summaryrefslogtreecommitdiff
path: root/src/video_core/renderer_vulkan
diff options
context:
space:
mode:
authorGravatar bunnei2021-08-06 21:45:24 -0700
committerGravatar GitHub2021-08-06 21:45:24 -0700
commit268b5764c70a8300d24c32985dee595046a1e2e1 (patch)
tree178317fbc7f34549a93b8e28d9f0b6857aa104c8 /src/video_core/renderer_vulkan
parentMerge pull request #6799 from ameerj/vp9-fixes (diff)
parentastc_decoder: Reduce workgroup size (diff)
downloadyuzu-268b5764c70a8300d24c32985dee595046a1e2e1.tar.gz
yuzu-268b5764c70a8300d24c32985dee595046a1e2e1.tar.xz
yuzu-268b5764c70a8300d24c32985dee595046a1e2e1.zip
Merge pull request #6791 from ameerj/astc-opt
astc_decoder: Various performance and memory optimizations
Diffstat (limited to 'src/video_core/renderer_vulkan')
-rw-r--r--src/video_core/renderer_vulkan/vk_compute_pass.cpp97
-rw-r--r--src/video_core/renderer_vulkan/vk_compute_pass.h5
2 files changed, 8 insertions, 94 deletions
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.cpp b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
index 561cf5e11..3e96c0f60 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pass.cpp
+++ b/src/video_core/renderer_vulkan/vk_compute_pass.cpp
@@ -30,16 +30,12 @@
30namespace Vulkan { 30namespace Vulkan {
31 31
32using Tegra::Texture::SWIZZLE_TABLE; 32using Tegra::Texture::SWIZZLE_TABLE;
33using Tegra::Texture::ASTC::ASTC_ENCODINGS_VALUES;
34using namespace Tegra::Texture::ASTC;
35 33
36namespace { 34namespace {
37 35
38constexpr u32 ASTC_BINDING_INPUT_BUFFER = 0; 36constexpr u32 ASTC_BINDING_INPUT_BUFFER = 0;
39constexpr u32 ASTC_BINDING_ENC_BUFFER = 1; 37constexpr u32 ASTC_BINDING_OUTPUT_IMAGE = 1;
40constexpr u32 ASTC_BINDING_SWIZZLE_BUFFER = 2; 38constexpr size_t ASTC_NUM_BINDINGS = 2;
41constexpr u32 ASTC_BINDING_OUTPUT_IMAGE = 3;
42constexpr size_t ASTC_NUM_BINDINGS = 4;
43 39
44template <size_t size> 40template <size_t size>
45inline constexpr VkPushConstantRange COMPUTE_PUSH_CONSTANT_RANGE{ 41inline constexpr VkPushConstantRange COMPUTE_PUSH_CONSTANT_RANGE{
@@ -75,7 +71,7 @@ constexpr DescriptorBankInfo INPUT_OUTPUT_BANK_INFO{
75 .score = 2, 71 .score = 2,
76}; 72};
77 73
78constexpr std::array<VkDescriptorSetLayoutBinding, 4> ASTC_DESCRIPTOR_SET_BINDINGS{{ 74constexpr std::array<VkDescriptorSetLayoutBinding, ASTC_NUM_BINDINGS> ASTC_DESCRIPTOR_SET_BINDINGS{{
79 { 75 {
80 .binding = ASTC_BINDING_INPUT_BUFFER, 76 .binding = ASTC_BINDING_INPUT_BUFFER,
81 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER, 77 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
@@ -84,20 +80,6 @@ constexpr std::array<VkDescriptorSetLayoutBinding, 4> ASTC_DESCRIPTOR_SET_BINDIN
84 .pImmutableSamplers = nullptr, 80 .pImmutableSamplers = nullptr,
85 }, 81 },
86 { 82 {
87 .binding = ASTC_BINDING_ENC_BUFFER,
88 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
89 .descriptorCount = 1,
90 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
91 .pImmutableSamplers = nullptr,
92 },
93 {
94 .binding = ASTC_BINDING_SWIZZLE_BUFFER,
95 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
