shader_recompiler: gl_PrimitiveId scatter for geometry-stage stream compaction (#4330)

Guest shaders that reserve a compacted output slot via subgroup ballot
-> popcount -> atomic add cannot preserve the originating primitives'
relative order once run through the single-invocation subgroup
fallback for stages without subgroup support. gl_PrimitiveId is
already hardware-guaranteed unique and monotonically increasing per
primitive, so CompactionFallbackPass rewrites that pattern into a
gl_PrimitiveId-indexed scatter instead.

Confirmed fixing corrupted/order-scrambled compacted draws in NieR
Automata on a device without geometry-stage subgroup support.

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/4330

src/core/file_sys/ips_layer.cpp

@@ -123,13 +123,11 @@ static IPSwitchRecord EscapeStringSequences(std::string_view sv) {
     for (auto it = sv.cbegin(); it != sv.cend(); ) {
         if (*it == '\\' && it + 1 < sv.cend()) {
             switch (it[1]) {
-            case 'a': r.data[r.count] = '\a'; break;
-            case 'b': r.data[r.count] = '\b'; break;
-            case 'e': r.data[r.count] = '\e'; break;
-            case 'f': r.data[r.count] = '\f'; break;
             case 'n': r.data[r.count] = '\n'; break;
-            case 'r': r.data[r.count] = '\r'; break;
             case 't': r.data[r.count] = '\t'; break;
+            case 'b': r.data[r.count] = '\b'; break;
+            case 'r': r.data[r.count] = '\r'; break;
+            case 'e': r.data[r.count] = '\e'; break;
             case 'v': r.data[r.count] = '\v'; break;
             case '?': r.data[r.count] = '\?'; break;
             default: r.data[r.count] = it[1]; break;
@@ -144,60 +142,40 @@ static IPSwitchRecord EscapeStringSequences(std::string_view sv) {
     return r;
 }
 
