{layer} * depth + z) * slice_texels);
+            for (u32 y_index = 0; y_index < rows; ++y_index) {
+                auto decompress_stride = [data, width, height, block_width, block_height, out_slice,
+                                          cols, layout, slice_offset, y_index] {
+                    const u32 y = y_index * block_height;
+                    const u32 gob_y = y_index >> GOB_SIZE_Y_SHIFT;
+                    const size_t offset_y =
+                        (gob_y >> layout.gob_height) * size_t{layout.block_size} +
+                        ((gob_y & layout.gob_height_mask) << GOB_SIZE_SHIFT);
+                    const u32 swizzled_y = ((y_index & 1) << 4) | ((y_index & 6) << 5);
+                    for (u32 x_index = 0; x_index < cols; ++x_index) {
+                        const u32 byte_x = x_index * 16;
+                        const size_t offset_x = size_t{byte_x >> GOB_SIZE_X_SHIFT}
+                                                << layout.x_shift;
+                        const u32 swizzled_x = ((x_index & 1) << 5) | ((x_index & 2) << 7);
+                        const size_t src_offset = slice_offset + offset_y + offset_x +
+                                                  (swizzled_x | swizzled_y);
+                        DecodeAndStoreBlock(data, src_offset, x_index * block_width, y, width,
+                                            height, block_width, block_height, out_slice);
+                    }
+                };
+                workers.QueueWork(std::move(decompress_stride));
+            }
+        }
+    }
+    workers.WaitForRequests();
+}
+
 } // namespace Tegra::Texture::ASTC
@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later

@@ -5,7 +8,23 @@

 namespace Tegra::Texture::ASTC {

+struct BlockLinearLayout {
+    uint32_t layer_stride;
+    uint32_t slice_size;
+    uint32_t block_size;
+    uint32_t x_shift;
+    uint32_t gob_height;
+    uint32_t gob_height_mask;
+    uint32_t gob_depth;
+    uint32_t gob_depth_mask;
+};
+
 void Decompress(std::span<const uint8_t> data, uint32_t width, uint32_t height, uint32_t depth,
                uint32_t block_width, uint32_t block_height, std::span<uint8_t> output);

+void DecompressBlockLinear(std::span<const uint8_t> data, uint32_t width, uint32_t height,
+                           uint32_t depth, uint32_t layers, uint32_t block_width,
+                           uint32_t block_height, const BlockLinearLayout& layout,
+                           std::span<uint8_t> output);
+
 } // namespace Tegra::Texture::ASTC
@@ -30,26 +30,6 @@ namespace {

 // Helpers translating MemoryUsage to flags/usage

-    [[maybe_unused]] VkMemoryPropertyFlags MemoryUsagePropertyFlags(MemoryUsage usage) {
-        switch (usage) {
-            case MemoryUsage::DeviceLocal:
-                return VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
-            case MemoryUsage::Upload:
-                return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                        VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-            case MemoryUsage::Download:
-                return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                        VK_MEMORY_PROPERTY_HOST_COHERENT_BIT |
-                        VK_MEMORY_PROPERTY_HOST_CACHED_BIT;
-            case MemoryUsage::Stream:
-                return VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT |
-                        VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT |
-                        VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-        }
-        ASSERT_MSG(false, "Invalid memory usage={}", usage);
-        return VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;
-    }
-
    [[nodiscard]] VkMemoryPropertyFlags MemoryUsagePreferredVmaFlags(MemoryUsage usage) {
        if (usage == MemoryUsage::Download) {
            return VK_MEMORY_PROPERTY_HOST_CACHED_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT;