Shader OpCapability fallback + fp16 ban adjustments

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -332,6 +332,9 @@ void DefineEntryPoint(const IR::Program& program, EmitContext& ctx, Id main) {
 void SetupDenormControl(const Profile& profile, const IR::Program& program, EmitContext& ctx,
                         Id main_func) {
     const Info& info{program.info};
+    if (profile.has_broken_fp16_float_controls && info.uses_fp16) {
+        return;
+    }
     if (info.uses_fp32_denorms_flush && info.uses_fp32_denorms_preserve) {
         LOG_DEBUG(Shader_SPIRV, "Fp32 denorm flush and preserve on the same shader");
     } else if (info.uses_fp32_denorms_flush) {
@@ -432,7 +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_atomic.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
 
@@ -15,7 +18,7 @@ Id SharedPointer(EmitContext& ctx, Id offset, u32 index_offset = 0) {
     if (index_offset > 0) {
         index = ctx.OpIAdd(ctx.U32[1], index, ctx.Const(index_offset));
     }
-    return ctx.profile.support_explicit_workgroup_layout
+    return ctx.uses_explicit_workgroup_layout
                ? ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, ctx.u32_zero_value, index)
                : ctx.OpAccessChain(ctx.shared_u32, ctx.shared_memory_u32, index);
 }
@@ -155,7 +158,7 @@ Id EmitSharedAtomicExchange32(EmitContext& ctx, Id offset, Id value) {
 }
 
 Id EmitSharedAtomicExchange64(EmitContext& ctx, Id offset, Id value) {
-    if (ctx.profile.support_int64_atomics && ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.profile.support_shared_int64_atomics && ctx.uses_explicit_workgroup_layout) {
         const Id shift_id{ctx.Const(3U)};
         const Id index{ctx.OpShiftRightArithmetic(ctx.U32[1], offset, shift_id)};
         const Id pointer{

src/shader_recompiler/backend/spirv/emit_spirv_memory.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
 

src/shader_recompiler/backend/spirv/emit_spirv_shared_memory.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
 
@@ -28,7 +31,7 @@ std::pair<Id, Id> ExtractArgs(EmitContext& ctx, Id offset, u32 mask, u32 count)
 } // Anonymous namespace
 
 Id EmitLoadSharedU8(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{
             ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
         return ctx.OpUConvert(ctx.U32[1], ctx.OpLoad(ctx.U8, pointer));
@@ -39,7 +42,7 @@ Id EmitLoadSharedU8(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedS8(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{
             ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
         return ctx.OpSConvert(ctx.U32[1], ctx.OpLoad(ctx.U8, pointer));
@@ -50,7 +53,7 @@ Id EmitLoadSharedS8(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedU16(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
         return ctx.OpUConvert(ctx.U32[1], ctx.OpLoad(ctx.U16, pointer));
     } else {
@@ -60,7 +63,7 @@ Id EmitLoadSharedU16(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedS16(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
         return ctx.OpSConvert(ctx.U32[1], ctx.OpLoad(ctx.U16, pointer));
     } else {
@@ -70,7 +73,7 @@ Id EmitLoadSharedS16(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedU32(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32, ctx.shared_memory_u32, offset, 2)};
         return ctx.OpLoad(ctx.U32[1], pointer);
     } else {
@@ -79,7 +82,7 @@ Id EmitLoadSharedU32(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedU64(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32x2, ctx.shared_memory_u32x2, offset, 3)};
         return ctx.OpLoad(ctx.U32[2], pointer);
     } else {
@@ -94,7 +97,7 @@ Id EmitLoadSharedU64(EmitContext& ctx, Id offset) {
 }
 
 Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32x4, ctx.shared_memory_u32x4, offset, 4)};
         return ctx.OpLoad(ctx.U32[4], pointer);
     }
@@ -110,7 +113,7 @@ Id EmitLoadSharedU128(EmitContext& ctx, Id offset) {
 }
 
 void EmitWriteSharedU8(EmitContext& ctx, Id offset, Id value) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{
             ctx.OpAccessChain(ctx.shared_u8, ctx.shared_memory_u8, ctx.u32_zero_value, offset)};
         ctx.OpStore(pointer, ctx.OpUConvert(ctx.U8, value));
@@ -120,7 +123,7 @@ void EmitWriteSharedU8(EmitContext& ctx, Id offset, Id value) {
 }
 
 void EmitWriteSharedU16(EmitContext& ctx, Id offset, Id value) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u16, ctx.shared_memory_u16, offset, 1)};
         ctx.OpStore(pointer, ctx.OpUConvert(ctx.U16, value));
     } else {
@@ -130,7 +133,7 @@ void EmitWriteSharedU16(EmitContext& ctx, Id offset, Id value) {
 
 void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value) {
     Id pointer{};
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         pointer = Pointer(ctx, ctx.shared_u32, ctx.shared_memory_u32, offset, 2);
     } else {
         const Id shift{ctx.Const(2U)};
@@ -141,7 +144,7 @@ void EmitWriteSharedU32(EmitContext& ctx, Id offset, Id value) {
 }
 
 void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32x2, ctx.shared_memory_u32x2, offset, 3)};
         ctx.OpStore(pointer, value);
         return;
@@ -156,7 +159,7 @@ void EmitWriteSharedU64(EmitContext& ctx, Id offset, Id value) {
 }
 
 void EmitWriteSharedU128(EmitContext& ctx, Id offset, Id value) {
-    if (ctx.profile.support_explicit_workgroup_layout) {
+    if (ctx.uses_explicit_workgroup_layout) {
         const Id pointer{Pointer(ctx, ctx.shared_u32x4, ctx.shared_memory_u32x4, offset, 4)};
         ctx.OpStore(pointer, value);
         return;

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

@@ -371,7 +371,7 @@ Id CasFunction(EmitContext& ctx, Operation operation, Id value_type) {
 Id CasLoop(EmitContext& ctx, Operation operation, Id array_pointer, Id element_pointer,
            Id value_type, Id memory_type, spv::Scope scope) {
     const bool is_shared{scope == spv::Scope::Workgroup};
-    const bool is_struct{!is_shared || ctx.profile.support_explicit_workgroup_layout};
+    const bool is_struct{!is_shared || ctx.uses_explicit_workgroup_layout};
     const Id cas_func{CasFunction(ctx, operation, value_type)};
     const Id zero{ctx.u32_zero_value};
     const Id scope_id{ctx.Const(static_cast<u32>(scope))};
@@ -620,7 +620,11 @@ void EmitContext::DefineSharedMemory(const IR::Program& program) {
 
         return std::make_tuple(variable, element_pointer, pointer);
     }};
-    if (profile.support_explicit_workgroup_layout) {
+    uses_explicit_workgroup_layout =
+        profile.support_explicit_workgroup_layout &&
+        (!program.info.uses_int8 || profile.support_workgroup_layout_8bit_access) &&
+        (!program.info.uses_int16 || profile.support_workgroup_layout_16bit_access);
+    if (uses_explicit_workgroup_layout) {
         AddExtension("SPV_KHR_workgroup_memory_explicit_layout");
         AddCapability(spv::Capability::WorkgroupMemoryExplicitLayoutKHR);
         if (program.info.uses_int8) {
@@ -938,6 +942,10 @@ void EmitContext::DefineGlobalMemoryFunctions(const Info& info) {
     if (!info.uses_global_memory || !profile.support_int64) {
         return;
     }
+    if (!profile.support_descriptor_aliasing) {
+        DefineGlobalMemoryFunctionsU32Fallback(info);
+        return;
+    }
     using DefPtr = Id StorageDefinitions::*;
     const Id zero{u32_zero_value};
     const auto define_body{[&](DefPtr ssbo_member, Id addr, Id element_pointer, u32 shift,
@@ -1015,6 +1023,111 @@ void EmitContext::DefineGlobalMemoryFunctions(const Info& info) {
         define(&StorageDefinitions::U32x4, storage_types.U32x4, U32[4], sizeof(u32[4]));
 }
 
