From 9764c13d6d2977903f407761b27d847c0056e1c4 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Wed, 30 Dec 2020 02:25:23 -0300 Subject: video_core: Rewrite the texture cache The current texture cache has several points that hurt maintainability and performance. It's easy to break unrelated parts of the cache when doing minor changes. The cache can easily forget valuable information about the cached textures by CPU writes or simply by its normal usage.The current texture cache has several points that hurt maintainability and performance. It's easy to break unrelated parts of the cache when doing minor changes. The cache can easily forget valuable information about the cached textures by CPU writes or simply by its normal usage. This commit aims to address those issues. --- src/video_core/compatible_formats.h | 23 ++--------------------- 1 file changed, 2 insertions(+), 21 deletions(-) (limited to 'src/video_core/compatible_formats.h') diff --git a/src/video_core/compatible_formats.h b/src/video_core/compatible_formats.h index 51766349b..b5eb03bea 100644 --- a/src/video_core/compatible_formats.h +++ b/src/video_core/compatible_formats.h @@ -4,31 +4,12 @@ #pragma once -#include -#include -#include - #include "video_core/surface.h" namespace VideoCore::Surface { -class FormatCompatibility { -public: - using Table = std::array, MaxPixelFormat>; - - explicit FormatCompatibility(); - - bool TestView(PixelFormat format_a, PixelFormat format_b) const noexcept { - return view[static_cast(format_a)][static_cast(format_b)]; - } - - bool TestCopy(PixelFormat format_a, PixelFormat format_b) const noexcept { - return copy[static_cast(format_a)][static_cast(format_b)]; - } +bool IsViewCompatible(PixelFormat format_a, PixelFormat format_b); -private: - Table view; - Table copy; -}; +bool IsCopyCompatible(PixelFormat format_a, PixelFormat format_b); } // namespace VideoCore::Surface -- cgit v1.2.3 From 7d904fef2e6ac9ce8a3df71e758a36d39b8f69e5 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Mon, 4 Jan 2021 01:56:44 -0300 Subject: gl_texture_cache: Avoid format views on Intel and AMD Intel and AMD proprietary drivers are incapable of rendering to texture views of different formats than the original texture. Avoid creating these at a cache level. This will consume more memory, emulating them with copies. --- src/video_core/compatible_formats.cpp | 9 +++++---- src/video_core/compatible_formats.h | 2 +- src/video_core/renderer_opengl/gl_device.cpp | 4 ++++ src/video_core/renderer_opengl/gl_device.h | 5 +++++ src/video_core/renderer_opengl/gl_texture_cache.cpp | 2 ++ src/video_core/renderer_vulkan/vk_texture_cache.h | 5 +++++ src/video_core/texture_cache/image_base.cpp | 4 +++- src/video_core/texture_cache/image_view_base.cpp | 2 +- src/video_core/texture_cache/texture_cache.h | 15 ++++++++++----- src/video_core/texture_cache/util.cpp | 13 +++++++------ src/video_core/texture_cache/util.h | 8 +++++--- 11 files changed, 48 insertions(+), 21 deletions(-) (limited to 'src/video_core/compatible_formats.h') diff --git a/src/video_core/compatible_formats.cpp b/src/video_core/compatible_formats.cpp index 1619d8664..acf2668dc 100644 --- a/src/video_core/compatible_formats.cpp +++ b/src/video_core/compatible_formats.cpp @@ -10,9 +10,7 @@ #include "video_core/surface.h" namespace VideoCore::Surface { - namespace { - using Table = std::array, MaxPixelFormat>; // Compatibility table taken from Table 3.X.2 in: @@ -233,10 +231,13 @@ constexpr Table MakeCopyTable() { EnableRange(copy, COPY_CLASS_64_BITS); return copy; } - } // Anonymous namespace -bool IsViewCompatible(PixelFormat format_a, PixelFormat format_b) { +bool IsViewCompatible(PixelFormat format_a, PixelFormat format_b, bool broken_views) { + if (broken_views) { + // If format views are broken, only accept formats that are identical. + return format_a == format_b; + } static constexpr Table TABLE = MakeViewTable(); return IsSupported(TABLE, format_a, format_b); } diff --git a/src/video_core/compatible_formats.h b/src/video_core/compatible_formats.h index b5eb03bea..9a0522988 100644 --- a/src/video_core/compatible_formats.h +++ b/src/video_core/compatible_formats.h @@ -8,7 +8,7 @@ namespace VideoCore::Surface { -bool IsViewCompatible(PixelFormat format_a, PixelFormat format_b); +bool IsViewCompatible(PixelFormat format_a, PixelFormat format_b, bool broken_views); bool IsCopyCompatible(PixelFormat format_a, PixelFormat format_b); diff --git a/src/video_core/renderer_opengl/gl_device.cpp b/src/video_core/renderer_opengl/gl_device.cpp index b24179d59..81b71edfb 100644 --- a/src/video_core/renderer_opengl/gl_device.cpp +++ b/src/video_core/renderer_opengl/gl_device.cpp @@ -208,6 +208,7 @@ Device::Device() const bool is_nvidia = vendor == "NVIDIA Corporation"; const bool is_amd = vendor == "ATI Technologies Inc."; + const bool is_intel = vendor == "Intel"; bool disable_fast_buffer_sub_data = false; if (is_nvidia && version == "4.6.0 NVIDIA 443.24") { @@ -231,6 +232,7 @@ Device::Device() has_variable_aoffi = TestVariableAoffi(); has_component_indexing_bug = is_amd; has_precise_bug = TestPreciseBug(); + has_broken_texture_view_formats = is_amd || is_intel; has_nv_viewport_array2 = GLAD_GL_NV_viewport_array2; has_vertex_buffer_unified_memory = GLAD_GL_NV_vertex_buffer_unified_memory; has_debugging_tool_attached = IsDebugToolAttached(extensions); @@ -248,6 +250,8 @@ Device::Device() LOG_INFO(Render_OpenGL, "Renderer_VariableAOFFI: {}", has_variable_aoffi); LOG_INFO(Render_OpenGL, "Renderer_ComponentIndexingBug: {}", has_component_indexing_bug); LOG_INFO(Render_OpenGL, "Renderer_PreciseBug: {}", has_precise_bug); + LOG_INFO(Render_OpenGL, "Renderer_BrokenTextureViewFormats: {}", + has_broken_texture_view_formats); if (Settings::values.use_assembly_shaders.GetValue() && !use_assembly_shaders) { LOG_ERROR(Render_OpenGL, "Assembly shaders enabled but not supported"); diff --git a/src/video_core/renderer_opengl/gl_device.h b/src/video_core/renderer_opengl/gl_device.h index 13e66846c..3e79d1e37 100644 --- a/src/video_core/renderer_opengl/gl_device.h +++ b/src/video_core/renderer_opengl/gl_device.h @@ -96,6 +96,10 @@ public: return has_precise_bug; } + bool HasBrokenTextureViewFormats() const { + return has_broken_texture_view_formats; + } + bool HasFastBufferSubData() const { return has_fast_buffer_sub_data; } @@ -137,6 +141,7 @@ private: bool has_variable_aoffi{}; bool has_component_indexing_bug{}; bool has_precise_bug{}; + bool has_broken_texture_view_formats{}; bool has_fast_buffer_sub_data{}; bool has_nv_viewport_array2{}; bool has_debugging_tool_attached{}; diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp index 710874311..546cb6d00 100644 --- a/src/video_core/renderer_opengl/gl_texture_cache.cpp +++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp @@ -430,6 +430,8 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, ProgramManager& format_properties[i].emplace(format, properties); } } + has_broken_texture_view_formats = device.HasBrokenTextureViewFormats(); + null_image_1d_array.Create(GL_TEXTURE_1D_ARRAY); null_image_cube_array.Create(GL_TEXTURE_CUBE_MAP_ARRAY); null_image_3d.Create(GL_TEXTURE_3D); diff --git a/src/video_core/renderer_vulkan/vk_texture_cache.h b/src/video_core/renderer_vulkan/vk_texture_cache.h index 576515bcc..e5599de5e 100644 --- a/src/video_core/renderer_vulkan/vk_texture_cache.h +++ b/src/video_core/renderer_vulkan/vk_texture_cache.h @@ -104,6 +104,11 @@ struct TextureCacheRuntime { } void InsertUploadMemoryBarrier() {} + + bool