astc_decoder: Compute offset swizzles in-shader

Alleviates the dependency on the swizzle table and a uniform which is constant for all ASTC texture sizes.

src/video_core/host_shaders/astc_decoder.comp

@@ -10,8 +10,7 @@
 #define END_PUSH_CONSTANTS };
 #define UNIFORM(n)
 #define BINDING_INPUT_BUFFER 0
-#define BINDING_SWIZZLE_BUFFER 1
-#define BINDING_OUTPUT_IMAGE 2
+#define BINDING_OUTPUT_IMAGE 1
 
 #else // ^^^ Vulkan ^^^ // vvv OpenGL vvv
 
@@ -19,7 +18,6 @@
 #define END_PUSH_CONSTANTS
 #define UNIFORM(n) layout(location = n) uniform
 #define BINDING_INPUT_BUFFER 0
-#define BINDING_SWIZZLE_BUFFER 1
 #define BINDING_OUTPUT_IMAGE 0
 
 #endif
@@ -28,13 +26,11 @@ layout(local_size_x = 32, local_size_y = 32, local_size_z = 1) in;
 
 BEGIN_PUSH_CONSTANTS
 UNIFORM(1) uvec2 block_dims;
-
-UNIFORM(2) uint bytes_per_block_log2;
-UNIFORM(3) uint layer_stride;
-UNIFORM(4) uint block_size;
-UNIFORM(5) uint x_shift;
-UNIFORM(6) uint block_height;
-UNIFORM(7) uint block_height_mask;
+UNIFORM(2) uint layer_stride;
+UNIFORM(3) uint block_size;
+UNIFORM(4) uint x_shift;
+UNIFORM(5) uint block_height;
+UNIFORM(6) uint block_height_mask;
 END_PUSH_CONSTANTS
 
 struct EncodingData {
@@ -53,35 +49,17 @@ struct TexelWeightParams {
     bool void_extent_hdr;
 };
 
-// Swizzle data
-layout(binding = BINDING_SWIZZLE_BUFFER, std430) readonly buffer SwizzleTable {
-    uint swizzle_table[];
-};
-
 layout(binding = BINDING_INPUT_BUFFER, std430) readonly buffer InputBufferU32 {
     uvec4 astc_data[];
 };
 
 layout(binding = BINDING_OUTPUT_IMAGE, rgba8) uniform writeonly image2DArray dest_image;
 
-const uint GOB_SIZE_X = 64;
-const uint GOB_SIZE_Y = 8;
-const uint GOB_SIZE_Z = 1;
-const uint GOB_SIZE = GOB_SIZE_X * GOB_SIZE_Y * GOB_SIZE_Z;
-
 const uint GOB_SIZE_X_SHIFT = 6;
 const uint GOB_SIZE_Y_SHIFT = 3;
-const uint GOB_SIZE_Z_SHIFT = 0;
-const uint GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT + GOB_SIZE_Z_SHIFT;
+const uint GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT;
 
-const uvec2 SWIZZLE_MASK = uvec2(GOB_SIZE_X - 1, GOB_SIZE_Y - 1);
-
-const int BLOCK_SIZE_IN_BYTES = 16;
-
-const int BLOCK_INFO_ERROR = 0;
-const int BLOCK_INFO_VOID_EXTENT_HDR = 1;
-const int BLOCK_INFO_VOID_EXTENT_LDR = 2;
-const int BLOCK_INFO_NORMAL = 3;
+const uint BYTES_PER_BLOCK_LOG2 = 4;
 
 const int JUST_BITS = 0;
 const int QUINT = 1;
@@ -168,8 +146,10 @@ int texel_vector_index = 0;
 uint unquantized_texel_weights[2][144];
 
 uint SwizzleOffset(uvec2 pos) {
-    pos = pos & SWIZZLE_MASK;
-    return swizzle_table[pos.y * 64 + pos.x];
+    uint x = pos.x;
+    uint y = pos.y;
+    return ((x % 64) / 32) * 256 + ((y % 8) / 2) * 64 + ((x % 32) / 16) * 32 +
+                          (y % 2) * 16 + (x % 16);
 }
 
 // Replicates low num_bits such that [(to_bit - 1):(to_bit - 1 - from_bit)]
@@ -1253,7 +1233,7 @@ void DecompressBlock(ivec3 coord) {
 
 void main() {
     uvec3 pos = gl_GlobalInvocationID;
-    pos.x <<= bytes_per_block_log2;
+    pos.x <<= BYTES_PER_BLOCK_LOG2;
 
