spirv: Support OpenGL uniform buffers and change bindings

src/shader_recompiler/backend/spirv/emit_context.cpp

@@ -441,8 +441,13 @@ void VectorTypes::Define(Sirit::Module& sirit_ctx, Id base_type, std::string_vie
     }
 }
 
-EmitContext::EmitContext(const Profile& profile_, IR::Program& program, u32& binding)
+EmitContext::EmitContext(const Profile& profile_, IR::Program& program, Bindings& binding)
     : Sirit::Module(profile_.supported_spirv), profile{profile_}, stage{program.stage} {
+    const bool is_unified{profile.unified_descriptor_binding};
+    u32& uniform_binding{is_unified ? binding.unified : binding.uniform_buffer};
+    u32& storage_binding{is_unified ? binding.unified : binding.storage_buffer};
+    u32& texture_binding{is_unified ? binding.unified : binding.texture};
+    u32& image_binding{is_unified ? binding.unified : binding.image};
     AddCapability(spv::Capability::Shader);
     DefineCommonTypes(program.info);
     DefineCommonConstants();
@@ -450,12 +455,12 @@ EmitContext::EmitContext(const Profile& profile_, IR::Program& program, u32& bin
     DefineLocalMemory(program);
     DefineSharedMemory(program);
     DefineSharedMemoryFunctions(program);
-    DefineConstantBuffers(program.info, binding);
+    DefineConstantBuffers(program.info, uniform_binding);
-    DefineStorageBuffers(program.info, binding);
+    DefineStorageBuffers(program.info, storage_binding);
-    DefineTextureBuffers(program.info, binding);
+    DefineTextureBuffers(program.info, texture_binding);
-    DefineImageBuffers(program.info, binding);
+    DefineImageBuffers(program.info, image_binding);
-    DefineTextures(program.info, binding);
+    DefineTextures(program.info, texture_binding);
-    DefineImages(program.info, binding);
+    DefineImages(program.info, image_binding);
     DefineAttributeMemAccess(program.info);
     DefineGlobalMemoryFunctions(program.info);
     DefineLabels(program);
@@ -489,6 +494,20 @@ Id EmitContext::Def(const IR::Value& value) {
     }
 }
 
+Id EmitContext::BitOffset8(const IR::Value& offset) {
+    if (offset.IsImmediate()) {
+        return Const((offset.U32() % 4) * 8);
+    }
+    return OpBitwiseAnd(U32[1], OpShiftLeftLogical(U32[1], Def(offset), Const(3u)), Const(24u));
+}
+
+Id EmitContext::BitOffset16(const IR::Value& offset) {
+    if (offset.IsImmediate()) {
+        return Const(((offset.U32() / 2) % 2) * 16);
+    }
+    return OpBitwiseAnd(U32[1], OpShiftLeftLogical(U32[1], Def(offset), Const(3u)), Const(16u));
+}
+
 void EmitContext::DefineCommonTypes(const Info& info) {
     void_id = TypeVoid();
 
@@ -496,6 +515,7 @@ void EmitContext::DefineCommonTypes(const Info& info) {
 
     F32.Define(*this, TypeFloat(32), "f32");
     U32.Define(*this, TypeInt(32, false), "u32");
+    S32.Define(*this, TypeInt(32, true), "s32");
 
     private_u32 = Name(TypePointer(spv::StorageClass::Private, U32[1]), "private_u32");
 
@@ -889,13 +909,16 @@ void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) {
     if (info.constant_buffer_descriptors.empty()) {
         return;
     }
+    if (profile.support_descriptor_aliasing) {
         if (True(info.used_constant_buffer_types & IR::Type::U8)) {
             DefineConstBuffers(*this, info, &UniformDefinitions::U8, binding, U8, 'u', sizeof(u8));
             DefineConstBuffers(*this, info, &UniformDefinitions::S8, binding, S8, 's', sizeof(s8));
         }
         if (True(info.used_constant_buffer_types & IR::Type::U16)) {
-        DefineConstBuffers(*this, info, &UniformDefinitions::U16, binding, U16, 'u', sizeof(u16));
+            DefineConstBuffers(*this, info, &UniformDefinitions::U16, binding, U16, 'u',
-        DefineConstBuffers(*this, info, &UniformDefinitions::S16, binding, S16, 's', sizeof(s16));
+                               sizeof(u16));
+            DefineConstBuffers(*this, info, &UniformDefinitions::S16, binding, S16, 's',
+                               sizeof(s16));
         }
         if (True(info.used_constant_buffer_types & IR::Type::U32)) {
             DefineConstBuffers(*this, info, &UniformDefinitions::U32, binding, U32[1], 'u',
@@ -909,10 +932,15 @@ void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) {
             DefineConstBuffers(*this, info, &UniformDefinitions::U32x2, binding, U32[2], 'u',
                                sizeof(u32[2]));
         }
+        binding += static_cast<u32>(info.constant_buffer_descriptors.size());
+    } else {
+        DefineConstBuffers(*this, info, &UniformDefinitions::U32x4, binding, U32[4], 'u',
+                           sizeof(u32[4]));
         for (const ConstantBufferDescriptor& desc : info.constant_buffer_descriptors) {
             binding += desc.count;
         }
     }
+}
 
 void EmitContext::DefineStorageBuffers(const Info& info, u32& binding) {
     if (info.storage_buffers_descriptors.empty()) {
@@ -920,35 +948,37 @@ void EmitContext::DefineStorageBuffers(const Info& info, u32& binding) {
     }
     AddExtension("SPV_KHR_storage_buffer_storage_class");
 