+void EmitContext::DefineGlobalMemoryFunctionsU32Fallback(const Info& info) {
+    const Id zero{u32_zero_value};
+    const auto define_body{[&](Id addr, u32 num_words, auto&& callback) {
+        AddLabel();
+        const size_t num_buffers{info.storage_buffers_descriptors.size()};
+        for (size_t index = 0; index < num_buffers; ++index) {
+            if (!info.nvn_buffer_used[index]) {
+                continue;
+            }
+            const auto& ssbo{info.storage_buffers_descriptors[index]};
+            const u32 addr_word{ssbo.cbuf_offset / 4};
+            const u32 addr_lo_comp{addr_word % 4};
+            const Id cbuf{cbufs[ssbo.cbuf_index].U32x4};
+            const Id addr_vec_pointer{
+                OpAccessChain(uniform_types.U32x4, cbuf, zero, Const(addr_word / 4))};
+            const Id addr_vec{OpLoad(U32[4], addr_vec_pointer)};
+            const Id addr_lo{OpCompositeExtract(U32[1], addr_vec, addr_lo_comp)};
+            const Id addr_hi{OpCompositeExtract(U32[1], addr_vec, addr_lo_comp + 1U)};
+            const Id unaligned_addr{
+                OpBitcast(U64, OpCompositeConstruct(U32[2], addr_lo, addr_hi))};
+
+            const u64 ssbo_align_mask{~(profile.min_ssbo_alignment - 1U)};
+            const Id ssbo_addr{OpBitwiseAnd(U64, unaligned_addr, Constant(U64, ssbo_align_mask))};
+
+            const u32 size_word{addr_word + 2};
+            Id size_vec{addr_vec};
+            if (size_word / 4 != addr_word / 4) {
+                const Id size_vec_pointer{
+                    OpAccessChain(uniform_types.U32x4, cbuf, zero, Const(size_word / 4))};
+                size_vec = OpLoad(U32[4], size_vec_pointer);
+            }
+            const Id ssbo_size{
+                OpUConvert(U64, OpCompositeExtract(U32[1], size_vec, size_word % 4))};
+            const Id ssbo_end{OpIAdd(U64, ssbo_addr, ssbo_size)};
+            const Id cond{OpLogicalAnd(U1, OpUGreaterThanEqual(U1, addr, ssbo_addr),
+                                       OpULessThan(U1, addr, ssbo_end))};
+            const Id then_label{OpLabel()};
+            const Id else_label{OpLabel()};
+            OpSelectionMerge(else_label, spv::SelectionControlMask::MaskNone);
+            OpBranchConditional(cond, then_label, else_label);
+            AddLabel(then_label);
+            const Id ssbo_id{ssbos[index].U32};
+            const Id ssbo_offset{OpUConvert(U32[1], OpISub(U64, addr, ssbo_addr))};
+            const Id base_word{OpShiftRightLogical(U32[1], ssbo_offset, Const(2U))};
+            std::array<Id, 4> word_pointers{};
+            for (u32 word = 0; word < num_words; ++word) {
+                const Id word_index{word == 0 ? base_word
+                                              : OpIAdd(U32[1], base_word, Const(word))};
+                word_pointers[word] =
+                    OpAccessChain(storage_types.U32.element, ssbo_id, zero, word_index);
+            }
+            callback(word_pointers);
+            AddLabel(else_label);
+        }
+    }};
+    const auto define_load{[&](Id type, u32 num_words) {
+        const Id function_type{TypeFunction(type, U64)};
+        const Id func_id{OpFunction(type, spv::FunctionControlMask::MaskNone, function_type)};
+        const Id addr{OpFunctionParameter(U64)};
+        define_body(addr, num_words, [&](const std::array<Id, 4>& pointers) {
+            std::array<Id, 4> words{};
+            for (u32 word = 0; word < num_words; ++word) {
+                words[word] = OpLoad(U32[1], pointers[word]);
+            }
+            switch (num_words) {
+            case 1:
+                OpReturnValue(words[0]);
+                break;
+            case 2:
+                OpReturnValue(OpCompositeConstruct(type, words[0], words[1]));
+                break;
+            default:
+                OpReturnValue(
+                    OpCompositeConstruct(type, words[0], words[1], words[2], words[3]));
+                break;
+            }
+        });
+        OpReturnValue(ConstantNull(type));
+        OpFunctionEnd();
+        return func_id;
+    }};
+    const auto define_write{[&](Id type, u32 num_words) {
+        const Id function_type{TypeFunction(void_id, U64, type)};
+        const Id func_id{OpFunction(void_id, spv::FunctionControlMask::MaskNone, function_type)};
+        const Id addr{OpFunctionParameter(U64)};
+        const Id data{OpFunctionParameter(type)};
+        define_body(addr, num_words, [&](const std::array<Id, 4>& pointers) {
+            for (u32 word = 0; word < num_words; ++word) {
+                const Id value{num_words == 1 ? data : OpCompositeExtract(U32[1], data, word)};
+                OpStore(pointers[word], value);
+            }
+            OpReturn();
+        });
+        OpReturn();
+        OpFunctionEnd();
+        return func_id;
+    }};
+    load_global_func_u32 = define_load(U32[1], 1);
+    load_global_func_u32x2 = define_load(U32[2], 2);
+    load_global_func_u32x4 = define_load(U32[4], 4);
+    write_global_func_u32 = define_write(U32[1], 1);
+    write_global_func_u32x2 = define_write(U32[2], 2);
+    write_global_func_u32x4 = define_write(U32[4], 4);
+}
+
 void EmitContext::DefineRescalingInput(const Info& info) {
     if (!info.uses_rescaling_uniform) {
         return;
@@ -1438,7 +1551,7 @@ 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);
@@ -1452,9 +1565,10 @@ 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/backend/spirv/spirv_emit_context.h

