Re-do geometry compaction pass

src/shader_recompiler/CMakeLists.txt

@@ -172,6 +172,7 @@ 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/geometry_compaction_pass.cpp
     ir_opt/global_memory_to_storage_buffer_pass.cpp
     ir_opt/identity_removal_pass.cpp
     ir_opt/layer_pass.cpp

src/shader_recompiler/frontend/maxwell/translate_program.cpp

@@ -393,6 +393,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
     Optimization::PositionPass(env, program);
 
     Optimization::GlobalMemoryToStorageBufferPass(program, normalized_host_info);
+    Optimization::GeometryCompactionPass(program, normalized_host_info);
     Optimization::TexturePass(env, program, normalized_host_info);
 
     if (Settings::values.resolution_info.active || Settings::values.rescale_hack.GetValue()) {

src/shader_recompiler/host_translate_info.h

@@ -39,6 +39,7 @@ struct HostTranslateInfo {
                                                 ///< passthrough shaders
     bool support_conditional_barrier{}; ///< True when the device supports barriers in conditional
                                         ///< control flow
+    bool single_lane_geometry_subgroups{};
 
     void ApplyDescriptorLimitPolicy() noexcept {
         if (min_ssbo_alignment == 0) {

src/shader_recompiler/ir_opt/geometry_compaction_pass.cpp

@@ -0,0 +1,129 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include <boost/container/small_vector.hpp>
+
+#include "shader_recompiler/frontend/ir/ir_emitter.h"
+#include "shader_recompiler/host_translate_info.h"
+#include "shader_recompiler/ir_opt/passes.h"
+
+namespace Shader::Optimization {
+namespace {
+constexpr int MAX_BALLOT_DEPTH = 8;
+
+struct SlotAtomic {
+    IR::Block* block;
+    IR::Inst* atomic;
+    IR::Inst* ballot;
+};
+
+IR::Inst* FindBallot(const IR::Value& value, int depth) {
+    if (depth > MAX_BALLOT_DEPTH || value.IsImmediate()) {
+        return nullptr;
+    }
+    IR::Inst* const inst{value.InstRecursive()};
+    if (inst->GetOpcode() == IR::Opcode::SubgroupBallot) {
+        return inst;
+    }
+    if (inst->GetOpcode() == IR::Opcode::Phi) {
+        return nullptr;
+    }
+    for (size_t index = 0; index < inst->NumArgs(); ++index) {
+        if (IR::Inst* const ballot{FindBallot(inst->Arg(index), depth + 1)}) {
+            return ballot;
+        }
+    }
+    return nullptr;
+}
+
+bool IsShuffle(IR::Opcode opcode) {
+    switch (opcode) {
+    case IR::Opcode::ShuffleIndex:
+    case IR::Opcode::ShuffleUp:
+    case IR::Opcode::ShuffleDown:
+    case IR::Opcode::ShuffleButterfly:
+        return true;
+    default:
+        return false;
+    }
+}
+
+IR::U32 PrimitiveSlot(IR::IREmitter& ir, u32 invocations, const IR::U32& amount) {
+    IR::U32 key{ir.GetAttributeU32(IR::Attribute::PrimitiveId)};
+    if (invocations > 1) {
+        key = ir.IAdd(ir.IMul(key, ir.Imm32(invocations)), ir.InvocationId());
+    }
+    return ir.IMul(key, amount);
+}
+
+bool MergesAtomic(const IR::Inst& phi, const IR::Block& block, const IR::Inst& atomic) {
+    for (size_t index = 0; index < phi.NumArgs(); ++index) {
+        const IR::Value arg{phi.Arg(index)};
+        if (phi.PhiBlock(index) == &block && !arg.IsImmediate() &&
+            arg.InstRecursive() == &atomic) {
+            return true;
+        }
+    }
+    return false;
+}
+
+void RewriteMergedSlots(IR::Block& block, IR::Inst& atomic, u32 invocations) {
+    for (IR::Block* const successor : block.ImmSuccessors()) {
