astc_decoder: Reimplement Layers

Reimplements the approach to decoding layers in the compute shader. Fixes multilayer astc decoding when using Vulkan.

src/video_core/host_shaders/astc_decoder.comp

@@ -39,17 +39,15 @@ layout(local_size_x = 32, local_size_y = 32, local_size_z = 1) in;
 BEGIN_PUSH_CONSTANTS
 UNIFORM(0) uvec2 num_image_blocks;
 UNIFORM(1) uvec2 block_dims;
-UNIFORM(2) uint layer;
-
-UNIFORM(3) uvec3 origin;
-UNIFORM(4) ivec3 destination;
-UNIFORM(5) uint bytes_per_block_log2;
-UNIFORM(6) uint layer_stride;
-UNIFORM(7) uint block_size;
-UNIFORM(8) uint x_shift;
-UNIFORM(9) uint block_height;
-UNIFORM(10) uint block_height_mask;
 
+UNIFORM(2) uvec3 origin;
+UNIFORM(3) ivec3 destination;
+UNIFORM(4) uint bytes_per_block_log2;
+UNIFORM(5) uint layer_stride;
+UNIFORM(6) uint block_size;
+UNIFORM(7) uint x_shift;
+UNIFORM(8) uint block_height;
+UNIFORM(9) uint block_height_mask;
 END_PUSH_CONSTANTS
 
 uint current_index = 0;
@@ -82,7 +80,7 @@ layout(binding = BINDING_SWIZZLE_BUFFER, std430) readonly buffer SwizzleTable {
     uint swizzle_table[];
 };
 
-layout(binding = BINDING_INPUT_BUFFER, std430) buffer InputBufferU32 {
+layout(binding = BINDING_INPUT_BUFFER, std430) readonly buffer InputBufferU32 {
     uint astc_data[];
 };
 
@@ -104,7 +102,7 @@ layout(binding = BINDING_BYTE_TO_16_BUFFER, std430) readonly buffer REPLICATE_BY
     uint REPLICATE_BYTE_TO_16_TABLE[];
 };
 
-layout(binding = BINDING_OUTPUT_IMAGE, rgba8) uniform writeonly image2D dest_image;
+layout(binding = BINDING_OUTPUT_IMAGE, rgba8) uniform writeonly image2DArray dest_image;
 
 const uint GOB_SIZE_X = 64;
 const uint GOB_SIZE_Y = 8;
@@ -1086,10 +1084,9 @@ TexelWeightParams DecodeBlockInfo(uint block_index) {
 void FillError(ivec3 coord) {
     for (uint j = 0; j < block_dims.y; j++) {
         for (uint i = 0; i < block_dims.x; i++) {
-            imageStore(dest_image, coord.xy + ivec2(i, j), vec4(1.0, 1.0, 0.0, 1.0));
+            imageStore(dest_image, coord + ivec3(i, j, 0), vec4(1.0, 1.0, 0.0, 1.0));
         }
     }
-    return;
 }
 
 void FillVoidExtentLDR(ivec3 coord, uint block_index) {
@@ -1107,7 +1104,7 @@ void FillVoidExtentLDR(ivec3 coord, uint block_index) {
     float b = float(b_u) / 65535.0f;
     for (uint j = 0; j < block_dims.y; j++) {
         for (uint i = 0; i < block_dims.x; i++) {
-            imageStore(dest_image, coord.xy + ivec2(i, j), vec4(r, g, b, a));
+            imageStore(dest_image, coord + ivec3(i, j, 0), vec4(r, g, b, a));
         }
     }
 }
@@ -1264,7 +1261,7 @@ void DecompressBlock(ivec3 coord, uint block_index) {
             }
             vec4 Cf = vec4((C0 * (uvec4(64) - weight_vec) + C1 * weight_vec + uvec4(32)) >> 6);
             p = (Cf / 65535.0);
-            imageStore(dest_image, coord.xy + ivec2(i, j), p.gbar);
+            imageStore(dest_image, coord + ivec3(i, j, 0), p.gbar);
         }
     }
 }
@@ -1279,7 +1276,7 @@ void main() {
     const uint block_y = pos.y >> GOB_SIZE_Y_SHIFT;
 