-    if (True(info.used_storage_buffer_types & IR::Type::U8)) {
+    const IR::Type used_types{profile.support_descriptor_aliasing ? info.used_storage_buffer_types
+                                                                  : IR::Type::U32};
+    if (True(used_types & IR::Type::U8)) {
         DefineSsbos(*this, storage_types.U8, &StorageDefinitions::U8, info, binding, U8,
                     sizeof(u8));
         DefineSsbos(*this, storage_types.S8, &StorageDefinitions::S8, info, binding, S8,
                     sizeof(u8));
     }
-    if (True(info.used_storage_buffer_types & IR::Type::U16)) {
+    if (True(used_types & IR::Type::U16)) {
         DefineSsbos(*this, storage_types.U16, &StorageDefinitions::U16, info, binding, U16,
                     sizeof(u16));
         DefineSsbos(*this, storage_types.S16, &StorageDefinitions::S16, info, binding, S16,
                     sizeof(u16));
     }
-    if (True(info.used_storage_buffer_types & IR::Type::U32)) {
+    if (True(used_types & IR::Type::U32)) {
         DefineSsbos(*this, storage_types.U32, &StorageDefinitions::U32, info, binding, U32[1],
                     sizeof(u32));
     }
-    if (True(info.used_storage_buffer_types & IR::Type::F32)) {
+    if (True(used_types & IR::Type::F32)) {
         DefineSsbos(*this, storage_types.F32, &StorageDefinitions::F32, info, binding, F32[1],
                     sizeof(f32));
     }
-    if (True(info.used_storage_buffer_types & IR::Type::U64)) {
+    if (True(used_types & IR::Type::U64)) {
         DefineSsbos(*this, storage_types.U64, &StorageDefinitions::U64, info, binding, U64,
                     sizeof(u64));
     }
-    if (True(info.used_storage_buffer_types & IR::Type::U32x2)) {
+    if (True(used_types & IR::Type::U32x2)) {
         DefineSsbos(*this, storage_types.U32x2, &StorageDefinitions::U32x2, info, binding, U32[2],
                     sizeof(u32[2]));
     }
-    if (True(info.used_storage_buffer_types & IR::Type::U32x4)) {
+    if (True(used_types & IR::Type::U32x4)) {
         DefineSsbos(*this, storage_types.U32x4, &StorageDefinitions::U32x4, info, binding, U32[4],
                     sizeof(u32[4]));
     }

src/shader_recompiler/backend/spirv/emit_context.h

@@ -17,6 +17,14 @@ namespace Shader::Backend::SPIRV {
 
 using Sirit::Id;
 
+struct Bindings {
+    u32 unified{};
+    u32 uniform_buffer{};
+    u32 storage_buffer{};
+    u32 texture{};
+    u32 image{};
+};
+
 class VectorTypes {
 public:
     void Define(Sirit::Module& sirit_ctx, Id base_type, std::string_view name);
@@ -62,6 +70,7 @@ struct UniformDefinitions {
     Id U32{};
     Id F32{};
     Id U32x2{};
+    Id U32x4{};
 };
 
 struct StorageTypeDefinition {
@@ -101,11 +110,14 @@ struct GenericElementInfo {
 
 class EmitContext final : public Sirit::Module {
 public:
-    explicit EmitContext(const Profile& profile, IR::Program& program, u32& binding);
+    explicit EmitContext(const Profile& profile, IR::Program& program, Bindings& binding);
     ~EmitContext();
 
     [[nodiscard]] Id Def(const IR::Value& value);
 
+    [[nodiscard]] Id BitOffset8(const IR::Value& offset);
+    [[nodiscard]] Id BitOffset16(const IR::Value& offset);
+
     Id Const(u32 value) {
         return Constant(U32[1], value);
     }
@@ -139,6 +151,7 @@ public:
     Id U64{};
     VectorTypes F32;
     VectorTypes U32;
+    VectorTypes S32;
     VectorTypes F16;
     VectorTypes F64;
 

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -368,7 +368,7 @@ void PatchPhiNodes(IR::Program& program, EmitContext& ctx) {
 }
 } // Anonymous namespace
 
-std::vector<u32> EmitSPIRV(const Profile& profile, IR::Program& program, u32& binding) {
+std::vector<u32> EmitSPIRV(const Profile& profile, IR::Program& program, Bindings& binding) {
     EmitContext ctx{profile, program, binding};
     const Id main{DefineMain(ctx, program)};
     DefineEntryPoint(program, ctx, main);

