From 6ee316cb8f4d6f28b4dcad7a8a3d447fdf70d3bb Mon Sep 17 00:00:00 2001 From: Fernando Sahmkow Date: Sun, 5 Apr 2020 15:26:16 -0400 Subject: [PATCH] Texture Cache: Use vAddr instead of physical memory for caching. --- .../renderer_opengl/gl_rasterizer.cpp | 7 +- .../renderer_vulkan/vk_rasterizer.cpp | 7 +- src/video_core/texture_cache/surface_base.cpp | 42 ++---- src/video_core/texture_cache/surface_base.h | 34 ++--- src/video_core/texture_cache/texture_cache.h | 121 +++++++++--------- 5 files changed, 81 insertions(+), 130 deletions(-) diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp index 2617388307..a25e7486d1 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer.cpp +++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp @@ -662,7 +662,7 @@ void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) { return; } CacheAddr cache_addr = ToCacheAddr(system.Memory().GetPointer(addr)); - texture_cache.FlushRegion(cache_addr, size); + texture_cache.FlushRegion(addr, size); buffer_cache.FlushRegion(cache_addr, size); query_cache.FlushRegion(cache_addr, size); } @@ -673,7 +673,7 @@ void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) { return; } CacheAddr cache_addr = ToCacheAddr(system.Memory().GetPointer(addr)); - texture_cache.InvalidateRegion(cache_addr, size); + texture_cache.InvalidateRegion(addr, size); shader_cache.InvalidateRegion(cache_addr, size); buffer_cache.InvalidateRegion(cache_addr, size); query_cache.InvalidateRegion(cache_addr, size); @@ -718,8 +718,7 @@ bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& config, MICROPROFILE_SCOPE(OpenGL_CacheManagement); - const auto surface{ - texture_cache.TryFindFramebufferSurface(system.Memory().GetPointer(framebuffer_addr))}; + const auto surface{texture_cache.TryFindFramebufferSurface(framebuffer_addr)}; if (!surface) { return {}; } diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index 0ca72eb451..cc76d96ea0 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -500,7 +500,7 @@ void RasterizerVulkan::FlushRegion(VAddr addr, u64 size) { return; } CacheAddr cache_addr = ToCacheAddr(system.Memory().GetPointer(addr)); - texture_cache.FlushRegion(cache_addr, size); + texture_cache.FlushRegion(addr, size); buffer_cache.FlushRegion(cache_addr, size); query_cache.FlushRegion(cache_addr, size); } @@ -510,7 +510,7 @@ void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size) { return; } CacheAddr cache_addr = ToCacheAddr(system.Memory().GetPointer(addr)); - texture_cache.InvalidateRegion(cache_addr, size); + texture_cache.InvalidateRegion(addr, size); pipeline_cache.InvalidateRegion(cache_addr, size); buffer_cache.InvalidateRegion(cache_addr, size); query_cache.InvalidateRegion(cache_addr, size); @@ -548,8 +548,7 @@ bool RasterizerVulkan::AccelerateDisplay(const Tegra::FramebufferConfig& config, return false; } - const u8* host_ptr{system.Memory().GetPointer(framebuffer_addr)}; - const auto surface{texture_cache.TryFindFramebufferSurface(host_ptr)}; + const auto surface{texture_cache.TryFindFramebufferSurface(framebuffer_addr)}; if (!surface) { return false; } diff --git a/src/video_core/texture_cache/surface_base.cpp b/src/video_core/texture_cache/surface_base.cpp index 6fe8151354..7af0e792c4 100644 --- a/src/video_core/texture_cache/surface_base.cpp +++ b/src/video_core/texture_cache/surface_base.cpp @@ -190,22 +190,11 @@ void SurfaceBaseImpl::LoadBuffer(Tegra::MemoryManager& memory_manager, MICROPROFILE_SCOPE(GPU_Load_Texture); auto& staging_buffer = staging_cache.GetBuffer(0); u8* host_ptr; - is_continuous = memory_manager.IsBlockContinuous(gpu_addr, guest_memory_size); - - // Handle continuouty - if (is_continuous) { - // Use physical memory directly - host_ptr = memory_manager.GetPointer(gpu_addr); - if (!host_ptr) { - return; - } - } else { - // Use an extra temporal buffer - auto& tmp_buffer = staging_cache.GetBuffer(1); - tmp_buffer.resize(guest_memory_size); - host_ptr = tmp_buffer.data(); - memory_manager.ReadBlockUnsafe(gpu_addr, host_ptr, guest_memory_size); - } + // Use an extra temporal buffer + auto& tmp_buffer = staging_cache.GetBuffer(1); + tmp_buffer.resize(guest_memory_size); + host_ptr = tmp_buffer.data(); + memory_manager.ReadBlockUnsafe(gpu_addr, host_ptr, guest_memory_size); if (params.is_tiled) { ASSERT_MSG(params.block_width == 0, "Block width is defined as {} on texture target {}", @@ -257,19 +246,10 @@ void SurfaceBaseImpl::FlushBuffer(Tegra::MemoryManager& memory_manager, auto& staging_buffer = staging_cache.GetBuffer(0); u8* host_ptr; - // Handle continuouty - if (is_continuous) { - // Use physical memory directly - host_ptr = memory_manager.GetPointer(gpu_addr); - if (!host_ptr) { - return; - } - } else { - // Use an extra temporal buffer - auto& tmp_buffer = staging_cache.GetBuffer(1); - tmp_buffer.resize(guest_memory_size); - host_ptr = tmp_buffer.data(); - } + // Use an extra temporal buffer + auto& tmp_buffer = staging_cache.GetBuffer(1); + tmp_buffer.resize(guest_memory_size); + host_ptr = tmp_buffer.data(); if (params.is_tiled) { ASSERT_MSG(params.block_width == 0, "Block width is defined as {}", params.block_width); @@ -300,9 +280,7 @@ void SurfaceBaseImpl::FlushBuffer(Tegra::MemoryManager& memory_manager, } } } - if (!is_continuous) { - memory_manager.WriteBlockUnsafe(gpu_addr, host_ptr, guest_memory_size); - } + memory_manager.WriteBlockUnsafe(gpu_addr, host_ptr, guest_memory_size); } } // namespace VideoCommon diff --git a/src/video_core/texture_cache/surface_base.h b/src/video_core/texture_cache/surface_base.h index d7882a0318..a39a8661bc 100644 --- a/src/video_core/texture_cache/surface_base.h +++ b/src/video_core/texture_cache/surface_base.h @@ -68,8 +68,8 @@ public: return gpu_addr; } - bool Overlaps(const CacheAddr start, const CacheAddr end) const { - return (cache_addr < end) && (cache_addr_end > start); + bool Overlaps(const VAddr start, const VAddr end) const { + return (cpu_addr < end) && (cpu_addr_end > start); } bool IsInside(const GPUVAddr other_start, const GPUVAddr other_end) { @@ -86,21 +86,13 @@ public: return cpu_addr; } + VAddr GetCpuAddrEnd() const { + return cpu_addr_end; + } + void SetCpuAddr(const VAddr new_addr) { cpu_addr = new_addr; - } - - CacheAddr GetCacheAddr() const { - return cache_addr; - } - - CacheAddr GetCacheAddrEnd() const { - return cache_addr_end; - } - - void SetCacheAddr(const CacheAddr new_addr) { - cache_addr = new_addr; - cache_addr_end = new_addr + guest_memory_size; + cpu_addr_end = new_addr + guest_memory_size; } const SurfaceParams& GetSurfaceParams() const { @@ -119,14 +111,6 @@ public: return mipmap_sizes[level]; } - void