+[[nodiscard]] static inline std::array<u8, 32> ReadNSOBuildId(std::string_view const s) {
+    std::array<u8, 32> r{};
+    for (std::size_t i = 0; i < s.size(); ++i)
+        r[i / 2] |= u8(u8(Common::ToHexNibble(s[i])) << u8((i % 2) * 4));
+    return r;
+}
+
 void IPSwitchCompiler::Parse(std::span<u8 const> bytes) {
     LOG_INFO(Loader, "IPSwitchCompiler: '{}'", patch_text->GetName());
-    bool is_little_endian = true;
+    bool is_little_endian = false;
     s64 offset_shift = 0;
     //bool print_values = false;
+
     auto const parse_line = [&](std::string_view const line) {
         // Keep in mind lines have trimmed spaces (at the end & start)!
         LOG_INFO(Loader, "<{}>", line);
-        // IPSwitch is case insensitive
+        if (line.starts_with("@stop")) {
-        // Yes this is how the logic goes for the main reference parsers!
+            return false; // Force stop
-        if (line.size() > 2 && line[0] == '@') {
+        } else if (line.starts_with("@nsobid-")) { // NSO Build ID Specifier
-            switch (line[1]) {
+            nso_build_id = ReadNSOBuildId(line.substr(8));
-            // yes, @nsobid too -- NSO Build ID Specifier
+        } else if (line.starts_with("@enabled")) {
-            case 'n':
+            patches.push_back({{}, true}); //enabled patch
-            case 'N':
+        } else if (line.starts_with("@disabled")) {
-                nso_build_id = Common::HexStringToArray<0x20>(fmt::format("{:0<64}", line.substr(8)));
+            patches.push_back({{}, false}); //disabled patch
-                break;
+        } else if (line.starts_with("@flag offset_shift ")) {
-            // @stop
-            case 's':
-            case 'S':
-                return false;
-            // @enabled
-            case 'e':
-            case 'E':
-                patches.push_back({{}, true});
-                break;
-            // @disabled
-            case 'd':
-            case 'D':
-                patches.push_back({{}, false});
-                break;
-            // @flag
-            case 'f':
-            case 'F': {
-                if (line.starts_with("@flag offset_shift")) {
             offset_shift = std::strtoll(line.data() + 19, nullptr, 0);  // Offset Shift Flag
+        } else if (line.starts_with("@little-endian")) {
+            is_little_endian = true;  // Set values to read as little endian
+        } else if (line.starts_with("@big-endian")) {
+            is_little_endian = false;  // Set values to read as big endian
         } else if (line.starts_with("@flag print_values")) {
             //print_values = true; // Force printing of applied values
-                }
+        } else if (line.starts_with("@")) {
-                break;
-            }
-            case 'l':
-            case 'L':
-                is_little_endian = true;
-                break;
-            // IPS parsers dont support big endian no more, we do due to backcompat
-            case 'b':
-            case 'B':
-                is_little_endian = false;
-                break;
-            default:
             LOG_WARNING(Loader, "Unknown flag {}", line);
-                break;
-            }
         } else {
             size_t offset = size_t(std::strtoul(line.data(), nullptr, 16));
             offset += size_t(offset_shift);
@@ -219,7 +197,6 @@ void IPSwitchCompiler::Parse(std::span<u8 const> bytes) {
                 auto const start = line.cbegin() + first_space + 1;
                 auto const end = line.cend();
                 if (start <= line.cend() && end <= line.cend()) {
-                    // Actually IPS wants ordering from {lsb, ..., msb} -- so LE and BE are inverted, fun!
                     auto const hs = Common::HexStringToVector({start, end}, is_little_endian);
                     std::memcpy(r.data.data(), hs.data(), hs.size());
                     r.count = hs.size();
@@ -254,8 +231,7 @@ void IPSwitchCompiler::Parse(std::span<u8 const> bytes) {
             char quote = '\0';
             auto const sline_start = p;
             for (; p < sline.cend(); ) {
-                // we dont check for "//", IPS checks for '/' only...
+                if ((!quote && p + 1 < sline.cend() && p[0] == '/' && p[1] == '/')
-                if ((!quote && p[0] == '/')
                 || (!quote && p[0] == '#')) {
                     break;
                 } else if (p[0] == '\"' || p[0] == '\'') {
@@ -290,8 +266,6 @@ VirtualFile IPSwitchCompiler::Apply(const VirtualFile& in) const {
                     if (record.first + replace_size > in_data.size())
                         replace_size = in_data.size() - record.first;
                     std::memcpy(in_data.data() + record.first, record.second.data.data(), replace_size);
-                } else {
-                    LOG_WARNING(Loader, "record offs={:x},size={:x}", record.first, record.second.data.size());
                 }
             }
         }

src/shader_recompiler/CMakeLists.txt

@@ -217,6 +217,7 @@ add_library(shader_recompiler STATIC
     frontend/maxwell/translate_program.h
     host_translate_info.h
     ir_opt/collect_shader_info_pass.cpp
+    ir_opt/compaction_fallback_pass.cpp
     ir_opt/conditional_barrier_pass.cpp
     ir_opt/constant_propagation_pass.cpp
     ir_opt/dead_code_elimination_pass.cpp

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -440,8 +440,7 @@ void SetupCapabilities(const Profile& profile, const Info& info, EmitContext& ct
     }
     if ((info.uses_subgroup_vote || info.uses_subgroup_invocation_id ||
          info.uses_subgroup_shuffles) &&
-        profile.support_vote &&
+        profile.support_vote && profile.SupportsSubgroupStage(ctx.stage)) {
-        (ctx.stage != Stage::Geometry || profile.support_subgroup_in_geometry_stage)) {
         ctx.AddCapability(spv::Capability::GroupNonUniformBallot);
         ctx.AddCapability(spv::Capability::GroupNonUniformShuffle);
         if (!profile.warp_size_potentially_larger_than_guest) {

src/shader_recompiler/backend/spirv/emit_spirv_warp.cpp

@@ -13,18 +13,6 @@ Id SubgroupScope(EmitContext& ctx) {
     return ctx.Const(static_cast<u32>(spv::Scope::Subgroup));
 }
 
