vulkan dims specialization

src/video_core/host_shaders/CMakeLists.txt

@@ -13,6 +13,11 @@ set(GLSL_INCLUDES
     ${FIDELITYFX_FILES}
 )
 
+set(ASTC_INCLUDES
+    # astc_decoder_glsl_includes.h
+    astc_decoder_spv_includes.h
+)
+
 set(SHADER_FILES
     astc_decoder.comp
     blit_color_float.frag
@@ -95,9 +100,60 @@ if (NOT GLSLANG_ERROR STREQUAL "")
     set(QUIET_FLAG "")
 endif()
 
+macro(ASTC_GEN)
+    # paired list of valid astc block dimensions
+    set(ASTC_WIDTHS  4 5 5 6 6 6 8 8 8 10 10 10 10 12 12)
+    set(ASTC_HEIGHTS 4 4 5 4 5 6 5 6 8  5  6  8 10 10 12)
+    list(LENGTH ASTC_WIDTHS NUM_ASTC_FORMATS)
+    math(EXPR NUM_ASTC_FORMATS "${NUM_ASTC_FORMATS}-1")
+    foreach(i RANGE ${NUM_ASTC_FORMATS})
+        list(GET ASTC_WIDTHS ${i} ASTC_WIDTH)
+        list(GET ASTC_HEIGHTS ${i} ASTC_HEIGHT)
+
+        # Vulkan SPIR-V Specialization
+
+        string(TOUPPER ${SHADER_NAME}_${ASTC_WIDTH}x${ASTC_HEIGHT}_SPV SPIRV_VARIABLE_NAME)
+        set(SPIRV_HEADER_FILE ${SHADER_DIR}/${SHADER_NAME}_${ASTC_WIDTH}x${ASTC_HEIGHT}_spv.h)
+        add_custom_command(
+            OUTPUT
+                ${SPIRV_HEADER_FILE}
+            COMMAND
+                ${GLSLANGVALIDATOR} -V -DBLOCK_WIDTH=${ASTC_WIDTH} -DBLOCK_HEIGHT=${ASTC_HEIGHT} ${QUIET_FLAG} ${GLSL_FLAGS} --variable-name ${SPIRV_VARIABLE_NAME} -o ${SPIRV_HEADER_FILE} ${SOURCE_FILE}
+            MAIN_DEPENDENCY
+                ${SOURCE_FILE}
+        )
+        set(SHADER_HEADERS ${SHADER_HEADERS} ${SPIRV_HEADER_FILE})
+
+        # GLSL Specialization
+        # Disabled as there was no noticeable performance uplift specializing the shaders for OGL
+
+        # set(SOURCE_HEADER_FILE ${SHADER_DIR}/${SHADER_NAME}_${ASTC_WIDTH}x${ASTC_HEIGHT}.h)
+        # set(SHADER_DEFINES "#define BLOCK_WIDTH ${ASTC_WIDTH}" "#define BLOCK_HEIGHT ${ASTC_HEIGHT}")
+        # set(DEFINES_LINE_NUMBER 14)
+        # string(TOUPPER ${SHADER_NAME}_${ASTC_WIDTH}x${ASTC_HEIGHT} GLSL_VARIABLE_NAME)
+        # add_custom_command(
+        #     OUTPUT
+        #         ${SOURCE_HEADER_FILE}
+        #     COMMAND
+        #         ${CMAKE_COMMAND} -P ${HEADER_GENERATOR} ${SOURCE_FILE} ${SOURCE_HEADER_FILE} ${INPUT_FILE} "${SHADER_DEFINES}" ${DEFINES_LINE_NUMBER} ${GLSL_VARIABLE_NAME}
+        #     MAIN_DEPENDENCY
+        #         ${SOURCE_FILE}
+        #     DEPENDS
+        #         ${INPUT_FILE}
+        #         ${SOURCE_FILE}
+        # )
+        # set(SHADER_HEADERS ${SHADER_HEADERS} ${SOURCE_HEADER_FILE})
+    endforeach()
+endmacro()
+
 foreach(FILENAME IN ITEMS ${SHADER_FILES})
     string(REPLACE "." "_" SHADER_NAME ${FILENAME})
     set(SOURCE_FILE ${CMAKE_CURRENT_SOURCE_DIR}/${FILENAME})
+
+    if (${FILENAME} MATCHES "astc_decoder.comp")
+        ASTC_GEN()
+    endif()
+
     # Skip generating source headers on Vulkan exclusive files
     if (NOT ${FILENAME} MATCHES "vulkan.*")
         set(SOURCE_HEADER_FILE ${SHADER_DIR}/${SHADER_NAME}.h)
@@ -151,6 +207,7 @@ endforeach()
 
