From c448b3af2ffe06f036e06b610c87db8dab74b766 Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Thu, 22 Dec 2022 16:52:53 -0500 Subject: [PATCH 1/3] texture_cache: Use pre-allocated buffer for texture uploads --- src/video_core/texture_cache/texture_cache.h | 26 ++++++++++++++----- .../texture_cache/texture_cache_base.h | 5 +++- src/video_core/texture_cache/util.cpp | 9 +++---- src/video_core/texture_cache/util.h | 1 + 4 files changed, 28 insertions(+), 13 deletions(-) diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h index 8e68a2e53..fccf4316d 100644 --- a/src/video_core/texture_cache/texture_cache.h +++ b/src/video_core/texture_cache/texture_cache.h @@ -39,6 +39,12 @@ TextureCache

::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface& sampler_descriptor.mipmap_filter.Assign(Tegra::Texture::TextureMipmapFilter::Linear); sampler_descriptor.cubemap_anisotropy.Assign(1); + // These values were chosen based on typical peak swizzle data sizes seen in some titles + static constexpr size_t SWIZZLE_DATA_BUFFER_INITIAL_CAPACITY = 8_MiB; + static constexpr size_t UNSWIZZLE_DATA_BUFFER_INITIAL_CAPACITY = 1_MiB; + swizzle_data_buffer.reserve(SWIZZLE_DATA_BUFFER_INITIAL_CAPACITY); + unswizzle_data_buffer.reserve(UNSWIZZLE_DATA_BUFFER_INITIAL_CAPACITY); + // Make sure the first index is reserved for the null resources // This way the null resource becomes a compile time constant void(slot_images.insert(NullImageParams{})); @@ -734,13 +740,21 @@ void TextureCache

::UploadImageContents(Image& image, StagingBuffer& staging) gpu_memory->ReadBlockUnsafe(gpu_addr, mapped_span.data(), mapped_span.size_bytes()); const auto uploads = FullUploadSwizzles(image.info); runtime.AccelerateImageUpload(image, staging, uploads); - } else if (True(image.flags & ImageFlagBits::Converted)) { - std::vector unswizzled_data(image.unswizzled_size_bytes); - auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, unswizzled_data); - ConvertImage(unswizzled_data, image.info, mapped_span, copies); + return; + } + const size_t guest_size_bytes = image.guest_size_bytes; + swizzle_data_buffer.resize(guest_size_bytes); + gpu_memory->ReadBlockUnsafe(gpu_addr, swizzle_data_buffer.data(), guest_size_bytes); + + if (True(image.flags & ImageFlagBits::Converted)) { + unswizzle_data_buffer.resize(image.unswizzled_size_bytes); + auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer, + unswizzle_data_buffer); + ConvertImage(unswizzle_data_buffer, image.info, mapped_span, copies); image.UploadMemory(staging, copies); } else { - const auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, mapped_span); + const auto copies = + UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer, mapped_span); image.UploadMemory(staging, copies); } } @@ -910,7 +924,7 @@ void TextureCache