96 .descriptorCount = 1,
97 .stageFlags = VK_SHADER_STAGE_COMPUTE_BIT,
98 .pImmutableSamplers = nullptr,
99 },
100 {
101 .binding = ASTC_BINDING_OUTPUT_IMAGE, 83 .binding = ASTC_BINDING_OUTPUT_IMAGE,
102 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE, 84 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_IMAGE,
103 .descriptorCount = 1, 85 .descriptorCount = 1,
@@ -108,12 +90,12 @@ constexpr std::array<VkDescriptorSetLayoutBinding, 4> ASTC_DESCRIPTOR_SET_BINDIN
108 90
109constexpr DescriptorBankInfo ASTC_BANK_INFO{ 91constexpr DescriptorBankInfo ASTC_BANK_INFO{
110 .uniform_buffers = 0, 92 .uniform_buffers = 0,
111 .storage_buffers = 3, 93 .storage_buffers = 1,
112 .texture_buffers = 0, 94 .texture_buffers = 0,
113 .image_buffers = 0, 95 .image_buffers = 0,
114 .textures = 0, 96 .textures = 0,
115 .images = 1, 97 .images = 1,
116 .score = 4, 98 .score = 2,
117}; 99};
118 100
119constexpr VkDescriptorUpdateTemplateEntryKHR INPUT_OUTPUT_DESCRIPTOR_UPDATE_TEMPLATE{ 101constexpr VkDescriptorUpdateTemplateEntryKHR INPUT_OUTPUT_DESCRIPTOR_UPDATE_TEMPLATE{
@@ -136,22 +118,6 @@ constexpr std::array<VkDescriptorUpdateTemplateEntryKHR, ASTC_NUM_BINDINGS>
136 .stride = sizeof(DescriptorUpdateEntry), 118 .stride = sizeof(DescriptorUpdateEntry),
137 }, 119 },
138 { 120 {
139 .dstBinding = ASTC_BINDING_ENC_BUFFER,
140 .dstArrayElement = 0,
141 .descriptorCount = 1,
142 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
143 .offset = ASTC_BINDING_ENC_BUFFER * sizeof(DescriptorUpdateEntry),
144 .stride = sizeof(DescriptorUpdateEntry),
145 },
146 {
147 .dstBinding = ASTC_BINDING_SWIZZLE_BUFFER,
148 .dstArrayElement = 0,
149 .descriptorCount = 1,
150 .descriptorType = VK_DESCRIPTOR_TYPE_STORAGE_BUFFER,
151 .offset = ASTC_BINDING_SWIZZLE_BUFFER * sizeof(DescriptorUpdateEntry),
152 .stride = sizeof(DescriptorUpdateEntry),
153 },
154 {
155 .dstBinding = ASTC_BINDING_OUTPUT_IMAGE, 121 .dstBinding = ASTC_BINDING_OUTPUT_IMAGE,
156 .dstArrayElement = 0, 122 .dstArrayElement = 0,
157 .descriptorCount = 1, 123 .descriptorCount = 1,
@@ -163,7 +129,6 @@ constexpr std::array<VkDescriptorUpdateTemplateEntryKHR, ASTC_NUM_BINDINGS>
163 129
164struct AstcPushConstants { 130struct AstcPushConstants {
165 std::array<u32, 2> blocks_dims; 131 std::array<u32, 2> blocks_dims;
166 u32 bytes_per_block_log2;
167 u32 layer_stride; 132 u32 layer_stride;
168 u32 block_size; 133 u32 block_size;
169 u32 x_shift; 134 u32 x_shift;
@@ -354,46 +319,6 @@ ASTCDecoderPass::ASTCDecoderPass(const Device& device_, VKScheduler& scheduler_,
354 319
355ASTCDecoderPass::~ASTCDecoderPass() = default; 320ASTCDecoderPass::~ASTCDecoderPass() = default;
356 321
357void ASTCDecoderPass::MakeDataBuffer() {
358 constexpr size_t TOTAL_BUFFER_SIZE = sizeof(ASTC_ENCODINGS_VALUES) + sizeof(SWIZZLE_TABLE);
359 data_buffer = device.GetLogical().CreateBuffer(VkBufferCreateInfo{
360 .sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO,
361 .pNext = nullptr,
362 .flags = 0,
363 .size = TOTAL_BUFFER_SIZE,
364 .usage = VK_BUFFER_USAGE_STORAGE_BUFFER_BIT | VK_BUFFER_USAGE_TRANSFER_DST_BIT,