 set(SHADER_SOURCES ${SHADER_FILES})
 list(APPEND SHADER_SOURCES ${GLSL_INCLUDES})
+list(APPEND SHADER_SOURCES ${ASTC_INCLUDES})
 
 add_custom_target(host_shaders
     DEPENDS

src/video_core/host_shaders/astc_decoder.comp

@@ -24,7 +24,9 @@
 layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
 
 BEGIN_PUSH_CONSTANTS
+#ifndef BLOCK_WIDTH
 UNIFORM(1) uvec2 block_dims;
+#endif
 UNIFORM(2) uint layer_stride;
 UNIFORM(3) uint block_size;
 UNIFORM(4) uint x_shift;
@@ -75,7 +77,15 @@ int color_bitsread = 0;
 // At most will require BLOCK_WIDTH x BLOCK_HEIGHT x 2 in dual plane mode
 // So the maximum would be 144 (12 x 12) elements, x 2 for two planes
 #define DIVCEIL(number, divisor) (number + divisor - 1) / divisor
+
+#ifndef BLOCK_WIDTH
+#define BLOCK_WIDTH block_dims.x
+#define BLOCK_HEIGHT block_dims.y
 #define ARRAY_NUM_ELEMENTS 144
+#else
+#define ARRAY_NUM_ELEMENTS BLOCK_WIDTH * BLOCK_HEIGHT
+#endif
+
 #define VECTOR_ARRAY_SIZE DIVCEIL(ARRAY_NUM_ELEMENTS * 2, 4)
 uvec4 result_vector[VECTOR_ARRAY_SIZE];
 
@@ -265,7 +275,7 @@ uint Hash52(uint p) {
 }
 
 uint Select2DPartition(uint seed, uint x, uint y, uint partition_count) {
-    if ((block_dims.y * block_dims.x) < 32) {
+    if ((BLOCK_WIDTH * BLOCK_HEIGHT) < 32) {
         x <<= 1;
         y <<= 1;
     }
@@ -878,8 +888,8 @@ uint UnquantizeTexelWeight(EncodingData val) {
 uvec4 unquantized_texel_weights[VECTOR_ARRAY_SIZE];
 