@@ -310,6 +310,10 @@ public:
 
     Id local_memory{};
 
+    /// True when this shader's shared memory uses SPV_KHR_workgroup_memory_explicit_layout.
+    /// False when the host lacks the extension or a width this shader accesses natively.
+    bool uses_explicit_workgroup_layout{};
+
     Id shared_memory_u8{};
     Id shared_memory_u16{};
     Id shared_memory_u32{};
@@ -387,6 +391,7 @@ private:
     void DefineAttributeMemAccess(const Info& info);
     void DefineWriteStorageCasLoopFunction(const Info& info);
     void DefineGlobalMemoryFunctions(const Info& info);
+    void DefineGlobalMemoryFunctionsU32Fallback(const Info& info);
     void DefineRescalingInput(const Info& info);
     void DefineRescalingInputPushConstant();
     void DefineRescalingInputUniformConstant();

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{};
@@ -29,12 +31,16 @@ struct Profile {
     bool support_fp32_signed_zero_nan_preserve{};
     bool support_fp64_signed_zero_nan_preserve{};
     bool support_explicit_workgroup_layout{};
+    bool support_workgroup_layout_8bit_access{};
+    bool support_workgroup_layout_16bit_access{};
     bool support_vote{};
+    u32 supported_subgroup_stages{0x7F};
     bool support_viewport_index_layer_non_geometry{};
     bool support_viewport_mask{};
     bool support_typeless_image_loads{};
     bool support_demote_to_helper_invocation{};
     bool support_int64_atomics{};
+    bool support_shared_int64_atomics{};
     bool support_derivative_control{};
     bool support_geometry_shader_passthrough{};
     bool support_native_ndc{};
@@ -93,6 +99,10 @@ 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

@@ -378,6 +378,20 @@ 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,
@@ -402,7 +416,12 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .support_fp64_signed_zero_nan_preserve =
             float_control.shaderSignedZeroInfNanPreserveFloat64 != VK_FALSE,
         .support_explicit_workgroup_layout = device.IsKhrWorkgroupMemoryExplicitLayoutSupported(),
+        .support_workgroup_layout_8bit_access =
+            device.IsWorkgroupMemoryExplicitLayout8BitSupported(),
+        .support_workgroup_layout_16bit_access =
+            device.IsWorkgroupMemoryExplicitLayout16BitSupported(),
         .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(),
@@ -410,6 +429,7 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .support_demote_to_helper_invocation =
             device.IsExtShaderDemoteToHelperInvocationSupported(),
         .support_int64_atomics = device.IsExtShaderAtomicInt64Supported(),
+        .support_shared_int64_atomics = device.IsSharedInt64AtomicsSupported(),
         .support_derivative_control = true,
         .support_geometry_shader_passthrough = device.IsNvGeometryShaderPassthroughSupported(),
         .support_native_ndc = device.IsExtDepthClipControlSupported(),
@@ -433,7 +453,8 @@ PipelineCache::PipelineCache(Tegra::MaxwellDeviceMemoryManager& device_memory_,
         .has_broken_spirv_position_input = driver_id == false,
         .has_broken_unsigned_image_offsets = false,
         .has_broken_signed_operations = false,
-        .has_broken_fp16_float_controls = driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY,
+        .has_broken_fp16_float_controls = driver_id == VK_DRIVER_ID_NVIDIA_PROPRIETARY ||
+                                          driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY,
         .ignore_nan_fp_comparisons = false,
         .has_broken_spirv_subgroup_mask_vector_extract_dynamic = false,
         .has_broken_robust =
@@ -904,12 +925,8 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline(
     }
 