365 .sharingMode = VK_SHARING_MODE_EXCLUSIVE,
366 .queueFamilyIndexCount = 0,
367 .pQueueFamilyIndices = nullptr,
368 });
369 data_buffer_commit = memory_allocator.Commit(data_buffer, MemoryUsage::Upload);
370
371 const auto staging_ref = staging_buffer_pool.Request(TOTAL_BUFFER_SIZE, MemoryUsage::Upload);
372 std::memcpy(staging_ref.mapped_span.data(), &ASTC_ENCODINGS_VALUES,
373 sizeof(ASTC_ENCODINGS_VALUES));
374 // Tack on the swizzle table at the end of the buffer
375 std::memcpy(staging_ref.mapped_span.data() + sizeof(ASTC_ENCODINGS_VALUES), &SWIZZLE_TABLE,
376 sizeof(SWIZZLE_TABLE));
377
378 scheduler.Record([src = staging_ref.buffer, offset = staging_ref.offset, dst = *data_buffer,
379 TOTAL_BUFFER_SIZE](vk::CommandBuffer cmdbuf) {
380 static constexpr VkMemoryBarrier write_barrier{
381 .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER,
382 .pNext = nullptr,
383 .srcAccessMask = VK_ACCESS_TRANSFER_WRITE_BIT,
384 .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
385 };
386 const VkBufferCopy copy{
387 .srcOffset = offset,
388 .dstOffset = 0,
389 .size = TOTAL_BUFFER_SIZE,
390 };
391 cmdbuf.CopyBuffer(src, dst, copy);
392 cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT,
393 0, write_barrier);
394 });
395}
396
397void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map, 322void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
398 std::span<const VideoCommon::SwizzleParameters> swizzles) { 323 std::span<const VideoCommon::SwizzleParameters> swizzles) {
399 using namespace VideoCommon::Accelerated; 324 using namespace VideoCommon::Accelerated;
@@ -402,9 +327,6 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
402 VideoCore::Surface::DefaultBlockHeight(image.info.format), 327 VideoCore::Surface::DefaultBlockHeight(image.info.format),
403 }; 328 };
404 scheduler.RequestOutsideRenderPassOperationContext(); 329 scheduler.RequestOutsideRenderPassOperationContext();
405 if (!data_buffer) {
406 MakeDataBuffer();
407 }
408 const VkPipeline vk_pipeline = *pipeline; 330 const VkPipeline vk_pipeline = *pipeline;
409 const VkImageAspectFlags aspect_mask = image.AspectMask(); 331 const VkImageAspectFlags aspect_mask = image.AspectMask();
410 const VkImage vk_image = image.Handle(); 332 const VkImage vk_image = image.Handle();
@@ -436,16 +358,13 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
436 }); 358 });
437 for (const VideoCommon::SwizzleParameters& swizzle : swizzles) { 359 for (const VideoCommon::SwizzleParameters& swizzle : swizzles) {
438 const size_t input_offset = swizzle.buffer_offset + map.offset; 360 const size_t input_offset = swizzle.buffer_offset + map.offset;
439 const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 32U); 361 const u32 num_dispatches_x = Common::DivCeil(swizzle.num_tiles.width, 8U);
440 const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 32U); 362 const u32 num_dispatches_y = Common::DivCeil(swizzle.num_tiles.height, 8U);
441 const u32 num_dispatches_z = image.info.resources.layers; 363 const u32 num_dispatches_z = image.info.resources.layers;