 void UnquantizeTexelWeights(uvec2 size, bool is_dual_plane) {
-    const uint Ds = uint((block_dims.x * 0.5f + 1024) / (block_dims.x - 1));
+    const uint Ds = uint((BLOCK_WIDTH * 0.5f + 1024) / (BLOCK_WIDTH - 1));
-    const uint Dt = uint((block_dims.y * 0.5f + 1024) / (block_dims.y - 1));
+    const uint Dt = uint((BLOCK_HEIGHT * 0.5f + 1024) / (BLOCK_HEIGHT - 1));
     const uint num_planes = is_dual_plane ? 2 : 1;
     const uint area = size.x * size.y;
     const uint loop_count = min(result_index, area * num_planes);
@@ -890,8 +900,8 @@ void UnquantizeTexelWeights(uvec2 size, bool is_dual_plane) {
             UnquantizeTexelWeight(GetEncodingFromVector(itr));
     }
     for (uint plane = 0; plane < num_planes; ++plane) {
-        for (uint t = 0; t < block_dims.y; t++) {
+        for (uint t = 0; t < BLOCK_HEIGHT; t++) {
-            for (uint s = 0; s < block_dims.x; s++) {
+            for (uint s = 0; s < BLOCK_WIDTH; s++) {
                 const uint cs = Ds * s;
                 const uint ct = Dt * t;
                 const uint gs = (cs * (size.x - 1) + 32) >> 6;
@@ -934,7 +944,7 @@ void UnquantizeTexelWeights(uvec2 size, bool is_dual_plane) {
                     VectorIndicesFromBase(offset_base);
                     p.w = result_vector[array_index][vector_index];
                 }
-                const uint offset = (t * block_dims.x + s) + ARRAY_NUM_ELEMENTS * plane;
+                const uint offset = (t * BLOCK_WIDTH + s) + ARRAY_NUM_ELEMENTS * plane;
                 const uint array_index = offset / 4;
                 const uint vector_index = offset % 4;
                 unquantized_texel_weights[array_index][vector_index] = (uint(dot(p, w)) + 8) >> 4;
@@ -976,8 +986,8 @@ int FindLayout(uint mode) {
 
 
 void FillError(ivec3 coord) {
-    for (uint j = 0; j < block_dims.y; j++) {
+    for (uint j = 0; j < BLOCK_HEIGHT; j++) {
-        for (uint i = 0; i < block_dims.x; i++) {
+        for (uint i = 0; i < BLOCK_WIDTH; i++) {
             imageStore(dest_image, coord + ivec3(i, j, 0), vec4(0.0, 0.0, 0.0, 0.0));
         }
     }
@@ -993,8 +1003,8 @@ void FillVoidExtentLDR(ivec3 coord) {
     const float r = float(r_u) / 65535.0f;
     const float g = float(g_u) / 65535.0f;
     const float b = float(b_u) / 65535.0f;
-    for (uint j = 0; j < block_dims.y; j++) {
+    for (uint j = 0; j < BLOCK_HEIGHT; j++) {
-        for (uint i = 0; i < block_dims.x; i++) {
+        for (uint i = 0; i < BLOCK_WIDTH; i++) {
             imageStore(dest_image, coord + ivec3(i, j, 0), vec4(r, g, b, a));
         }
     }
@@ -1089,7 +1099,7 @@ void DecompressBlock(ivec3 coord) {
         return;
     }
     const uvec2 size_params = DecodeBlockSize(mode);
-    if ((size_params.x > block_dims.x) || (size_params.y > block_dims.y)) {
+    if ((size_params.x > BLOCK_WIDTH) || (size_params.y > BLOCK_HEIGHT)) {
         FillError(coord);
         return;
     }
@@ -1218,21 +1228,21 @@ void DecompressBlock(ivec3 coord) {
     DecodeIntegerSequence(max_weight, GetNumWeightValues(size_params, dual_plane));
 
     UnquantizeTexelWeights(size_params, dual_plane);
-    for (uint j = 0; j < block_dims.y; j++) {
+    for (uint j = 0; j < BLOCK_HEIGHT; j++) {
-        for (uint i = 0; i < block_dims.x; i++) {
+        for (uint i = 0; i < BLOCK_WIDTH; i++) {
             uint local_partition = 0;
             if (num_partitions > 1) {
                 local_partition = Select2DPartition(partition_index, i, j, num_partitions);
             }
             const uvec4 C0 = ReplicateByteTo16(endpoints0[local_partition]);
             const uvec4 C1 = ReplicateByteTo16(endpoints1[local_partition]);
-            const uint weight_offset = (j * block_dims.x + i);
+            const uint weight_offset = (j * BLOCK_WIDTH + i);
             const uint array_index = weight_offset / 4;
             const uint vector_index = weight_offset % 4;
             const uint primary_weight = unquantized_texel_weights[array_index][vector_index];
             uvec4 weight_vec = uvec4(primary_weight);
             if (dual_plane) {
-                const uint secondary_weight_offset = (j * block_dims.x + i) + ARRAY_NUM_ELEMENTS;
+                const uint secondary_weight_offset = (j * BLOCK_WIDTH + i) + ARRAY_NUM_ELEMENTS;
                 const uint secondary_array_index = secondary_weight_offset / 4;
                 const uint secondary_vector_index = secondary_weight_offset % 4;
                 const uint secondary_weight =
@@ -1270,7 +1280,7 @@ void main() {
     offset += (pos.x >> GOB_SIZE_X_SHIFT) << x_shift;
     offset += swizzle;
 
