From c9c3dca3efe737ab4d0b1d9e582e7e89b91d9283 Mon Sep 17 00:00:00 2001 From: pineappleEA Date: Tue, 24 Aug 2021 07:14:25 +0200 Subject: [PATCH] early-access version 2010 --- README.md | 2 +- src/shader_recompiler/CMakeLists.txt | 1 - .../maxwell/structured_control_flow.cpp | 119 +++++++++++++++++- .../maxwell/structured_control_flow.h | 9 +- .../frontend/maxwell/translate_program.cpp | 7 +- src/shader_recompiler/host_translate_info.h | 6 +- src/shader_recompiler/ir_opt/passes.h | 3 +- .../renderer_opengl/gl_shader_cache.cpp | 2 +- .../renderer_vulkan/vk_pipeline_cache.cpp | 2 +- 9 files changed, 131 insertions(+), 20 deletions(-) diff --git a/README.md b/README.md index a405ba571..a8592f508 100755 --- a/README.md +++ b/README.md @@ -1,7 +1,7 @@ yuzu emulator early access ============= -This is the source code for early-access 2009. +This is the source code for early-access 2010. ## Legal Notice diff --git a/src/shader_recompiler/CMakeLists.txt b/src/shader_recompiler/CMakeLists.txt index c20a583d9..b5b7e5e83 100755 --- a/src/shader_recompiler/CMakeLists.txt +++ b/src/shader_recompiler/CMakeLists.txt @@ -216,7 +216,6 @@ add_library(shader_recompiler STATIC ir_opt/constant_propagation_pass.cpp ir_opt/dead_code_elimination_pass.cpp ir_opt/dual_vertex_pass.cpp - ir_opt/get_attribute_reorder_pass.cpp ir_opt/global_memory_to_storage_buffer_pass.cpp ir_opt/identity_removal_pass.cpp ir_opt/lower_fp16_to_fp32.cpp diff --git a/src/shader_recompiler/frontend/maxwell/structured_control_flow.cpp b/src/shader_recompiler/frontend/maxwell/structured_control_flow.cpp index 8b3e0a15c..0eb156034 100755 --- a/src/shader_recompiler/frontend/maxwell/structured_control_flow.cpp +++ b/src/shader_recompiler/frontend/maxwell/structured_control_flow.cpp @@ -20,6 +20,7 @@ #include "shader_recompiler/frontend/maxwell/decode.h" #include "shader_recompiler/frontend/maxwell/structured_control_flow.h" #include "shader_recompiler/frontend/maxwell/translate/translate.h" +#include "shader_recompiler/host_translate_info.h" #include "shader_recompiler/object_pool.h" namespace Shader::Maxwell { @@ -652,7 +653,8 @@ class TranslatePass { public: TranslatePass(ObjectPool& inst_pool_, ObjectPool& block_pool_, ObjectPool& stmt_pool_, Environment& env_, Statement& root_stmt, - IR::AbstractSyntaxList& syntax_list_) + IR::AbstractSyntaxList& syntax_list_, + [[maybe_unused]] const HostTranslateInfo& host_info) : stmt_pool{stmt_pool_}, inst_pool{inst_pool_}, block_pool{block_pool_}, env{env_}, syntax_list{syntax_list_} { Visit(root_stmt, nullptr, nullptr); @@ -660,6 +662,11 @@ public: IR::Block& first_block{*syntax_list.front().data.block}; IR::IREmitter ir(first_block, first_block.begin()); ir.Prologue(); + // TODO(ameerj): Enable only on devices that require the reorder after testing + // if (uses_demote_to_helper && host_info.needs_demote_reorder ) { + if (uses_demote_to_helper) { + DemoteCombinationPass(); + } } private: @@ -809,7 +816,14 @@ private: } case StatementType::Return: { ensure_block(); - IR::IREmitter{*current_block}.Epilogue(); + IR::Block* return_block{block_pool.Create(inst_pool)}; + IR::IREmitter{*return_block}.Epilogue(); + current_block->AddBranch(return_block); + + auto& merge{syntax_list.emplace_back()}; + merge.type = IR::AbstractSyntaxNode::Type::Block; + merge.data.block = return_block; + current_block = nullptr; syntax_list.emplace_back().type = IR::AbstractSyntaxNode::Type::Return; break; @@ -824,6 +838,7 @@ private: auto& merge{syntax_list.emplace_back()}; merge.type = IR::AbstractSyntaxNode::Type::Block; merge.data.block = demote_block; + uses_demote_to_helper = true; break; } case StatementType::Unreachable: { @@ -855,11 +870,106 @@ private: return block_pool.Create(inst_pool); } + void