     // Read as soon as possible due to its latency
     const uint swizzle = SwizzleOffset(pos.xy);

src/video_core/renderer_opengl/util_shaders.cpp

@@ -68,7 +68,6 @@ UtilShaders::~UtilShaders() = default;
 void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
                              std::span<const VideoCommon::SwizzleParameters> swizzles) {
     static constexpr GLuint BINDING_INPUT_BUFFER = 0;
-    static constexpr GLuint BINDING_SWIZZLE_BUFFER = 1;
     static constexpr GLuint BINDING_OUTPUT_IMAGE = 0;
 
     const Extent2D tile_size{
@@ -76,10 +75,9 @@ void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
         .height = VideoCore::Surface::DefaultBlockHeight(image.info.format),
     };
     program_manager.BindComputeProgram(astc_decoder_program.handle);
-    glBindBufferBase(GL_SHADER_STORAGE_BUFFER, BINDING_SWIZZLE_BUFFER, swizzle_table_buffer.handle);
-
     glFlushMappedNamedBufferRange(map.buffer, map.offset, image.guest_size_bytes);
     glUniform2ui(1, tile_size.width, tile_size.height);
+
     // Ensure buffer data is valid before dispatching
     glFlush();
     for (const SwizzleParameters& swizzle : swizzles) {
@@ -90,13 +88,13 @@ void UtilShaders::ASTCDecode(Image& image, const ImageBufferMap& map,
         const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
         ASSERT(params.origin == (std::array<u32, 3>{0, 0, 0}));
         ASSERT(params.destination == (std::array<s32, 3>{0, 0, 0}));
+        ASSERT(params.bytes_per_block_log2 == 4);
 
-        glUniform1ui(2, params.bytes_per_block_log2);
-        glUniform1ui(3, params.layer_stride);
-        glUniform1ui(4, params.block_size);
-        glUniform1ui(5, params.x_shift);
-        glUniform1ui(6, params.block_height);
-        glUniform1ui(7, params.block_height_mask);
+        glUniform1ui(2, params.layer_stride);
+        glUniform1ui(3, params.block_size);
+        glUniform1ui(4, params.x_shift);
+        glUniform1ui(5, params.block_height);
+        glUniform1ui(6, params.block_height_mask);
 
         // ASTC texture data
         glBindBufferRange(GL_SHADER_STORAGE_BUFFER, BINDING_INPUT_BUFFER, map.buffer, input_offset,

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -34,9 +34,8 @@ using Tegra::Texture::SWIZZLE_TABLE;
 namespace {
 
 constexpr u32 ASTC_BINDING_INPUT_BUFFER = 0;
-constexpr u32 ASTC_BINDING_SWIZZLE_BUFFER = 1;
-constexpr u32 ASTC_BINDING_OUTPUT_IMAGE = 2;
-constexpr size_t ASTC_NUM_BINDINGS = 3;
+constexpr u32 ASTC_BINDING_OUTPUT_IMAGE = 1;
+constexpr size_t ASTC_NUM_BINDINGS = 2;
 
 template <size_t size>
 inline constexpr VkPushConstantRange COMPUTE_PUSH_CONSTANT_RANGE{
@@ -80,13 +79,6 @@ constexpr std::array<VkDescriptorSetLayoutBinding, ASTC_NUM_BINDINGS> ASTC_DESCR
         .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
         .pImmutableSamplers = nullptr,
     },
-    {
-        .binding = ASTC_BINDING_SWIZZLE_BUFFER,
-        .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
-        .descriptorCount = 1,
-        .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
-        .pImmutableSamplers = nullptr,
-    },
     {
         .binding = ASTC_BINDING_OUTPUT_IMAGE,
         .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
@@ -98,12 +90,12 @@ constexpr std::array<VkDescriptorSetLayoutBinding, ASTC_NUM_BINDINGS> ASTC_DESCR
 
 constexpr DescriptorBankInfo ASTC_BANK_INFO{
     .uniform_buffers = 0,
-    .storage_buffers = 2,
+    .storage_buffers = 1,
     .texture_buffers = 0,
     .image_buffers = 0,
     .textures = 0,
     .images = 1,
-    .score = 3,
+    .score = 2,
 };
 
 constexpr VkDescriptorUpdateTemplateEntryKHR INPUT_OUTPUT_DESCRIPTOR_UPDATE_TEMPLATE{
@@ -125,14 +117,6 @@ constexpr std::array<VkDescriptorUpdateTemplateEntryKHR, ASTC_NUM_BINDINGS>
             .offset = ASTC_BINDING_INPUT_BUFFER * sizeof(DescriptorUpdateEntry),
             .stride = sizeof(DescriptorUpdateEntry),
         },
-        {
-            .dstBinding = ASTC_BINDING_SWIZZLE_BUFFER,
-            .dstArrayElement = 0,
-            .descriptorCount = 1,
-            .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
-            .offset = ASTC_BINDING_SWIZZLE_BUFFER * sizeof(DescriptorUpdateEntry),
-            .stride = sizeof(DescriptorUpdateEntry),
-        },
         {
             .dstBinding = ASTC_BINDING_OUTPUT_IMAGE,
             .dstArrayElement = 0,
@@ -145,7 +129,6 @@ constexpr std::array<VkDescriptorUpdateTemplateEntryKHR, ASTC_NUM_BINDINGS>
 