     uint offset = 0;
-    offset += layer * layer_stride;
+    offset += pos.z * layer_stride;
     offset += (block_y >> block_height) * block_size;
     offset += (block_y & block_height_mask) << GOB_SIZE_SHIFT;
     offset += (pos.x >> GOB_SIZE_X_SHIFT) << x_shift;
@@ -1287,7 +1284,7 @@ void main() {
 
     const ivec3 coord = ivec3(gl_GlobalInvocationID * uvec3(block_dims, 1.0));
     uint block_index =
-        layer * num_image_blocks.x * num_image_blocks.y + pos.y * num_image_blocks.x + pos.x;
+        pos.z * num_image_blocks.x * num_image_blocks.y + pos.y * num_image_blocks.x + pos.x;
     current_index = 0;
     bitsread = 0;
     for (int i = 0; i < 16; i++) {

src/video_core/renderer_opengl/util_shaders.cpp

@@ -110,7 +110,6 @@ void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
     static constexpr GLuint BINDING_OUTPUT_IMAGE = 0;
     static constexpr GLuint LOC_NUM_IMAGE_BLOCKS = 0;
     static constexpr GLuint LOC_BLOCK_DIMS = 1;
-    static constexpr GLuint LOC_LAYER = 2;
 
     const Extent3D tile_size = {
         VideoCore::Surface::DefaultBlockWidth(image.info.format),
@@ -130,35 +129,31 @@ void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
 
     glFlushMappedNamedBufferRange(map.buffer, map.offset, image.guest_size_bytes);
     glUniform2ui(LOC_BLOCK_DIMS, tile_size.width, tile_size.height);
-
-    for (u32 layer = 0; layer < image.info.resources.layers; layer++) {
     for (const SwizzleParameters& swizzle : swizzles) {
-            glBindImageTexture(BINDING_OUTPUT_IMAGE, image.StorageHandle(), swizzle.level, GL_FALSE,
+        glBindImageTexture(BINDING_OUTPUT_IMAGE, image.StorageHandle(), swizzle.level, GL_TRUE, 0,
-                               layer, GL_WRITE_ONLY, GL_RGBA8);
+                           GL_WRITE_ONLY, GL_RGBA8);
         const size_t input_offset = swizzle.buffer_offset + map.offset;
         const auto num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
         const auto num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
 
         glUniform2ui(LOC_NUM_IMAGE_BLOCKS, swizzle.num_tiles.width, swizzle.num_tiles.height);
-            glUniform1ui(LOC_LAYER, layer);
 
         // To unswizzle the ASTC data
         const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
-            glUniform3uiv(3, 1, params.origin.data());
+        glUniform3uiv(2, 1, params.origin.data());
-            glUniform3iv(4, 1, params.destination.data());
+        glUniform3iv(3, 1, params.destination.data());
-            glUniform1ui(5, params.bytes_per_block_log2);
+        glUniform1ui(4, params.bytes_per_block_log2);
-            glUniform1ui(6, params.layer_stride);
+        glUniform1ui(5, params.layer_stride);
-            glUniform1ui(7, params.block_size);
+        glUniform1ui(6, params.block_size);
-            glUniform1ui(8, params.x_shift);
+        glUniform1ui(7, params.x_shift);
-            glUniform1ui(9, params.block_height);
+        glUniform1ui(8, params.block_height);
-            glUniform1ui(10, params.block_height_mask);
+        glUniform1ui(9, params.block_height_mask);
 
         // ASTC texture data
-            glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.buffer,
+        glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.buffer, input_offset,
-                              input_offset, image.guest_size_bytes - swizzle.buffer_offset);
+                          image.guest_size_bytes - swizzle.buffer_offset);
 
-            glDispatchCompute(num_dispatches_x, num_dispatches_y, 1);
+        glDispatchCompute(num_dispatches_x, num_dispatches_y, image.info.resources.layers);
-        }
     }
     program_manager.RestoreGuestCompute();
 }

