[vulkan] Subgroups gating correction

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -435,7 +435,7 @@ 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.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

@@ -1,3 +1,6 @@
+// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
+// SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: Copyright 2021 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
@@ -10,7 +13,14 @@ Id SubgroupScope(EmitContext& ctx) {
     return ctx.Const(static_cast<u32>(spv::Scope::Subgroup));
 }
 
+bool StageSupportsSubgroups(EmitContext& ctx) {
+    return ctx.profile.SupportsSubgroupStage(ctx.stage);
+}
+
 Id GetThreadId(EmitContext& ctx) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return ctx.u32_zero_value;
+    }
     return ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id);
 }
 
@@ -68,6 +78,9 @@ Id GetMaxThreadId(EmitContext& ctx, Id thread_id, Id clamp, Id segmentation_mask
 }
 
 Id SelectValue(EmitContext& ctx, Id in_range, Id value, Id src_thread_id) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return value;
+    }
     return ctx.OpSelect(
         ctx.U32[1], in_range,
         ctx.OpGroupNonUniformShuffle(ctx.U32[1], SubgroupScope(ctx), value, src_thread_id), value);
@@ -89,6 +102,9 @@ Id EmitLaneId(EmitContext& ctx) {
 }
 
 Id EmitVoteAll(EmitContext& ctx, Id pred) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return pred;
+    }
     if (!ctx.profile.warp_size_potentially_larger_than_guest) {
         return ctx.OpGroupNonUniformAll(ctx.U1, SubgroupScope(ctx), pred);
     }
@@ -102,6 +118,9 @@ Id EmitVoteAll(EmitContext& ctx, Id pred) {
 }
 
 Id EmitVoteAny(EmitContext& ctx, Id pred) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return pred;
+    }
     if (!ctx.profile.warp_size_potentially_larger_than_guest) {
         return ctx.OpGroupNonUniformAny(ctx.U1, SubgroupScope(ctx), pred);
     }
@@ -115,6 +134,9 @@ Id EmitVoteAny(EmitContext& ctx, Id pred) {
 }
 
 Id EmitVoteEqual(EmitContext& ctx, Id pred) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return ctx.true_value;
+    }
     if (!ctx.profile.warp_size_potentially_larger_than_guest) {
         return ctx.OpGroupNonUniformAllEqual(ctx.U1, SubgroupScope(ctx), pred);
     }
@@ -129,6 +151,9 @@ Id EmitVoteEqual(EmitContext& ctx, Id pred) {
 }
 
 Id EmitSubgroupBallot(EmitContext& ctx, Id pred) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return ctx.OpSelect(ctx.U32[1], pred, ctx.Const(1u), ctx.u32_zero_value);
+    }
     const Id ballot{ctx.OpGroupNonUniformBallot(ctx.U32[4], SubgroupScope(ctx), pred)};
     if (!ctx.profile.warp_size_potentially_larger_than_guest) {
         return ctx.OpCompositeExtract(ctx.U32[1], ballot, 0U);
@@ -137,22 +162,37 @@ Id EmitSubgroupBallot(EmitContext& ctx, Id pred) {
 }
 
 Id EmitSubgroupEqMask(EmitContext& ctx) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return ctx.Const(1u);
+    }
     return LoadMask(ctx, ctx.subgroup_mask_eq);
 }
 
 Id EmitSubgroupLtMask(EmitContext& ctx) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return ctx.u32_zero_value;
+    }
     return LoadMask(ctx, ctx.subgroup_mask_lt);
 }
 
 Id EmitSubgroupLeMask(EmitContext& ctx) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return ctx.Const(1u);
+    }
     return LoadMask(ctx, ctx.subgroup_mask_le);
 }
 
 Id EmitSubgroupGtMask(EmitContext& ctx) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return ctx.u32_zero_value;
+    }
     return LoadMask(ctx, ctx.subgroup_mask_gt);
 }
 
 Id EmitSubgroupGeMask(EmitContext& ctx) {
+    if (!StageSupportsSubgroups(ctx)) {
+        return ctx.Const(1u);
+    }
     return LoadMask(ctx, ctx.subgroup_mask_ge);
 }
 
@@ -222,7 +262,7 @@ Id EmitShuffleButterfly(EmitContext& ctx, IR::Inst* inst, Id value, Id index, Id
 
 Id EmitFSwizzleAdd(EmitContext& ctx, Id op_a, Id op_b, Id swizzle) {
     const Id three{ctx.Const(3U)};
-    Id mask{ctx.OpLoad(ctx.U32[1], ctx.subgroup_local_invocation_id)};
+    Id mask{GetThreadId(ctx)};
     mask = ctx.OpBitwiseAnd(ctx.U32[1], mask, three);
     mask = ctx.OpShiftLeftLogical(ctx.U32[1], mask, ctx.Const(1U));
     mask = ctx.OpShiftRightLogical(ctx.U32[1], swizzle, mask);

