diff options
Diffstat (limited to 'src/video_core')
20 files changed, 828 insertions, 317 deletions
diff --git a/src/video_core/engines/fermi_2d.cpp b/src/video_core/engines/fermi_2d.cpp index 912e785b9..74e44c7fe 100644 --- a/src/video_core/engines/fermi_2d.cpp +++ b/src/video_core/engines/fermi_2d.cpp @@ -47,9 +47,12 @@ void Fermi2D::HandleSurfaceCopy() { u32 dst_bytes_per_pixel = RenderTargetBytesPerPixel(regs.dst.format); if (!rasterizer.AccelerateSurfaceCopy(regs.src, regs.dst)) { - // TODO(bunnei): The below implementation currently will not get hit, as - // AccelerateSurfaceCopy tries to always copy and will always return success. This should be - // changed once we properly support flushing. + rasterizer.FlushRegion(source_cpu, src_bytes_per_pixel * regs.src.width * regs.src.height); + // We have to invalidate the destination region to evict any outdated surfaces from the + // cache. We do this before actually writing the new data because the destination address + // might contain a dirty surface that will have to be written back to memory. + rasterizer.InvalidateRegion(dest_cpu, + dst_bytes_per_pixel * regs.dst.width * regs.dst.height); if (regs.src.linear == regs.dst.linear) { // If the input layout and the output layout are the same, just perform a raw copy. @@ -62,14 +65,16 @@ void Fermi2D::HandleSurfaceCopy() { u8* dst_buffer = Memory::GetPointer(dest_cpu); if (!regs.src.linear && regs.dst.linear) { // If the input is tiled and the output is linear, deswizzle the input and copy it over. - Texture::CopySwizzledData(regs.src.width, regs.src.height, src_bytes_per_pixel, - dst_bytes_per_pixel, src_buffer, dst_buffer, true, - regs.src.BlockHeight()); + Texture::CopySwizzledData(regs.src.width, regs.src.height, regs.src.depth, + src_bytes_per_pixel, dst_bytes_per_pixel, src_buffer, + dst_buffer, true, regs.src.BlockHeight(), + regs.src.BlockDepth()); } else { // If the input is linear and the output is tiled, swizzle the input and copy it over. - Texture::CopySwizzledData(regs.src.width, regs.src.height, src_bytes_per_pixel, - dst_bytes_per_pixel, dst_buffer, src_buffer, false, - regs.dst.BlockHeight()); + Texture::CopySwizzledData(regs.src.width, regs.src.height, regs.src.depth, + src_bytes_per_pixel, dst_bytes_per_pixel, dst_buffer, + src_buffer, false, regs.dst.BlockHeight(), + regs.dst.BlockDepth()); } } } diff --git a/src/video_core/engines/kepler_memory.cpp b/src/video_core/engines/kepler_memory.cpp index 66ae6332d..585290d9f 100644 --- a/src/video_core/engines/kepler_memory.cpp +++ b/src/video_core/engines/kepler_memory.cpp @@ -5,10 +5,14 @@ #include "common/logging/log.h" #include "core/memory.h" #include "video_core/engines/kepler_memory.h" +#include "video_core/rasterizer_interface.h" namespace Tegra::Engines { -KeplerMemory::KeplerMemory(MemoryManager& memory_manager) : memory_manager(memory_manager) {} +KeplerMemory::KeplerMemory(VideoCore::RasterizerInterface& rasterizer, + MemoryManager& memory_manager) + : memory_manager(memory_manager), rasterizer{rasterizer} {} + KeplerMemory::~KeplerMemory() = default; void KeplerMemory::WriteReg(u32 method, u32 value) { @@ -37,6 +41,11 @@ void KeplerMemory::ProcessData(u32 data) { VAddr dest_address = *memory_manager.GpuToCpuAddress(address + state.write_offset * sizeof(u32)); + // We have to invalidate the destination region to evict any outdated surfaces from the cache. + // We do this before actually writing the new data because the destination address might contain + // a dirty surface that will have to be written back to memory. + rasterizer.InvalidateRegion(dest_address, sizeof(u32)); + Memory::Write32(dest_address, data); state.write_offset++; diff --git a/src/video_core/engines/kepler_memory.h b/src/video_core/engines/kepler_memory.h index b0d0078cf..bf4a13cff 100644 --- a/src/video_core/engines/kepler_memory.h +++ b/src/video_core/engines/kepler_memory.h @@ -11,6 +11,10 @@ #include "common/common_types.h" #include "video_core/memory_manager.h" +namespace VideoCore { +class RasterizerInterface; +} + namespace Tegra::Engines { #define KEPLERMEMORY_REG_INDEX(field_name) \ @@ -18,7 +22,7 @@ namespace Tegra::Engines { class KeplerMemory final { public: - KeplerMemory(MemoryManager& memory_manager); + KeplerMemory(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager); ~KeplerMemory(); /// Write the value to the register identified by method. @@ -72,6 +76,7 @@ public: private: MemoryManager& memory_manager; + VideoCore::RasterizerInterface& rasterizer; void ProcessData(u32 data); }; diff --git a/src/video_core/engines/maxwell_3d.h b/src/video_core/engines/maxwell_3d.h index c8d1b6478..c8af1c6b6 100644 --- a/src/video_core/engines/maxwell_3d.h +++ b/src/video_core/engines/maxwell_3d.h @@ -448,7 +448,10 @@ public: BitField<8, 3, u32> block_depth; BitField<12, 1, InvMemoryLayout> type; } memory_layout; - u32 array_mode; + union { + BitField<0, 16, u32> array_mode; + BitField<16, 1, u32> volume; + }; u32 layer_stride; u32 base_layer; INSERT_PADDING_WORDS(7); diff --git a/src/video_core/engines/maxwell_dma.cpp b/src/video_core/engines/maxwell_dma.cpp index aa7481b8c..103cd110e 100644 --- a/src/video_core/engines/maxwell_dma.cpp +++ b/src/video_core/engines/maxwell_dma.cpp @@ -4,12 +4,14 @@ #include "core/memory.h" #include "video_core/engines/maxwell_dma.h" +#include "video_core/rasterizer_interface.h" #include "video_core/textures/decoders.h" namespace Tegra { namespace Engines { -MaxwellDMA::MaxwellDMA(MemoryManager& memory_manager) : memory_manager(memory_manager) {} +MaxwellDMA::MaxwellDMA(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager) + : memory_manager(memory_manager), rasterizer{rasterizer} {} void MaxwellDMA::WriteReg(u32 method, u32 value) { ASSERT_MSG(method < Regs::NUM_REGS, @@ -44,36 +46,79 @@ void MaxwellDMA::HandleCopy() { ASSERT(regs.exec.query_mode == Regs::QueryMode::None); ASSERT(regs.exec.query_intr == Regs::QueryIntr::None); ASSERT(regs.exec.copy_mode == Regs::CopyMode::Unk2); - ASSERT(regs.src_params.pos_x == 0); - ASSERT(regs.src_params.pos_y == 0); ASSERT(regs.dst_params.pos_x == 0); ASSERT(regs.dst_params.pos_y == 0); - if (regs.exec.is_dst_linear == regs.exec.is_src_linear) { - std::size_t copy_size = regs.x_count; + if (!regs.exec.is_dst_linear && !regs.exec.is_src_linear) { + // If both the source and the destination are in block layout, assert. + UNREACHABLE_MSG("Tiled->Tiled DMA transfers are not yet implemented"); + return; + } + if (regs.exec.is_dst_linear && regs.exec.is_src_linear) { // When the enable_2d bit is disabled, the copy is performed as if we were copying a 1D - // buffer of length `x_count`, otherwise we copy a 2D buffer of size (x_count, y_count). - if (regs.exec.enable_2d) { - copy_size = copy_size * regs.y_count; + // buffer of length `x_count`, otherwise we copy a 2D image of dimensions (x_count, + // y_count). + if (!regs.exec.enable_2d) { + Memory::CopyBlock(dest_cpu, source_cpu, regs.x_count); + return; } - Memory::CopyBlock(dest_cpu, source_cpu, copy_size); + // If both the source and the destination are in linear layout, perform a line-by-line + // copy. We're going to take a subrect of size (x_count, y_count) from the source + // rectangle. There is no need to manually flush/invalidate the regions because + // CopyBlock does that for us. + for (u32 line = 0; line < regs.y_count; ++line) { + const VAddr source_line = source_cpu + line * regs.src_pitch; + const VAddr dest_line = dest_cpu + line * regs.dst_pitch; + Memory::CopyBlock(dest_line, source_line, regs.x_count); + } return; } ASSERT(regs.exec.enable_2d == 1); + + std::size_t copy_size = regs.x_count * regs.y_count; + + const auto FlushAndInvalidate = [&](u32 src_size, u32 dst_size) { + // TODO(Subv): For now, manually flush the regions until we implement GPU-accelerated + // copying. + rasterizer.FlushRegion(source_cpu, src_size); + + // We have to invalidate the destination region to evict any outdated surfaces from the + // cache. We do this before actually writing the new data because the destination address + // might contain a dirty surface that will have to be written back to memory. + rasterizer.InvalidateRegion(dest_cpu, dst_size); + }; + u8* src_buffer = Memory::GetPointer(source_cpu); u8* dst_buffer = Memory::GetPointer(dest_cpu); if (regs.exec.is_dst_linear && !regs.exec.is_src_linear) { + ASSERT(regs.src_params.size_z == 1); // If the input is tiled and the output is linear, deswizzle the input and copy it over. - Texture::CopySwizzledData(regs.src_params.size_x, regs.src_params.size_y, 1, 1, src_buffer, - dst_buffer, true, regs.src_params.BlockHeight()); + + u32 src_bytes_per_pixel = regs.src_pitch / regs.src_params.size_x; + + FlushAndInvalidate(regs.src_pitch * regs.src_params.size_y, + copy_size * src_bytes_per_pixel); + + Texture::UnswizzleSubrect(regs.x_count, regs.y_count, regs.dst_pitch, + regs.src_params.size_x, src_bytes_per_pixel, source_cpu, dest_cpu, + regs.src_params.BlockHeight(), regs.src_params.pos_x, + regs.src_params.pos_y); } else { + ASSERT(regs.dst_params.size_z == 1); + ASSERT(regs.src_pitch == regs.x_count); + + u32 src_bpp = regs.src_pitch / regs.x_count; + + FlushAndInvalidate(regs.src_pitch * regs.y_count, + regs.dst_params.size_x * regs.dst_params.size_y * src_bpp); + // If the input is linear and the output is tiled, swizzle the input and copy it over. - Texture::CopySwizzledData(regs.dst_params.size_x, regs.dst_params.size_y, 1, 1, dst_buffer, - src_buffer, false, regs.dst_params.BlockHeight()); + Texture::SwizzleSubrect(regs.x_count, regs.y_count, regs.src_pitch, regs.dst_params.size_x, + src_bpp, dest_cpu, source_cpu, regs.dst_params.BlockHeight()); } } diff --git a/src/video_core/engines/maxwell_dma.h b/src/video_core/engines/maxwell_dma.h index 311ccb616..5f3704f05 100644 --- a/src/video_core/engines/maxwell_dma.h +++ b/src/video_core/engines/maxwell_dma.h @@ -12,11 +12,15 @@ #include "video_core/gpu.h" #include "video_core/memory_manager.h" +namespace VideoCore { +class RasterizerInterface; +} + namespace Tegra::Engines { class MaxwellDMA final { public: - explicit MaxwellDMA(MemoryManager& memory_manager); + explicit MaxwellDMA(VideoCore::RasterizerInterface& rasterizer, MemoryManager& memory_manager); ~MaxwellDMA() = default; /// Write the value to the register identified