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -31,6 +31,7 @@ namespace Vulkan {
 
 using Tegra::Texture::SWIZZLE_TABLE;
 using Tegra::Texture::ASTC::EncodingsValues;
+using namespace Tegra::Texture::ASTC;
 
 namespace {
 
@@ -214,7 +215,6 @@ std::array<VkDescriptorUpdateTemplateEntryKHR, 8> BuildASTCPassDescriptorUpdateT
 struct AstcPushConstants {
     std::array<u32, 2> num_image_blocks;
     std::array<u32, 2> blocks_dims;
-    u32 layer;
     VideoCommon::Accelerated::BlockLinearSwizzle2DParams params;
 };
 
@@ -226,6 +226,7 @@ struct AstcBufferData {
     decltype(REPLICATE_8_BIT_TO_8_TABLE) replicate_8_to_8 = REPLICATE_8_BIT_TO_8_TABLE;
     decltype(REPLICATE_BYTE_TO_16_TABLE) replicate_byte_to_16 = REPLICATE_BYTE_TO_16_TABLE;
 } constexpr ASTC_BUFFER_DATA;
+
 } // Anonymous namespace
 
 VKComputePass::VKComputePass(const Device& device, VKDescriptorPool& descriptor_pool,
@@ -403,7 +404,6 @@ std::pair<VkBuffer, VkDeviceSize> QuadIndexedPass::Assemble(
     return {staging.buffer, staging.offset};
 }
 