+        for (IR::Inst& phi : successor->Instructions()) {
+            if (phi.GetOpcode() != IR::Opcode::Phi || !MergesAtomic(phi, block, atomic)) {
+                continue;
+            }
+            for (size_t index = 0; index < phi.NumArgs(); ++index) {
+                const IR::Value arg{phi.Arg(index)};
+                if (arg.IsImmediate() || !IsShuffle(arg.InstRecursive()->GetOpcode())) {
+                    continue;
+                }
+                IR::IREmitter ir{*phi.PhiBlock(index)};
+                const IR::U32 amount{arg.InstRecursive()->Arg(0)};
+                phi.SetArg(index, PrimitiveSlot(ir, invocations, amount));
+            }
+        }
+    }
+}
+
+void RewriteAtomic(IR::Block& block, IR::Inst& atomic, u32 invocations) {
+    const auto insert_point{IR::Block::InstructionList::s_iterator_to(atomic)};
+    IR::IREmitter ir{block, insert_point};
+    const IR::U32 amount{atomic.Arg(2)};
+    const IR::U32 slot{PrimitiveSlot(ir, invocations, amount)};
+    block.PrependNewInst(insert_point, IR::Opcode::StorageAtomicUMax32,
+                         {atomic.Arg(0), atomic.Arg(1), ir.IAdd(slot, amount)});
+    atomic.ReplaceUsesWith(slot);
+}
+
+void NeutralizePredicate(IR::Inst& ballot) {
+    const IR::Value pred{ballot.Arg(0)};
+    if (!pred.IsImmediate()) {
+        pred.InstRecursive()->ReplaceUsesWith(IR::Value{true});
+    }
+}
+}
+
+void GeometryCompactionPass(IR::Program& program, const HostTranslateInfo& host_info) {
+    if (program.stage != Stage::Geometry || !host_info.single_lane_geometry_subgroups) {
+        return;
+    }
+    boost::container::small_vector<SlotAtomic, 4> slot_atomics;
+    for (IR::Block* const block : program.post_order_blocks) {
+        for (IR::Inst& inst : block->Instructions()) {
+            if (inst.GetOpcode() != IR::Opcode::StorageAtomicIAdd32) {
+                continue;
+            }
+            if (IR::Inst* const ballot{FindBallot(inst.Arg(2), 0)}) {
+                slot_atomics.push_back({block, &inst, ballot});
+            }
+        }
+    }
+    for (const SlotAtomic& slot_atomic : slot_atomics) {
+        RewriteMergedSlots(*slot_atomic.block, *slot_atomic.atomic, program.invocations);
+        RewriteAtomic(*slot_atomic.block, *slot_atomic.atomic, program.invocations);
+        NeutralizePredicate(*slot_atomic.ballot);
+    }
+}
+
+}

src/shader_recompiler/ir_opt/passes.h

@@ -16,6 +16,7 @@ void CollectShaderInfoPass(Environment& env, IR::Program& program);
 void ConditionalBarrierPass(IR::Program& program);
 void ConstantPropagationPass(Environment& env, IR::Program& program);
 void DeadCodeEliminationPass(IR::Program& program);
+void GeometryCompactionPass(IR::Program& program, const HostTranslateInfo& host_info);
 void GlobalMemoryToStorageBufferPass(IR::Program& program, const HostTranslateInfo& host_info);
 void IdentityRemovalPass(IR::Program& program);
 void LowerFp64ToFp32(IR::Program& program);

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -489,6 +489,7 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .support_viewport_mask = device.IsNvViewportArray2Supported(),
         .support_geometry_shader_passthrough = device.IsNvGeometryShaderPassthroughSupported(),
         .support_conditional_barrier = device.SupportsConditionalBarriers(),
+        .single_lane_geometry_subgroups = !profile.SupportsSubgroupStage(Shader::Stage::Geometry),
     };
     host_info.ApplyDescriptorLimitPolicy();