by method. @@ -43,6 +47,10 @@ public: u32 BlockHeight() const { return 1 << block_height; } + + u32 BlockDepth() const { + return 1 << block_depth; + } }; static_assert(sizeof(Parameters) == 24, "Parameters has wrong size"); @@ -129,6 +137,8 @@ public: MemoryManager& memory_manager; private: + VideoCore::RasterizerInterface& rasterizer; + /// Performs the copy from the source buffer to the destination buffer as configured in the /// registers. void HandleCopy(); diff --git a/src/video_core/engines/shader_bytecode.h b/src/video_core/engines/shader_bytecode.h index 39ae065de..e3d67ff87 100644 --- a/src/video_core/engines/shader_bytecode.h +++ b/src/video_core/engines/shader_bytecode.h @@ -267,7 +267,7 @@ enum class ControlCode : u64 { GTU = 12, NEU = 13, GEU = 14, - // + T = 15, OFF = 16, LO = 17, SFF = 18, diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp index 9ba7e3533..83c7e5b0b 100644 --- a/src/video_core/gpu.cpp +++ b/src/video_core/gpu.cpp @@ -27,8 +27,8 @@ GPU::GPU(VideoCore::RasterizerInterface& rasterizer) { maxwell_3d = std::make_unique<Engines::Maxwell3D>(rasterizer, *memory_manager); fermi_2d = std::make_unique<Engines::Fermi2D>(rasterizer, *memory_manager); maxwell_compute = std::make_unique<Engines::MaxwellCompute>(); - maxwell_dma = std::make_unique<Engines::MaxwellDMA>(*memory_manager); - kepler_memory = std::make_unique<Engines::KeplerMemory>(*memory_manager); + maxwell_dma = std::make_unique<Engines::MaxwellDMA>(rasterizer, *memory_manager); + kepler_memory = std::make_unique<Engines::KeplerMemory>(rasterizer, *memory_manager); } GPU::~GPU() = default; diff --git a/src/video_core/memory_manager.cpp b/src/video_core/memory_manager.cpp index ca923d17d..022d4ab74 100644 --- a/src/video_core/memory_manager.cpp +++ b/src/video_core/memory_manager.cpp @@ -87,6 +87,16 @@ GPUVAddr MemoryManager::UnmapBuffer(GPUVAddr gpu_addr, u64 size) { return gpu_addr; } +GPUVAddr MemoryManager::GetRegionEnd(GPUVAddr region_start) const { + for (const auto& region : mapped_regions) { + const GPUVAddr region_end{region.gpu_addr + region.size}; + if (region_start >= region.gpu_addr && region_start < region_end) { + return region_end; + } + } + return {}; +} + boost::optional<GPUVAddr> MemoryManager::FindFreeBlock(u64 size, u64 align) { GPUVAddr gpu_addr = 0; u64 free_space = 0; diff --git a/src/video_core/memory_manager.h b/src/video_core/memory_manager.h index 86765e72a..caf80093f 100644 --- a/src/video_core/memory_manager.h +++ b/src/video_core/memory_manager.h @@ -26,6 +26,7 @@ public: GPUVAddr MapBufferEx(VAddr cpu_addr, u64 size); GPUVAddr MapBufferEx(VAddr cpu_addr, GPUVAddr gpu_addr, u64 size); GPUVAddr UnmapBuffer(GPUVAddr gpu_addr, u64 size); + GPUVAddr GetRegionEnd(GPUVAddr region_start) const; boost::optional<VAddr> GpuToCpuAddress(GPUVAddr gpu_addr); std::vector<GPUVAddr> CpuToGpuAddress(VAddr cpu_addr) const; diff --git a/src/video_core/rasterizer_cache.h b/src/video_core/rasterizer_cache.h index 083b283b0..0a3b3951e 100644 --- a/src/video_core/rasterizer_cache.h +++ b/src/video_core/rasterizer_cache.h @@ -11,32 +11,77 @@ #include "common/common_types.h" #include "core/core.h" +#include "core/settings.h" #include "video_core/rasterizer_interface.h" #include "video_core/renderer_base.h" +class RasterizerCacheObject { +public: + /// Gets the address of the shader in guest memory, required for cache management + virtual VAddr GetAddr() const = 0; + + /// Gets the size of the shader in guest memory, required for cache management + virtual std::size_t GetSizeInBytes() const = 0; + + /// Wriets any cached resources back to memory + virtual void Flush() = 0; + + /// Sets whether the cached object should be considered registered + void SetIsRegistered(bool registered) { + is_registered = registered; + } + + /// Returns true if the cached object is registered + bool IsRegistered() const { + return is_registered; + } + + /// Returns true if the cached object is dirty + bool IsDirty() const { + return is_dirty; + } + + /// Returns ticks from when this cached object was last modified + u64 GetLastModifiedTicks() const { + return last_modified_ticks; + } + + /// Marks an object as recently modified, used to specify whether it is clean or dirty + template <class T> + void MarkAsModified(bool dirty, T& cache) { + is_dirty = dirty; + last_modified_ticks = cache.GetModifiedTicks(); + } + +private: + bool is_registered{}; ///< Whether the object is currently registered with the cache + bool is_dirty{}; ///< Whether the object is dirty (out of sync with guest memory) + u64 last_modified_ticks{}; ///< When the object was last modified, used for in-order flushing +}; + template <class T> class RasterizerCache : NonCopyable { + friend class RasterizerCacheObject; + public: + /// Write any cached resources overlapping the specified region back to memory + void FlushRegion(Tegra::GPUVAddr addr, size_t size) { + const auto& objects{GetSortedObjectsFromRegion(addr, size)}; + for (auto& object : objects) { + FlushObject(object); + } + } + /// Mark the specified region as being invalidated void InvalidateRegion(VAddr addr, u64 size) { - if (size == 0) - return; - - const ObjectInterval interval{addr, addr + size}; - for (auto& pair : boost::make_iterator_range(object_cache.equal_range(interval))) { - for (auto& cached_object : pair.second) { - if (!cached_object) - continue; - - remove_objects.emplace(cached_object); + const auto& objects{GetSortedObjectsFromRegion(addr, size)}; + for (auto& object : objects) { + if (!object->IsRegistered()) { + // Skip duplicates + continue; } + Unregister(object); } - - for (auto& remove_object : remove_objects) { - Unregister(remove_object); - } - - remove_objects.clear(); } /// Invalidates everything in the cache @@ -62,6 +107,7 @@ protected: /// Register an object into the cache void Register(const T& object) { + object->SetIsRegistered(true); object_cache.add({GetInterval(object), ObjectSet{object}}); auto& rasterizer = Core::System::GetInstance().Renderer().Rasterizer(); rasterizer.UpdatePagesCachedCount(object->GetAddr(), object->GetSizeInBytes(), 1); @@ -69,12 +115,57 @@ protected: /// Unregisters an object from the cache void Unregister(const T& object) { + object->SetIsRegistered(false); auto& rasterizer = Core::System::GetInstance().Renderer().Rasterizer(); rasterizer.UpdatePagesCachedCount(object->GetAddr(), object->GetSizeInBytes(), -1); + + // Only flush if use_accurate_gpu_emulation is enabled, as it incurs a performance hit + if (Settings::values.use_accurate_gpu_emulation) { + FlushObject(object); + } + object_cache.subtract({GetInterval(object), ObjectSet{object}}); } + /// Returns a ticks counter used for tracking when cached objects were last modified + u64 GetModifiedTicks() { + return ++modified_ticks; + } + private: + /// Returns a list of cached objects from the specified memory region, ordered by access time + std::vector<T> GetSortedObjectsFromRegion(VAddr addr, u64 size) { + if (size == 0) { + return {}; + } + + std::vector<T> objects; + const ObjectInterval interval{addr, addr + size}; + for (auto& pair : boost::make_iterator_range(object_cache.equal_range(interval))) { + for (auto& cached_object : pair.second) { + if (!cached_object) { + continue; + } + objects.push_back(cached_object); + } + } + + std::sort(objects.begin(), objects.end(), [](const T& a, const T& b) -> bool { + return a->GetLastModifiedTicks() < b->GetLastModifiedTicks(); + }); + + return objects; + } + + /// Flushes the specified object, updating appropriate cache state as needed + void FlushObject(const T& object) { + if (!object->IsDirty()) { + return; + } + object->Flush(); + object->MarkAsModified(false, *this); + } + using ObjectSet = std::set<T>; using ObjectCache = boost::icl::interval_map<VAddr, ObjectSet>; using ObjectInterval = typename ObjectCache::interval_type; @@ -84,6 +175,6 @@ private: object->GetAddr() + object->GetSizeInBytes()); } - ObjectCache object_cache; - ObjectSet remove_objects; + ObjectCache object_cache; ///< Cache of objects + u64 modified_ticks{}; ///< Counter of cache state ticks, used for in-order flushing }; diff --git a/src/video_core/renderer_opengl/gl_buffer_cache.h b/src/video_core/renderer_opengl/gl_buffer_cache.h index 965976334..be29dc8be 100644 --- a/src/video_core/renderer_opengl/gl_buffer_cache.h +++ b/src/video_core/renderer_opengl/gl_buffer_cache.h @@ -15,15 +15,18 @@ namespace OpenGL { -struct CachedBufferEntry final { - VAddr GetAddr() const { +struct CachedBufferEntry final : public RasterizerCacheObject { + VAddr GetAddr() const override { return addr; } - std::size_t GetSizeInBytes() const { + std::size_t GetSizeInBytes() const override { return size; } + // We do not have to flush this cache as things in it are never modified by us. + void Flush() override {} + VAddr addr; std::size_t size; GLintptr offset; diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp index 84582c777..3daccf82f 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer.cpp +++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp @@ -286,7 +286,8 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) { &ubo, sizeof(ubo), static_cast<std::size_t>(uniform_buffer_alignment)); // Bind the buffer - glBindBufferRange(GL_UNIFORM_BUFFER, stage, buffer_cache.GetHandle(), offset, sizeof(ubo)); + glBindBufferRange(GL_UNIFORM_BUFFER, static_cast<GLuint>(stage), buffer_cache.GetHandle(), + offset, static_cast<GLsizeiptr>(sizeof(ubo))); Shader shader{shader_cache.GetStageProgram(program)}; @@ -423,6 +424,13 @@ void RasterizerOpenGL::ConfigureFramebuffers(bool using_color_fb, bool using_dep // Used when just a single color attachment is enabled, e.g. for clearing a color buffer Surface color_surface = res_cache.GetColorBufferSurface(*single_color_target, preserve_contents); + + if (color_surface) { + // Assume that a surface will be written to if it is used as a framebuffer, even if + // the shader doesn't actually write to it. + color_surface->MarkAsModified(true, res_cache); + } + glFramebufferTexture2D( GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + static_cast<GLenum>(*single_color_target), GL_TEXTURE_2D, @@ -433,6 +441,13 @@ void RasterizerOpenGL::ConfigureFramebuffers(bool using_color_fb, bool using_dep std::array<GLenum, Maxwell::NumRenderTargets> buffers; for (std::size_t index = 0; index < Maxwell::NumRenderTargets; ++index) { Surface color_surface = res_cache.GetColorBufferSurface(index, preserve_contents); + + if (color_surface) { + // Assume that a surface will be written to if it is used as a framebuffer, even + // if the shader doesn't actually write to it. + color_surface->MarkAsModified(true, res_cache); + } + buffers[index] = GL_COLOR_ATTACHMENT0 + regs.rt_control.GetMap(index); glFramebufferTexture2D( GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0 + static_cast<GLenum>(index), @@ -452,6 +467,10 @@ void RasterizerOpenGL::ConfigureFramebuffers(bool using_color_fb, bool using_dep } if (depth_surface) { + // Assume that a surface will be written to if it is used as a framebuffer, even if + // the shader doesn't actually write to it. + depth_surface->MarkAsModified(true, res_cache); + if (regs.stencil_enable) { // Attach both depth and stencil glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, @@ -616,7 +635,14 @@ void RasterizerOpenGL::DrawArrays() { void RasterizerOpenGL::FlushAll() {} -void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) {} +void RasterizerOpenGL::FlushRegion(VAddr addr, u64 size) { + MICROPROFILE_SCOPE(OpenGL_CacheManagement); + + if (Settings::values.use_accurate_gpu_emulation) { + // Only flush if use_accurate_gpu_emulation is enabled, as it incurs a performance hit + res_cache.FlushRegion(addr, size); + } +} void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) { MICROPROFILE_SCOPE(OpenGL_CacheManagement); @@ -626,12 +652,19 @@ void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size) { } void RasterizerOpenGL::FlushAndInvalidateRegion(VAddr addr, u64 size) { + FlushRegion(addr, size); InvalidateRegion(addr, size); } bool RasterizerOpenGL::AccelerateSurfaceCopy(const Tegra::Engines::Fermi2D::Regs::Surface& src, const Tegra::Engines::Fermi2D::Regs::Surface& dst) { MICROPROFILE_SCOPE(OpenGL_Blits); + + if (Settings::values.use_accurate_gpu_emulation) { + // Skip the accelerated copy and perform a slow but more accurate copy + return false; + } + res_cache.FermiCopySurface(src, dst); return true; } diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp index 65a220c41..9c8925383 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp +++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp @@ -34,16 +34,53 @@ struct FormatTuple { bool compressed; }; -static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) { - auto& gpu{Core::System::GetInstance().GPU()}; - const auto cpu_addr{gpu.MemoryManager().GpuToCpuAddress(gpu_addr)}; - return cpu_addr ? *cpu_addr : 0; +static bool IsPixelFormatASTC(PixelFormat format) { + switch (format) { + case PixelFormat::ASTC_2D_4X4: + case PixelFormat::ASTC_2D_5X4: + case PixelFormat::ASTC_2D_8X8: + case PixelFormat::ASTC_2D_8X5: + return true; + default: + return false; + } +} + +static std::pair<u32, u32> GetASTCBlockSize(PixelFormat format) { + switch (format) { + case PixelFormat::ASTC_2D_4X4: + return {4, 4}; + case PixelFormat::ASTC_2D_5X4: + return {5, 4}; + case PixelFormat::ASTC_2D_8X8: + return {8, 8}; + case PixelFormat::ASTC_2D_8X5: + return {8, 5}; + default: + LOG_CRITICAL(HW_GPU, "Unhandled format: {}", static_cast<u32>(format)); + UNREACHABLE(); + } +} + +void SurfaceParams::InitCacheParameters(Tegra::GPUVAddr gpu_addr_) { + auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()}; + const auto cpu_addr{memory_manager.GpuToCpuAddress(gpu_addr_)}; + + addr = cpu_addr ? *cpu_addr : 0; + gpu_addr = gpu_addr_; + size_in_bytes = SizeInBytesRaw(); + + if (IsPixelFormatASTC(pixel_format)) { + // ASTC is uncompressed in software, in emulated as RGBA8 + size_in_bytes_gl = width * height * depth * 4; + } else { + size_in_bytes_gl = SizeInBytesGL(); + } } /*static*/ SurfaceParams SurfaceParams::CreateForTexture( const Tegra::Texture::FullTextureInfo& config, const GLShader::SamplerEntry& entry) { SurfaceParams params{}; - params.addr = TryGetCpuAddr(config.tic.Address()); params.is_tiled = config.tic.IsTiled(); params.block_width = params.is_tiled ? config.tic.BlockWidth() : 0, params.block_height = params.is_tiled ? config.tic.BlockHeight() : 0, @@ -87,18 +124,18 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) { break; } - params.size_in_bytes_total = params.SizeInBytesTotal(); - params.size_in_bytes_2d = params.SizeInBytes2D(); params.max_mip_level = config.tic.max_mip_level + 1; params.rt = {}; + params.InitCacheParameters(config.tic.Address()); + return params; } /*static*/ SurfaceParams SurfaceParams::CreateForFramebuffer(std::size_t index) { const auto& config{Core::System::GetInstance().GPU().Maxwell3D().regs.rt[index]}; SurfaceParams params{}; - params.addr = TryGetCpuAddr(config.Address()); + params.is_tiled = config.memory_layout.type == Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout::BlockLinear; params.block_width = 1 << config.memory_layout.block_width; @@ -112,16 +149,17 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) { params.unaligned_height = config.height; params.target = SurfaceTarget::Texture2D; params.depth = 1; - params.size_in_bytes_total = params.SizeInBytesTotal(); - params.size_in_bytes_2d = params.SizeInBytes2D(); params.max_mip_level = 0; // Render target specific parameters, not used for caching params.rt.index = static_cast<u32>(index); params.rt.array_mode = config.array_mode; params.rt.layer_stride = config.layer_stride; + params.rt.volume = config.volume; params.rt.base_layer = config.base_layer; + params.InitCacheParameters(config.Address()); + return params; } @@ -130,7 +168,7 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) { u32 block_width, u32 block_height, u32 block_depth, Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout type) { SurfaceParams params{}; - params.addr = TryGetCpuAddr(zeta_address); + params.is_tiled = type == Tegra::Engines::Maxwell3D::Regs::InvMemoryLayout::BlockLinear; params.block_width = 1 << std::min(block_width, 5U); params.block_height = 1 << std::min(block_height, 5U); @@ -143,18 +181,18 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) { params.unaligned_height = zeta_height; params.target = SurfaceTarget::Texture2D; params.depth = 1; - params.size_in_bytes_total = params.SizeInBytesTotal(); - params.size_in_bytes_2d = params.SizeInBytes2D(); params.max_mip_level = 0; params.rt = {}; + params.InitCacheParameters(zeta_address); + return params; } /*static*/ SurfaceParams SurfaceParams::CreateForFermiCopySurface( const Tegra::Engines::Fermi2D::Regs::Surface& config) { SurfaceParams params{}; - params.addr = TryGetCpuAddr(config.Address()); + params.is_tiled = !config.linear; params.block_width = params.is_tiled ? std::min(config.BlockWidth(), 32U) : 0, params.block_height = params.is_tiled ? std::min(config.BlockHeight(), 32U) : 0, @@ -167,11 +205,11 @@ static VAddr TryGetCpuAddr(Tegra::GPUVAddr gpu_addr) { params.unaligned_height = config.height; params.target = SurfaceTarget::Texture2D; params.depth = 1; - params.size_in_bytes_total = params.SizeInBytesTotal(); - params.size_in_bytes_2d = params.SizeInBytes2D(); params.max_mip_level = 0; params.rt = {}; + params.InitCacheParameters(config.Address()); + return params; } @@ -231,6 +269,8 @@ static constexpr std::array<FormatTuple, SurfaceParams::MaxPixelFormat> tex_form {GL_RG32UI, GL_RG_INTEGER, GL_UNSIGNED_INT, ComponentType::UInt, false}, // RG32UI {GL_R32UI, GL_RED_INTEGER, GL_UNSIGNED_INT, ComponentType::UInt, false}, // R32UI {GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_8X8 + {GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_8X5 + {GL_RGBA8, GL_RGBA, GL_UNSIGNED_BYTE, ComponentType::UNorm, false}, // ASTC_2D_5X4 // Depth formats {GL_DEPTH_COMPONENT32F, GL_DEPTH_COMPONENT, GL_FLOAT, ComponentType::Float, false}, // Z32F @@ -274,28 +314,6 @@ static const FormatTuple& GetFormatTuple(PixelFormat pixel_format, ComponentType return format; } -static bool IsPixelFormatASTC(PixelFormat format) { - switch (format) { - case PixelFormat::ASTC_2D_4X4: - case PixelFormat::ASTC_2D_8X8: - return true; - default: - return false; - } -} - -static std::pair<u32, u32> GetASTCBlockSize(PixelFormat format) { - switch (format) { - case PixelFormat::ASTC_2D_4X4: - return {4, 4}; - case PixelFormat::ASTC_2D_8X8: - return {8, 8}; - default: - LOG_CRITICAL(HW_GPU, "Unhandled format: {}", static_cast<u32>(format)); - UNREACHABLE(); - } -} - MathUtil::Rectangle<u32> SurfaceParams::GetRect() const { u32 actual_height{unaligned_height}; if (IsPixelFormatASTC(pixel_format)) { @@ -323,29 +341,27 @@ static bool IsFormatBCn(PixelFormat format) { } template <bool morton_to_gl, PixelFormat format> -void MortonCopy(u32 stride, u32 block_height, u32 height, u8* gl_buffer, std::size_t gl_buffer_size, - VAddr addr) { - constexpr u32 bytes_per_pixel = SurfaceParams::GetFormatBpp(format) / CHAR_BIT; - constexpr u32 gl_bytes_per_pixel = CachedSurface::GetGLBytesPerPixel(format); +void MortonCopy(u32 stride, u32 block_height, u32 height, u32 block_depth, u32 depth, u8* gl_buffer, + std::size_t gl_buffer_size, VAddr addr) { + constexpr u32 bytes_per_pixel = SurfaceParams::GetBytesPerPixel(format); + + // With the BCn formats (DXT and DXN), each 4x4 tile is swizzled instead of just individual + // pixel values. + const u32 tile_size{IsFormatBCn(format) ? 4U : 1U}; if (morton_to_gl) { - // With the BCn formats (DXT and DXN), each 4x4 tile is swizzled instead of just individual - // pixel values. - const u32 tile_size{IsFormatBCn(format) ? 