-using namespace Tegra::Texture::ASTC;
 ASTCDecoderPass::ASTCDecoderPass(const Device& device_, VKScheduler& scheduler_,
                                  VKDescriptorPool& descriptor_pool_,
                                  StagingBufferPool& staging_buffer_pool_,
@@ -464,63 +464,20 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
     if (!data_buffer) {
         MakeDataBuffer();
     }
-    const std::array<u32, 2> block_dims{tile_size.width, tile_size.height};
+    const VkImageAspectFlags aspect_mask = image.AspectMask();
-    for (s32 layer = 0; layer < image.info.resources.layers; layer++) {
+    const VkImage vk_image = image.Handle();
-        for (const VideoCommon::SwizzleParameters& swizzle : swizzles) {
+    const bool is_initialized = image.ExchangeInitialization();
-            const size_t input_offset = swizzle.buffer_offset + map.offset;
+    scheduler.Record([vk_image, aspect_mask, is_initialized](vk::CommandBuffer cmdbuf) {
-            const auto num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
-            const auto num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
-            const std::array num_image_blocks{swizzle.num_tiles.width, swizzle.num_tiles.height};
-            const u32 layer_image_size =
-                image.guest_size_bytes - static_cast<u32>(swizzle.buffer_offset);
-
-            update_descriptor_queue.Acquire();
-            update_descriptor_queue.AddBuffer(*data_buffer,
-                                              offsetof(AstcBufferData, swizzle_table_buffer),
-                                              sizeof(AstcBufferData::swizzle_table_buffer));
-            update_descriptor_queue.AddBuffer(map.buffer, input_offset, image.guest_size_bytes);
-            update_descriptor_queue.AddBuffer(*data_buffer,
-                                              offsetof(AstcBufferData, encoding_values),
-                                              sizeof(AstcBufferData::encoding_values));
-            update_descriptor_queue.AddBuffer(*data_buffer,
-                                              offsetof(AstcBufferData, replicate_6_to_8),
-                                              sizeof(AstcBufferData::replicate_6_to_8));
-            update_descriptor_queue.AddBuffer(*data_buffer,
-                                              offsetof(AstcBufferData, replicate_7_to_8),
-                                              sizeof(AstcBufferData::replicate_7_to_8));
-            update_descriptor_queue.AddBuffer(*data_buffer,
-                                              offsetof(AstcBufferData, replicate_8_to_8),
-                                              sizeof(AstcBufferData::replicate_8_to_8));
-            update_descriptor_queue.AddBuffer(*data_buffer,
-                                              offsetof(AstcBufferData, replicate_byte_to_16),
-                                              sizeof(AstcBufferData::replicate_byte_to_16));
-            update_descriptor_queue.AddImage(image.StorageImageView());
-
-            const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue);
-            // To unswizzle the ASTC data
-            const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
-            scheduler.Record([layout = *layout, pipeline = *pipeline, buffer = map.buffer,
-                              num_dispatches_x, num_dispatches_y, layer_image_size,
-                              num_image_blocks, block_dims, layer, params, set,
-                              image = image.Handle(), input_offset,
-                              aspect_mask = image.AspectMask()](vk::CommandBuffer cmdbuf) {
-                const AstcPushConstants uniforms{num_image_blocks, block_dims, layer, params};
-
-                cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, pipeline);
-                cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, layout, 0, set, {});
-                cmdbuf.PushConstants(layout, VK_SHADER_STAGE_COMPUTE_BIT, uniforms);
-                cmdbuf.Dispatch(num_dispatches_x, num_dispatches_y, 1);
-
         const VkImageMemoryBarrier image_barrier{
             .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
             .pNext = nullptr,
             .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
-                    .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
-                    .oldLayout = VK_IMAGE_LAYOUT_UNDEFINED,
+            .oldLayout = is_initialized ? VK_IMAGE_LAYOUT_GENERAL : VK_IMAGE_LAYOUT_UNDEFINED,
             .newLayout = VK_IMAGE_LAYOUT_GENERAL,
             .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
             .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
-                    .image = image,
+            .image = vk_image,
             .subresourceRange{
                 .aspectMask = aspect_mask,
                 .baseMipLevel = 0,
@@ -529,11 +486,72 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
                 .layerCount = VK_REMAINING_ARRAY_LAYERS,
             },
         };
-                cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
+        cmdbuf.PipelineBarrier(0, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, image_barrier);
-                                       VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT, 0, image_barrier);
+    });
+    const std::array<u32, 2> block_dims{tile_size.width, tile_size.height};
+    for (const VideoCommon::SwizzleParameters& swizzle : swizzles) {
+        const size_t input_offset = swizzle.buffer_offset + map.offset;
+        const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U);
+        const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U);
+        const u32 num_dispatches_z = image.info.resources.layers;
+        const std::array num_image_blocks{swizzle.num_tiles.width, swizzle.num_tiles.height};
+        const u32 layer_image_size =
+            image.guest_size_bytes - static_cast<u32>(swizzle.buffer_offset);
+
+        update_descriptor_queue.Acquire();
+        update_descriptor_queue.AddBuffer(*data_buffer,
+                                          offsetof(AstcBufferData, swizzle_table_buffer),
+                                          sizeof(AstcBufferData::swizzle_table_buffer));
+        update_descriptor_queue.AddBuffer(map.buffer, input_offset, layer_image_size);
+        update_descriptor_queue.AddBuffer(*data_buffer, offsetof(AstcBufferData, encoding_values),
+                                          sizeof(AstcBufferData::encoding_values));
+        update_descriptor_queue.AddBuffer(*data_buffer, offsetof(AstcBufferData, replicate_6_to_8),
+                                          sizeof(AstcBufferData::replicate_6_to_8));
+        update_descriptor_queue.AddBuffer(*data_buffer, offsetof(AstcBufferData, replicate_7_to_8),
+                                          sizeof(AstcBufferData::replicate_7_to_8));
+        update_descriptor_queue.AddBuffer(*data_buffer, offsetof(AstcBufferData, replicate_8_to_8),
+                                          sizeof(AstcBufferData::replicate_8_to_8));
+        update_descriptor_queue.AddBuffer(*data_buffer,
+                                          offsetof(AstcBufferData, replicate_byte_to_16),
+                                          sizeof(AstcBufferData::replicate_byte_to_16));
+        update_descriptor_queue.AddImage(image.StorageImageView(swizzle.level));
+
+        const VkDescriptorSet set = CommitDescriptorSet(update_descriptor_queue);
+        const VkPipelineLayout vk_layout = *layout;
+        const VkPipeline vk_pipeline = *pipeline;
+        // To unswizzle the ASTC data
+        const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
+        scheduler.Record([vk_layout, vk_pipeline, buffer = map.buffer, num_dispatches_x,
+                          num_dispatches_y, num_dispatches_z, num_image_blocks, block_dims, params,
+                          set, input_offset](vk::CommandBuffer cmdbuf) {
+            const AstcPushConstants uniforms{num_image_blocks, block_dims, params};
+            cmdbuf.BindPipeline(VK_PIPELINE_BIND_POINT_COMPUTE, vk_pipeline);
+            cmdbuf.BindDescriptorSets(VK_PIPELINE_BIND_POINT_COMPUTE, vk_layout, 0, set, {});
+            cmdbuf.PushConstants(vk_layout, VK_SHADER_STAGE_COMPUTE_BIT, uniforms);
+            cmdbuf.Dispatch(num_dispatches_x, num_dispatches_y, num_dispatches_z);
         });
     }
-    }
+    scheduler.Record([vk_image, aspect_mask](vk::CommandBuffer cmdbuf) {
+        const VkImageMemoryBarrier image_barrier{
+            .sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER,
+            .pNext = nullptr,
+            .srcAccessMask = VK_ACCESS_SHADER_WRITE_BIT,
+            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT | VK_ACCESS_SHADER_WRITE_BIT,
+            .oldLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .newLayout = VK_IMAGE_LAYOUT_GENERAL,
+            .srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
+            .image = vk_image,
+            .subresourceRange{
+                .aspectMask = aspect_mask,
+                .baseMipLevel = 0,
+                .levelCount = VK_REMAINING_MIP_LEVELS,
+                .baseArrayLayer = 0,
+                .layerCount = VK_REMAINING_ARRAY_LAYERS,
+            },
+        };
+        cmdbuf.PipelineBarrier(0, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT, 0, image_barrier);
+    });
 }
 