     auto program{TranslateProgram(pools.inst, pools.block, env, cfg, host_info)};
-    const VkDriverIdKHR driver_id = device.GetDriverID();
-    const bool needs_shared_mem_clamp =
-        driver_id == VK_DRIVER_ID_QUALCOMM_PROPRIETARY ||
-        driver_id == VK_DRIVER_ID_ARM_PROPRIETARY;
     const u32 max_shared_memory = device.GetMaxComputeSharedMemorySize();
-    if (needs_shared_mem_clamp && program.shared_memory_size > max_shared_memory) {
+    if (program.shared_memory_size > max_shared_memory) {
         LOG_WARNING(Render_Vulkan,
                     "Compute shader 0x{:016x} requests {}KB shared memory but device max is {}KB - clamping",
                     key.unique_hash,

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -1255,6 +1255,18 @@ void RasterizerVulkan::UpdateDepthBias(Tegra::Engines::Maxwell3D::Regs& regs) {
         }
     }
 
+    const bool is_float_depth =
+        regs.zeta.format == Tegra::DepthFormat::Z32_FLOAT ||
+        regs.zeta.format == Tegra::DepthFormat::Z32_FLOAT_X24S8_UINT;
+    if (is_float_depth && units != 0.0f && !device.IsExtDepthBiasControlSupported()) {
+        static bool logged_float_bias_warning = false;
+        if (!logged_float_bias_warning) {
+            logged_float_bias_warning = true;
+            LOG_WARNING(Render_Vulkan,
+                        "Depth bias on a float depth target without VK_EXT_depth_bias_control");
+        }
+    }
+
     scheduler.Record([constant = units, clamp = regs.depth_bias_clamp,
                       factor = regs.slope_scale_depth_bias, this](vk::CommandBuffer cmdbuf) {
         if (device.IsExtDepthBiasControlSupported()) {

src/video_core/texture_cache/texture_cache.h

@@ -1915,13 +1915,16 @@ void TextureCache<P>::TrimInactiveSamplers(size_t budget) {
         ankerl::unordered_dense::set<SamplerId> active_sampler_ids;
         for (auto const& e : channel_state->sampler_ids)
             active_sampler_ids.insert(e.second);
+        if constexpr (requires { runtime.Finish(); }) {
+            runtime.Finish();
+        }
         // Elements in the map must be necesarily valid
         size_t removed = 0;
         for (auto it = channel_state->samplers.begin(); it != channel_state->samplers.end();) {
             const SamplerId sampler_id = it->second;
             if (!sampler_id || sampler_id == CORRUPT_ID) {
                 it = channel_state->samplers.erase(it);
-            } else if (std::ranges::find(active_sampler_ids, sampler_id) != active_sampler_ids.end()) {
+            } else if (active_sampler_ids.contains(sampler_id)) {
                 ++it;
             } else {
                 slot_samplers.erase(sampler_id);

src/video_core/vulkan_common/vulkan_device.cpp

@@ -501,14 +501,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
         LOG_WARNING(Render_Vulkan,
                     "Qualcomm drivers have slow push descriptor implementation");
         RemoveExtension(extensions.push_descriptor, VK_KHR_PUSH_DESCRIPTOR_EXTENSION_NAME);
-        LOG_WARNING(Render_Vulkan,
-                    "Disabling shader float controls and 64-bit integer features on Qualcomm proprietary drivers");
-        RemoveExtension(extensions.shader_float_controls, VK_KHR_SHADER_FLOAT_CONTROLS_EXTENSION_NAME);
-        RemoveExtensionFeature(extensions.shader_atomic_int64, features.shader_atomic_int64,
-                               VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
-        features.shader_atomic_int64.shaderBufferInt64Atomics = false;
-        features.shader_atomic_int64.shaderSharedInt64Atomics = false;
-        features.features.shaderInt64 = false;
 