 struct AstcPushConstants {
     std::array<u32, 2> blocks_dims;
-    u32 bytes_per_block_log2;
     u32 layer_stride;
     u32 block_size;
     u32 x_shift;
@@ -336,42 +319,6 @@ ASTCDecoderPass::ASTCDecoderPass(const Device& device_, VKScheduler& scheduler_,
 
 ASTCDecoderPass::~ASTCDecoderPass() = default;
 
-void ASTCDecoderPass::MakeDataBuffer() {
-    constexpr size_t TOTAL_BUFFER_SIZE = sizeof(SWIZZLE_TABLE);
-    data_buffer = device.GetLogical().CreateBuffer(VkBufferCreateInfo{
-        .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
-        .pNext = nullptr,
-        .flags = 0,
-        .size = TOTAL_BUFFER_SIZE,
-        .usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
-        .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
-        .queueFamilyIndexCount = 0,
-        .pQueueFamilyIndices = nullptr,
-    });
-    data_buffer_commit = memory_allocator.Commit(data_buffer, MemoryUsage::Upload);
-
-    const auto staging_ref = staging_buffer_pool.Request(TOTAL_BUFFER_SIZE, MemoryUsage::Upload);
-    std::memcpy(staging_ref.mapped_span.data(), &SWIZZLE_TABLE, sizeof(SWIZZLE_TABLE));
-
-    scheduler.Record([src = staging_ref.buffer, offset = staging_ref.offset, dst = *data_buffer,
-                      TOTAL_BUFFER_SIZE](vk::CommandBuffer cmdbuf) {
-        static constexpr VkMemoryBarrier write_barrier{
-            .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
-            .pNext = nullptr,
-            .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
-            .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
-        };
-        const VkBufferCopy copy{
-            .srcOffset = offset,
-            .dstOffset = 0,
-            .size = TOTAL_BUFFER_SIZE,
-        };
-        cmdbuf.CopyBuffer(src, dst, copy);
-        cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
-                               0, write_barrier);
-    });
-}
-
 void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
                                std::span<const VideoCommon::SwizzleParameters> swizzles) {
     using namespace VideoCommon::Accelerated;
@@ -380,9 +327,6 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         VideoCore::Surface::DefaultBlockHeight(image.info.format),
     };
     scheduler.RequestOutsideRenderPassOperationContext();
-    if (!data_buffer) {
-        MakeDataBuffer();
-    }
     const VkPipeline vk_pipeline = *pipeline;
     const VkImageAspectFlags aspect_mask = image.AspectMask();
     const VkImage vk_image = image.Handle();
@@ -421,7 +365,6 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         update_descriptor_queue.Acquire();
         update_descriptor_queue.AddBuffer(map.buffer, input_offset,
                                           image.guest_size_bytes - swizzle.buffer_offset);
-        update_descriptor_queue.AddBuffer(*data_buffer, 0, sizeof(SWIZZLE_TABLE));
         update_descriptor_queue.AddImage(image.StorageImageView(swizzle.level));
         const void* const descriptor_data{update_descriptor_queue.UpdateData()};
 
@@ -429,11 +372,11 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
         ASSERT(params.origin == (std::array<u32, 3>{0, 0, 0}));
         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](vk::CommandBuffer cmdbuf) {
             const AstcPushConstants uniforms{
                 .blocks_dims = block_dims,
-                .bytes_per_block_log2 = params.bytes_per_block_log2,
                 .layer_stride = params.layer_stride,
                 .block_size = params.block_size,
                 .x_shift = params.x_shift,

src/video_core/renderer_vulkan/vk_compute_pass.h

@@ -96,15 +96,10 @@ public:
                   std::span<const VideoCommon::SwizzleParameters> swizzles);
 
 private:
-    void MakeDataBuffer();
-
     VKScheduler& scheduler;
     StagingBufferPool& staging_buffer_pool;
     VKUpdateDescriptorQueue& update_descriptor_queue;
     MemoryAllocator& memory_allocator;
-
-    vk::Buffer data_buffer;
-    MemoryCommit data_buffer_commit;
 };
 
 } // namespace Vulkan