-// Some mobile GPUs (e.g. Adreno/Turnip) only advertise subgroup ballot/shuffle support for the
-// fragment and compute stages (VkPhysicalDeviceSubgroupProperties::supportedStages), even though
-// they support these operations elsewhere. Guest shaders that use VOTE/SHFL in a geometry program
-// would otherwise emit GroupNonUniform* SPIR-V the driver never declared support for in that
-// stage. There is no barrier in the geometry stage, so a real cross-invocation emulation can't be
-// made correct; instead, treat the current invocation as if it were alone in its subgroup. This is
-// semantically wrong for guest code that relies on genuine cross-lane communication, but it is
-// well-defined, valid SPIR-V that doesn't depend on unsupported hardware capabilities.
-bool NeedsGeometrySubgroupFallback(EmitContext& ctx) {
-    return ctx.stage == Stage::Geometry && !ctx.profile.support_subgroup_in_geometry_stage;
-}
-
 bool StageSupportsSubgroups(EmitContext& ctx) {
     return ctx.profile.SupportsSubgroupStage(ctx.stage);
 }
@@ -106,9 +94,6 @@ Id AddPartitionBase(EmitContext& ctx, Id thread_id) {
 } // Anonymous namespace
 
 Id EmitLaneId(EmitContext& ctx) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return ctx.u32_zero_value;
-    }
     const Id id{GetThreadId(ctx)};
     if (!ctx.profile.warp_size_potentially_larger_than_guest) {
         return id;
@@ -117,9 +102,6 @@ Id EmitLaneId(EmitContext& ctx) {
 }
 
 Id EmitVoteAll(EmitContext& ctx, Id pred) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return pred;
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return pred;
     }
@@ -136,9 +118,6 @@ Id EmitVoteAll(EmitContext& ctx, Id pred) {
 }
 
 Id EmitVoteAny(EmitContext& ctx, Id pred) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return pred;
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return pred;
     }
@@ -155,9 +134,6 @@ Id EmitVoteAny(EmitContext& ctx, Id pred) {
 }
 
 Id EmitVoteEqual(EmitContext& ctx, Id pred) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return ctx.true_value;
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return ctx.true_value;
     }
@@ -175,16 +151,6 @@ Id EmitVoteEqual(EmitContext& ctx, Id pred) {
 }
 
 Id EmitSubgroupBallot(EmitContext& ctx, Id pred) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        // Reflect only this invocation's own predicate. There is no way to observe other
-        // invocations' predicates without real subgroup hardware support in this stage, so this
-        // is a best-effort approximation: it keeps any branch gated on "did anyone match" live
-        // (rather than letting the SPIR-V optimizer prove it dead, which previously caused
-        // indirect draws fed by this shader to see indexCount=instanceCount=0), but any downstream
-        // math that assumes a real cross-lane population count (e.g. popcount-based compaction
-        // offsets) will not be correct.
-        return ctx.OpSelect(ctx.U32[1], pred, ctx.Const(1U), ctx.u32_zero_value);
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return ctx.OpSelect(ctx.U32[1], pred, ctx.Const(1u), ctx.u32_zero_value);
     }
@@ -196,9 +162,6 @@ Id EmitSubgroupBallot(EmitContext& ctx, Id pred) {
 }
 
 Id EmitSubgroupEqMask(EmitContext& ctx) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return ctx.Const(1U);
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return ctx.Const(1u);
     }
@@ -206,9 +169,6 @@ Id EmitSubgroupEqMask(EmitContext& ctx) {
 }
 
 Id EmitSubgroupLtMask(EmitContext& ctx) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return ctx.u32_zero_value;
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return ctx.u32_zero_value;
     }
@@ -216,9 +176,6 @@ Id EmitSubgroupLtMask(EmitContext& ctx) {
 }
 