 #if defined(__ANDROID__) && defined(ARCHITECTURE_arm64)
         // BCn patching only safe on Android 9+ (API 28+). Older versions crash on driver load.
@@ -561,7 +553,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
             features.shader_float16_int8.shaderFloat16 = false;
         }
 
-        // Mali/ NVIDIA proprietary drivers: Shader stencil export not supported
         // Use hardware depth/stencil blits instead when available
         if (!extensions.shader_stencil_export) {
             LOG_INFO(Render_Vulkan,
@@ -574,8 +565,8 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
 
             if (!is_blit_depth24_stencil8_supported && !is_blit_depth32_stencil8_supported) {
                 LOG_WARNING(Render_Vulkan,
-                            "NVIDIA: Neither shader export nor hardware blits available for "
-                            "depth/stencil. Performance may be degraded.");
+                            "Neither shader export nor hardware blits available for "
+                            "depth/stencil.");
             }
         }
     }
@@ -653,21 +644,13 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
 
     if (is_turnip || is_qualcomm) {
         LOG_WARNING(Render_Vulkan, "Driver requires higher-than-reported binding limits");
-        properties.properties.limits.maxVertexInputBindings = 32;
+        properties.properties.limits.maxVertexInputBindings =
+            (std::max)(properties.properties.limits.maxVertexInputBindings, 32U);
     }
 
     const auto dyna_state = Settings::values.dyna_state.GetValue();
-
-    // Base dynamic states (VIEWPORT, SCISSOR, DEPTH_BIAS, etc.) are ALWAYS active in vk_graphics_pipeline.cpp
-    // This slider controls EXTENDED dynamic states with accumulative levels per Vulkan specs:
-    //   Level 0 = Core Dynamic States only (Vulkan 1.0)
-    //   Level 1 = Core + VK_EXT_extended_dynamic_state
-    //   Level 2 = Core + VK_EXT_extended_dynamic_state + VK_EXT_extended_dynamic_state2
-    //   Level 3 = Core + VK_EXT_extended_dynamic_state + VK_EXT_extended_dynamic_state2 + VK_EXT_extended_dynamic_state3
-
     switch (dyna_state) {
     case Settings::ExtendedDynamicState::Disabled:
-        // Level 0: Disable all extended dynamic state extensions
         RemoveExtensionFeature(extensions.extended_dynamic_state, features.extended_dynamic_state,
                               VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME);
         RemoveExtensionFeature(extensions.extended_dynamic_state2, features.extended_dynamic_state2,
@@ -678,7 +661,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
         dynamic_state3_enables = false;
         break;
     case Settings::ExtendedDynamicState::EDS1:
-        // Level 1: Enable EDS1, disable EDS2 and EDS3
         RemoveExtensionFeature(extensions.extended_dynamic_state2, features.extended_dynamic_state2,
                               VK_EXT_EXTENDED_DYNAMIC_STATE_2_EXTENSION_NAME);
         RemoveExtensionFeature(extensions.extended_dynamic_state3, features.extended_dynamic_state3,
@@ -687,7 +669,6 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
         dynamic_state3_enables = false;
         break;
     case Settings::ExtendedDynamicState::EDS2:
-        // Level 2: Enable EDS1 + EDS2, disable EDS3
         RemoveExtensionFeature(extensions.extended_dynamic_state3, features.extended_dynamic_state3,
                               VK_EXT_EXTENDED_DYNAMIC_STATE_3_EXTENSION_NAME);
         dynamic_state3_blending = false;
@@ -695,12 +676,9 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
         break;
     case Settings::ExtendedDynamicState::EDS3:
     default:
-        // Level 3: Enable all (EDS1 + EDS2 + EDS3)
         break;
     }
 