::InvalidateScale(Image& image) { } template -u64 TextureCache

::GetScaledImageSizeBytes(ImageBase& image) { +u64 TextureCache

::GetScaledImageSizeBytes(const ImageBase& image) { const u64 scale_up = static_cast(Settings::values.resolution_info.up_scale * Settings::values.resolution_info.up_scale); const u64 down_shift = static_cast(Settings::values.resolution_info.down_shift + diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h index 587339a31..67e8acf25 100644 --- a/src/video_core/texture_cache/texture_cache_base.h +++ b/src/video_core/texture_cache/texture_cache_base.h @@ -368,7 +368,7 @@ private: void InvalidateScale(Image& image); bool ScaleUp(Image& image); bool ScaleDown(Image& image); - u64 GetScaledImageSizeBytes(ImageBase& image); + u64 GetScaledImageSizeBytes(const ImageBase& image); Runtime& runtime; @@ -417,6 +417,9 @@ private: std::unordered_map image_allocs_table; + std::vector swizzle_data_buffer; + std::vector unswizzle_data_buffer; + u64 modification_tick = 0; u64 frame_tick = 0; }; diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp index e8c908b42..4488fa9da 100644 --- a/src/video_core/texture_cache/util.cpp +++ b/src/video_core/texture_cache/util.cpp @@ -765,8 +765,9 @@ bool IsValidEntry(const Tegra::MemoryManager& gpu_memory, const TICEntry& config } std::vector UnswizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, - const ImageInfo& info, std::span output) { - const size_t guest_size_bytes = CalculateGuestSizeInBytes(info); + const ImageInfo& info, std::span input, + std::span output) { + const size_t guest_size_bytes = input.size_bytes(); const u32 bpp_log2 = BytesPerBlockLog2(info.format); const Extent3D size = info.size; @@ -789,10 +790,6 @@ std::vector UnswizzleImage(Tegra::MemoryManager& gpu_memory, GP .image_extent = size, }}; } - const auto input_data = std::make_unique(guest_size_bytes); - gpu_memory.ReadBlockUnsafe(gpu_addr, input_data.get(), guest_size_bytes); - const std::span input(input_data.get(), guest_size_bytes); - const LevelInfo level_info = MakeLevelInfo(info); const s32 num_layers = info.resources.layers; const s32 num_levels = info.resources.levels; diff --git a/src/video_core/texture_cache/util.h b/src/video_core/texture_cache/util.h index 5e28f4ab3..ddf0b3b06 100644 --- a/src/video_core/texture_cache/util.h +++ b/src/video_core/texture_cache/util.h @@ -59,6 +59,7 @@ struct OverlapResult { [[nodiscard]] std::vector UnswizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info, + std::span input, std::span output); [[nodiscard]] BufferCopy UploadBufferCopy(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, From 1209d428f14ffc6f1c8b6049b0431c30b26138d6 Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Thu, 22 Dec 2022 17:01:39 -0500 Subject: [PATCH 2/3] texture_cache: Use pre-allocated buffer for texture downloads --- src/video_core/texture_cache/texture_cache.h | 9 ++++++--- src/video_core/texture_cache/util.cpp | 11 ++++++----- src/video_core/texture_cache/util.h | 3 ++- 3 files changed, 14 insertions(+), 9 deletions(-) diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h index fccf4316d..6d7d8226f 100644 --- a/src/video_core/texture_cache/texture_cache.h +++ b/src/video_core/texture_cache/texture_cache.h @@ -96,7 +96,8 @@ void TextureCache

::RunGarbageCollector() { const auto copies = FullDownloadCopies(image.info); image.DownloadMemory(map, copies); runtime.Finish(); - SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span); + SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span, + swizzle_data_buffer); } if (True(image.flags & ImageFlagBits::Tracked)) { UntrackImage(image, image_id); @@ -467,7 +468,8 @@ void TextureCache

::DownloadMemory(VAddr cpu_addr, size_t size) { const auto copies = FullDownloadCopies(image.info); image.DownloadMemory(map, copies); runtime.Finish(); - SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span); + SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span, + swizzle_data_buffer); } } @@ -678,7 +680,8 @@ void TextureCache