442 364
443 update_descriptor_queue.Acquire(); 365 update_descriptor_queue.Acquire();
444 update_descriptor_queue.AddBuffer(map.buffer, input_offset, 366 update_descriptor_queue.AddBuffer(map.buffer, input_offset,
445 image.guest_size_bytes - swizzle.buffer_offset); 367 image.guest_size_bytes - swizzle.buffer_offset);
446 update_descriptor_queue.AddBuffer(*data_buffer, 0, sizeof(ASTC_ENCODINGS_VALUES));
447 update_descriptor_queue.AddBuffer(*data_buffer, sizeof(ASTC_ENCODINGS_VALUES),
448 sizeof(SWIZZLE_TABLE));
449 update_descriptor_queue.AddImage(image.StorageImageView(swizzle.level)); 368 update_descriptor_queue.AddImage(image.StorageImageView(swizzle.level));
450 const void* const descriptor_data{update_descriptor_queue.UpdateData()}; 369 const void* const descriptor_data{update_descriptor_queue.UpdateData()};
451 370
@@ -453,11 +372,11 @@ void ASTCDecoderPass::Assemble(Image& image, const StagingBufferRef& map,
453 const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info); 372 const auto params = MakeBlockLinearSwizzle2DParams(swizzle, image.info);
454 ASSERT(params.origin == (std::array<u32, 3>{0, 0, 0})); 373 ASSERT(params.origin == (std::array<u32, 3>{0, 0, 0}));
455 ASSERT(params.destination == (std::array<s32, 3>{0, 0, 0})); 374 ASSERT(params.destination == (std::array<s32, 3>{0, 0, 0}));
375 ASSERT(params.bytes_per_block_log2 == 4);
456 scheduler.Record([this, num_dispatches_x, num_dispatches_y, num_dispatches_z, block_dims, 376 scheduler.Record([this, num_dispatches_x, num_dispatches_y, num_dispatches_z, block_dims,
457 params, descriptor_data](vk::CommandBuffer cmdbuf) { 377 params, descriptor_data](vk::CommandBuffer cmdbuf) {
458 const AstcPushConstants uniforms{ 378 const AstcPushConstants uniforms{
459 .blocks_dims = block_dims, 379 .blocks_dims = block_dims,
460 .bytes_per_block_log2 = params.bytes_per_block_log2,
461 .layer_stride = params.layer_stride, 380 .layer_stride = params.layer_stride,
462 .block_size = params.block_size, 381 .block_size = params.block_size,
463 .x_shift = params.x_shift, 382 .x_shift = params.x_shift,
diff --git a/src/video_core/renderer_vulkan/vk_compute_pass.h b/src/video_core/renderer_vulkan/vk_compute_pass.h
index 114aef2bd..c7b92cce0 100644
--- a/src/video_core/renderer_vulkan/vk_compute_pass.h
+++ b/src/video_core/renderer_vulkan/vk_compute_pass.h
@@ -96,15 +96,10 @@ public:
96 std::span<const VideoCommon::SwizzleParameters> swizzles); 96 std::span<const VideoCommon::SwizzleParameters> swizzles);
97 97
98private: 98private:
99 void MakeDataBuffer();
100
101 VKScheduler& scheduler; 99 VKScheduler& scheduler;
102 StagingBufferPool& staging_buffer_pool; 100 StagingBufferPool& staging_buffer_pool;
103 VKUpdateDescriptorQueue& update_descriptor_queue; 101 VKUpdateDescriptorQueue& update_descriptor_queue;
104 MemoryAllocator& memory_allocator; 102 MemoryAllocator& memory_allocator;
105
106 vk::Buffer data_buffer;
107 MemoryCommit data_buffer_commit;
108}; 103};
109 104
110} // namespace Vulkan 105} // namespace Vulkan