-    // VK_EXT_vertex_input_dynamic_state is independent from EDS
-    // It can be enabled even without extended_dynamic_state
     if (!Settings::values.vertex_input_dynamic_state.GetValue()) {
         RemoveExtensionFeature(extensions.vertex_input_dynamic_state, features.vertex_input_dynamic_state, VK_EXT_VERTEX_INPUT_DYNAMIC_STATE_EXTENSION_NAME);
     }
@@ -1282,8 +1260,7 @@ void Device::RemoveUnsuitableExtensions() {
                                        VK_EXT_IMAGE_ROBUSTNESS_EXTENSION_NAME);
 
     // VK_KHR_shader_atomic_int64
-    extensions.shader_atomic_int64 = features.shader_atomic_int64.shaderBufferInt64Atomics &&
-                                     features.shader_atomic_int64.shaderSharedInt64Atomics;
+    extensions.shader_atomic_int64 = features.shader_atomic_int64.shaderBufferInt64Atomics;
     RemoveExtensionFeatureIfUnsuitable(extensions.shader_atomic_int64, features.shader_atomic_int64,
                                        VK_KHR_SHADER_ATOMIC_INT64_EXTENSION_NAME);
 
@@ -1332,10 +1309,7 @@ void Device::RemoveUnsuitableExtensions() {
 
     // VK_KHR_workgroup_memory_explicit_layout
     extensions.workgroup_memory_explicit_layout =
-        features.features.shaderInt16 &&
         features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout &&
-        features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout8BitAccess &&
-        features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout16BitAccess &&
         features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayoutScalarBlockLayout;
     RemoveExtensionFeatureIfUnsuitable(extensions.workgroup_memory_explicit_layout,
                                        features.workgroup_memory_explicit_layout,

src/video_core/vulkan_common/vulkan_device.h

@@ -403,6 +403,11 @@ FN_MAX_LIMIT_LIST
         return properties.subgroup_properties.supportedOperations & feature;
     }
 
+    /// Returns the shader stages that support subgroup operations.
+    VkShaderStageFlags GetSubgroupSupportedStages() const {
+        return properties.subgroup_properties.supportedStages;
+    }
+
     /// Returns the maximum number of push descriptors.
     u32 MaxPushDescriptors() const {
         return properties.push_descriptor.maxPushDescriptors;
@@ -483,6 +488,18 @@ FN_MAX_LIMIT_LIST
         return extensions.workgroup_memory_explicit_layout;
     }
 
+    /// Returns true if the device supports 8-bit accesses to workgroup explicit layout memory.
+    bool IsWorkgroupMemoryExplicitLayout8BitSupported() const {
+        return extensions.workgroup_memory_explicit_layout &&
+               features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout8BitAccess;
+    }
+
+    /// Returns true if the device supports 16-bit accesses to workgroup explicit layout memory.
+    bool IsWorkgroupMemoryExplicitLayout16BitSupported() const {
+        return extensions.workgroup_memory_explicit_layout && features.features.shaderInt16 &&
+               features.workgroup_memory_explicit_layout.workgroupMemoryExplicitLayout16BitAccess;
+    }
+
     /// Returns true if the device supports VK_KHR_image_format_list.
     bool IsKhrImageFormatListSupported() const {
         return extensions.image_format_list || instance_version >= VK_API_VERSION_1_2;
@@ -711,9 +728,16 @@ FN_MAX_LIMIT_LIST
                features.provoking_vertex.transformFeedbackPreservesProvokingVertex;
     }
 
-    /// Returns true if the device supports VK_KHR_shader_atomic_int64.
+    /// Returns true if the device supports int64 atomics on storage buffers.
     bool IsExtShaderAtomicInt64Supported() const {
-        return extensions.shader_atomic_int64;
+        return extensions.shader_atomic_int64 &&
+               features.shader_atomic_int64.shaderBufferInt64Atomics;
+    }
+
+    /// Returns true if the device supports int64 atomics on workgroup (shared) memory.
+    bool IsSharedInt64AtomicsSupported() const {
+        return extensions.shader_atomic_int64 &&
+               features.shader_atomic_int64.shaderSharedInt64Atomics;
     }
 
     bool IsExtConditionalRendering() const {