::PopAsyncFlushes() { for (const ImageId image_id : download_ids) { const ImageBase& image = slot_images[image_id]; const auto copies = FullDownloadCopies(image.info); - SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span); + SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span, + swizzle_data_buffer); download_map.offset += image.unswizzled_size_bytes; download_span = download_span.subspan(image.unswizzled_size_bytes); } diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp index 4488fa9da..7999a7f06 100644 --- a/src/video_core/texture_cache/util.cpp +++ b/src/video_core/texture_cache/util.cpp @@ -505,7 +505,7 @@ void SwizzlePitchLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info, const BufferImageCopy& copy, - std::span input) { + std::span input, std::vector& tmp_buffer) { const Extent3D size = info.size; const LevelInfo level_info = MakeLevelInfo(info); const Extent2D tile_size = DefaultBlockSize(info.format); @@ -534,8 +534,8 @@ void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr tile_size.height, info.tile_width_spacing); const size_t subresource_size = sizes[level]; - const auto dst_data = std::make_unique(subresource_size); - const std::span dst(dst_data.get(), subresource_size); + tmp_buffer.resize(subresource_size); + const std::span dst(tmp_buffer); for (s32 layer = 0; layer < info.resources.layers; ++layer) { const std::span src = input.subspan(host_offset); @@ -977,13 +977,14 @@ std::vector FullUploadSwizzles(const ImageInfo& info) { } void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info, - std::span copies, std::span memory) { + std::span copies, std::span memory, + std::vector& tmp_buffer) { const bool is_pitch_linear = info.type == ImageType::Linear; for (const BufferImageCopy& copy : copies) { if (is_pitch_linear) { SwizzlePitchLinearImage(gpu_memory, gpu_addr, info, copy, memory); } else { - SwizzleBlockLinearImage(gpu_memory, gpu_addr, info, copy, memory); + SwizzleBlockLinearImage(gpu_memory, gpu_addr, info, copy, memory, tmp_buffer); } } } diff --git a/src/video_core/texture_cache/util.h b/src/video_core/texture_cache/util.h index ddf0b3b06..2c991b4d2 100644 --- a/src/video_core/texture_cache/util.h +++ b/src/video_core/texture_cache/util.h @@ -77,7 +77,8 @@ void ConvertImage(std::span input, const ImageInfo& info, std::span FullUploadSwizzles(const ImageInfo& info); void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info, - std::span copies, std::span memory); + std::span copies, std::span memory, + std::vector& tmp_buffer); [[nodiscard]] bool IsBlockLinearSizeCompatible(const ImageInfo& new_info, const ImageInfo& overlap_info, u32 new_level, From 7584d36922669808b7c8a62667380453687b0ad9 Mon Sep 17 00:00:00 2001 From: ameerj <52414509+ameerj@users.noreply.github.com> Date: Sun, 25 Dec 2022 15:47:41 -0500 Subject: [PATCH 3/3] texture_cache: Use Common::ScratchBuffer for swizzle buffers --- src/video_core/texture_cache/texture_cache.h | 8 ++++---- src/video_core/texture_cache/texture_cache_base.h | 5 +++-- src/video_core/texture_cache/util.cpp | 6 +++--- src/video_core/texture_cache/util.h | 3 ++- 4 files changed, 12 insertions(+), 10 deletions(-) diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h index 6d7d8226f..27c82cd20 100644 --- a/src/video_core/texture_cache/texture_cache.h +++ b/src/video_core/texture_cache/texture_cache.h @@ -42,8 +42,8 @@ TextureCache

::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface& // These values were chosen based on typical peak swizzle data sizes seen in some titles static constexpr size_t SWIZZLE_DATA_BUFFER_INITIAL_CAPACITY = 8_MiB; static constexpr size_t UNSWIZZLE_DATA_BUFFER_INITIAL_CAPACITY = 1_MiB; - swizzle_data_buffer.reserve(SWIZZLE_DATA_BUFFER_INITIAL_CAPACITY); - unswizzle_data_buffer.reserve(UNSWIZZLE_DATA_BUFFER_INITIAL_CAPACITY); + swizzle_data_buffer.resize_destructive(SWIZZLE_DATA_BUFFER_INITIAL_CAPACITY); + unswizzle_data_buffer.resize_destructive(UNSWIZZLE_DATA_BUFFER_INITIAL_CAPACITY); // Make sure the first index is reserved for the null resources // This way the null resource becomes a compile time constant @@ -746,11 +746,11 @@ void TextureCache