 Id EmitSubgroupLeMask(EmitContext& ctx) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return ctx.Const(1U);
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return ctx.Const(1u);
     }
@@ -226,9 +183,6 @@ Id EmitSubgroupLeMask(EmitContext& ctx) {
 }
 
 Id EmitSubgroupGtMask(EmitContext& ctx) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return ctx.u32_zero_value;
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return ctx.u32_zero_value;
     }
@@ -236,9 +190,6 @@ Id EmitSubgroupGtMask(EmitContext& ctx) {
 }
 
 Id EmitSubgroupGeMask(EmitContext& ctx) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        return ctx.Const(1U);
-    }
     if (!StageSupportsSubgroups(ctx)) {
         return ctx.Const(1u);
     }
@@ -247,10 +198,6 @@ Id EmitSubgroupGeMask(EmitContext& ctx) {
 
 Id EmitShuffleIndex(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
                     Id segmentation_mask) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
-        SetInBoundsFlag(inst, ctx.false_value);
-        return value;
-    }
     const Id not_seg_mask{ctx.OpNot(ctx.U32[1], segmentation_mask)};
     const Id thread_id{EmitLaneId(ctx)};
     const Id min_thread_id{ComputeMinThreadId(ctx, thread_id, segmentation_mask)};
@@ -270,7 +217,7 @@ Id EmitShuffleIndex(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id cla
 
 Id EmitShuffleUp(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
                  Id segmentation_mask) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
+    if (!StageSupportsSubgroups(ctx)) {
         SetInBoundsFlag(inst, ctx.false_value);
         return value;
     }
@@ -289,7 +236,7 @@ Id EmitShuffleUp(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
 
 Id EmitShuffleDown(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
                    Id segmentation_mask) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
+    if (!StageSupportsSubgroups(ctx)) {
         SetInBoundsFlag(inst, ctx.false_value);
         return value;
     }
@@ -308,7 +255,7 @@ Id EmitShuffleDown(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clam
 
 Id EmitShuffleButterfly(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id clamp,
                         Id segmentation_mask) {
-    if (NeedsGeometrySubgroupFallback(ctx)) {
+    if (!StageSupportsSubgroups(ctx)) {
         SetInBoundsFlag(inst, ctx.false_value);
         return value;
     }

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -1455,14 +1455,13 @@ void EmitContext::DefineInputs(const IR::Program& program) {
     if (info.uses_is_helper_invocation) {
         is_helper_invocation = DefineInput(*this, U1, false, spv::BuiltIn::HelperInvocation);
     }
-    if (info.uses_subgroup_mask &&
+    if (info.uses_subgroup_mask && profile.SupportsSubgroupStage(stage)) {
-        (stage != Stage::Geometry || profile.support_subgroup_in_geometry_stage)) {
         subgroup_mask_eq = DefineInput(*this, U32[4], false, spv::BuiltIn::SubgroupEqMaskKHR);
         subgroup_mask_lt = DefineInput(*this, U32[4], false, spv::BuiltIn::SubgroupLtMaskKHR);
         subgroup_mask_le = DefineInput(*this, U32[4], false, spv::BuiltIn::SubgroupLeMaskKHR);
         subgroup_mask_gt = DefineInput(*this, U32[4], false, spv::BuiltIn::SubgroupGtMaskKHR);
         subgroup_mask_ge = DefineInput(*this, U32[4], false, spv::BuiltIn::SubgroupGeMaskKHR);
-        if (profile.support_explicit_workgroup_layout) {
+        if (stage == Stage::Fragment) {
             Decorate(subgroup_mask_eq, spv::Decoration::Flat);
             Decorate(subgroup_mask_lt, spv::Decoration::Flat);
             Decorate(subgroup_mask_le, spv::Decoration::Flat);
@@ -1470,11 +1469,10 @@ void EmitContext::DefineInputs(const IR::Program& program) {
             Decorate(subgroup_mask_ge, spv::Decoration::Flat);
         }
     }
-    if (info.uses_fswzadd ||
+    if ((info.uses_fswzadd || info.uses_subgroup_invocation_id || info.uses_subgroup_shuffles ||
-        ((info.uses_subgroup_invocation_id || info.uses_subgroup_shuffles ||
          (profile.warp_size_potentially_larger_than_guest &&
           (info.uses_subgroup_vote || info.uses_subgroup_mask))) &&
-         (stage != Stage::Geometry || profile.support_subgroup_in_geometry_stage))) {
+        profile.SupportsSubgroupStage(stage)) {
         AddCapability(spv::Capability::GroupNonUniform);
         subgroup_local_invocation_id =
             DefineInput(*this, U32[1], false, spv::BuiltIn::SubgroupLocalInvocationId);