-    const ivec3 coord = ivec3(gl_GlobalInvocationID * uvec3(block_dims, 1));
+    const ivec3 coord = ivec3(gl_GlobalInvocationID * uvec3(BLOCK_WIDTH, BLOCK_HEIGHT, 1));
     if (any(greaterThanEqual(coord, imageSize(dest_image)))) {
         return;
     }

src/video_core/host_shaders/astc_decoder_spv_includes.h

@@ -0,0 +1,20 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#include "video_core/host_shaders/astc_decoder_comp_10x10_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_10x5_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_10x6_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_10x8_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_12x10_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_12x12_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_4x4_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_5x4_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_5x5_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_6x5_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_6x6_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_8x5_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_8x6_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_8x8_spv.h"
+#include "video_core/host_shaders/astc_decoder_comp_spv.h"

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -11,7 +11,7 @@
 #include "common/assert.h"
 #include "common/common_types.h"
 #include "common/div_ceil.h"
-#include "video_core/host_shaders/astc_decoder_comp_spv.h"
+#include "video_core/host_shaders/astc_decoder_spv_includes.h"
 #include "video_core/host_shaders/vulkan_quad_indexed_comp_spv.h"
 #include "video_core/host_shaders/vulkan_uint8_comp_spv.h"
 #include "video_core/renderer_vulkan/vk_compute_pass.h"
@@ -124,13 +124,62 @@ constexpr std::array<VkDescriptorUpdateTemplateEntry, ASTC_NUM_BINDINGS>
     }};
 
 struct AstcPushConstants {
-    std::array<u32, 2> blocks_dims;
     u32 layer_stride;
     u32 block_size;
     u32 x_shift;
     u32 block_height;
     u32 block_height_mask;
 };
+
+size_t AstcFormatIndex(VideoCore::Surface::PixelFormat format) {
+    switch (format) {
+    case VideoCore::Surface::PixelFormat::ASTC_2D_4X4_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_4X4_UNORM:
+        return 0;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_5X4_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_5X4_UNORM:
+        return 1;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_5X5_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_5X5_UNORM:
+        return 2;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_6X5_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_6X5_UNORM:
+        return 3;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_6X6_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_6X6_UNORM:
+        return 4;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_8X5_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_8X5_UNORM:
+        return 5;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_8X6_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_8X6_UNORM:
+        return 6;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_8X8_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_8X8_UNORM:
+        return 7;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_10X5_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_10X5_UNORM:
+        return 8;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_10X6_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_10X6_UNORM:
+        return 9;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_10X8_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_10X8_UNORM:
+        return 10;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_10X10_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_10X10_UNORM:
+        return 11;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_12X10_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_12X10_UNORM:
+        return 12;
+    case VideoCore::Surface::PixelFormat::ASTC_2D_12X12_SRGB:
+    case VideoCore::Surface::PixelFormat::ASTC_2D_12X12_UNORM:
+        return 13;
+    default:
+        UNREACHABLE();
+        return 0;
+    }
+}
 } // Anonymous namespace
 