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -1447,7 +1447,7 @@ 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)) {
         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);
@@ -1461,9 +1461,10 @@ void EmitContext::DefineInputs(const IR::Program& program) {
             Decorate(subgroup_mask_ge, spv::Decoration::Flat);
         }
     }
-    if (info.uses_fswzadd || info.uses_subgroup_invocation_id || info.uses_subgroup_shuffles ||
-        (profile.warp_size_potentially_larger_than_guest &&
-         (info.uses_subgroup_vote || info.uses_subgroup_mask))) {
+    if ((info.uses_fswzadd || info.uses_subgroup_invocation_id || info.uses_subgroup_shuffles ||
+         (profile.warp_size_potentially_larger_than_guest &&
+          (info.uses_subgroup_vote || info.uses_subgroup_mask))) &&
+        profile.SupportsSubgroupStage(stage)) {
         AddCapability(spv::Capability::GroupNonUniform);
         subgroup_local_invocation_id =
             DefineInput(*this, U32[1], false, spv::BuiltIn::SubgroupLocalInvocationId);

src/shader_recompiler/profile.h

@@ -10,6 +10,8 @@
 
 namespace Shader {
 
+enum class Stage : u32;
+
 struct Profile {
     u32 supported_spirv{0x00010000};
     bool unified_descriptor_binding{};
@@ -32,6 +34,7 @@ struct Profile {
     bool support_fp64_signed_zero_nan_preserve{};
     bool support_explicit_workgroup_layout{};
     bool support_vote{};
+    u32 supported_subgroup_stages{0x7F};
     bool support_viewport_index_layer_non_geometry{};
     bool support_viewport_mask{};
     bool support_typeless_image_loads{};
@@ -95,6 +98,10 @@ struct Profile {
 
     u64 min_ssbo_alignment{};
     u32 max_user_clip_distances{};
+
+    bool SupportsSubgroupStage(Stage stage) const {
+        return (supported_subgroup_stages & (1u << static_cast<u32>(stage))) != 0;
+    }
 };
 
 } // namespace Shader

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -353,6 +353,20 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
       serialization_thread(1, "VkPipelineSerialization") {
     const auto& float_control{device.FloatControlProperties()};
     const VkDriverId driver_id{device.GetDriverID()};
+    const VkShaderStageFlags subgroup_stages{device.GetSubgroupSupportedStages()};
+    const auto subgroup_stage_bit{[subgroup_stages](VkShaderStageFlags flag, Shader::Stage stage) {
+        return (subgroup_stages & flag) != 0 ? (1u << static_cast<u32>(stage)) : 0u;
+    }};
+    const u32 supported_subgroup_stages{
+        subgroup_stage_bit(VK_SHADER_STAGE_VERTEX_BIT, Shader::Stage::VertexA) |
+        subgroup_stage_bit(VK_SHADER_STAGE_VERTEX_BIT, Shader::Stage::VertexB) |
+        subgroup_stage_bit(VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT,
+                           Shader::Stage::TessellationControl) |
+        subgroup_stage_bit(VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT,
+                           Shader::Stage::TessellationEval) |
+        subgroup_stage_bit(VK_SHADER_STAGE_GEOMETRY_BIT, Shader::Stage::Geometry) |
+        subgroup_stage_bit(VK_SHADER_STAGE_FRAGMENT_BIT, Shader::Stage::Fragment) |
+        subgroup_stage_bit(VK_SHADER_STAGE_COMPUTE_BIT, Shader::Stage::Compute)};
     profile = Shader::Profile{
         .supported_spirv = device.SupportedSpirvVersion(),
         .unified_descriptor_binding = true,
@@ -382,6 +396,7 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
             float_control.shaderSignedZeroInfNanPreserveFloat64 != VK_FALSE,
         .support_explicit_workgroup_layout = device.IsKhrWorkgroupMemoryExplicitLayoutSupported(),
         .support_vote = device.IsSubgroupFeatureSupported(VK_SUBGROUP_FEATURE_VOTE_BIT),
+        .supported_subgroup_stages = supported_subgroup_stages,
         .support_viewport_index_layer_non_geometry =
             device.IsExtShaderViewportIndexLayerSupported(),
         .support_viewport_mask = device.IsNvViewportArray2Supported(),

src/video_core/vulkan_common/vulkan_device.h

@@ -430,6 +430,10 @@ FN_MAX_LIMIT_LIST
         return properties.subgroup_properties.supportedOperations & feature;
     }
 
+    VkShaderStageFlags GetSubgroupSupportedStages() const {
+        return properties.subgroup_properties.supportedStages;
+    }
+
     /// Returns the maximum number of push descriptors.
     u32 MaxPushDescriptors() const {
         return properties.push_descriptor.maxPushDescriptors;