diff options
Diffstat (limited to '')
-rw-r--r-- | src/video_core/renderer_opengl/gl_shader_cache.cpp | 359 |
1 files changed, 212 insertions, 147 deletions
diff --git a/src/video_core/renderer_opengl/gl_shader_cache.cpp b/src/video_core/renderer_opengl/gl_shader_cache.cpp index ac8a9e6b7..42ca3b1bd 100644 --- a/src/video_core/renderer_opengl/gl_shader_cache.cpp +++ b/src/video_core/renderer_opengl/gl_shader_cache.cpp @@ -23,13 +23,13 @@ namespace OpenGL { using VideoCommon::Shader::ProgramCode; -// One UBO is always reserved for emulation values -constexpr u32 RESERVED_UBOS = 1; +// One UBO is always reserved for emulation values on staged shaders +constexpr u32 STAGE_RESERVED_UBOS = 1; struct UnspecializedShader { std::string code; GLShader::ShaderEntries entries; - Maxwell::ShaderProgram program_type; + ProgramType program_type; }; namespace { @@ -55,15 +55,17 @@ ProgramCode GetShaderCode(Tegra::MemoryManager& memory_manager, const GPUVAddr g } /// Gets the shader type from a Maxwell program type -constexpr GLenum GetShaderType(Maxwell::ShaderProgram program_type) { +constexpr GLenum GetShaderType(ProgramType program_type) { switch (program_type) { - case Maxwell::ShaderProgram::VertexA: - case Maxwell::ShaderProgram::VertexB: + case ProgramType::VertexA: + case ProgramType::VertexB: return GL_VERTEX_SHADER; - case Maxwell::ShaderProgram::Geometry: + case ProgramType::Geometry: return GL_GEOMETRY_SHADER; - case Maxwell::ShaderProgram::Fragment: + case ProgramType::Fragment: return GL_FRAGMENT_SHADER; + case ProgramType::Compute: + return GL_COMPUTE_SHADER; default: return GL_NONE; } @@ -100,18 +102,44 @@ constexpr std::tuple<const char*, const char*, u32> GetPrimitiveDescription(GLen } } +ProgramType GetProgramType(Maxwell::ShaderProgram program) { + switch (program) { + case Maxwell::ShaderProgram::VertexA: + return ProgramType::VertexA; + case Maxwell::ShaderProgram::VertexB: + return ProgramType::VertexB; + case Maxwell::ShaderProgram::TesselationControl: + return ProgramType::TessellationControl; + case Maxwell::ShaderProgram::TesselationEval: + return ProgramType::TessellationEval; + case Maxwell::ShaderProgram::Geometry: + return ProgramType::Geometry; + case Maxwell::ShaderProgram::Fragment: + return ProgramType::Fragment; + } + UNREACHABLE(); + return {}; +} + /// Calculates the size of a program stream std::size_t CalculateProgramSize(const GLShader::ProgramCode& program) { constexpr std::size_t start_offset = 10; + // This is the encoded version of BRA that jumps to itself. All Nvidia + // shaders end with one. + constexpr u64 self_jumping_branch = 0xE2400FFFFF07000FULL; + constexpr u64 mask = 0xFFFFFFFFFF7FFFFFULL; std::size_t offset = start_offset; std::size_t size = start_offset * sizeof(u64); while (offset < program.size()) { const u64 instruction = program[offset]; if (!IsSchedInstruction(offset, start_offset)) { - if (instruction == 0 || (instruction >> 52) == 0x50b) { + if ((instruction & mask) == self_jumping_branch) { // End on Maxwell's "nop" instruction break; } + if (instruction == 0) { + break; + } } size += sizeof(u64); offset++; @@ -121,11 +149,13 @@ std::size_t CalculateProgramSize(const GLShader::ProgramCode& program) { } /// Hashes one (or two) program streams -u64 GetUniqueIdentifier(Maxwell::ShaderProgram program_type, const ProgramCode& code, - const ProgramCode& code_b) { - u64 unique_identifier = - Common::CityHash64(reinterpret_cast<const char*>(code.data()), CalculateProgramSize(code)); - if (program_type != Maxwell::ShaderProgram::VertexA) { +u64 GetUniqueIdentifier(ProgramType program_type, const ProgramCode& code, + const ProgramCode& code_b, std::size_t size_a = 0, std::size_t size_b = 0) { + if (size_a == 0) { + size_a = CalculateProgramSize(code); + } + u64 unique_identifier = Common::CityHash64(reinterpret_cast<const char*>(code.data()), size_a); + if (program_type != ProgramType::VertexA) { return unique_identifier; } // VertexA programs include two programs @@ -133,46 +163,70 @@ u64 GetUniqueIdentifier(Maxwell::ShaderProgram program_type, const ProgramCode& std::size_t seed = 0; boost::hash_combine(seed, unique_identifier); - const u64 identifier_b = Common::CityHash64(reinterpret_cast<const char*>(code_b.data()), - CalculateProgramSize(code_b)); + if (size_b == 0) { + size_b = CalculateProgramSize(code_b); + } + const u64 identifier_b = + Common::CityHash64(reinterpret_cast<const char*>(code_b.data()), size_b); boost::hash_combine(seed, identifier_b); return static_cast<u64>(seed); } /// Creates an unspecialized program from code streams -GLShader::ProgramResult CreateProgram(const Device& device, Maxwell::ShaderProgram program_type, +GLShader::ProgramResult CreateProgram(const Device& device, ProgramType program_type, ProgramCode program_code, ProgramCode program_code_b) { GLShader::ShaderSetup setup(program_code); - if (program_type == Maxwell::ShaderProgram::VertexA) { + setup.program.size_a = CalculateProgramSize(program_code); + setup.program.size_b = 0; + if (program_type == ProgramType::VertexA) { // VertexB is always enabled, so when VertexA is enabled, we have two vertex shaders. // Conventional HW does not support this, so we combine VertexA and VertexB into one // stage here. setup.SetProgramB(program_code_b); + setup.program.size_b = CalculateProgramSize(program_code_b); } - setup.program.unique_identifier = - GetUniqueIdentifier(program_type, program_code, program_code_b); + setup.program.unique_identifier = GetUniqueIdentifier( + program_type, program_code, program_code_b, setup.program.size_a, setup.program.size_b); switch (program_type) { - case Maxwell::ShaderProgram::VertexA: - case Maxwell::ShaderProgram::VertexB: + case ProgramType::VertexA: + case ProgramType::VertexB: return GLShader::GenerateVertexShader(device, setup); - case Maxwell::ShaderProgram::Geometry: + case ProgramType::Geometry: return GLShader::GenerateGeometryShader(device, setup); - case Maxwell::ShaderProgram::Fragment: + case ProgramType::Fragment: return GLShader::GenerateFragmentShader(device, setup); + case ProgramType::Compute: + return GLShader::GenerateComputeShader(device, setup); default: - LOG_CRITICAL(HW_GPU, "Unimplemented program_type={}", static_cast<u32>(program_type)); - UNREACHABLE(); + UNIMPLEMENTED_MSG("Unimplemented program_type={}", static_cast<u32>(program_type)); return {}; } } CachedProgram SpecializeShader(const std::string& code, const GLShader::ShaderEntries& entries, - Maxwell::ShaderProgram program_type, BaseBindings base_bindings, - GLenum primitive_mode, bool hint_retrievable = false) { - std::string source = "#version 430 core\n" - "#extension GL_ARB_separate_shader_objects : enable\n\n"; - source += fmt::format("#define EMULATION_UBO_BINDING {}\n", base_bindings.cbuf++); + ProgramType program_type, const ProgramVariant& variant, + bool hint_retrievable = false) { + auto base_bindings{variant.base_bindings}; + const auto primitive_mode{variant.primitive_mode}; + const auto texture_buffer_usage{variant.texture_buffer_usage}; + + std::string