 } // namespace Vulkan

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -823,12 +823,15 @@ Image::Image(TextureCacheRuntime& runtime, const ImageInfo& info_, GPUVAddr gpu_
         .usage = VK_IMAGE_USAGE_STORAGE_BIT,
     };
     if (IsPixelFormatASTC(info.format) && !runtime.device.IsOptimalAstcSupported()) {
-        storage_image_view = runtime.device.GetLogical().CreateImageView(VkImageViewCreateInfo{
+        const auto& device = runtime.device.GetLogical();
+        storage_image_views.reserve(info.resources.levels);
+        for (s32 level = 0; level < info.resources.levels; ++level) {
+            storage_image_views.push_back(device.CreateImageView(VkImageViewCreateInfo{
                 .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
                 .pNext = &storage_image_view_usage_create_info,
                 .flags = 0,
                 .image = *image,
-            .viewType = VK_IMAGE_VIEW_TYPE_2D,
+                .viewType = VK_IMAGE_VIEW_TYPE_2D_ARRAY,
                 .format = VK_FORMAT_A8B8G8R8_UNORM_PACK32,
                 .components{
                     .r = VK_COMPONENT_SWIZZLE_IDENTITY,
@@ -838,12 +841,13 @@ Image::Image(TextureCacheRuntime& runtime, const ImageInfo& info_, GPUVAddr gpu_
                 },
                 .subresourceRange{
                     .aspectMask = aspect_mask,
-                .baseMipLevel = 0,
+                    .baseMipLevel = static_cast<u32>(level),
-                .levelCount = VK_REMAINING_MIP_LEVELS,
+                    .levelCount = 1,
                     .baseArrayLayer = 0,
                     .layerCount = VK_REMAINING_ARRAY_LAYERS,
                 },
-        });
+            }));
+        }
     }
 }
 

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -121,12 +121,17 @@ public:
         return *buffer;
     }
 
-    [[nodiscard]] VkImageCreateFlags AspectMask() const noexcept {
+    [[nodiscard]] VkImageAspectFlags AspectMask() const noexcept {
         return aspect_mask;
     }
 
-    [[nodiscard]] VkImageView StorageImageView() const noexcept {
+    [[nodiscard]] VkImageView StorageImageView(s32 level) const noexcept {
-        return *storage_image_view;
+        return *storage_image_views[level];
+    }
+
+    /// Returns true when the image is already initialized and mark it as initialized
+    [[nodiscard]] bool ExchangeInitialization() noexcept {
+        return std::exchange(initialized, true);
     }
 
 private:
@@ -135,7 +140,7 @@ private:
     vk::Buffer buffer;
     MemoryCommit commit;
     vk::ImageView image_view;
-    vk::ImageView storage_image_view;
+    std::vector<vk::ImageView> storage_image_views;
     VkImageAspectFlags aspect_mask = 0;
     bool initialized = false;
 };