MarkAsContinuous(const bool is_continuous) { - this->is_continuous = is_continuous; - } - - bool IsContinuous() const { - return is_continuous; - } - bool IsLinear() const { return !params.is_tiled; } @@ -175,10 +159,8 @@ protected: std::size_t guest_memory_size; std::size_t host_memory_size; GPUVAddr gpu_addr{}; - CacheAddr cache_addr{}; - CacheAddr cache_addr_end{}; VAddr cpu_addr{}; - bool is_continuous{}; + VAddr cpu_addr_end{}; bool is_converted{}; std::vector mipmap_sizes; diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h index c8f8d659d8..88fe3e25fe 100644 --- a/src/video_core/texture_cache/texture_cache.h +++ b/src/video_core/texture_cache/texture_cache.h @@ -52,11 +52,9 @@ using RenderTargetConfig = Tegra::Engines::Maxwell3D::Regs::RenderTargetConfig; template class TextureCache { - using IntervalMap = boost::icl::interval_map>; - using IntervalType = typename IntervalMap::interval_type; public: - void InvalidateRegion(CacheAddr addr, std::size_t size) { + void InvalidateRegion(VAddr addr, std::size_t size) { std::lock_guard lock{mutex}; for (const auto& surface : GetSurfacesInRegion(addr, size)) { @@ -76,7 +74,7 @@ public: guard_samplers = new_guard; } - void FlushRegion(CacheAddr addr, std::size_t size) { + void FlushRegion(VAddr addr, std::size_t size) { std::lock_guard lock{mutex}; auto surfaces = GetSurfacesInRegion(addr, size); @@ -99,9 +97,9 @@ public: return GetNullSurface(SurfaceParams::ExpectedTarget(entry)); } - const auto host_ptr{system.GPU().MemoryManager().GetPointer(gpu_addr)}; - const auto cache_addr{ToCacheAddr(host_ptr)}; - if (!cache_addr) { + const std::optional cpu_addr = + system.GPU().MemoryManager().GpuToCpuAddress(gpu_addr); + if (!cpu_addr) { return GetNullSurface(SurfaceParams::ExpectedTarget(entry)); } @@ -110,7 +108,7 @@ public: } const auto params{SurfaceParams::CreateForTexture(format_lookup_table, tic, entry)}; - const auto [surface, view] = GetSurface(gpu_addr, cache_addr, params, true, false); + const auto [surface, view] = GetSurface(gpu_addr, *cpu_addr, params, true, false); if (guard_samplers) { sampled_textures.push_back(surface); } @@ -124,13 +122,13 @@ public: if (!gpu_addr) { return GetNullSurface(SurfaceParams::ExpectedTarget(entry)); } - const auto host_ptr{system.GPU().MemoryManager().GetPointer(gpu_addr)}; - const auto cache_addr{ToCacheAddr(host_ptr)}; - if (!cache_addr) { + const std::optional cpu_addr = + system.GPU().MemoryManager().GpuToCpuAddress(gpu_addr); + if (!cpu_addr) { return GetNullSurface(SurfaceParams::ExpectedTarget(entry)); } const auto params{SurfaceParams::CreateForImage(format_lookup_table, tic, entry)}; - const auto [surface, view] = GetSurface(gpu_addr, cache_addr, params, true, false); + const auto [surface, view] = GetSurface(gpu_addr, *cpu_addr, params, true, false); if (guard_samplers) { sampled_textures.push_back(surface); } @@ -159,14 +157,14 @@ public: SetEmptyDepthBuffer(); return {}; } - const auto host_ptr{system.GPU().MemoryManager().GetPointer(gpu_addr)}; - const auto cache_addr{ToCacheAddr(host_ptr)}; - if (!cache_addr) { + const std::optional cpu_addr = + system.GPU().MemoryManager().GpuToCpuAddress(gpu_addr); + if (!cpu_addr) { SetEmptyDepthBuffer(); return {}; } const auto depth_params{SurfaceParams::CreateForDepthBuffer(system)}; - auto