source = R"(#version 430 core +#extension GL_ARB_separate_shader_objects : enable +#extension GL_ARB_shader_viewport_layer_array : enable +#extension GL_EXT_shader_image_load_formatted : enable +#extension GL_NV_gpu_shader5 : enable +#extension GL_NV_shader_thread_group : enable +#extension GL_NV_shader_thread_shuffle : enable +)"; + if (program_type == ProgramType::Compute) { + source += "#extension GL_ARB_compute_variable_group_size : require\n"; + } + source += '\n'; + + if (program_type != ProgramType::Compute) { + source += fmt::format("#define EMULATION_UBO_BINDING {}\n", base_bindings.cbuf++); + } for (const auto& cbuf : entries.const_buffers) { source += @@ -186,15 +240,34 @@ CachedProgram SpecializeShader(const std::string& code, const GLShader::ShaderEn source += fmt::format("#define SAMPLER_BINDING_{} {}\n", sampler.GetIndex(), base_bindings.sampler++); } + for (const auto& image : entries.images) { + source += + fmt::format("#define IMAGE_BINDING_{} {}\n", image.GetIndex(), base_bindings.image++); + } + + // Transform 1D textures to texture samplers by declaring its preprocessor macros. + for (std::size_t i = 0; i < texture_buffer_usage.size(); ++i) { + if (!texture_buffer_usage.test(i)) { + continue; + } + source += fmt::format("#define SAMPLER_{}_IS_BUFFER\n", i); + } + if (texture_buffer_usage.any()) { + source += '\n'; + } - if (program_type == Maxwell::ShaderProgram::Geometry) { + if (program_type == ProgramType::Geometry) { const auto [glsl_topology, debug_name, max_vertices] = GetPrimitiveDescription(primitive_mode); - source += "layout (" + std::string(glsl_topology) + ") in;\n"; + source += "layout (" + std::string(glsl_topology) + ") in;\n\n"; source += "#define MAX_VERTEX_INPUT " + std::to_string(max_vertices) + '\n'; } + if (program_type == ProgramType::Compute) { + source += "layout (local_size_variable) in;\n"; + } + source += '\n'; source += code; OGLShader shader; @@ -221,131 +294,97 @@ std::set<GLenum> GetSupportedFormats() { } // Anonymous namespace -CachedShader::CachedShader(const Device& device, VAddr cpu_addr, u64 unique_identifier, - Maxwell::ShaderProgram program_type, ShaderDiskCacheOpenGL& disk_cache, - const PrecompiledPrograms& precompiled_programs, - ProgramCode&& program_code, ProgramCode&& program_code_b, u8* host_ptr) - : RasterizerCacheObject{host_ptr}, host_ptr{host_ptr}, cpu_addr{cpu_addr}, - unique_identifier{unique_identifier}, program_type{program_type}, disk_cache{disk_cache}, - precompiled_programs{precompiled_programs} { - const std::size_t code_size{CalculateProgramSize(program_code)}; - const std::size_t code_size_b{program_code_b.empty() ? 0 - : CalculateProgramSize(program_code_b)}; - GLShader::ProgramResult program_result{ - CreateProgram(device, program_type, program_code, program_code_b)}; - if (program_result.first.empty()) { +CachedShader::CachedShader(const ShaderParameters& params, ProgramType program_type, + GLShader::ProgramResult result) + : RasterizerCacheObject{params.host_ptr}, cpu_addr{params.cpu_addr}, + unique_identifier{params.unique_identifier}, program_type{program_type}, + disk_cache{params.disk_cache}, precompiled_programs{params.precompiled_programs}, + entries{result.second}, code{std::move(result.first)}, shader_length{entries.shader_length} {} + +Shader CachedShader::CreateStageFromMemory(const ShaderParameters& params, + Maxwell::ShaderProgram program_type, + ProgramCode&& program_code, + ProgramCode&& program_code_b) { + const auto