From 6888d776fffb3d5e105eddc271a2d6231abf0922 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Mon, 6 Jan 2020 21:55:06 -0300 Subject: vk_pipeline_cache: Initial implementation Given a pipeline key, this cache returns a pipeline abstraction (for graphics or compute). --- .../renderer_vulkan/vk_pipeline_cache.cpp | 352 +++++++++++++++++++++ 1 file changed, 352 insertions(+) (limited to 'src/video_core/renderer_vulkan/vk_pipeline_cache.cpp') diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp index 9bc027cbf..48e23d4cd 100644 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp @@ -2,16 +2,368 @@ // Licensed under GPLv2 or any later version // Refer to the license.txt file included. +#include #include +#include #include +#include "common/microprofile.h" +#include "core/core.h" +#include "core/memory.h" +#include "video_core/engines/kepler_compute.h" +#include "video_core/engines/maxwell_3d.h" +#include "video_core/memory_manager.h" #include "video_core/renderer_vulkan/declarations.h" +#include "video_core/renderer_vulkan/fixed_pipeline_state.h" +#include "video_core/renderer_vulkan/maxwell_to_vk.h" +#include "video_core/renderer_vulkan/vk_compute_pipeline.h" +#include "video_core/renderer_vulkan/vk_descriptor_pool.h" #include "video_core/renderer_vulkan/vk_device.h" +#include "video_core/renderer_vulkan/vk_graphics_pipeline.h" #include "video_core/renderer_vulkan/vk_pipeline_cache.h" +#include "video_core/renderer_vulkan/vk_rasterizer.h" +#include "video_core/renderer_vulkan/vk_renderpass_cache.h" +#include "video_core/renderer_vulkan/vk_resource_manager.h" +#include "video_core/renderer_vulkan/vk_scheduler.h" #include "video_core/renderer_vulkan/vk_update_descriptor.h" +#include "video_core/shader/compiler_settings.h" namespace Vulkan { +MICROPROFILE_DECLARE(Vulkan_PipelineCache); + +using Tegra::Engines::ShaderType; + +namespace { + +constexpr VideoCommon::Shader::CompilerSettings compiler_settings{ + VideoCommon::Shader::CompileDepth::FullDecompile}; + +/// Gets the address for the specified shader stage program +GPUVAddr GetShaderAddress(Core::System& system, Maxwell::ShaderProgram program) { + const auto& gpu{system.GPU().Maxwell3D()}; + const auto& shader_config{gpu.regs.shader_config[static_cast(program)]}; + return gpu.regs.code_address.CodeAddress() + shader_config.offset; +} + +/// Gets if the current instruction offset is a scheduler instruction +constexpr bool IsSchedInstruction(std::size_t offset, std::size_t main_offset) { + // Sched instructions appear once every 4 instructions. + constexpr std::size_t SchedPeriod = 4; + const std::size_t absolute_offset = offset - main_offset; + return (absolute_offset % SchedPeriod) == 0; +} + +/// Calculates the size of a program stream +std::size_t CalculateProgramSize(const ProgramCode& program, bool is_compute) { + const std::size_t start_offset = is_compute ? 0 : 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; + while (offset < program.size()) { + const u64 instruction = program[offset]; + if (!IsSchedInstruction(offset, start_offset)) { + if ((instruction & mask) == self_jumping_branch) { + // End on Maxwell's "nop" instruction + break; + } + if (instruction == 0) { + break; + } + } + ++offset; + } + // The last instruction is included in the program size + return std::min(offset + 1, program.size()); +} + +/// Gets the shader program code from memory for the specified address +ProgramCode GetShaderCode(Tegra::MemoryManager& memory_manager, const GPUVAddr gpu_addr, + const u8* host_ptr, bool is_compute) { + ProgramCode program_code(VideoCommon::Shader::MAX_PROGRAM_LENGTH); + ASSERT_OR_EXECUTE(host_ptr != nullptr, { + std::fill(program_code.begin(), program_code.end(), 0); + return program_code; + }); + memory_manager.ReadBlockUnsafe(gpu_addr, program_code.data(), + program_code.size() * sizeof(u64)); + program_code.resize(CalculateProgramSize(program_code, is_compute)); + return program_code; +} + +constexpr std::size_t GetStageFromProgram(std::size_t program) { + return program == 0 ? 