 ComputePass::ComputePass(const Device& device_, DescriptorPool& descriptor_pool,
@@ -312,19 +361,53 @@ ASTCDecoderPass::ASTCDecoderPass(const Device& device_, Scheduler& scheduler_,
                   COMPUTE_PUSH_CONSTANT_RANGE<sizeof(AstcPushConstants)>, ASTC_DECODER_COMP_SPV),
       scheduler{scheduler_}, staging_buffer_pool{staging_buffer_pool_},
       compute_pass_descriptor_queue{compute_pass_descriptor_queue_}, memory_allocator{
-                                                                         memory_allocator_} {}
+                                                                         memory_allocator_} {
+    // These must match the order found in AstcFormatIndex
+    static constexpr std::array<std::span<const u32>, 14> ASTC_SHADERS{
+        ASTC_DECODER_COMP_4X4_SPV,   ASTC_DECODER_COMP_5X4_SPV,   ASTC_DECODER_COMP_5X5_SPV,
+        ASTC_DECODER_COMP_6X5_SPV,   ASTC_DECODER_COMP_6X6_SPV,   ASTC_DECODER_COMP_8X5_SPV,
+        ASTC_DECODER_COMP_8X6_SPV,   ASTC_DECODER_COMP_8X8_SPV,   ASTC_DECODER_COMP_10X5_SPV,
+        ASTC_DECODER_COMP_10X6_SPV,  ASTC_DECODER_COMP_10X8_SPV,  ASTC_DECODER_COMP_10X10_SPV,
+        ASTC_DECODER_COMP_12X10_SPV, ASTC_DECODER_COMP_12X12_SPV,
+    };
+    for (size_t index = 0; index < ASTC_SHADERS.size(); ++index) {
+        const auto& code = ASTC_SHADERS[index];
+        const auto module_ = device.GetLogical().CreateShaderModule({
+            .sType = VK_STRUCTURE_TYPE_SHADER_MODULE_CREATE_INFO,
+            .pNext = nullptr,
+            .flags = 0,
+            .codeSize = static_cast<u32>(code.size_bytes()),
+            .pCode = code.data(),
+        });
+        device.SaveShader(code);
+        astc_pipelines[index] = device.GetLogical().CreateComputePipeline({
+            .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
+            .pNext = nullptr,
+            .flags = 0,
+            .stage{
+                .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
+                .pNext = nullptr,
+                .flags = 0,
+                .stage = VK_SHADER_STAGE_COMPUTE_BIT,
+                .module = *module_,
+                .pName = "main",
+                .pSpecializationInfo = nullptr,
+            },
+            .layout = *layout,
+            .basePipelineHandle = nullptr,
+            .basePipelineIndex = 0,
+        });
+    }
+}
 
 ASTCDecoderPass::~ASTCDecoderPass() = default;
 
 void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
                                std::span<const VideoCommon::SwizzleParameters> swizzles) {
     using namespace VideoCommon::Accelerated;
-    const std::array<u32, 2> block_dims{
+
-        VideoCore::Surface::DefaultBlockWidth(image.info.format),
-        VideoCore::Surface::DefaultBlockHeight(image.info.format),
-    };
     scheduler.RequestOutsideRenderPassOperationContext();
-    const VkPipeline vk_pipeline = *pipeline;
+    const VkPipeline vk_pipeline = *astc_pipelines[AstcFormatIndex(image.info.format)];
     const VkImageAspectFlags aspect_mask = image.AspectMask();
     const VkImage vk_image = image.Handle();
     const bool is_initialized = image.ExchangeInitialization();
@@ -371,10 +454,9 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
         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,
+        scheduler.Record([this, num_dispatches_x, num_dispatches_y, num_dispatches_z, params,
-                          params, descriptor_data](vk::CommandBuffer cmdbuf) {
+                          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,

src/video_core/renderer_vulkan/vk_compute_pass.h

@@ -95,6 +95,8 @@ public:
                   std::span<const VideoCommon::SwizzleParameters> swizzles);
 
 private:
+    std::array<vk::Pipeline, 14> astc_pipelines;
+
     Scheduler& scheduler;
     StagingBufferPool& staging_buffer_pool;
     ComputePassDescriptorQueue& compute_pass_descriptor_queue;