DemoteCombinationPass() { + using Type = IR::AbstractSyntaxNode::Type; + std::vector demote_blocks; + std::vector demote_conds; + u32 num_epilogues{}; + for (const IR::AbstractSyntaxNode& node : syntax_list) { + if (node.type != Type::Block) { + continue; + } + for (const IR::Inst& inst : node.data.block->Instructions()) { + const IR::Opcode op{inst.GetOpcode()}; + if (op == IR::Opcode::DemoteToHelperInvocation) { + demote_blocks.push_back(node.data.block); + break; + } + if (op == IR::Opcode::Epilogue) { + ++num_epilogues; + } + } + } + if (demote_blocks.size() == 0) { + return; + } + if (num_epilogues > 1) { + LOG_DEBUG(Shader, "Combining demotes with more than one return is not implemented."); + return; + } + s64 last_iterator_offset{}; + auto& asl{syntax_list}; + for (const IR::Block* demote_block : demote_blocks) { + const auto start_it{asl.begin() + last_iterator_offset}; + auto asl_it{std::find_if(start_it, asl.end(), [&](const IR::AbstractSyntaxNode& asn) { + return asn.type == Type::If && asn.data.if_node.body == demote_block; + })}; + if (asl_it == asl.end()) { + // Demote without a conditional branch. + // No need to proceed since all fragment instances will be demoted regardless. + return; + } + const IR::Block* const end_if = asl_it->data.if_node.merge; + demote_conds.push_back(asl_it->data.if_node.cond); + last_iterator_offset = std::distance(asl.begin(), asl_it); + + asl_it = asl.erase(asl_it); + asl_it = std::find_if(asl_it, asl.end(), [&](const IR::AbstractSyntaxNode& asn) { + return asn.type == Type::Block && asn.data.block == demote_block; + }); + + asl_it = asl.erase(asl_it); + asl_it = std::find_if(asl_it, asl.end(), [&](const IR::AbstractSyntaxNode& asn) { + return asn.type == Type::EndIf && asn.data.end_if.merge == end_if; + }); + asl_it = asl.erase(asl_it); + } + const auto epilogue_func{[](const IR::AbstractSyntaxNode& asn) { + if (asn.type != Type::Block) { + return false; + } + for (const auto& inst : asn.data.block->Instructions()) { + if (inst.GetOpcode() == IR::Opcode::Epilogue) { + return true; + } + } + return false; + }}; + const auto reverse_it{std::find_if(asl.rbegin(), asl.rend(), epilogue_func)}; + const auto return_block_it{(reverse_it + 1).base()}; + + IR::IREmitter ir{*(return_block_it - 1)->data.block}; + IR::U1 cond(IR::Value(false)); + for (const auto& demote_cond : demote_conds) { + cond = ir.LogicalOr(cond, demote_cond); + } + cond.Inst()->DestructiveAddUsage(1); + + IR::AbstractSyntaxNode demote_if_node{}; + demote_if_node.type = Type::If; + demote_if_node.data.if_node.cond = cond; + demote_if_node.data.if_node.body = demote_blocks[0]; + demote_if_node.data.if_node.merge = return_block_it->data.block; + + IR::AbstractSyntaxNode demote_node{}; + demote_node.type = Type::Block; + demote_node.data.block = demote_blocks[0]; + + IR::AbstractSyntaxNode demote_endif_node{}; + demote_endif_node.type = Type::EndIf; + demote_endif_node.data.end_if.merge = return_block_it->data.block; + + asl.insert(return_block_it, demote_endif_node); + asl.insert(return_block_it, demote_node); + asl.insert(return_block_it, demote_if_node); + } + ObjectPool& stmt_pool; ObjectPool& inst_pool; ObjectPool& block_pool; Environment& env; IR::AbstractSyntaxList& syntax_list; + bool uses_demote_to_helper{}; // TODO: C++20 Remove this when all compilers support constexpr std::vector #if __cpp_lib_constexpr_vector >= 201907 @@ -871,12 +981,13 @@ private: } // Anonymous namespace IR::AbstractSyntaxList BuildASL(ObjectPool& inst_pool, ObjectPool& block_pool, - Environment& env, Flow::CFG& cfg) { + Environment& env, Flow::CFG& cfg, + const HostTranslateInfo& host_info) { ObjectPool stmt_pool{64}; GotoPass goto_pass{cfg, stmt_pool}; Statement& root{goto_pass.RootStatement()}; IR::AbstractSyntaxList