4U : 1U}; const std::vector<u8> data = Tegra::Texture::UnswizzleTexture( - addr, tile_size, bytes_per_pixel, stride, height, block_height); + addr, tile_size, bytes_per_pixel, stride, height, depth, block_height, block_depth); const std::size_t size_to_copy{std::min(gl_buffer_size, data.size())}; memcpy(gl_buffer, data.data(), size_to_copy); } else { - // TODO(bunnei): Assumes the default rendering GOB size of 16 (128 lines). We should - // check the configuration for this and perform more generic un/swizzle - LOG_WARNING(Render_OpenGL, "need to use correct swizzle/GOB parameters!"); - VideoCore::MortonCopyPixels128(stride, height, bytes_per_pixel, gl_bytes_per_pixel, - Memory::GetPointer(addr), gl_buffer, morton_to_gl); + Tegra::Texture::CopySwizzledData(stride / tile_size, height / tile_size, depth, + bytes_per_pixel, bytes_per_pixel, Memory::GetPointer(addr), + gl_buffer, false, block_height, block_depth); } } -static constexpr std::array<void (*)(u32, u32, u32, u8*, std::size_t, VAddr), +static constexpr std::array<void (*)(u32, u32, u32, u32, u32, u8*, std::size_t, VAddr), SurfaceParams::MaxPixelFormat> morton_to_gl_fns = { // clang-format off @@ -395,6 +411,8 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, std::size_t, VAddr), MortonCopy<true, PixelFormat::RG32UI>, MortonCopy<true, PixelFormat::R32UI>, MortonCopy<true, PixelFormat::ASTC_2D_8X8>, + MortonCopy<true, PixelFormat::ASTC_2D_8X5>, + MortonCopy<true, PixelFormat::ASTC_2D_5X4>, MortonCopy<true, PixelFormat::Z32F>, MortonCopy<true, PixelFormat::Z16>, MortonCopy<true, PixelFormat::Z24S8>, @@ -403,7 +421,7 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, std::size_t, VAddr), // clang-format on }; -static constexpr std::array<void (*)(u32, u32, u32, u8*, std::size_t, VAddr), +static constexpr std::array<void (*)(u32, u32, u32, u32, u32, u8*, std::size_t, VAddr), SurfaceParams::MaxPixelFormat> gl_to_morton_fns = { // clang-format off @@ -420,17 +438,16 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, std::size_t, VAddr), MortonCopy<false, PixelFormat::RGBA16UI>, MortonCopy<false, PixelFormat::R11FG11FB10F>, MortonCopy<false, PixelFormat::RGBA32UI>, - // TODO(Subv): Swizzling DXT1/DXT23/DXT45/DXN1/DXN2/BC7U/BC6H_UF16/BC6H_SF16/ASTC_2D_4X4 - // formats are not supported - nullptr, - nullptr, - nullptr, - nullptr, - nullptr, - nullptr, - nullptr, - nullptr, - nullptr, + MortonCopy<false, PixelFormat::DXT1>, + MortonCopy<false, PixelFormat::DXT23>, + MortonCopy<false, PixelFormat::DXT45>, + MortonCopy<false, PixelFormat::DXN1>, + MortonCopy<false, PixelFormat::DXN2UNORM>, + MortonCopy<false, PixelFormat::DXN2SNORM>, + MortonCopy<false, PixelFormat::BC7U>, + MortonCopy<false, PixelFormat::BC6H_UF16>, + MortonCopy<false, PixelFormat::BC6H_SF16>, + // TODO(Subv): Swizzling ASTC formats are not supported nullptr, MortonCopy<false, PixelFormat::G8R8U>, MortonCopy<false, PixelFormat::G8R8S>, @@ -455,6 +472,8 @@ static constexpr std::array<void (*)(u32, u32, u32, u8*, std::size_t, VAddr), MortonCopy<false, PixelFormat::RG32UI>, MortonCopy<false, PixelFormat::R32UI>, nullptr, + nullptr, + nullptr, MortonCopy<false, PixelFormat::Z32F>, MortonCopy<false, PixelFormat::Z16>, MortonCopy<false, PixelFormat::Z24S8>, @@ -614,22 +633,21 @@ static void CopySurface(const Surface& src_surface, const Surface& dst_surface, auto source_format = GetFormatTuple(src_params.pixel_format, src_params.component_type); auto dest_format = GetFormatTuple(dst_params.pixel_format, dst_params.component_type); - std::size_t buffer_size = - std::max(src_params.size_in_bytes_total, dst_params.size_in_bytes_total); + std::size_t buffer_size = std::max(src_params.size_in_bytes, dst_params.size_in_bytes); glBindBuffer(GL_PIXEL_PACK_BUFFER, copy_pbo_handle); glBufferData(GL_PIXEL_PACK_BUFFER, buffer_size, nullptr, GL_STREAM_DRAW_ARB); if (source_format.compressed) { glGetCompressedTextureImage(src_surface->Texture().handle, src_attachment, - static_cast<GLsizei>(src_params.size_in_bytes_total), nullptr); + static_cast<GLsizei>(src_params.size_in_bytes), nullptr); } else { glGetTextureImage(src_surface->Texture().handle, src_attachment, source_format.format, - source_format.type, static_cast<GLsizei>(src_params.size_in_bytes_total), + source_format.type, static_cast<GLsizei>(src_params.size_in_bytes), nullptr); } // If the new texture is bigger than the previous one, we need to fill in the rest with data // from the CPU. - if (src_params.size_in_bytes_total < dst_params.size_in_bytes_total) { + if (src_params.size_in_bytes < dst_params.size_in_bytes) { // Upload the rest of the memory. if (dst_params.is_tiled) { // TODO(Subv): We might have to de-tile the subtexture and re-tile it with the rest @@ -639,12 +657,12 @@ static void CopySurface(const Surface& src_surface, const Surface& dst_surface, LOG_DEBUG(HW_GPU, "Trying to upload extra texture data from the CPU during " "reinterpretation but the texture is tiled."); } - std::size_t remaining_size = - dst_params.size_in_bytes_total - src_params.size_in_bytes_total; + std::size_t remaining_size = dst_params.size_in_bytes - src_params.size_in_bytes; std::vector<u8> data(remaining_size); - Memory::ReadBlock(dst_params.addr + src_params.size_in_bytes_total, data.data(), - data.size()); - glBufferSubData(GL_PIXEL_PACK_BUFFER, src_params.size_in_bytes_total, remaining_size, + std::memcpy(data.data(), Memory::GetPointer(dst_params.addr + src_params.size_in_bytes), + data.size()); + + glBufferSubData(GL_PIXEL_PACK_BUFFER, src_params.size_in_bytes, remaining_size, data.data()); } @@ -690,7 +708,8 @@ static void CopySurface(const Surface& src_surface, const Surface& dst_surface, } CachedSurface::CachedSurface(const SurfaceParams& params) - : params(params), gl_target(SurfaceTargetToGL(params.target)) { + : params(params), gl_target(SurfaceTargetToGL(params.target)), + cached_size_in_bytes(params.size_in_bytes) { texture.Create(); const auto& rect{params.GetRect()}; @@ -740,9 +759,21 @@ CachedSurface::CachedSurface(const SurfaceParams& params) VideoCore::LabelGLObject(GL_TEXTURE, texture.handle, params.addr, SurfaceParams::SurfaceTargetName(params.target)); + + // Clamp size to mapped GPU memory region + // TODO(bunnei): Super Mario Odyssey maps a 0x40000 byte region and then uses it for a 0x80000 + // R32F render buffer. We do not yet know if this is a game bug or something else, but this + // check is necessary to prevent flushing from overwriting unmapped memory. + + auto& memory_manager{Core::System::GetInstance().GPU().MemoryManager()}; + const u64 max_size{memory_manager.GetRegionEnd(params.gpu_addr) - params.gpu_addr}; + if (cached_size_in_bytes > max_size) { + LOG_ERROR(HW_GPU, "Surface size {} exceeds region size {}", params.size_in_bytes, max_size); + cached_size_in_bytes = max_size; + } } -static void ConvertS8Z24ToZ24S8(std::vector<u8>& data, u32 width, u32 height) { +static void ConvertS8Z24ToZ24S8(std::vector<u8>& data, u32 width, u32 height, bool reverse) { union S8Z24 { BitField<0, 24, u32> z24; BitField<24, 8, u32> s8; @@ -755,22 +786,29 @@ static void ConvertS8Z24ToZ24S8(std::vector<u8>& data, u32 width, u32 height) { }; static_assert(sizeof(Z24S8) == 4, "Z24S8 is incorrect size"); - S8Z24 input_pixel{}; - Z24S8 output_pixel{}; - constexpr auto bpp{CachedSurface::GetGLBytesPerPixel(PixelFormat::S8Z24)}; + S8Z24 s8z24_pixel{}; + Z24S8 z24s8_pixel{}; + constexpr auto bpp{SurfaceParams::GetBytesPerPixel(PixelFormat::S8Z24)}; for (std::size_t y = 0; y < height; ++y) { for (std::size_t x = 0; x < width; ++x) { const std::size_t offset{bpp * (y * width + x)}; - std::memcpy(&input_pixel, &data[offset], sizeof(S8Z24)); - output_pixel.s8.Assign(input_pixel.s8); - output_pixel.z24.Assign(input_pixel.z24); - std::memcpy(&data[offset], &output_pixel, sizeof(Z24S8)); + if (reverse) { + std::memcpy(&z24s8_pixel, &data[offset], sizeof(Z24S8)); + s8z24_pixel.s8.Assign(z24s8_pixel.s8); + s8z24_pixel.z24.Assign(z24s8_pixel.z24); + std::memcpy(&data[offset], &s8z24_pixel, sizeof(S8Z24)); + } else { + std::memcpy(&s8z24_pixel, &data[offset], sizeof(S8Z24)); + z24s8_pixel.s8.Assign(s8z24_pixel.s8); + z24s8_pixel.z24.Assign(s8z24_pixel.z24); + std::memcpy(&data[offset], &z24s8_pixel, sizeof(Z24S8)); + } } } } static void ConvertG8R8ToR8G8(std::vector<u8>& data, u32 width, u32 height) { - constexpr auto bpp{CachedSurface::GetGLBytesPerPixel(PixelFormat::G8R8U)}; + constexpr auto bpp{SurfaceParams::GetBytesPerPixel(PixelFormat::G8R8U)}; for (std::size_t y = 0; y < height; ++y) { for (std::size_t x = 0; x < width; ++x) { const std::size_t offset{bpp * (y * width + x)}; @@ -790,7 +828,9 @@ static void ConvertFormatAsNeeded_LoadGLBuffer(std::vector<u8>& data, PixelForma u32 width, u32 height) { switch (pixel_format) { case PixelFormat::ASTC_2D_4X4: - case PixelFormat::ASTC_2D_8X8: { + case PixelFormat::ASTC_2D_8X8: + case PixelFormat::ASTC_2D_8X5: + case PixelFormat::ASTC_2D_5X4: { // Convert ASTC pixel formats to RGBA8, as most desktop GPUs do not support ASTC. u32 block_width{}; u32 block_height{}; @@ -800,7 +840,7 @@ static void ConvertFormatAsNeeded_LoadGLBuffer(std::vector<u8>& data, PixelForma } case PixelFormat::S8Z24: // Convert the S8Z24 depth format to Z24S8, as OpenGL does not support S8Z24. - ConvertS8Z24ToZ24S8(data, width, height); + ConvertS8Z24ToZ24S8(data, width, height, false); break; case PixelFormat::G8R8U: @@ -811,54 +851,54 @@ static void ConvertFormatAsNeeded_LoadGLBuffer(std::vector<u8>& data, PixelForma } } +/** + * Helper function to perform software conversion (as needed) when flushing a buffer from OpenGL to + * Switch memory. This is for Maxwell pixel formats that cannot be represented as-is in OpenGL or + * with typical desktop GPUs. + */ +static void ConvertFormatAsNeeded_FlushGLBuffer(std::vector<u8>& data, PixelFormat pixel_format, + u32 width, u32 height) { + switch (pixel_format) { + case PixelFormat::G8R8U: + case PixelFormat::G8R8S: + case PixelFormat::ASTC_2D_4X4: + case PixelFormat::ASTC_2D_8X8: { + LOG_CRITICAL(HW_GPU, "Conversion of format {} after texture flushing is not implemented", + static_cast<u32>(pixel_format)); + UNREACHABLE(); + break; + } + case PixelFormat::S8Z24: + // Convert the Z24S8 depth format to S8Z24, as OpenGL does not support S8Z24. + ConvertS8Z24ToZ24S8(data, width, height, true); + break; + } +} + MICROPROFILE_DEFINE(OpenGL_SurfaceLoad, "OpenGL", "Surface Load", MP_RGB(128, 64, 192)); void CachedSurface::LoadGLBuffer() { - ASSERT(params.type != SurfaceType::Fill); - - const u8* const texture_src_data = Memory::GetPointer(params.addr); - - ASSERT(texture_src_data); - - const u32 bytes_per_pixel = GetGLBytesPerPixel(params.pixel_format); - const u32 copy_size = params.width * params.height * bytes_per_pixel; - const std::size_t total_size = copy_size * params.depth; - MICROPROFILE_SCOPE(OpenGL_SurfaceLoad); + gl_buffer.resize(params.size_in_bytes_gl); if (params.is_tiled) { - gl_buffer.resize(total_size); + u32 depth = params.depth; + u32 block_depth = params.block_depth; ASSERT_MSG(params.block_width == 1, "Block width is defined as {} on texture type {}", params.block_width, static_cast<u32>(params.target)); - ASSERT_MSG(params.block_depth == 1, "Block depth is defined as {} on texture type {}", - params.block_depth, static_cast<u32>(params.target)); - // TODO(bunnei): This only unswizzles and copies a 2D texture - we do not yet know how to do - // this for 3D textures, etc. - switch (params.target) { - case SurfaceParams::SurfaceTarget::Texture2D: - // Pass impl. to the fallback code below - break; - case SurfaceParams::SurfaceTarget::Texture2DArray: - case SurfaceParams::SurfaceTarget::TextureCubemap: - for (std::size_t index = 0; index < params.depth; ++index) { - const std::size_t offset{index * copy_size}; - morton_to_gl_fns[static_cast<std::size_t>(params.pixel_format)]( - params.width, params.block_height, params.height, gl_buffer.data() + offset, - copy_size, params.addr + offset); - } - break; - default: - LOG_CRITICAL(HW_GPU, "Unimplemented tiled load for target={}", - static_cast<u32>(params.target)); - UNREACHABLE(); + if (params.target == SurfaceParams::SurfaceTarget::Texture2D) { + // TODO(Blinkhawk): Eliminate this condition once all texture types are implemented. + depth = 1U; + block_depth = 1U; } morton_to_gl_fns[static_cast<std::size_t>(params.pixel_format)]( - params.width, params.block_height, params.height, gl_buffer.data(), copy_size, - params.addr); + params.width, params.block_height, params.height, block_depth, depth, gl_buffer.data(), + gl_buffer.size(), params.addr); } else { - const u8* const texture_src_data_end{texture_src_data + total_size}; + const auto texture_src_data{Memory::GetPointer(params.addr)}; + const auto texture_src_data_end{texture_src_data + params.size_in_bytes_gl}; gl_buffer.assign(texture_src_data, texture_src_data_end); } @@ -867,7 +907,44 @@ void CachedSurface::LoadGLBuffer() { MICROPROFILE_DEFINE(OpenGL_SurfaceFlush, "OpenGL", "Surface Flush", MP_RGB(128, 192, 64)); void CachedSurface::FlushGLBuffer() { - ASSERT_MSG(false, "Unimplemented"); + MICROPROFILE_SCOPE(OpenGL_SurfaceFlush); + + ASSERT_MSG(!IsPixelFormatASTC(params.pixel_format), "Unimplemented"); + + // OpenGL temporary buffer needs to be big enough to store raw texture size + gl_buffer.resize(GetSizeInBytes()); + + const FormatTuple& tuple = GetFormatTuple(params.pixel_format, params.component_type); + // Ensure no bad interactions with GL_UNPACK_ALIGNMENT + ASSERT(params.width * SurfaceParams::GetBytesPerPixel(params.pixel_format) % 4 == 0); + glPixelStorei(GL_PACK_ROW_LENGTH, static_cast<GLint>(params.width)); + ASSERT(!tuple.compressed); + glBindBuffer(GL_PIXEL_PACK_BUFFER, 0); + glGetTextureImage(texture.handle, 0, tuple.format, tuple.type, gl_buffer.size(), + gl_buffer.data()); + glPixelStorei(GL_PACK_ROW_LENGTH, 0); + ConvertFormatAsNeeded_FlushGLBuffer(gl_buffer, params.pixel_format, params.width, + params.height); + ASSERT(params.type != SurfaceType::Fill); + const u8* const texture_src_data = Memory::GetPointer(params.addr); + ASSERT(texture_src_data); + if (params.is_tiled) { + u32 depth = params.depth; + u32 block_depth = params.block_depth; + + ASSERT_MSG(params.block_width == 1, "Block width is defined as {} on texture type {}", + params.block_width, static_cast<u32>(params.target)); + + if (params.target == SurfaceParams::SurfaceTarget::Texture2D) { + // TODO(Blinkhawk): Eliminate this condition once all texture types are implemented. + depth = 1U; + } + gl_to_morton_fns[static_cast<size_t>(params.pixel_format)]( + params.width, params.block_height, params.height, block_depth, depth, gl_buffer.data(), + gl_buffer.size(), GetAddr()); + } else { + std::memcpy(Memory::GetPointer(GetAddr()), gl_buffer.data(), GetSizeInBytes()); + } } MICROPROFILE_DEFINE(OpenGL_TextureUL, "OpenGL", "Texture Upload", MP_RGB(128, 64, 192)); @@ -877,9 +954,6 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle MICROPROFILE_SCOPE(OpenGL_TextureUL); - ASSERT(gl_buffer.size() == static_cast<std::size_t>(params.width) * params.height * - GetGLBytesPerPixel(params.pixel_format) * params.depth); - const auto& rect{params.GetRect()}; // Load data from memory to the surface @@ -888,7 +962,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle std::size_t buffer_offset = static_cast<std::size_t>(static_cast<std::size_t>(y0) * params.width + static_cast<std::size_t>(x0)) * - GetGLBytesPerPixel(params.pixel_format); + SurfaceParams::GetBytesPerPixel(params.pixel_format); const FormatTuple& tuple = GetFormatTuple(params.pixel_format, params.component_type); const GLuint target_tex = texture.handle; @@ -904,7 +978,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle cur_state.Apply(); // Ensure no bad interactions with GL_UNPACK_ALIGNMENT - ASSERT(params.width * GetGLBytesPerPixel(params.pixel_format) % 4 == 0); + ASSERT(params.width * SurfaceParams::GetBytesPerPixel(params.pixel_format) % 4 == 0); glPixelStorei(GL_UNPACK_ROW_LENGTH, static_cast<GLint>(params.width)); glActiveTexture(GL_TEXTURE0); @@ -914,7 +988,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle glCompressedTexImage2D( SurfaceTargetToGL(params.target), 0, tuple.internal_format, static_cast<GLsizei>(params.width), static_cast<GLsizei>(params.height), 0, - static_cast<GLsizei>(params.size_in_bytes_2d), &gl_buffer[buffer_offset]); + static_cast<GLsizei>(params.size_in_bytes_gl), &gl_buffer[buffer_offset]); break; case SurfaceParams::SurfaceTarget::Texture3D: case SurfaceParams::SurfaceTarget::Texture2DArray: @@ -922,16 +996,16 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle SurfaceTargetToGL(params.target), 0, tuple.internal_format, static_cast<GLsizei>(params.width), static_cast<GLsizei>(params.height), static_cast<GLsizei>(params.depth), 0, - static_cast<GLsizei>(params.size_in_bytes_total), &gl_buffer[buffer_offset]); + static_cast<GLsizei>(params.size_in_bytes_gl), &gl_buffer[buffer_offset]); break; case SurfaceParams::SurfaceTarget::TextureCubemap: for (std::size_t face = 0; face < params.depth; ++face) { glCompressedTexImage2D(static_cast<GLenum>(GL_TEXTURE_CUBE_MAP_POSITIVE_X + face), 0, tuple.internal_format, static_cast<GLsizei>(params.width), static_cast<GLsizei>(params.height), 0, - static_cast<GLsizei>(params.size_in_bytes_2d), + static_cast<GLsizei>(params.SizeInBytesCubeFaceGL()), &gl_buffer[buffer_offset]); - buffer_offset += params.size_in_bytes_2d; + buffer_offset += params.SizeInBytesCubeFace(); } break; default: @@ -941,7 +1015,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle glCompressedTexImage2D( GL_TEXTURE_2D, 0, tuple.internal_format, static_cast<GLsizei>(params.width), static_cast<GLsizei>(params.height), 0, - static_cast<GLsizei>(params.size_in_bytes_2d), &gl_buffer[buffer_offset]); + static_cast<GLsizei>(params.size_in_bytes_gl), &gl_buffer[buffer_offset]); } } else { @@ -970,7 +1044,7 @@ void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle y0, static_cast<GLsizei>(rect.GetWidth()), static_cast<GLsizei>(rect.GetHeight()), tuple.format, tuple.type, &gl_buffer[buffer_offset]); - buffer_offset += params.size_in_bytes_2d; + buffer_offset += params.SizeInBytesCubeFace(); } break; default: @@ -1032,10 +1106,7 @@ Surface RasterizerCacheOpenGL::GetColorBufferSurface(std::size_t index, bool pre void RasterizerCacheOpenGL::LoadSurface(const Surface& surface) { surface->LoadGLBuffer(); surface->UploadGLTexture(read_framebuffer.handle, draw_framebuffer.handle); -} - -void RasterizerCacheOpenGL::FlushSurface(const Surface& surface) { - surface->FlushGLBuffer(); + surface->MarkAsModified(false, *this); } Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool preserve_contents) { @@ -1052,8 +1123,8 @@ Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool pres } else if (preserve_contents) { // If surface parameters changed and we care about keeping the previous data, recreate // the surface from the old one - Unregister(surface); Surface new_surface{RecreateSurface(surface, params)}; + Unregister(surface); Register(new_surface); return new_surface; } else { @@ -1104,6 +1175,14 @@ void RasterizerCacheOpenGL::FermiCopySurface( FastCopySurface(GetSurface(src_params, true), GetSurface(dst_params, false)); } +void RasterizerCacheOpenGL::AccurateCopySurface(const Surface& src_surface, + const Surface& dst_surface) { + const auto& src_params{src_surface->GetSurfaceParams()}; + const auto& dst_params{dst_surface->GetSurfaceParams()}; + FlushRegion(src_params.addr, dst_params.size_in_bytes); + LoadSurface(dst_surface); +} + Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface, const SurfaceParams& new_params) { // Verify surface is compatible for blitting @@ -1112,6 +1191,12 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface, // Get a new surface with the new parameters, and blit the previous surface to it Surface new_surface{GetUncachedSurface(new_params)}; + // With use_accurate_gpu_emulation enabled, do an accurate surface copy + if (Settings::values.use_accurate_gpu_emulation) { + AccurateCopySurface(old_surface, new_surface); + return new_surface; + } + // For compatible surfaces, we can just do fast glCopyImageSubData based copy if (old_params.target == new_params.target && old_params.type == new_params.type && old_params.depth == new_params.depth && old_params.depth == 1 && @@ -1123,11 +1208,10 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface, // If the format is the same, just do a framebuffer blit. This is significantly faster than // using PBOs. The is also likely less accurate, as textures will be converted rather than - // reinterpreted. When use_accurate_framebuffers setting is enabled, perform a more accurate + // reinterpreted. When use_accurate_gpu_emulation setting is enabled, perform a more accurate // surface copy, where pixels are reinterpreted as a new format (without conversion). This // code path uses OpenGL PBOs and is quite slow. - const bool is_blit{old_params.pixel_format == new_params.pixel_format || - !Settings::values.use_accurate_framebuffers}; + const bool is_blit{old_params.pixel_format == new_params.pixel_format}; switch (new_params.target) { case SurfaceParams::SurfaceTarget::Texture2D: @@ -1137,6 +1221,9 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& old_surface, CopySurface(old_surface, new_surface, copy_pbo.handle); } break; + case SurfaceParams::SurfaceTarget::Texture3D: + AccurateCopySurface(old_surface, new_surface); + break; case