code_size{CalculateProgramSize(program_code)}; + const auto code_size_b{CalculateProgramSize(program_code_b)}; + auto result{ + CreateProgram(params.device, GetProgramType(program_type), program_code, program_code_b)}; + if (result.first.empty()) { // TODO(Rodrigo): Unimplemented shader stages hit here, avoid using these for now - return; + return {}; } - code = program_result.first; - entries = program_result.second; - shader_length = entries.shader_length; + params.disk_cache.SaveRaw(ShaderDiskCacheRaw( + params.unique_identifier, GetProgramType(program_type), + static_cast<u32>(code_size / sizeof(u64)), static_cast<u32>(code_size_b / sizeof(u64)), + std::move(program_code), std::move(program_code_b))); - const ShaderDiskCacheRaw raw(unique_identifier, program_type, - static_cast<u32>(code_size / sizeof(u64)), - static_cast<u32>(code_size_b / sizeof(u64)), - std::move(program_code), std::move(program_code_b)); - disk_cache.SaveRaw(raw); + return std::shared_ptr<CachedShader>( + new CachedShader(params, GetProgramType(program_type), std::move(result))); } -CachedShader::CachedShader(VAddr cpu_addr, u64 unique_identifier, - Maxwell::ShaderProgram program_type, ShaderDiskCacheOpenGL& disk_cache, - const PrecompiledPrograms& precompiled_programs, - GLShader::ProgramResult result, u8* host_ptr) - : RasterizerCacheObject{host_ptr}, cpu_addr{cpu_addr}, unique_identifier{unique_identifier}, - program_type{program_type}, disk_cache{disk_cache}, precompiled_programs{ - precompiled_programs} { - code = std::move(result.first); - entries = result.second; - shader_length = entries.shader_length; +Shader CachedShader::CreateStageFromCache(const ShaderParameters& params, + Maxwell::ShaderProgram program_type, + GLShader::ProgramResult result) { + return std::shared_ptr<CachedShader>( + new CachedShader(params, GetProgramType(program_type), std::move(result))); } -std::tuple<GLuint, BaseBindings> CachedShader::GetProgramHandle(GLenum primitive_mode, - BaseBindings base_bindings) { - GLuint handle{}; - if (program_type == Maxwell::ShaderProgram::Geometry) { - handle = GetGeometryShader(primitive_mode, base_bindings); - } else { - const auto [entry, is_cache_miss] = programs.try_emplace(base_bindings); - auto& program = entry->second; - if (is_cache_miss) { - program = TryLoadProgram(primitive_mode, base_bindings); - if (!program) { - program = - SpecializeShader(code, entries, program_type, base_bindings, primitive_mode); - disk_cache.SaveUsage(GetUsage(primitive_mode, base_bindings)); - } +Shader CachedShader::CreateKernelFromMemory(const ShaderParameters& params, ProgramCode&& code) { + auto result{CreateProgram(params.device, ProgramType::Compute, code, {})}; - LabelGLObject(GL_PROGRAM, program->handle, cpu_addr); - } + const auto code_size{CalculateProgramSize(code)}; + params.disk_cache.SaveRaw(ShaderDiskCacheRaw(params.unique_identifier, ProgramType::Compute, + static_cast<u32>(code_size / sizeof(u64)), 0, + std::move(code), {})); - handle = program->handle; - } - - base_bindings.cbuf += static_cast<u32>(entries.const_buffers.size()) + RESERVED_UBOS; - base_bindings.gmem += static_cast<u32>(entries.global_memory_entries.size()); - base_bindings.sampler += static_cast<u32>(entries.samplers.size()); + return std::shared_ptr<CachedShader>( + new CachedShader(params, ProgramType::Compute, std::move(result))); +} - return {handle, base_bindings}; +Shader CachedShader::CreateKernelFromCache(const ShaderParameters& params, + GLShader::ProgramResult result) { + return