src/shader_recompiler/frontend/maxwell/translate_program.cpp

@@ -299,6 +299,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
     Optimization::PositionPass(env, program);
 
     Optimization::GlobalMemoryToStorageBufferPass(program, normalized_host_info);
+    Optimization::CompactionFallbackPass(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

@@ -38,6 +38,7 @@ struct HostTranslateInfo {
                                                 ///< passthrough shaders
     bool support_conditional_barrier{}; ///< True when the device supports barriers in conditional
                                         ///< control flow
+    bool support_subgroup_in_geometry_stage{};
 
     void ApplyDescriptorLimitPolicy() noexcept {
         if (min_ssbo_alignment == 0) {

src/shader_recompiler/ir_opt/compaction_fallback_pass.cpp

@@ -0,0 +1,69 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#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_SEARCH_DEPTH = 8;
+
+bool DependsOnSubgroupBallot(const IR::Value& value, int depth) {
+    if (depth > MAX_BALLOT_SEARCH_DEPTH || value.IsImmediate()) {
+        return false;
+    }
+    IR::Inst* const producer{value.Inst()};
+    if (producer->GetOpcode() == IR::Opcode::SubgroupBallot) {
+        return true;
+    }
+    if (producer->GetOpcode() == IR::Opcode::Phi) {
+        return false;
+    }
+    const size_t num_args{producer->NumArgs()};
+    for (size_t index = 0; index < num_args; ++index) {
+        if (DependsOnSubgroupBallot(producer->Arg(index), depth + 1)) {
+            return true;
+        }
+    }
+    return false;
+}
+
+void RewriteAtomic(IR::Block& block, IR::Inst& inst) {
+    const auto insert_point{IR::Block::InstructionList::s_iterator_to(inst)};
+    IR::IREmitter ir{block, insert_point};
+
+    const IR::U32 amount{inst.Arg(2)};
+    const IR::U32 primitive_id{ir.GetAttributeU32(IR::Attribute::PrimitiveId)};
+    const IR::U32 new_index{ir.IMul(primitive_id, amount)};
+    const IR::U32 new_atomic_value{ir.IAdd(new_index, amount)};
+
+    const IR::Value umax_result{&*block.PrependNewInst(
+        insert_point, IR::Opcode::StorageAtomicUMax32,
+        {inst.Arg(0), inst.Arg(1), new_atomic_value})};
+    static_cast<void>(umax_result);
+
+    inst.ReplaceUsesWith(new_index);
+}
+
+} // Anonymous namespace
+
+void CompactionFallbackPass(IR::Program& program, const HostTranslateInfo& host_info) {
+    if (program.stage != Stage::Geometry || host_info.support_subgroup_in_geometry_stage) {
+        return;
+    }
+    for (IR::Block* const block : program.post_order_blocks) {
+        for (IR::Inst& inst : block->Instructions()) {
+            if (inst.GetOpcode() != IR::Opcode::StorageAtomicIAdd32) {
+                continue;
+            }
+            if (!DependsOnSubgroupBallot(inst.Arg(2), 0)) {
+                continue;
+            }
+            RewriteAtomic(*block, inst);
+        }
+    }
+}
+
+} // namespace Shader::Optimization