SurfaceParams::SurfaceTarget::TextureCubemap: { if (old_params.rt.array_mode != 1) { // TODO(bunnei): This is used by Breath of the Wild, I'm not sure how to implement this diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.h b/src/video_core/renderer_opengl/gl_rasterizer_cache.h index 66d98ad4e..0dd0d90a3 100644 --- a/src/video_core/renderer_opengl/gl_rasterizer_cache.h +++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.h @@ -18,6 +18,7 @@ #include "video_core/rasterizer_cache.h" #include "video_core/renderer_opengl/gl_resource_manager.h" #include "video_core/renderer_opengl/gl_shader_gen.h" +#include "video_core/textures/decoders.h" #include "video_core/textures/texture.h" namespace OpenGL { @@ -74,19 +75,21 @@ struct SurfaceParams { RG32UI = 43, R32UI = 44, ASTC_2D_8X8 = 45, + ASTC_2D_8X5 = 46, + ASTC_2D_5X4 = 47, MaxColorFormat, // Depth formats - Z32F = 46, - Z16 = 47, + Z32F = 48, + Z16 = 49, MaxDepthFormat, // DepthStencil formats - Z24S8 = 48, - S8Z24 = 49, - Z32FS8 = 50, + Z24S8 = 50, + S8Z24 = 51, + Z32FS8 = 52, MaxDepthStencilFormat, @@ -129,6 +132,8 @@ struct SurfaceParams { case Tegra::Texture::TextureType::Texture2D: case Tegra::Texture::TextureType::Texture2DNoMipmap: return SurfaceTarget::Texture2D; + case Tegra::Texture::TextureType::Texture3D: + return SurfaceTarget::Texture3D; case Tegra::Texture::TextureType::TextureCubemap: return SurfaceTarget::TextureCubemap; case Tegra::Texture::TextureType::Texture1DArray: @@ -220,6 +225,8 @@ struct SurfaceParams { 1, // RG32UI 1, // R32UI 4, // ASTC_2D_8X8 + 4, // ASTC_2D_8X5 + 4, // ASTC_2D_5X4 1, // Z32F 1, // Z16 1, // Z24S8 @@ -282,6 +289,8 @@ struct SurfaceParams { 64, // RG32UI 32, // R32UI 16, // ASTC_2D_8X8 + 32, // ASTC_2D_8X5 + 32, // ASTC_2D_5X4 32, // Z32F 16, // Z16 32, // Z24S8 @@ -553,8 +562,12 @@ struct SurfaceParams { return PixelFormat::BC6H_SF16; case Tegra::Texture::TextureFormat::ASTC_2D_4X4: return PixelFormat::ASTC_2D_4X4; + case Tegra::Texture::TextureFormat::ASTC_2D_5X4: + return PixelFormat::ASTC_2D_5X4; case Tegra::Texture::TextureFormat::ASTC_2D_8X8: return PixelFormat::ASTC_2D_8X8; + case Tegra::Texture::TextureFormat::ASTC_2D_8X5: + return PixelFormat::ASTC_2D_8X5; case Tegra::Texture::TextureFormat::R16_G16: switch (component_type) { case Tegra::Texture::ComponentType::FLOAT: @@ -691,21 +704,42 @@ struct SurfaceParams { return SurfaceType::Invalid; } + /// Returns the sizer in bytes of the specified pixel format + static constexpr u32 GetBytesPerPixel(PixelFormat pixel_format) { + if (pixel_format == SurfaceParams::PixelFormat::Invalid) { + return 0; + } + return GetFormatBpp(pixel_format) / CHAR_BIT; + } + /// Returns the rectangle corresponding to this surface MathUtil::Rectangle<u32> GetRect() const; - /// Returns the size of this surface as a 2D texture in bytes, adjusted for compression - std::size_t SizeInBytes2D() const { + /// Returns the total size of this surface in bytes, adjusted for compression + std::size_t SizeInBytesRaw(bool ignore_tiled = false) const { const u32 compression_factor{GetCompressionFactor(pixel_format)}; - ASSERT(width % compression_factor == 0); - ASSERT(height % compression_factor == 0); - return (width / compression_factor) * (height / compression_factor) * - GetFormatBpp(pixel_format) / CHAR_BIT; + const u32 bytes_per_pixel{GetBytesPerPixel(pixel_format)}; + const size_t uncompressed_size{ + Tegra::Texture::CalculateSize((ignore_tiled ? false : is_tiled), bytes_per_pixel, width, + height, depth, block_height, block_depth)}; + + // Divide by compression_factor^2, as height and width are factored by this + return uncompressed_size / (compression_factor * compression_factor); } - /// Returns the total size of this surface in bytes, adjusted for compression - std::size_t SizeInBytesTotal() const { - return SizeInBytes2D() * depth; + /// Returns the size of this surface as an OpenGL texture in bytes + std::size_t SizeInBytesGL() const { + return SizeInBytesRaw(true); + } + + /// Returns the size of this surface as a cube face in bytes + std::size_t SizeInBytesCubeFace() const { + return size_in_bytes / 6; + } + + /// Returns the size of this surface as an OpenGL cube face in bytes + std::size_t SizeInBytesCubeFaceGL() const { + return size_in_bytes_gl / 6; } /// Creates SurfaceParams from a texture configuration @@ -732,7 +766,9 @@ struct SurfaceParams { other.depth); } - VAddr addr; + /// Initializes parameters for caching, should be called after everything has been initialized + void InitCacheParameters(Tegra::GPUVAddr gpu_addr); + bool is_tiled; u32 block_width; u32 block_height; @@ -744,15 +780,20 @@ struct SurfaceParams { u32 height; u32 depth; u32 unaligned_height; - std::size_t size_in_bytes_total; - std::size_t size_in_bytes_2d; SurfaceTarget target; u32 max_mip_level; + // Parameters used for caching + VAddr addr; + Tegra::GPUVAddr gpu_addr; + std::size_t size_in_bytes; + std::size_t size_in_bytes_gl; + // Render target specific parameters, not used in caching struct { u32 index; u32 array_mode; + u32 volume; u32 layer_stride; u32 base_layer; } rt; @@ -765,7 +806,8 @@ struct SurfaceReserveKey : Common::HashableStruct<OpenGL::SurfaceParams> { static SurfaceReserveKey Create(const OpenGL::SurfaceParams& params) { SurfaceReserveKey res; res.state = params; - res.state.rt = {}; // Ignore rt config in caching + res.state.gpu_addr = {}; // Ignore GPU vaddr in caching + res.state.rt = {}; // Ignore rt config in caching return res; } }; @@ -780,16 +822,20 @@ struct hash<SurfaceReserveKey> { namespace OpenGL { -class CachedSurface final { +class CachedSurface final : public RasterizerCacheObject { public: CachedSurface(const SurfaceParams& params); - VAddr GetAddr() const { + VAddr GetAddr() const override { return params.addr; } - std::size_t GetSizeInBytes() const { - return params.size_in_bytes_total; + std::size_t GetSizeInBytes() const override { + return cached_size_in_bytes; + } + + void Flush() override { + FlushGLBuffer(); } const OGLTexture& Texture() const { @@ -800,13 +846,6 @@ public: return gl_target; } - static constexpr unsigned int GetGLBytesPerPixel(SurfaceParams::PixelFormat format) { - if (format == SurfaceParams::PixelFormat::Invalid) - return 0; - - return SurfaceParams::GetFormatBpp(format) / CHAR_BIT; - } - const SurfaceParams& GetSurfaceParams() const { return params; } @@ -823,6 +862,7 @@ private: std::vector<u8> gl_buffer; SurfaceParams params; GLenum gl_target; + std::size_t cached_size_in_bytes; }; class RasterizerCacheOpenGL final : public RasterizerCache<Surface> { @@ -839,9 +879,6 @@ public: /// Get the color surface based on the framebuffer configuration and the specified render target Surface GetColorBufferSurface(std::size_t index, bool preserve_contents); - /// Flushes the surface to Switch memory - void FlushSurface(const Surface& surface); - /// Tries to find a framebuffer using on the provided CPU address Surface TryFindFramebufferSurface(VAddr addr) const; @@ -865,6 +902,9 @@ private: /// Tries to get a reserved surface for the specified parameters Surface TryGetReservedSurface(const SurfaceParams& params); + /// Performs a slow but accurate surface copy, flushing to RAM and reinterpreting the data + void AccurateCopySurface(const Surface& src_surface, const Surface& dst_surface); + /// 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 /// destroyed when used with different surface parameters. diff --git a/src/video_core/renderer_opengl/gl_shader_cache.h b/src/video_core/renderer_opengl/gl_shader_cache.h index 7bb287f56..a210f1731 100644 --- a/src/video_core/renderer_opengl/gl_shader_cache.h +++ b/src/video_core/renderer_opengl/gl_shader_cache.h @@ -19,20 +19,21 @@ class CachedShader; using Shader = std::shared_ptr<CachedShader>; using Maxwell = Tegra::Engines::Maxwell3D::Regs; -class CachedShader final { +class CachedShader final : public RasterizerCacheObject { public: CachedShader(VAddr addr, Maxwell::ShaderProgram program_type); - /// Gets the address of the shader in guest memory, required for cache management - VAddr GetAddr() const { + VAddr GetAddr() const override { return addr; } - /// Gets the size of the shader in guest memory, required for cache management - std::size_t GetSizeInBytes() const { + std::size_t GetSizeInBytes() const override { return GLShader::MAX_PROGRAM_CODE_LENGTH * sizeof(u64); } + // We do not have to flush this cache as things in it are never modified by us. + void Flush() override {} + /// Gets the shader entries for the shader const GLShader::ShaderEntries& GetShaderEntries() const { return entries; diff --git a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp index 23349b1a1..e050b063a 100644 --- a/src/video_core/renderer_opengl/gl_shader_decompiler.cpp +++ b/src/video_core/renderer_opengl/gl_shader_decompiler.cpp @@ -1233,6 +1233,7 @@ private: case Tegra::Shader::TextureType::Texture2D: { return 2; } + case Tegra::Shader::TextureType::Texture3D: case Tegra::Shader::TextureType::TextureCube: { return 3; } @@ -1527,7 +1528,6 @@ private: break; } - case OpCode::Type::Shift: { std::string op_a = regs.GetRegisterAsInteger(instr.gpr8, 0, true); std::string op_b; @@ -1569,7 +1569,6 @@ private: } break; } - case OpCode::Type::ArithmeticIntegerImmediate: { std::string op_a = regs.GetRegisterAsInteger(instr.gpr8); std::string op_b = std::to_string(instr.alu.imm20_32.Value()); @@ -2262,9 +2261,9 @@ private: break; } case OpCode::Id::TEX: { - ASSERT_MSG(instr.tex.array == 0, "TEX arrays unimplemented"); Tegra::Shader::TextureType texture_type{instr.tex.texture_type}; std::string coord; + const bool is_array = instr.tex.array != 0; ASSERT_MSG(!instr.tex.UsesMiscMode(Tegra::Shader::TextureMiscMode::NODEP), "NODEP is not implemented"); @@ -2279,21 +2278,59 @@ private: switch (num_coordinates) { case 1: { - const std::string x = regs.GetRegisterAsFloat(instr.gpr8); - coord = "float coords = " + x + ';'; + if (is_array) { + const std::string index = regs.GetRegisterAsInteger(instr.gpr8); + const std::string x = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); + coord = "vec2 coords = vec2(" + x + ", " + index + ");"; + } else { + const std::string x = regs.GetRegisterAsFloat(instr.gpr8); + coord = "float coords = " + x + ';'; + } break; } case 2: { - const std::string x = regs.GetRegisterAsFloat(instr.gpr8); - const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); - coord = "vec2 coords = vec2(" + x + ", " + y + ");"; + if (is_array) { + const std::string index = regs.GetRegisterAsInteger(instr.gpr8); + const std::string x = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); + const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 2); + coord = "vec3 coords = vec3(" + x + ", " + y + ", " + index + ");"; + } else { + const std::string x = regs.GetRegisterAsFloat(instr.gpr8); + const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); + coord = "vec2 coords = vec2(" + x + ", " + y + ");"; + } break; } case 3: { - const std::string x = regs.GetRegisterAsFloat(instr.gpr8); - const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); - const std::string z = regs.GetRegisterAsFloat(instr.gpr20); - coord = "vec3 coords = vec3(" + x + ", " + y + ", " + z + ");"; + if (depth_compare) { + if (is_array) { + const std::string index = regs.GetRegisterAsInteger(instr.gpr8); + const std::string x = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); + const std::string y = regs.GetRegisterAsFloat(instr.gpr20); + const std::string z = regs.GetRegisterAsFloat(instr.gpr20.Value() + 1); + coord = "vec4 coords = vec4(" + x + ", " + y + ", " + z + ", " + index + + ");"; + } else { + const std::string x = regs.GetRegisterAsFloat(instr.gpr8); + const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); + const std::string z = regs.GetRegisterAsFloat(instr.gpr20); + coord = "vec3 coords = vec3(" + x + ", " + y + ", " + z + ");"; + } + } else { + if (is_array) { + const std::string index = regs.GetRegisterAsInteger(instr.gpr8); + const std::string x = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); + const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 2); + const std::string z = regs.GetRegisterAsFloat(instr.gpr8.Value() + 3); + coord = "vec4 coords = vec4(" + x + ", " + y + ", " + z + ", " + index + + ");"; + } else { + const std::string x = regs.GetRegisterAsFloat(instr.gpr8); + const std::string y = regs.GetRegisterAsFloat(instr.gpr8.Value() + 1); + const std::string z = regs.GetRegisterAsFloat(instr.gpr8.Value() + 2); + coord = "vec3 coords = vec3(" + x + ", " + y + ", " + z + ");"; + } + } break; } default: @@ -2312,7 +2349,7 @@ private: std::string op_c; const std::string sampler = - GetSampler(instr.sampler, texture_type, false, depth_compare); + GetSampler(instr.sampler, texture_type, is_array, depth_compare); // Add an extra scope and declare the texture coords inside to prevent // overwriting them in case they are used as outputs of the texs instruction. @@ -2332,10 +2369,13 @@ private: } case Tegra::Shader::TextureProcessMode::LB: case Tegra::Shader::TextureProcessMode::LBA: { - if (num_coordinates <= 2) { - op_c = regs.GetRegisterAsFloat(instr.gpr20); + if (depth_compare) { + if (is_array) + op_c = regs.GetRegisterAsFloat(instr.gpr20.Value() + 2); + else + op_c = regs.GetRegisterAsFloat(instr.gpr20.Value() + 1); } else { - op_c = regs.GetRegisterAsFloat(instr.gpr20.Value() + 1); + op_c = regs.GetRegisterAsFloat(instr.gpr20); } // TODO: Figure if A suffix changes the equation at all. texture = "texture(" + sampler + ", coords, " + op_c + ')'; @@ -2478,6 +2518,8 @@ private: ASSERT_MSG(!instr.tlds.UsesMiscMode(Tegra::Shader::TextureMiscMode::MZ), "MZ is not implemented"); + u32 op_c_offset = 0; + switch (texture_type) { case Tegra::Shader::TextureType::Texture1D: { const std::string x = regs.GetRegisterAsInteger(instr.gpr8); @@ -2492,6 +2534,7 @@ private: const std::string x = regs.GetRegisterAsInteger(instr.gpr8); const std::string y = regs.GetRegisterAsInteger(instr.gpr20); coord = "ivec2 coords = ivec2(" + x + ", " + y + ");"; + op_c_offset = 1; } break; } @@ -2503,13 +2546,14 @@ private: const std::string sampler = GetSampler(instr.sampler, texture_type, is_array, false); std::string texture = "texelFetch(" + sampler + ", coords, 0)"; - const std::string op_c = regs.GetRegisterAsInteger(instr.gpr20.Value() + 1); switch (instr.tlds.GetTextureProcessMode()) { case Tegra::Shader::TextureProcessMode::LZ: { texture = "texelFetch(" + sampler + ", coords, 0)"; break; } case Tegra::Shader::TextureProcessMode::LL: { + const std::string op_c = + regs.GetRegisterAsInteger(instr.gpr20.Value() + op_c_offset); texture = "texelFetch(" + sampler + ", coords, " + op_c + ')'; break; } @@ -2895,14 +2939,14 @@ private: const std::string pred = GetPredicateCondition(instr.csetp.pred39, instr.csetp.neg_pred39 != 0); const std::string combiner = GetPredicateCombiner(instr.csetp.op); - const std::string controlCode = regs.GetControlCode(instr.csetp.cc); + const std::string control_code = regs.GetControlCode(instr.csetp.cc); if (instr.csetp.pred3 != static_cast<u64>(Pred::UnusedIndex)) { SetPredicate(instr.csetp.pred3, - '(' + controlCode + ") " + combiner + " (" + pred + ')'); + '(' + control_code + ") " + combiner + " (" + pred + ')'); } if (instr.csetp.pred0 != static_cast<u64>(Pred::UnusedIndex)) { SetPredicate(instr.csetp.pred0, - "!(" + controlCode + ") " + combiner + " (" + pred + ')'); + "!(" + control_code + ") " + combiner + " (" + pred + ')'); } break; } diff --git a/src/video_core/textures/decoders.cpp b/src/video_core/textures/decoders.cpp index 0d2456b56..f1b40e7f5 100644 --- a/src/video_core/textures/decoders.cpp +++ b/src/video_core/textures/decoders.cpp @@ -40,72 +40,146 @@ struct alignas(64) SwizzleTable { constexpr auto legacy_swizzle_table = SwizzleTable<8, 64, 1>(); constexpr auto fast_swizzle_table = SwizzleTable<8, 4, 16>(); -static void LegacySwizzleData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel, - u8* swizzled_data, u8* unswizzled_data, bool unswizzle, - u32 block_height) { +/** + * This function manages ALL the GOBs(Group of Bytes) Inside a single block. + * Instead of going gob by gob, we map the coordinates inside a block and manage from + * those. Block_Width is assumed to be 1. + */ +void PreciseProcessBlock(u8* swizzled_data, u8* unswizzled_data, const bool unswizzle, + const u32 x_start, const u32 y_start, const u32 z_start, const u32 x_end, + const u32 y_end, const u32 z_end, const u32 tile_offset, + const u32 xy_block_size, const u32 layer_z, const u32 stride_x, + const u32 bytes_per_pixel, const u32 out_bytes_per_pixel) { std::array<u8*, 2> data_ptrs; - const std::size_t stride = width * bytes_per_pixel; - const std::size_t gobs_in_x = 64; - const std::size_t gobs_in_y = 8; - const std::size_t gobs_size = gobs_in_x * gobs_in_y; - const std::size_t image_width_in_gobs{(stride + gobs_in_x - 1) / gobs_in_x}; - for (std::size_t y = 0; y < height; ++y) { - const std::size_t gob_y_address = - (y / (gobs_in_y * block_height)) * gobs_size * block_height * image_width_in_gobs + - (y % (gobs_in_y * block_height) / gobs_in_y) * gobs_size; - const auto& table = legacy_swizzle_table[y % gobs_in_y]; - for (std::size_t x = 0; x < width; ++x) { - const std::size_t gob_address = - gob_y_address + (x * bytes_per_pixel / gobs_in_x) * gobs_size * block_height; - const std::size_t x2 = x * bytes_per_pixel; - const std::size_t swizzle_offset = gob_address + table[x2 % gobs_in_x]; - const std::size_t pixel_index = (x + y * width) * out_bytes_per_pixel; - - data_ptrs[unswizzle] = swizzled_data + swizzle_offset; - data_ptrs[!unswizzle] = unswizzled_data + pixel_index; - - std::memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel); + u32 z_address = tile_offset; + const u32 gob_size_x = 64; + const u32 gob_size_y = 8; + const u32 gob_size_z = 1; + const u32 gob_size = gob_size_x * gob_size_y * gob_size_z; + for (u32 z = z_start; z < z_end; z++) { + u32 y_address = z_address; + u32 pixel_base = layer_z * z + y_start * stride_x; + for (u32 y = y_start; y < y_end; y++) { + const auto& table = legacy_swizzle_table[y % gob_size_y]; + for (u32 x = x_start; x < x_end; x++) { + const u32 swizzle_offset{y_address + table[x * bytes_per_pixel % gob_size_x]}; + const u32 pixel_index{x * out_bytes_per_pixel + pixel_base}; + data_ptrs[unswizzle] = swizzled_data + swizzle_offset; + data_ptrs[!unswizzle] = unswizzled_data + pixel_index; + std::memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel); + } + pixel_base += stride_x; + if ((y + 1) % gob_size_y == 0) + y_address += gob_size; } + z_address += xy_block_size; } } -static void FastSwizzleData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel, - u8* swizzled_data, u8* unswizzled_data, bool unswizzle, - u32 block_height) { +/** + * This function manages ALL the GOBs(Group of Bytes) Inside a single block. + * Instead of going gob by gob, we map the coordinates inside a block and manage from + * those. Block_Width is assumed to be 1. + */ +void FastProcessBlock(u8* swizzled_data, u8* unswizzled_data, const bool unswizzle, + const u32 x_start, const u32 y_start, const u32 z_start, const u32 x_end, + const u32 y_end, const u32 z_end, const u32 tile_offset, + const u32 xy_block_size, const u32 layer_z, const u32 stride_x, + const u32 bytes_per_pixel, const u32 out_bytes_per_pixel) { std::array<u8*, 2> data_ptrs; - const std::size_t stride{width * bytes_per_pixel}; - const std::size_t gobs_in_x = 64; - const std::size_t gobs_in_y = 8; - const std::size_t gobs_size = gobs_in_x * gobs_in_y; - const std::size_t image_width_in_gobs{(stride + gobs_in_x - 1) / gobs_in_x}; - const std::size_t copy_size{16}; - for (std::size_t y = 0; y < height; ++y) { - const std::size_t initial_gob = - (y / (gobs_in_y * block_height)) * gobs_size * block_height * image_width_in_gobs + - (y % (gobs_in_y * block_height) / gobs_in_y) * gobs_size; - const std::size_t pixel_base{y * width * out_bytes_per_pixel}; - const auto& table = fast_swizzle_table[y % gobs_in_y]; - for (std::size_t xb = 0; xb < stride; xb += copy_size) { - const std::size_t gob_address{initial_gob + - (xb / gobs_in_x) * gobs_size * block_height}; - const std::size_t swizzle_offset{gob_address + table[(xb / 16) % 4]}; - const std::size_t out_x = xb * out_bytes_per_pixel / bytes_per_pixel; - const std::size_t pixel_index{out_x + pixel_base}; - data_ptrs[unswizzle] = swizzled_data + swizzle_offset; - data_ptrs[!unswizzle] = unswizzled_data + pixel_index; - std::memcpy(data_ptrs[0], data_ptrs[1], copy_size); + u32 z_address = tile_offset; + const u32 x_startb = x_start * bytes_per_pixel; + const u32 x_endb = x_end * bytes_per_pixel; + const u32 copy_size = 16; + const u32 gob_size_x = 64; + const u32 gob_size_y = 8; + const u32 gob_size_z = 1; + const u32 gob_size = gob_size_x * gob_size_y * gob_size_z; + for (u32 z = z_start; z < z_end; z++) { + u32 y_address = z_address; + u32 pixel_base = layer_z * z + y_start * stride_x; + for (u32 y = y_start; y < y_end; y++) { + const auto& table = fast_swizzle_table[y % gob_size_y]; + for (u32 xb = x_startb; xb < x_endb; xb += copy_size) { + const u32 swizzle_offset{y_address + table[(xb / copy_size) % 4]}; + const u32 out_x = xb * out_bytes_per_pixel / bytes_per_pixel; + const u32 pixel_index{out_x + pixel_base}; + data_ptrs[unswizzle] = swizzled_data + swizzle_offset; + data_ptrs[!unswizzle] = unswizzled_data + pixel_index; + std::memcpy(data_ptrs[0], data_ptrs[1], copy_size); + } + pixel_base += stride_x; + if ((y + 1) % gob_size_y == 0) + y_address += gob_size; } + z_address += xy_block_size; } } -void CopySwizzledData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel, - u8* swizzled_data, u8* unswizzled_data, bool unswizzle, u32 block_height) { +/** + * This function unswizzles or swizzles a texture by mapping Linear to BlockLinear Textue. + * The body of this function takes care of splitting the swizzled texture into blocks, + * and managing the extents of it. Once all the parameters of a single block are obtained, + * the function calls 'ProcessBlock' to process that particular Block. + * + * Documentation for the memory layout and decoding can be found at: + * https://envytools.readthedocs.io/en/latest/hw/memory/g80-surface.html#blocklinear-surfaces + */ +template <bool fast> +void SwizzledData(u8* swizzled_data, u8* unswizzled_data, const bool unswizzle, const u32 width, + const u32 height, const u32 depth, const u32 bytes_per_pixel, + const u32 out_bytes_per_pixel, const u32 block_height, const u32 block_depth) { + auto div_ceil = [](const u32 x, const u32 y) { return ((x + y - 1) / y); }; + const u32 stride_x = width * out_bytes_per_pixel; + const u32 layer_z = height * stride_x; + const u32 gob_x_bytes = 64; + const u32 gob_elements_x = gob_x_bytes / bytes_per_pixel; + const u32 gob_elements_y = 8; + const u32 gob_elements_z = 1; + const u32 block_x_elements = gob_elements_x; + const u32 block_y_elements = gob_elements_y * block_height; + const u32 block_z_elements = gob_elements_z * block_depth; + const u32 blocks_on_x = div_ceil(width, block_x_elements); + const u32 blocks_on_y = div_ceil(height, block_y_elements); + const u32 blocks_on_z = div_ceil(depth, block_z_elements); + const u32 blocks = blocks_on_x * blocks_on_y * blocks_on_z; + const u32 gob_size = gob_x_bytes * gob_elements_y * gob_elements_z; + const u32 xy_block_size = gob_size * block_height; + const u32 block_size = xy_block_size * block_depth; + u32 tile_offset = 0; + for (u32 zb = 0; zb < blocks_on_z; zb++) { + const u32 z_start = zb * block_z_elements; + const u32 z_end = std::min(depth, z_start + block_z_elements); + for (u32 yb = 0; yb < blocks_on_y; yb++) { + const u32 y_start = yb * block_y_elements; + const u32 y_end = std::min(height, y_start + block_y_elements); + for (u32 xb = 0; xb < blocks_on_x; xb++) { + const u32 x_start = xb * block_x_elements; + const u32 x_end = std::min(width, x_start + block_x_elements); + if (fast) { + FastProcessBlock(swizzled_data, unswizzled_data, unswizzle, x_start, y_start, + z_start, x_end, y_end, z_end, tile_offset, xy_block_size, + layer_z, stride_x, bytes_per_pixel, out_bytes_per_pixel); + } else { + PreciseProcessBlock(swizzled_data, unswizzled_data, unswizzle, x_start, y_start, + z_start, x_end, y_end, z_end, tile_offset, xy_block_size, + layer_z, stride_x, bytes_per_pixel, out_bytes_per_pixel); + } + tile_offset += block_size; + } + } + } +} + +void CopySwizzledData(u32 width, u32 height, u32 depth, u32 bytes_per_pixel, + u32 out_bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data, + bool unswizzle, u32 block_height, u32 block_depth) { if (bytes_per_pixel % 3 != 0 && (width * bytes_per_pixel) % 16 == 0) { - FastSwizzleData(width, height, bytes_per_pixel, out_bytes_per_pixel, swizzled_data, - unswizzled_data, unswizzle, block_height); + SwizzledData<true>(swizzled_data, unswizzled_data, unswizzle, width, height, depth, + bytes_per_pixel, out_bytes_per_pixel, block_height, block_depth); } else { - LegacySwizzleData(width, height, bytes_per_pixel, out_bytes_per_pixel, swizzled_data, - unswizzled_data, unswizzle, block_height); + SwizzledData<false>(swizzled_data, unswizzled_data, unswizzle, width, height, depth, + bytes_per_pixel, out_bytes_per_pixel, block_height, block_depth); } } @@ -126,7 +200,9 @@ u32 BytesPerPixel(TextureFormat format) { case TextureFormat::R32_G32_B32: return 12; case TextureFormat::ASTC_2D_4X4: + case TextureFormat::ASTC_2D_5X4: case TextureFormat::ASTC_2D_8X8: + case TextureFormat::ASTC_2D_8X5: case TextureFormat::A8R8G8B8: case TextureFormat::A2B10G10R10: case TextureFormat::BF10GF11RF11: @@ -153,13 +229,54 @@ u32 BytesPerPixel(TextureFormat format) { } std::vector<u8> UnswizzleTexture(VAddr address, u32 tile_size, u32 bytes_per_pixel, u32 width, - u32 height, u32 block_height) { - std::vector<u8> unswizzled_data(width * height * bytes_per_pixel); - CopySwizzledData(width / tile_size, height / tile_size, bytes_per_pixel, bytes_per_pixel, - Memory::GetPointer(address), unswizzled_data.data(), true, block_height); + u32 height, u32 depth, u32 block_height, u32 block_depth) { + std::vector<u8> unswizzled_data(width * height * depth * bytes_per_pixel); + CopySwizzledData(width / tile_size, height / tile_size, depth, bytes_per_pixel, bytes_per_pixel, + Memory::GetPointer(address), unswizzled_data.data(), true, block_height, + block_depth); return unswizzled_data; } +void SwizzleSubrect(u32 subrect_width, u32 subrect_height, u32 source_pitch, u32 swizzled_width, + u32 bytes_per_pixel, VAddr swizzled_data, VAddr unswizzled_data, + u32 block_height) { + const u32 image_width_in_gobs{(swizzled_width * bytes_per_pixel + 63) / 64}; + for (u32 line = 0; line < subrect_height; ++line) { + const u32 gob_address_y = + (line / (8 * block_height)) * 512 * block_height * image_width_in_gobs + + (line % (8 * block_height) / 8) * 512; + const auto& table = legacy_swizzle_table[line % 8]; + for (u32 x = 0; x < subrect_width; ++x) { + const u32 gob_address = gob_address_y + (x * bytes_per_pixel / 64) * 512 * block_height; + const u32 swizzled_offset = gob_address + table[(x * bytes_per_pixel) % 64]; + const VAddr source_line = unswizzled_data + line * source_pitch + x * bytes_per_pixel; + const VAddr dest_addr = swizzled_data + swizzled_offset; + + Memory::CopyBlock(dest_addr, source_line, bytes_per_pixel); + } + } +} + +void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32 swizzled_width, + u32 bytes_per_pixel, VAddr swizzled_data, VAddr unswizzled_data, + u32 block_height, u32 offset_x, u32 offset_y) { + for (u32 line = 0; line < subrect_height; ++line) { + const u32 y2 = line + offset_y; + const u32 gob_address_y = + (y2 / (8 * block_height)) * 512 * block_height + (y2 % (8 * block_height) / 8) * 512; + const auto& table = legacy_swizzle_table[y2 % 8]; + for (u32 x = 0; x < subrect_width; ++x) { + const u32 x2 = (x + offset_x) * bytes_per_pixel; + const u32 gob_address = gob_address_y + (x2 / 64) * 512 * block_height; + const u32 swizzled_offset = gob_address + table[x2 % 64]; + const VAddr dest_line = unswizzled_data + line * dest_pitch + x * bytes_per_pixel; + const VAddr source_addr = swizzled_data + swizzled_offset; + + Memory::CopyBlock(dest_line, source_addr, bytes_per_pixel); + } + } +} + std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat format, u32 width, u32 height) { std::vector<u8> rgba_data; diff --git a/src/video_core/textures/decoders.h b/src/video_core/textures/decoders.h index 234d250af..4726f54a5 100644 --- a/src/video_core/textures/decoders.h +++ b/src/video_core/textures/decoders.h @@ -14,17 +14,14 @@ namespace Tegra::Texture { * Unswizzles a swizzled texture without changing its format. */ std::vector<u8> UnswizzleTexture(VAddr address, u32 tile_size, u32 bytes_per_pixel, u32 width, - u32 height, u32 block_height = TICEntry::DefaultBlockHeight); - -/** - * Unswizzles a swizzled depth texture without changing its format. - */ -std::vector<u8> UnswizzleDepthTexture(VAddr address, DepthFormat format, u32 width, u32 height, - u32 block_height = TICEntry::DefaultBlockHeight); + u32 height, u32 depth, + u32 block_height = TICEntry::DefaultBlockHeight, + u32 block_depth = TICEntry::DefaultBlockHeight); /// Copies texture data from a buffer and performs swizzling/unswizzling as necessary. -void CopySwizzledData(u32 width, u32 height, u32 bytes_per_pixel, u32 out_bytes_per_pixel, - u8* swizzled_data, u8* unswizzled_data, bool unswizzle, u32 block_height); +void CopySwizzledData(u32 width, u32 height, u32 depth, u32 bytes_per_pixel, + u32 out_bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data, + bool unswizzle, u32 block_height, u32 block_depth); /** * Decodes an unswizzled texture into a A8R8G8B8 texture. @@ -38,4 +35,13 @@ std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat std::size_t CalculateSize(bool tiled, u32 bytes_per_pixel, u32 width, u32 height, u32 depth, u32 block_height, u32 block_depth); +/// Copies an untiled subrectangle into a tiled surface. +void SwizzleSubrect(u32 subrect_width, u32 subrect_height, u32 source_pitch, u32 swizzled_width, + u32 bytes_per_pixel, VAddr swizzled_data, VAddr unswizzled_data, + u32 block_height); +/// Copies a tiled subrectangle into a linear surface. +void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32 swizzled_width, + u32 bytes_per_pixel, VAddr swizzled_data, VAddr unswizzled_data, + u32 block_height, u32 offset_x, u32 offset_y); + } // namespace Tegra::Texture diff --git a/src/video_core/textures/texture.h b/src/video_core/textures/texture.h index 58d17abcb..5947bd2b9 100644 --- a/src/video_core/textures/texture.h +++ b/src/video_core/textures/texture.h @@ -141,6 +141,7 @@ static_assert(sizeof(TextureHandle) == 4, "TextureHandle has wrong size"); struct TICEntry { static constexpr u32 DefaultBlockHeight = 16; + static constexpr u32 DefaultBlockDepth = 1; union { u32 raw; |