[vulkan] Re-wiring older ASTC decoding path + new helper function on wider ASTC textures

src/video_core/host_shaders/astc_decoder.comp

@@ -44,7 +44,7 @@ layout(binding = BINDING_INPUT_BUFFER, std430) readonly restrict buffer InputBuf
 };
 
 #ifdef VULKAN
-layout(binding = BINDING_OUTPUT_IMAGE, rgba32f) uniform writeonly restrict image2DArray dest_image;
+layout(binding = BINDING_OUTPUT_IMAGE) uniform writeonly restrict image2DArray dest_image;
 #else
 layout(binding = BINDING_OUTPUT_IMAGE, rgba8) uniform writeonly restrict image2DArray dest_image;
 #endif
@@ -137,17 +137,6 @@ uvec4 ReplicateByteTo16(uvec4 value) {
     return value * 0x101;
 }
 
-// sRGB EOTF (decode: encoded value -> linear). Only used on the Vulkan path, where the
-// destination image is a linear float format with no format-level sRGB tag of its own.
-// Vectorized (single vec3 overload, no per-component scalar ternary) to match the style
-// already used elsewhere in this file (see the HDR Mt piecewise selection) rather than the
-// float/vec3 overload pair this replaces.
-vec3 SRGBToLinear(vec3 c) {
-    const vec3 lo = c / 12.92;
-    const vec3 hi = pow(max((c + 0.055) / 1.055, vec3(0.0)), vec3(2.4));
-    return mix(hi, lo, lessThanEqual(c, vec3(0.04045)));
-}
-
 uint ReplicateBitTo7(uint value) {
     return value * 127;
 }
@@ -1405,9 +1394,6 @@ void DecompressBlock(ivec3 coord) {
                 const vec4 Cf =
                     vec4((C0 * (uvec4(64) - weight_vec) + C1 * weight_vec + uvec4(32)) / 64);
                 p = Cf / 65535.0f;
-#ifdef VULKAN
-                p.yzw = mix(p.yzw, SRGBToLinear(p.yzw), bvec3((block_height_mask & 0x80000000u) != 0u));
-#endif
             }
 
 #ifdef VULKAN
@@ -1436,7 +1422,7 @@ void main() {
     uint offset = 0;
     offset += pos.z * layer_stride;
     offset += (block_y >> block_height) * block_size;
-    offset += (block_y & (block_height_mask & 0x7FFFFFFFu)) << GOB_SIZE_SHIFT;
+    offset += (block_y & block_height_mask) << GOB_SIZE_SHIFT;
     offset += (pos.x >> GOB_SIZE_X_SHIFT) << x_shift;
     offset += swizzle;
 

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -564,7 +564,6 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         VideoCore::Surface::DefaultBlockWidth(image.info.format),
         VideoCore::Surface::DefaultBlockHeight(image.info.format),
     };
-    const bool is_srgb = VideoCore::Surface::IsPixelFormatSRGB(image.info.format);
     scheduler.RequestOutsideRenderPassOperationContext();
     const VkPipeline vk_pipeline = *pipeline;
     const VkImageAspectFlags aspect_mask = image.AspectMask();
@@ -614,15 +613,14 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         ASSERT(params.destination == (std::array<s32, 3>{0, 0, 0}));
         ASSERT(params.bytes_per_block_log2 == 4);
         scheduler.Record([this, num_dispatches_x, num_dispatches_y, num_dispatches_z, block_dims,
-                          params, descriptor_data, is_srgb](vk::CommandBuffer cmdbuf) {
+                          params, descriptor_data](vk::CommandBuffer cmdbuf) {
             const AstcPushConstants uniforms{
                 .blocks_dims = block_dims,
                 .layer_stride = params.layer_stride,
                 .block_size = params.block_size,
                 .x_shift = params.x_shift,
                 .block_height = params.block_height,
-                .block_height_mask =
+                .block_height_mask = params.block_height_mask,
-                    params.block_height_mask | (is_srgb ? 0x8000'0000u : 0u),
             };
             const VkDescriptorSet set = descriptor_allocator.Commit();
             device.GetLogical().UpdateDescriptorSet(set, *descriptor_template, descriptor_data);

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -129,11 +129,6 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
     return usage;
 }
 
-// Whether an ASTC texture will be decoded via the GPU compute pass (ASTCDecoderPass /
-// astc_decoder.comp) rather than the CPU ConvertImage path. Must stay the single source of
-// truth for this decision: both the destination image's real format (below) and the
-// AcceleratedUpload flag (Image::Image()) key off of it, and they must never disagree --
-// the compute shader's declared storage image format has to match what was actually allocated.
 [[nodiscard]] bool WillUseAcceleratedAstcDecode(const Device& device, const ImageInfo& info) {
     if (!IsPixelFormatASTC(info.format) || device.IsOptimalAstcSupported()) {
         return false;
@@ -146,13 +141,16 @@ constexpr VkBorderColor ConvertBorderColor(const std::array<float, 4>& color) {
           info.size.depth == 1;
 }
 