HasBrokenTextureViewFormats() const noexcept { + // No known Vulkan driver has broken image views + return false; + } }; class Image : public VideoCommon::ImageBase { diff --git a/src/video_core/texture_cache/image_base.cpp b/src/video_core/texture_cache/image_base.cpp index 448a05fcc..959b3f115 100644 --- a/src/video_core/texture_cache/image_base.cpp +++ b/src/video_core/texture_cache/image_base.cpp @@ -120,7 +120,9 @@ void AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_i if (lhs.info.type == ImageType::Linear) { base = SubresourceBase{.level = 0, .layer = 0}; } else { - base = FindSubresource(rhs.info, lhs, rhs.gpu_addr, OPTIONS); + // We are passing relaxed formats as an option, having broken views or not won't matter + static constexpr bool broken_views = false; + base = FindSubresource(rhs.info, lhs, rhs.gpu_addr, OPTIONS, broken_views); } if (!base) { LOG_ERROR(HW_GPU, "Image alias should have been flipped"); diff --git a/src/video_core/texture_cache/image_view_base.cpp b/src/video_core/texture_cache/image_view_base.cpp index 076a4bcfd..18f72e508 100644 --- a/src/video_core/texture_cache/image_view_base.cpp +++ b/src/video_core/texture_cache/image_view_base.cpp @@ -24,7 +24,7 @@ ImageViewBase::ImageViewBase(const ImageViewInfo& info, const ImageInfo& image_i .height = std::max(image_info.size.height >> range.base.level, 1u), .depth = std::max(image_info.size.depth >> range.base.level, 1u), } { - ASSERT_MSG(VideoCore::Surface::IsViewCompatible(image_info.format, info.format), + ASSERT_MSG(VideoCore::Surface::IsViewCompatible(image_info.format, info.format, false), "Image view format {} is incompatible with image format {}", info.format, image_info.format); const bool is_async = Settings::values.use_asynchronous_gpu_emulation.GetValue(); diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h index 968059842..ad86c50b4 100644 --- a/src/video_core/texture_cache/texture_cache.h +++ b/src/video_core/texture_cache/texture_cache.h @@ -883,6 +883,7 @@ ImageId TextureCache

::FindImage(const ImageInfo& info, GPUVAddr gpu_addr, if (!cpu_addr) { return ImageId{}; } + const bool broken_views = runtime.HasBrokenTextureViewFormats(); ImageId image_id; const auto lambda = [&](ImageId existing_image_id, ImageBase& existing_image) { if (info.type == ImageType::Linear || existing_image.info.type == ImageType::Linear) { @@ -892,11 +893,11 @@ ImageId TextureCache

::FindImage(const ImageInfo& info, GPUVAddr gpu_addr, if (existing_image.gpu_addr == gpu_addr && existing.type == info.type && existing.pitch == info.pitch && IsPitchLinearSameSize(existing, info, strict_size) && - IsViewCompatible(existing.format, info.format)) { + IsViewCompatible(existing.format, info.format, broken_views)) { image_id = existing_image_id; return true; } - } else if (IsSubresource(info, existing_image, gpu_addr, options)) { + } else if (IsSubresource(info, existing_image, gpu_addr, options, broken_views)) { image_id = existing_image_id; return true; } @@ -926,6 +927,7 @@ template ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VAddr cpu_addr) { ImageInfo new_info = info; const size_t size_bytes = CalculateGuestSizeInBytes(new_info); + const bool broken_views = runtime.HasBrokenTextureViewFormats(); std::vector overlap_ids; std::vector left_aliased_ids; std::vector right_aliased_ids; @@ -940,7 +942,9 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA } return; } - const auto solution = ResolveOverlap(new_info, gpu_addr, cpu_addr, overlap, true); + static constexpr bool strict_size = true; + const std::optional solution = + ResolveOverlap(new_info, gpu_addr, cpu_addr, overlap, strict_size, broken_views); if (solution) { gpu_addr = solution->gpu_addr; cpu_addr = solution->cpu_addr; @@ -950,9 +954,10 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA } static constexpr auto options = RelaxedOptions::Size | RelaxedOptions::Format; const ImageBase new_image_base(new_info, gpu_addr, cpu_addr); - if (IsSubresource(new_info, overlap, gpu_addr, options)) { + if (IsSubresource(new_info, overlap, gpu_addr, options, broken_views)) { left_aliased_ids.push_back(overlap_id); - } else if (IsSubresource(overlap.info, new_image_base, overlap.gpu_addr, options)) { + } else if (IsSubresource(overlap.info, new_image_base, overlap.gpu_addr, options, + broken_views)) { right_aliased_ids.push_back(overlap_id); } }); diff --git a/src/video_core/texture_cache/util.cpp b/src/video_core/texture_cache/util.cpp index 9ed1fc007..279932778 100644 --- a/src/video_core/texture_cache/util.cpp +++ b/src/video_core/texture_cache/util.cpp @@ -1069,13 +1069,13 @@ bool IsPitchLinearSameSize(const ImageInfo& lhs, const ImageInfo& rhs, bool stri std::optional ResolveOverlap(const ImageInfo& new_info, GPUVAddr gpu_addr, VAddr cpu_addr, const ImageBase& overlap, - bool strict_size) { + bool strict_size, bool broken_views) { ASSERT(new_info.type != ImageType::Linear); ASSERT(overlap.info.type != ImageType::Linear); if (!IsLayerStrideCompatible(new_info, overlap.info)) { return std::nullopt; } - if (!IsViewCompatible(overlap.info.format, new_info.format)) { + if (!IsViewCompatible(overlap.info.format, new_info.format, broken_views)) { return std::nullopt; } if (gpu_addr == overlap.gpu_addr) { @@ -1118,14 +1118,15 @@ bool IsLayerStrideCompatible(const ImageInfo& lhs, const ImageInfo& rhs) { } std::optional FindSubresource(const ImageInfo& candidate, const ImageBase& image, - GPUVAddr candidate_addr, RelaxedOptions options) { + GPUVAddr candidate_addr, RelaxedOptions options, + bool broken_views) { const std::optional base = image.TryFindBase(candidate_addr); if (!base) { return std::nullopt; } const ImageInfo& existing = image.info; if (False(options & RelaxedOptions::Format)) { - if (!IsViewCompatible(existing.format, candidate.format)) { + if (!IsViewCompatible(existing.format, candidate.format, broken_views)) { return std::nullopt; } } @@ -1162,8 +1163,8 @@ std::optional FindSubresource(const ImageInfo& candidate, const } bool IsSubresource(const ImageInfo& candidate, const ImageBase& image, GPUVAddr candidate_addr, - RelaxedOptions options) { - return FindSubresource(candidate, image, candidate_addr, options).has_value(); + RelaxedOptions options, bool broken_views) { + return FindSubresource(candidate, image, candidate_addr, options, broken_views).has_value(); } void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst, diff --git a/src/video_core/texture_cache/util.h b/src/video_core/texture_cache/util.h index dbbbd33cd..52a9207d6 100644 --- a/src/video_core/texture_cache/util.h +++ b/src/video_core/texture_cache/util.h @@ -87,17 +87,19 @@ void SwizzleImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr, const Ima [[nodiscard]] std::optional ResolveOverlap(const ImageInfo& new_info, GPUVAddr gpu_addr, VAddr cpu_addr, const ImageBase& overlap, - bool strict_size); + bool strict_size, bool broken_views); [[nodiscard]] bool IsLayerStrideCompatible(const ImageInfo& lhs, const ImageInfo& rhs); [[nodiscard]] std::optional FindSubresource(const ImageInfo& candidate, const ImageBase& image, GPUVAddr candidate_addr, - RelaxedOptions options); + RelaxedOptions options, + bool broken_views); [[nodiscard]] bool IsSubresource(const ImageInfo& candidate, const ImageBase& image, - GPUVAddr candidate_addr, RelaxedOptions options); + GPUVAddr candidate_addr, RelaxedOptions options, + bool broken_views); void DeduceBlitImages(ImageInfo& dst_info, ImageInfo& src_info, const ImageBase* dst, const ImageBase* src); -- cgit v1.2.3