std::shared_ptr<CachedShader>( + new CachedShader(params, ProgramType::Compute, std::move(result))); } -GLuint CachedShader::GetGeometryShader(GLenum primitive_mode, BaseBindings base_bindings) { - const auto [entry, is_cache_miss] = geometry_programs.try_emplace(base_bindings); - auto& programs = entry->second; +std::tuple<GLuint, BaseBindings> CachedShader::GetProgramHandle(const ProgramVariant& variant) { + const auto [entry, is_cache_miss] = programs.try_emplace(variant); + auto& program = entry->second; + if (is_cache_miss) { + program = TryLoadProgram(variant); + if (!program) { + program = SpecializeShader(code, entries, program_type, variant); + disk_cache.SaveUsage(GetUsage(variant)); + } - switch (primitive_mode) { - case GL_POINTS: - return LazyGeometryProgram(programs.points, base_bindings, primitive_mode); - case GL_LINES: - case GL_LINE_STRIP: - return LazyGeometryProgram(programs.lines, base_bindings, primitive_mode); - case GL_LINES_ADJACENCY: - case GL_LINE_STRIP_ADJACENCY: - return LazyGeometryProgram(programs.lines_adjacency, base_bindings, primitive_mode); - case GL_TRIANGLES: - case GL_TRIANGLE_STRIP: - case GL_TRIANGLE_FAN: - return LazyGeometryProgram(programs.triangles, base_bindings, primitive_mode); - case GL_TRIANGLES_ADJACENCY: - case GL_TRIANGLE_STRIP_ADJACENCY: - return LazyGeometryProgram(programs.triangles_adjacency, base_bindings, primitive_mode); - default: - UNREACHABLE_MSG("Unknown primitive mode."); - return LazyGeometryProgram(programs.points, base_bindings, primitive_mode); + LabelGLObject(GL_PROGRAM, program->handle, cpu_addr); } -} -GLuint CachedShader::LazyGeometryProgram(CachedProgram& target_program, BaseBindings base_bindings, - GLenum primitive_mode) { - if (target_program) { - return target_program->handle; + auto base_bindings = variant.base_bindings; + base_bindings.cbuf += static_cast<u32>(entries.const_buffers.size()); + if (program_type != ProgramType::Compute) { + base_bindings.cbuf += STAGE_RESERVED_UBOS; } - const auto [glsl_name, debug_name, vertices] = GetPrimitiveDescription(primitive_mode); - target_program = TryLoadProgram(primitive_mode, base_bindings); - if (!target_program) { - target_program = - SpecializeShader(code, entries, program_type, base_bindings, primitive_mode); - disk_cache.SaveUsage(GetUsage(primitive_mode, base_bindings)); - } - - LabelGLObject(GL_PROGRAM, target_program->handle, cpu_addr, debug_name); + base_bindings.gmem += static_cast<u32>(entries.global_memory_entries.size()); + base_bindings.sampler += static_cast<u32>(entries.samplers.size()); - return target_program->handle; -}; + return {program->handle, base_bindings}; +} -CachedProgram CachedShader::TryLoadProgram(GLenum primitive_mode, - BaseBindings base_bindings) const { - const auto found = precompiled_programs.find(GetUsage(primitive_mode, base_bindings)); +CachedProgram CachedShader::TryLoadProgram(const ProgramVariant& variant) const { + const auto found = precompiled_programs.find(GetUsage(variant)); if (found == precompiled_programs.end()) { return {}; } return found->second; } -ShaderDiskCacheUsage CachedShader::GetUsage(GLenum primitive_mode, - BaseBindings base_bindings) const { - return {unique_identifier, base_bindings, primitive_mode}; +ShaderDiskCacheUsage CachedShader::GetUsage(const ProgramVariant& variant) const { + ShaderDiskCacheUsage usage; + usage.unique_identifier = unique_identifier; + usage.variant = variant; + return usage; } ShaderCacheOpenGL::ShaderCacheOpenGL(RasterizerOpenGL& rasterizer, Core::System& system, @@ -411,8 +450,7 @@ void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading, } if (!shader) { shader = SpecializeShader(unspecialized.code, unspecialized.entries, - unspecialized.program_type, usage.bindings, - usage.primitive, true); + unspecialized.program_type, usage.variant, true); } std::scoped_lock lock(mutex); @@ -547,7 +585,7 @@ std::unordered_map<u64, UnspecializedShader> ShaderCacheOpenGL::GenerateUnspecia } Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) { - if (!system.GPU().Maxwell3D().dirty_flags.shaders) { + if (!system.GPU().Maxwell3D().dirty.shaders) { return last_shaders[static_cast<std::size_t>(program)]; } @@ -564,28 +602,55 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) { // No shader found - create a new one ProgramCode program_code{GetShaderCode(memory_manager, program_addr, host_ptr)}; ProgramCode program_code_b; - if (program == Maxwell::ShaderProgram::VertexA) { + const bool is_program_a{program == Maxwell::ShaderProgram::VertexA}; + if (is_program_a) { const GPUVAddr program_addr_b{GetShaderAddress(system, Maxwell::ShaderProgram::VertexB)}; program_code_b = GetShaderCode(memory_manager, program_addr_b, memory_manager.GetPointer(program_addr_b)); } - const u64 unique_identifier = GetUniqueIdentifier(program, program_code, program_code_b); - const VAddr cpu_addr{*memory_manager.GpuToCpuAddress(program_addr)}; + const auto unique_identifier = + GetUniqueIdentifier(GetProgramType(program), program_code, program_code_b); + const auto cpu_addr{*memory_manager.GpuToCpuAddress(program_addr)}; + const ShaderParameters params{disk_cache, precompiled_programs, device, cpu_addr, + host_ptr, unique_identifier}; + const auto found = precompiled_shaders.find(unique_identifier); - if (found != precompiled_shaders.end()) { - // Create a shader from the cache - shader = std::make_shared<CachedShader>(cpu_addr, unique_identifier, program, disk_cache, - precompiled_programs, found->second, host_ptr); + if (found == precompiled_shaders.end()) { + shader = CachedShader::CreateStageFromMemory(params, program, std::move(program_code), + std::move(program_code_b)); } else { - // Create a shader from guest memory - shader = std::make_shared<CachedShader>( - device, cpu_addr, unique_identifier, program, disk_cache, precompiled_programs, - std::move(program_code), std::move(program_code_b), host_ptr); + shader = CachedShader::CreateStageFromCache(params, program, found->second); } Register(shader); return last_shaders[static_cast<std::size_t>(program)] = shader; } +Shader ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) { + auto& memory_manager{system.GPU().MemoryManager()}; + const auto host_ptr{memory_manager.GetPointer(code_addr)}; + auto kernel = TryGet(host_ptr); + if (kernel) { + return kernel; + } + + // No kernel found - create a new one + auto code{GetShaderCode(memory_manager, code_addr, host_ptr)}; + const auto unique_identifier{GetUniqueIdentifier(ProgramType::Compute, code, {})}; + const auto cpu_addr{*memory_manager.GpuToCpuAddress(code_addr)}; + const ShaderParameters params{disk_cache, precompiled_programs, device, cpu_addr, + host_ptr, unique_identifier}; + + const auto found = precompiled_shaders.find(unique_identifier); + if (found == precompiled_shaders.end()) { + kernel = CachedShader::CreateKernelFromMemory(params, std::move(code)); + } else { + kernel = CachedShader::CreateKernelFromCache(params, found->second); + } + + Register(kernel); + return kernel; +} + } // namespace OpenGL |