+[[nodiscard]] bool WillUseWidenedAstcFormat(const Device& device, const ImageInfo& info) {
+    return WillUseAcceleratedAstcDecode(device, info) &&
+           !VideoCore::Surface::IsPixelFormatSRGB(info.format);
+}
+
 [[nodiscard]] VkImageCreateInfo MakeImageCreateInfo(const Device& device, const ImageInfo& info,
                                                     std::optional<VkFormat> format_override = {}) {
     auto format_info =
         MaxwellToVK::SurfaceFormat(device, FormatType::Optimal, false, info.format);
     if (format_override) {
-        // Accelerated ASTC HDR decode target: real (unclamped) values need a linear float
-        // format, not whatever the guest ASTC pixel format would normally map to.
         format_info.format = *format_override;
         format_info.attachable = false;
         format_info.storage = true;
@@ -1586,10 +1584,10 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
     : VideoCommon::ImageBase(info_, gpu_addr_, cpu_addr_), scheduler{&runtime_.scheduler},
       runtime{&runtime_},
       original_image(MakeImage(runtime_.device, runtime_.memory_allocator, info,
-                               WillUseAcceleratedAstcDecode(runtime_.device, info)
+                               WillUseWidenedAstcFormat(runtime_.device, info)
                                    ? std::span<const VkFormat>{}
                                    : runtime->ViewFormats(info.format),
-                               WillUseAcceleratedAstcDecode(runtime_.device, info)
+                               WillUseWidenedAstcFormat(runtime_.device, info)
                                    ? std::make_optional(VK_FORMAT_R32G32B32A32_SFLOAT)
                                    : std::nullopt)),
       aspect_mask(ImageAspectMask(info.format)) {
@@ -1622,7 +1620,7 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
         Settings::values.astc_recompression.GetValue() ==
             Settings::AstcRecompression::Uncompressed) {
         const auto& device = runtime->device.GetLogical();
-        const VkFormat storage_format = WillUseAcceleratedAstcDecode(runtime->device, info)
+        const VkFormat storage_format = WillUseWidenedAstcFormat(runtime->device, info)
                                             ? VK_FORMAT_R32G32B32A32_SFLOAT
                                             : VK_FORMAT_A8B8G8R8_UNORM_PACK32;
         for (s32 level = 0; level < info.resources.levels; ++level) {
@@ -1979,7 +1977,9 @@ VkImageView Image::StorageImageView(s32 level) noexcept {
         auto format_info =
             MaxwellToVK::SurfaceFormat(runtime->device, FormatType::Optimal, true, info.format);
         if (WillUseAcceleratedAstcDecode(runtime->device, info)) {
-            format_info.format = VK_FORMAT_R32G32B32A32_SFLOAT;
+            format_info.format = WillUseWidenedAstcFormat(runtime->device, info)
+                                     ? VK_FORMAT_R32G32B32A32_SFLOAT
+                                     : VK_FORMAT_A8B8G8R8_UNORM_PACK32;
         }
         view = MakeStorageView(runtime->device.GetLogical(), level, *(this->*current_image),
                                format_info.format);
@@ -2147,9 +2147,9 @@ ImageView::ImageView(TextureCacheRuntime& runtime, const VideoCommon::ImageViewI
             SanitizeDepthStencilSwizzle(swizzle, device->SupportsDepthStencilSwizzleOne());
         }
     }
-    uses_accelerated_astc_decode = WillUseAcceleratedAstcDecode(*device, image.info);
+    uses_widened_astc_format = WillUseWidenedAstcFormat(*device, image.info);
     auto format_info = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format);
-    if (uses_accelerated_astc_decode) {
+    if (uses_widened_astc_format) {
         format_info.format = VK_FORMAT_R32G32B32A32_SFLOAT;
     }
     const VkImageUsageFlags requested_view_usage = ImageUsageFlags(format_info, format);
@@ -2287,7 +2287,7 @@ VkImageView ImageView::StorageView(Shader::TextureType texture_type,
         if (image_format == Shader::ImageFormat::Typeless) {
             if (!typeless_storage_view) {
                 auto info = MaxwellToVK::SurfaceFormat(*device, FormatType::Optimal, true, format);
-                if (uses_accelerated_astc_decode) {
+                if (uses_widened_astc_format) {
                     info.format = VK_FORMAT_R32G32B32A32_SFLOAT;
                 }
                 typeless_storage_view = MakeView(info.format, VK_IMAGE_ASPECT_COLOR_BIT);

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -387,7 +387,7 @@ private:
     VkSampleCountFlagBits samples = VK_SAMPLE_COUNT_1_BIT;
     u32 buffer_size = 0;
 
-    bool uses_accelerated_astc_decode = false;
+    bool uses_widened_astc_format = false;
 };
 
 class ImageAlloc : public VideoCommon::ImageAllocBase {};