syntax_list; - TranslatePass{inst_pool, block_pool, stmt_pool, env, root, syntax_list}; + TranslatePass{inst_pool, block_pool, stmt_pool, env, root, syntax_list, host_info}; return syntax_list; } diff --git a/src/shader_recompiler/frontend/maxwell/structured_control_flow.h b/src/shader_recompiler/frontend/maxwell/structured_control_flow.h index 88b083649..e38158da3 100755 --- a/src/shader_recompiler/frontend/maxwell/structured_control_flow.h +++ b/src/shader_recompiler/frontend/maxwell/structured_control_flow.h @@ -11,10 +11,13 @@ #include "shader_recompiler/frontend/maxwell/control_flow.h" #include "shader_recompiler/object_pool.h" -namespace Shader::Maxwell { +namespace Shader { +struct HostTranslateInfo; +namespace Maxwell { [[nodiscard]] IR::AbstractSyntaxList BuildASL(ObjectPool& inst_pool, ObjectPool& block_pool, Environment& env, - Flow::CFG& cfg); + Flow::CFG& cfg, const HostTranslateInfo& host_info); -} // namespace Shader::Maxwell +} // namespace Maxwell +} // namespace Shader diff --git a/src/shader_recompiler/frontend/maxwell/translate_program.cpp b/src/shader_recompiler/frontend/maxwell/translate_program.cpp index b53073165..012d55357 100755 --- a/src/shader_recompiler/frontend/maxwell/translate_program.cpp +++ b/src/shader_recompiler/frontend/maxwell/translate_program.cpp @@ -130,7 +130,7 @@ void AddNVNStorageBuffers(IR::Program& program) { IR::Program TranslateProgram(ObjectPool& inst_pool, ObjectPool& block_pool, Environment& env, Flow::CFG& cfg, const HostTranslateInfo& host_info) { IR::Program program; - program.syntax_list = BuildASL(inst_pool, block_pool, env, cfg); + program.syntax_list = BuildASL(inst_pool, block_pool, env, cfg, host_info); program.blocks = GenerateBlocks(program.syntax_list); program.post_order_blocks = PostOrder(program.syntax_list.front()); program.stage = env.ShaderStage(); @@ -178,13 +178,10 @@ IR::Program TranslateProgram(ObjectPool& inst_pool, ObjectPool + #include "shader_recompiler/environment.h" #include "shader_recompiler/frontend/ir/basic_block.h" #include "shader_recompiler/frontend/ir/program.h" @@ -13,7 +15,6 @@ namespace Shader::Optimization { void CollectShaderInfoPass(Environment& env, IR::Program& program); void ConstantPropagationPass(IR::Program& program); void DeadCodeEliminationPass(IR::Program& program); -void GetAttributeReorderPass(IR::Program& program); void GlobalMemoryToStorageBufferPass(IR::Program& program); void IdentityRemovalPass(IR::Program& program); void LowerFp16ToFp32(IR::Program& program); diff --git a/src/video_core/renderer_opengl/gl_shader_cache.cpp b/src/video_core/renderer_opengl/gl_shader_cache.cpp index 371db6d63..b0e14182e 100755 --- a/src/video_core/renderer_opengl/gl_shader_cache.cpp +++ b/src/video_core/renderer_opengl/gl_shader_cache.cpp @@ -219,7 +219,7 @@ ShaderCache::ShaderCache(RasterizerOpenGL& rasterizer_, Core::Frontend::EmuWindo host_info{ .support_float16 = false, .support_int64 = device.HasShaderInt64(), - .needs_get_attribute_reorder = device.IsAmd(), + .needs_demote_reorder = device.IsAmd(), } { if (use_asynchronous_shaders) { workers = CreateWorkers(); diff --git a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp index d1a402547..77e15dfeb 100755 --- a/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp +++ b/src/video_core/renderer_vulkan/vk_pipeline_cache.cpp @@ -325,7 +325,7 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::Engines::Maxw host_info = Shader::HostTranslateInfo{ .support_float16 = device.IsFloat16Supported(), .support_int64 = device.IsShaderInt64Supported(), - .needs_get_attribute_reorder = driver_id == VK_DRIVER_ID_AMD_PROPRIETARY_KHR, + .needs_demote_reorder = driver_id == VK_DRIVER_ID_AMD_PROPRIETARY_KHR, }; }