src/shader_recompiler/ir_opt/passes.h

@@ -13,6 +13,7 @@ struct HostTranslateInfo;
 namespace Shader::Optimization {
 
 void CollectShaderInfoPass(Environment& env, IR::Program& program);
+void CompactionFallbackPass(IR::Program& program, const HostTranslateInfo& host_info);
 void ConditionalBarrierPass(IR::Program& program);
 void ConstantPropagationPass(Environment& env, IR::Program& program);
 void DeadCodeEliminationPass(IR::Program& program);

src/shader_recompiler/profile.h

@@ -41,12 +41,6 @@ struct Profile {
     bool support_quad_shuffles{};
     bool support_vote{};
     u32 supported_subgroup_stages{0x7F};
-    bool support_subgroup_in_geometry_stage{}; ///< True when the device advertises subgroup
-                                               ///< ballot/shuffle support for VK_SHADER_STAGE_GEOMETRY_BIT
-                                               ///< (VkPhysicalDeviceSubgroupProperties::supportedStages).
-                                               ///< Many mobile GPUs support subgroup ops only in
-                                               ///< fragment/compute; guest shaders using VOTE/SHFL in a
-                                               ///< geometry program need a non-subgroup fallback there.
     bool support_viewport_index_layer_non_geometry{};
     bool support_viewport_mask{};
     bool support_typeless_image_loads{};

src/video_core/renderer_opengl/gl_shader_cache.cpp

@@ -269,6 +269,7 @@ ShaderCache::ShaderCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .support_viewport_index_layer = device.HasVertexViewportLayer(),
         .support_geometry_shader_passthrough = device.HasGeometryShaderPassthrough(),
         .support_conditional_barrier = device.SupportsConditionalBarriers(),
+        .support_subgroup_in_geometry_stage = true,
       } {
     host_info.ApplyDescriptorLimitPolicy();
     if (use_asynchronous_shaders) {

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -410,11 +410,6 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .support_quad_shuffles = device.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_QUAD_BIT),
         .support_vote = device.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_VOTE_BIT),
         .supported_subgroup_stages = supported_subgroup_stages,
-        .support_subgroup_in_geometry_stage =
-            device.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_VOTE_BIT) &&
-            device.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_BALLOT_BIT) &&
-            device.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_SHUFFLE_BIT) &&
-            device.IsSubgroupFeatureSupportedInStage(VK_SHADER_STAGE_GEOMETRY_BIT),
         .support_viewport_index_layer_non_geometry =
             device.IsExtShaderViewportIndexLayerSupported(),
         .support_viewport_mask = device.IsNvViewportArray2Supported(),
@@ -484,6 +479,7 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .support_viewport_index_layer = device.IsExtShaderViewportIndexLayerSupported(),
         .support_geometry_shader_passthrough = device.IsNvGeometryShaderPassthroughSupported(),
         .support_conditional_barrier = device.SupportsConditionalBarriers(),
+        .support_subgroup_in_geometry_stage = profile.SupportsSubgroupStage(Shader::Stage::Geometry),
     };
     host_info.ApplyDescriptorLimitPolicy();
 

src/video_core/vulkan_common/vulkan_device.h

@@ -479,11 +479,6 @@ FN_MAX_LIMIT_LIST
         return properties.subgroup_properties.supportedStages;
     }
 
-    /// Returns true if the device supports subgroup ballot/shuffle in the given shader stage.
-    bool IsSubgroupFeatureSupportedInStage(VkShaderStageFlagBits stage) const {
-        return properties.subgroup_properties.supportedStages & stage;
-    }
-
     /// Returns the maximum number of push descriptors.
     u32 MaxPushDescriptors() const {
         return properties.push_descriptor.maxPushDescriptors;