0 : program - 1; +} + +constexpr ShaderType GetStageFromProgram(Maxwell::ShaderProgram program) { + return static_cast(GetStageFromProgram(static_cast(program))); +} + +ShaderType GetShaderType(Maxwell::ShaderProgram program) { + switch (program) { + case Maxwell::ShaderProgram::VertexB: + return ShaderType::Vertex; + case Maxwell::ShaderProgram::TesselationControl: + return ShaderType::TesselationControl; + case Maxwell::ShaderProgram::TesselationEval: + return ShaderType::TesselationEval; + case Maxwell::ShaderProgram::Geometry: + return ShaderType::Geometry; + case Maxwell::ShaderProgram::Fragment: + return ShaderType::Fragment; + default: + UNIMPLEMENTED_MSG("program={}", static_cast(program)); + return ShaderType::Vertex; + } +} + +u32 FillDescriptorLayout(const ShaderEntries& entries, + std::vector& bindings, + Maxwell::ShaderProgram program_type, u32 base_binding) { + const ShaderType stage = GetStageFromProgram(program_type); + const vk::ShaderStageFlags stage_flags = MaxwellToVK::ShaderStage(stage); + + u32 binding = base_binding; + const auto AddBindings = [&](vk::DescriptorType descriptor_type, std::size_t num_entries) { + for (std::size_t i = 0; i < num_entries; ++i) { + bindings.emplace_back(binding++, descriptor_type, 1, stage_flags, nullptr); + } + }; + AddBindings(vk::DescriptorType::eUniformBuffer, entries.const_buffers.size()); + AddBindings(vk::DescriptorType::eStorageBuffer, entries.global_buffers.size()); + AddBindings(vk::DescriptorType::eUniformTexelBuffer, entries.texel_buffers.size()); + AddBindings(vk::DescriptorType::eCombinedImageSampler, entries.samplers.size()); + AddBindings(vk::DescriptorType::eStorageImage, entries.images.size()); + return binding; +} + +} // Anonymous namespace + +CachedShader::CachedShader(Core::System& system, Tegra::Engines::ShaderType stage, + GPUVAddr gpu_addr, VAddr cpu_addr, u8* host_ptr, + ProgramCode program_code, u32 main_offset) + : RasterizerCacheObject{host_ptr}, gpu_addr{gpu_addr}, cpu_addr{cpu_addr}, + program_code{std::move(program_code)}, locker{stage, GetEngine(system, stage)}, + shader_ir{this->program_code, main_offset, compiler_settings, locker}, + entries{GenerateShaderEntries(shader_ir)} {} + +CachedShader::~CachedShader() = default; + +Tegra::Engines::ConstBufferEngineInterface& CachedShader::GetEngine( + Core::System& system, Tegra::Engines::ShaderType stage) { + if (stage == Tegra::Engines::ShaderType::Compute) { + return system.GPU().KeplerCompute(); + } else { + return system.GPU().Maxwell3D(); + } +} + +VKPipelineCache::VKPipelineCache(Core::System& system, RasterizerVulkan& rasterizer, + const VKDevice& device, VKScheduler& scheduler, + VKDescriptorPool& descriptor_pool, + VKUpdateDescriptorQueue& update_descriptor_queue) + : RasterizerCache{rasterizer}, system{system}, device{device}, scheduler{scheduler}, + descriptor_pool{descriptor_pool}, update_descriptor_queue{update_descriptor_queue}, + renderpass_cache(device) {} + +VKPipelineCache::~VKPipelineCache() = default; + +std::array VKPipelineCache::GetShaders() { + const auto& gpu = system.GPU().Maxwell3D(); + auto& dirty = system.GPU().Maxwell3D().dirty.shaders; + if (!dirty) { + return last_shaders; + } + dirty = false; + + std::array shaders; + for (std::size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) { + const auto& shader_config = gpu.regs.shader_config[index]; + const auto program{static_cast(index)}; + + // Skip stages that are not enabled + if (!gpu.regs.IsShaderConfigEnabled(index)) { + continue; + } + + auto& memory_manager{system.GPU().MemoryManager()}; + const GPUVAddr program_addr{GetShaderAddress(system, program)}; + const auto host_ptr{memory_manager.GetPointer(program_addr)}; + auto shader = TryGet(host_ptr); + if (!shader) { + // No shader found - create a new one + constexpr u32 stage_offset = 10; + const auto stage = static_cast(index == 0 ? 0 : index - 1); + auto code = GetShaderCode(memory_manager, program_addr, host_ptr, false); + + const std::optional cpu_addr = memory_manager.GpuToCpuAddress(program_addr); + ASSERT(cpu_addr); + + shader = std::make_shared(system, stage, program_addr, *cpu_addr, + host_ptr, std::move(code), stage_offset); + Register(shader); + } + shaders[index] = std::move(shader); + } + return last_shaders = shaders; +} + +VKGraphicsPipeline& VKPipelineCache::GetGraphicsPipeline(const GraphicsPipelineCacheKey& key) { + MICROPROFILE_SCOPE(Vulkan_PipelineCache); + + if (last_graphics_pipeline && last_graphics_key == key) { + return *last_graphics_pipeline; + } + last_graphics_key = key; + + const auto [pair, is_cache_miss] = graphics_cache.try_emplace(key); + auto& entry = pair->second; + if (is_cache_miss) { + LOG_INFO(Render_Vulkan, "Compile 0x{:016X}", key.Hash()); + const auto [program, bindings] = DecompileShaders(key); + entry = std::make_unique(device, scheduler, descriptor_pool, + update_descriptor_queue, renderpass_cache, key, + bindings, program); + } + return *(last_graphics_pipeline = entry.get()); +} + +VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCacheKey& key) { + MICROPROFILE_SCOPE(Vulkan_PipelineCache); + + const auto [pair, is_cache_miss] = compute_cache.try_emplace(key); + auto& entry = pair->second; + if (!is_cache_miss) { + return *entry; + } + LOG_INFO(Render_Vulkan, "Compile 0x{:016X}", key.Hash()); + + auto& memory_manager = system.GPU().MemoryManager(); + const auto program_addr = key.shader; + const auto host_ptr = memory_manager.GetPointer(program_addr); + + auto shader = TryGet(host_ptr); + if (!shader) { + // No shader found - create a new one + const auto cpu_addr = memory_manager.GpuToCpuAddress(program_addr); + ASSERT(cpu_addr); + + auto code = GetShaderCode(memory_manager, program_addr, host_ptr, true); + constexpr u32 kernel_main_offset = 0; + shader = std::make_shared(system, Tegra::Engines::ShaderType::Compute, + program_addr, *cpu_addr, host_ptr, std::move(code), + kernel_main_offset); + Register(shader); + } + + Specialization specialization; + specialization.workgroup_size = key.workgroup_size; + specialization.shared_memory_size = key.shared_memory_size; + + const SPIRVShader spirv_shader{ + Decompile(device, shader->GetIR(), ShaderType::Compute, specialization), + shader->GetEntries()}; + entry = std::make_unique(device, scheduler, descriptor_pool, + update_descriptor_queue, spirv_shader); + return *entry; +} + +void VKPipelineCache::Unregister(const Shader& shader) { + bool finished = false; + const auto Finish = [&] { + // TODO(Rodrigo): Instead of finishing here, wait for the fences that use this pipeline and + // flush. + if (finished) { + return; + } + finished = true; + scheduler.Finish(); + }; + + const GPUVAddr invalidated_addr = shader->GetGpuAddr(); + for (auto it = graphics_cache.begin(); it != graphics_cache.end();) { + auto& entry = it->first; + if (std::find(entry.shaders.begin(), entry.shaders.end(), invalidated_addr) == + entry.shaders.end()) { + ++it; + continue; + } + Finish(); + it = graphics_cache.erase(it); + } + for (auto it = compute_cache.begin(); it != compute_cache.end();) { + auto& entry = it->first; + if (entry.shader != invalidated_addr) { + ++it; + continue; + } + Finish(); + it = compute_cache.erase(it); + } + + RasterizerCache::Unregister(shader); +} + +std::pair> +VKPipelineCache::DecompileShaders(const GraphicsPipelineCacheKey& key) { + const auto& fixed_state = key.fixed_state; + auto& memory_manager = system.GPU().MemoryManager(); + const auto& gpu = system.GPU().Maxwell3D(); + + Specialization specialization; + specialization.primitive_topology = fixed_state.input_assembly.topology; + if (specialization.primitive_topology == Maxwell::PrimitiveTopology::Points) { + ASSERT(fixed_state.input_assembly.point_size != 0.0f); + specialization.point_size = fixed_state.input_assembly.point_size; + } + for (std::size_t i = 0; i < Maxwell::NumVertexAttributes; ++i) { + specialization.attribute_types[i] = fixed_state.vertex_input.attributes[i].type; + } + specialization.ndc_minus_one_to_one = fixed_state.rasterizer.ndc_minus_one_to_one; + specialization.tessellation.primitive = fixed_state.tessellation.primitive; + specialization.tessellation.spacing = fixed_state.tessellation.spacing; + specialization.tessellation.clockwise = fixed_state.tessellation.clockwise; + for (const auto& rt : key.renderpass_params.color_attachments) { + specialization.enabled_rendertargets.set(rt.index); + } + + SPIRVProgram program; + std::vector bindings; + + for (std::size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) { + const auto program_enum = static_cast(index); + + // Skip stages that are not enabled + if (!gpu.regs.IsShaderConfigEnabled(index)) { + continue; + } + + const GPUVAddr gpu_addr = GetShaderAddress(system, program_enum); + const auto host_ptr = memory_manager.GetPointer(gpu_addr); + const auto shader = TryGet(host_ptr); + ASSERT(shader); + + const std::size_t stage = index == 0 ? 0 : index - 1; // Stage indices are 0 - 5 + const auto program_type = GetShaderType(program_enum); + const auto& entries = shader->GetEntries(); + program[stage] = {Decompile(device, shader->GetIR(), program_type, specialization), + entries}; + + if (program_enum == Maxwell::ShaderProgram::VertexA) { + // VertexB was combined with VertexA, so we skip the VertexB iteration + ++index; + } + + const u32 old_binding = specialization.base_binding; + specialization.base_binding = + FillDescriptorLayout(entries, bindings, program_enum, specialization.base_binding); + ASSERT(old_binding + entries.NumBindings() == specialization.base_binding); + } + return {std::move(program), std::move(bindings)}; +} + void FillDescriptorUpdateTemplateEntries( const VKDevice& device, const ShaderEntries& entries, u32& binding, u32& offset, std::vector& template_entries) { -- cgit v1.2.3