surface_view = GetSurface(gpu_addr, cache_addr, depth_params, preserve_contents, true); + auto surface_view = GetSurface(gpu_addr, *cpu_addr, depth_params, preserve_contents, true); if (depth_buffer.target) depth_buffer.target->MarkAsRenderTarget(false, NO_RT); depth_buffer.target = surface_view.first; @@ -199,15 +197,15 @@ public: return {}; } - const auto host_ptr{system.GPU().MemoryManager().GetPointer(gpu_addr)}; - const auto cache_addr{ToCacheAddr(host_ptr)}; - if (!cache_addr) { + const std::optional cpu_addr = + system.GPU().MemoryManager().GpuToCpuAddress(gpu_addr); + if (!cpu_addr) { SetEmptyColorBuffer(index); return {}; } auto surface_view = - GetSurface(gpu_addr, cache_addr, SurfaceParams::CreateForFramebuffer(system, index), + GetSurface(gpu_addr, *cpu_addr, SurfaceParams::CreateForFramebuffer(system, index), preserve_contents, true); if (render_targets[index].target) render_targets[index].target->MarkAsRenderTarget(false, NO_RT); @@ -257,27 +255,26 @@ public: const GPUVAddr src_gpu_addr = src_config.Address(); const GPUVAddr dst_gpu_addr = dst_config.Address(); DeduceBestBlit(src_params, dst_params, src_gpu_addr, dst_gpu_addr); - const auto dst_host_ptr{system.GPU().MemoryManager().GetPointer(dst_gpu_addr)}; - const auto dst_cache_addr{ToCacheAddr(dst_host_ptr)}; - const auto src_host_ptr{system.GPU().MemoryManager().GetPointer(src_gpu_addr)}; - const auto src_cache_addr{ToCacheAddr(src_host_ptr)}; + const std::optional dst_cpu_addr = + system.GPU().MemoryManager().GpuToCpuAddress(dst_gpu_addr); + const std::optional src_cpu_addr = + system.GPU().MemoryManager().GpuToCpuAddress(src_gpu_addr); std::pair dst_surface = - GetSurface(dst_gpu_addr, dst_cache_addr, dst_params, true, false); + GetSurface(dst_gpu_addr, *dst_cpu_addr, dst_params, true, false); std::pair src_surface = - GetSurface(src_gpu_addr, src_cache_addr, src_params, true, false); + GetSurface(src_gpu_addr, *src_cpu_addr, src_params, true, false); ImageBlit(src_surface.second, dst_surface.second, copy_config); dst_surface.first->MarkAsModified(true, Tick()); } - TSurface TryFindFramebufferSurface(const u8* host_ptr) { - const CacheAddr cache_addr = ToCacheAddr(host_ptr); - if (!cache_addr) { + TSurface TryFindFramebufferSurface(VAddr addr) { + if (!addr) { return nullptr; } - const CacheAddr page = cache_addr >> registry_page_bits; + const VAddr page = addr >> registry_page_bits; std::vector& list = registry[page]; for (auto& surface : list) { - if (surface->GetCacheAddr() == cache_addr) { + if (surface->GetCpuAddr() == addr) { return surface; } } @@ -338,18 +335,14 @@ protected: void Register(TSurface surface) { const GPUVAddr gpu_addr = surface->GetGpuAddr(); - const CacheAddr cache_ptr = ToCacheAddr(system.GPU().MemoryManager().GetPointer(gpu_addr)); const std::size_t size = surface->GetSizeInBytes(); const std::optional cpu_addr = system.GPU().MemoryManager().GpuToCpuAddress(gpu_addr); - if (!cache_ptr || !cpu_addr) { + if (!cpu_addr) { LOG_CRITICAL(HW_GPU, "Failed to register surface with unmapped gpu_address 0x{:016x}", gpu_addr); return; } - const bool continuous = system.GPU().MemoryManager().IsBlockContinuous(gpu_addr, size); - surface->MarkAsContinuous(continuous); - surface->SetCacheAddr(cache_ptr); surface->SetCpuAddr(*cpu_addr); RegisterInnerCache(surface); surface->MarkAsRegistered(true); @@ -634,7 +627,7 @@ private: std::optional> Manage3DSurfaces(std::vector& overlaps, const SurfaceParams& params, const GPUVAddr gpu_addr, - const CacheAddr cache_addr, + const VAddr cpu_addr, bool preserve_contents) { if (params.target == SurfaceTarget::Texture3D) { bool failed = false; @@ -659,7 +652,7 @@ private: failed = true; break; } - const u32 offset = static_cast(surface->GetCacheAddr() - cache_addr); + const u32 offset = static_cast(surface->GetCpuAddr() - cpu_addr); const auto [x, y, z] = params.GetBlockOffsetXYZ(offset); modified |= surface->IsModified(); const CopyParams copy_params(0, 0, 0, 0, 0, z, 0, 0, params.width, params.height, @@ -679,7 +672,7 @@ private: } else { for (const auto& surface : overlaps) { if (!surface->MatchTarget(params.target)) { - if (overlaps.size() == 1 && surface->GetCacheAddr() == cache_addr) { + if (overlaps.size() == 1 && surface->GetCpuAddr() == cpu_addr) { if (Settings::values.use_accurate_gpu_emulation) { return std::nullopt; } @@ -688,7 +681,7 @@ private: } return std::nullopt; } - if (surface->GetCacheAddr() != cache_addr) { + if (surface->GetCpuAddr() != cpu_addr) { continue; } if (surface->MatchesStructure(params) == MatchStructureResult::FullMatch) { @@ -722,13 +715,13 @@ private: * left blank. * @param is_render Whether or not the surface is a render target. **/ - std::pair GetSurface(const GPUVAddr gpu_addr, const CacheAddr cache_addr, + std::pair GetSurface(const GPUVAddr gpu_addr, const VAddr cpu_addr, const SurfaceParams& params, bool preserve_contents, bool is_render) { // Step 1 // Check Level 1 Cache for a fast structural match. If candidate surface // matches at certain level we are pretty much done. - if (const auto iter = l1_cache.find(cache_addr); iter != l1_cache.end()) { + if (const auto iter = l1_cache.find(cpu_addr); iter != l1_cache.end()) { TSurface& current_surface = iter->second; const auto topological_result = current_surface->MatchesTopology(params); if (topological_result != MatchTopologyResult::FullMatch) { @@ -755,7 +748,7 @@ private: // Step 2 // Obtain all possible overlaps in the memory region const std::size_t candidate_size = params.GetGuestSizeInBytes(); - auto overlaps{GetSurfacesInRegion(cache_addr, candidate_size)}; + auto overlaps{GetSurfacesInRegion(cpu_addr, candidate_size)}; // If none are found, we are done. we just load the surface and create it. if (overlaps.empty()) { @@ -777,7 +770,7 @@ private: // Check if it's a 3D texture if (params.block_depth > 0) { auto surface = - Manage3DSurfaces(overlaps, params, gpu_addr, cache_addr, preserve_contents); + Manage3DSurfaces(overlaps, params, gpu_addr, cpu_addr, preserve_contents); if (surface) { return *surface; } @@ -852,16 +845,16 @@ private: * @param params The parameters on the candidate surface. **/ Deduction DeduceSurface(const GPUVAddr gpu_addr, const SurfaceParams& params) { - const auto host_ptr{system.GPU().MemoryManager().GetPointer(gpu_addr)}; - const auto cache_addr{ToCacheAddr(host_ptr)}; + const std::optional cpu_addr = + system.GPU().MemoryManager().GpuToCpuAddress(gpu_addr); - if (!cache_addr) { + if (!cpu_addr) { Deduction result{}; result.type = DeductionType::DeductionFailed; return result; } - if (const auto iter = l1_cache.find(cache_addr); iter != l1_cache.end()) { + if (const auto iter = l1_cache.find(*cpu_addr); iter != l1_cache.end()) { TSurface& current_surface = iter->second; const auto topological_result = current_surface->MatchesTopology(params); if (topological_result != MatchTopologyResult::FullMatch) { @@ -880,7 +873,7 @@ private: } const std::size_t candidate_size = params.GetGuestSizeInBytes(); - auto overlaps{GetSurfacesInRegion(cache_addr, candidate_size)}; + auto overlaps{GetSurfacesInRegion(*cpu_addr, candidate_size)}; if (overlaps.empty()) { Deduction result{}; @@ -1024,10 +1017,10 @@ private: } void RegisterInnerCache(TSurface& surface) { - const CacheAddr cache_addr = surface->GetCacheAddr(); - CacheAddr start = cache_addr >> registry_page_bits; - const CacheAddr end = (surface->GetCacheAddrEnd() - 1) >> registry_page_bits; - l1_cache[cache_addr] = surface; + const VAddr cpu_addr = surface->GetCpuAddr(); + VAddr start = cpu_addr >> registry_page_bits; + const VAddr end = (surface->GetCpuAddrEnd() - 1) >> registry_page_bits; + l1_cache[cpu_addr] = surface; while (start <= end) { registry[start].push_back(surface); start++; @@ -1035,10 +1028,10 @@ private: } void UnregisterInnerCache(TSurface& surface) { - const CacheAddr cache_addr = surface->GetCacheAddr(); - CacheAddr start = cache_addr >> registry_page_bits; - const CacheAddr end = (surface->GetCacheAddrEnd() - 1) >> registry_page_bits; - l1_cache.erase(cache_addr); + const VAddr cpu_addr = surface->GetCpuAddr(); + VAddr start = cpu_addr >> registry_page_bits; + const VAddr end = (surface->GetCpuAddrEnd() - 1) >> registry_page_bits; + l1_cache.erase(cpu_addr); while (start <= end) { auto& reg{registry[start]}; reg.erase(std::find(reg.begin(), reg.end(), surface)); @@ -1046,18 +1039,18 @@ private: } } - std::vector GetSurfacesInRegion(const CacheAddr cache_addr, const std::size_t size) { + std::vector GetSurfacesInRegion(const VAddr cpu_addr, const std::size_t size) { if (size == 0) { return {}; } - const CacheAddr cache_addr_end = cache_addr + size; - CacheAddr start = cache_addr >> registry_page_bits; - const CacheAddr end = (cache_addr_end - 1) >> registry_page_bits; + const VAddr cpu_addr_end = cpu_addr + size; + VAddr start = cpu_addr >> registry_page_bits; + const VAddr end = (cpu_addr_end - 1) >> registry_page_bits; std::vector surfaces; while (start <= end) { std::vector& list = registry[start]; for (auto& surface : list) { - if (!surface->IsPicked() && surface->Overlaps(cache_addr, cache_addr_end)) { + if (!surface->IsPicked() && surface->Overlaps(cpu_addr, cpu_addr_end)) { surface->MarkAsPicked(true); surfaces.push_back(surface); } @@ -1146,14 +1139,14 @@ private: // large in size. static constexpr u64 registry_page_bits{20}; static constexpr u64 registry_page_size{1 << registry_page_bits}; - std::unordered_map> registry; + std::unordered_map> registry; static constexpr u32 DEPTH_RT = 8; static constexpr u32 NO_RT = 0xFFFFFFFF; // The L1 Cache is used for fast texture lookup before checking the overlaps // This avoids calculating size and other stuffs. - std::unordered_map l1_cache; + std::unordered_map l1_cache; /// The surface reserve is a "backup" cache, this is where we put unique surfaces that have /// previously been used. This is to prevent surfaces from being constantly created and