ir: Add heuristic based LDS barrier pass (#1801)

* ir: Add heuristic based LDS barrier pass

* Attempts to insert barriers after zero-depth divergant conditional blocks in shaders that use shared memory

* lds_barriers: Limit to nvidia

* Intel has historically had problems with cs barriers, will debug other time

CMakeLists.txt

@@ -671,6 +671,7 @@ set(SHADER_RECOMPILER src/shader_recompiler/exception.h
                       src/shader_recompiler/ir/passes/resource_tracking_pass.cpp
                       src/shader_recompiler/ir/passes/ring_access_elimination.cpp
                       src/shader_recompiler/ir/passes/shader_info_collection_pass.cpp
+                      src/shader_recompiler/ir/passes/shared_memory_barrier_pass.cpp
                       src/shader_recompiler/ir/passes/ssa_rewrite_pass.cpp
                       src/shader_recompiler/ir/abstract_syntax_list.h
                       src/shader_recompiler/ir/attribute.cpp

src/shader_recompiler/ir/passes/ir_passes.h

@@ -6,6 +6,10 @@
 #include "shader_recompiler/ir/basic_block.h"
 #include "shader_recompiler/ir/program.h"
 
+namespace Shader {
+struct Profile;
+}
+
 namespace Shader::Optimization {
 
 void SsaRewritePass(IR::BlockList& program);
@@ -21,5 +25,6 @@ void RingAccessElimination(const IR::Program& program, const RuntimeInfo& runtim
 void TessellationPreprocess(IR::Program& program, RuntimeInfo& runtime_info);
 void HullShaderTransform(IR::Program& program, RuntimeInfo& runtime_info);
 void DomainShaderTransform(IR::Program& program, RuntimeInfo& runtime_info);
+void SharedMemoryBarrierPass(IR::Program& program, const Profile& profile);
 
 } // namespace Shader::Optimization

src/shader_recompiler/ir/passes/shared_memory_barrier_pass.cpp

@@ -0,0 +1,46 @@
+// SPDX-FileCopyrightText: Copyright 2024 shadPS4 Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include "shader_recompiler/ir/breadth_first_search.h"
+#include "shader_recompiler/ir/ir_emitter.h"
+#include "shader_recompiler/ir/program.h"
+#include "shader_recompiler/profile.h"
+
+namespace Shader::Optimization {
+
+void SharedMemoryBarrierPass(IR::Program& program, const Profile& profile) {
+    if (!program.info.uses_shared || !profile.needs_lds_barriers) {
+        return;
+    }
+    using Type = IR::AbstractSyntaxNode::Type;
+    u32 branch_depth{};
+    for (const IR::AbstractSyntaxNode& node : program.syntax_list) {
+        if (node.type == Type::EndIf) {
+            --branch_depth;
+            continue;
+        }
+        if (node.type != Type::If) {
+            continue;
+        }
+        u32 curr_depth = branch_depth++;
+        if (curr_depth != 0) {
+            continue;
+        }
+        const IR::U1 cond = node.data.if_node.cond;
+        const auto insert_barrier = IR::BreadthFirstSearch(cond, [](IR::Inst* inst) -> std::optional<bool> {
+            if (inst->GetOpcode() == IR::Opcode::GetAttributeU32 &&
+                inst->Arg(0).Attribute() == IR::Attribute::LocalInvocationId) {
+                return true;
+            }
+            return std::nullopt;
+        });
+        if (insert_barrier) {
+            IR::Block* const merge = node.data.if_node.merge;
+            auto insert_point = std::ranges::find_if_not(merge->Instructions(), IR::IsPhi);
+            IR::IREmitter ir{*merge, insert_point};
+            ir.Barrier();
+        }
+    }
+}
+
+} // namespace Shader::Optimization

src/shader_recompiler/profile.h

@@ -27,6 +27,7 @@ struct Profile {
     bool has_broken_spirv_clamp{};
     bool lower_left_origin_mode{};
     bool needs_manual_interpolation{};
+    bool needs_lds_barriers{};
     u64 min_ssbo_alignment{};
 };
 

src/shader_recompiler/recompiler.cpp

@@ -91,6 +91,7 @@ IR::Program TranslateProgram(std::span<const u32> code, Pools& pools, Info& info
     Shader::Optimization::IdentityRemovalPass(program.blocks);
     Shader::Optimization::DeadCodeEliminationPass(program);
     Shader::Optimization::CollectShaderInfoPass(program);
+    Shader::Optimization::SharedMemoryBarrierPass(program, profile);
 
     return program;
 }

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -204,6 +204,7 @@ PipelineCache::PipelineCache(const Instance& instance_, Scheduler& scheduler_,
         .supports_image_load_store_lod = instance_.IsImageLoadStoreLodSupported(),
         .needs_manual_interpolation = instance.IsFragmentShaderBarycentricSupported() &&
                                       instance.GetDriverID() == vk::DriverId::eNvidiaProprietary,
+        .needs_lds_barriers = instance.GetDriverID() == vk::DriverId::eNvidiaProprietary,
     };
     auto [cache_result, cache] = instance.GetDevice().createPipelineCacheUnique({});
     ASSERT_MSG(cache_result == vk::Result::eSuccess, "Failed to create pipeline cache: {}",