::UploadImageContents(Image& image, StagingBuffer& staging) return; } const size_t guest_size_bytes = image.guest_size_bytes; - swizzle_data_buffer.resize(guest_size_bytes); + swizzle_data_buffer.resize_destructive(guest_size_bytes); gpu_memory->ReadBlockUnsafe(gpu_addr, swizzle_data_buffer.data(), guest_size_bytes); if (True(image.flags & ImageFlagBits::Converted)) { - unswizzle_data_buffer.resize(image.unswizzled_size_bytes); + unswizzle_data_buffer.resize_destructive(image.unswizzled_size_bytes); auto copies = UnswizzleImage(*gpu_memory, gpu_addr, image.info, swizzle_data_buffer, unswizzle_data_buffer); ConvertImage(unswizzle_data_buffer, image.info, mapped_span, copies); diff --git a/src/video_core/texture_cache/texture_cache_base.h b/src/video_core/texture_cache/texture_cache_base.h index 67e8acf25..4fd677a80 100644 --- a/src/video_core/texture_cache/texture_cache_base.h +++ b/src/video_core/texture_cache/texture_cache_base.h @@ -17,6 +17,7 @@ #include "common/literals.h" #include "common/lru_cache.h" #include "common/polyfill_ranges.h" +#include "common/scratch_buffer.h" #include "video_core/compatible_formats.h" #include "video_core/control/channel_state_cache.h" #include "video_core/delayed_destruction_ring.h" @@ -417,8 +418,8 @@ private: std::unordered_map image_allocs_table; - std::vector swizzle_data_buffer; - std::vector unswizzle_data_buffer; + Common::ScratchBuffer swizzle_data_buffer; + Common::ScratchBuffer unswizzle_data_buffer; u64 modification_tick = 0; u64 frame_tick = 0; diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp index 7999a7f06..03acc68d9 100644 --- a/src/video_core/texture_cache/util.cpp +++ b/src/video_core/texture_cache/util.cpp @@ -505,7 +505,7 @@ void SwizzlePitchLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info, const BufferImageCopy& copy, - std::span input, std::vector& tmp_buffer) { + std::span input, Common::ScratchBuffer& tmp_buffer) { const Extent3D size = info.size; const LevelInfo level_info = MakeLevelInfo(info); const Extent2D tile_size = DefaultBlockSize(info.format); @@ -534,7 +534,7 @@ void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr tile_size.height, info.tile_width_spacing); const size_t subresource_size = sizes[level]; - tmp_buffer.resize(subresource_size); + tmp_buffer.resize_destructive(subresource_size); const std::span dst(tmp_buffer); for (s32 layer = 0; layer < info.resources.layers; ++layer) { @@ -978,7 +978,7 @@ std::vector FullUploadSwizzles(const ImageInfo& info) { void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const ImageInfo& info, std::span copies, std::span memory, - std::vector& tmp_buffer) { + Common::ScratchBuffer& tmp_buffer) { const bool is_pitch_linear = info.type == ImageType::Linear; for (const BufferImageCopy& copy : copies) { if (is_pitch_linear) { diff --git a/src/video_core/texture_cache/util.h b/src/video_core/texture_cache/util.h index 2c991b4d2..d103db8ae 100644 --- a/src/video_core/texture_cache/util.h +++ b/src/video_core/texture_cache/util.h @@ -7,6 +7,7 @@ #include #include "common/common_types.h" +#include "common/scratch_buffer.h" #include "video_core/surface.h" #include "video_core/texture_cache/image_base.h" @@ -78,7 +79,7 @@ void ConvertImage(std::span input, const ImageInfo& info, std::span copies, std::span memory, - std::vector& tmp_buffer); + Common::ScratchBuffer& tmp_buffer); [[nodiscard]] bool IsBlockLinearSizeCompatible(const ImageInfo& new_info, const ImageInfo& overlap_info, u32 new_level,