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -14,7 +14,7 @@
 namespace Shader::Backend::SPIRV {
 
 [[nodiscard]] std::vector<u32> EmitSPIRV(const Profile& profile, IR::Program& program,
-                                         u32& binding);
+                                         Bindings& binding);
 
 // Microinstruction emitters
 Id EmitPhi(EmitContext& ctx, IR::Inst* inst);

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -121,7 +121,7 @@ std::optional<OutAttr> OutputAttrPointer(EmitContext& ctx, IR::Attribute attr) {
 }
 
 Id GetCbuf(EmitContext& ctx, Id result_type, Id UniformDefinitions::*member_ptr, u32 element_size,
-           const IR::Value& binding, const IR::Value& offset) {
+           const IR::Value& binding, const IR::Value& offset, bool check_alignment = true) {
     if (!binding.IsImmediate()) {
         throw NotImplementedException("Constant buffer indexing");
     }
@@ -137,13 +137,31 @@ Id GetCbuf(EmitContext& ctx, Id result_type, Id UniformDefinitions::*member_ptr,
         const Id access_chain{ctx.OpAccessChain(uniform_type, cbuf, ctx.u32_zero_value, index)};
         return ctx.OpLoad(result_type, access_chain);
     }
-    if (offset.U32() % element_size != 0) {
+    if (check_alignment && offset.U32() % element_size != 0) {
         throw NotImplementedException("Unaligned immediate constant buffer load");
     }
     const Id imm_offset{ctx.Const(offset.U32() / element_size)};
     const Id access_chain{ctx.OpAccessChain(uniform_type, cbuf, ctx.u32_zero_value, imm_offset)};
     return ctx.OpLoad(result_type, access_chain);
 }
+
+Id GetCbufU32x4(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    return GetCbuf(ctx, ctx.U32[4], &UniformDefinitions::U32x4, sizeof(u32[4]), binding, offset,
+                   false);
+}
+
+Id GetCbufElement(EmitContext& ctx, Id vector, const IR::Value& offset, u32 index_offset) {
+    if (offset.IsImmediate()) {
+        const u32 element{(offset.U32() / 4) % 4 + index_offset};
+        return ctx.OpCompositeExtract(ctx.U32[1], vector, element);
+    }
+    const Id shift{ctx.OpShiftRightArithmetic(ctx.U32[1], ctx.Def(offset), ctx.Const(2u))};
+    Id element{ctx.OpBitwiseAnd(ctx.U32[1], shift, ctx.Const(3u))};
+    if (index_offset > 0) {
+        element = ctx.OpIAdd(ctx.U32[1], element, ctx.Const(index_offset));
+    }
+    return ctx.OpVectorExtractDynamic(ctx.U32[1], vector, element);
+}
 } // Anonymous namespace
 
 void EmitGetRegister(EmitContext&) {
@@ -179,40 +197,86 @@ void EmitGetIndirectBranchVariable(EmitContext&) {
 }
 
 Id EmitGetCbufU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    if (ctx.profile.support_descriptor_aliasing) {
         const Id load{GetCbuf(ctx, ctx.U8, &UniformDefinitions::U8, sizeof(u8), binding, offset)};
         return ctx.OpUConvert(ctx.U32[1], load);
+    } else {
+        const Id vector{GetCbufU32x4(ctx, binding, offset)};
+        const Id element{GetCbufElement(ctx, vector, offset, 0u)};
+        const Id bit_offset{ctx.BitOffset8(offset)};
+        return ctx.OpBitFieldUExtract(ctx.U32[1], element, bit_offset, ctx.Const(8u));
+    }
 }
 
 Id EmitGetCbufS8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    if (ctx.profile.support_descriptor_aliasing) {
         const Id load{GetCbuf(ctx, ctx.S8, &UniformDefinitions::S8, sizeof(s8), binding, offset)};
         return ctx.OpSConvert(ctx.U32[1], load);
+    } else {
+        const Id vector{GetCbufU32x4(ctx, binding, offset)};
+        const Id element{GetCbufElement(ctx, vector, offset, 0u)};
+        const Id bit_offset{ctx.BitOffset8(offset)};
+        return ctx.OpBitFieldSExtract(ctx.U32[1], element, bit_offset, ctx.Const(8u));
+    }
 }
 
 Id EmitGetCbufU16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
-    const Id load{GetCbuf(ctx, ctx.U16, &UniformDefinitions::U16, sizeof(u16), binding, offset)};
+    if (ctx.profile.support_descriptor_aliasing) {
+        const Id load{
+            GetCbuf(ctx, ctx.U16, &UniformDefinitions::U16, sizeof(u16), binding, offset)};
         return ctx.OpUConvert(ctx.U32[1], load);
+    } else {
+        const Id vector{GetCbufU32x4(ctx, binding, offset)};
+        const Id element{GetCbufElement(ctx, vector, offset, 0u)};
+        const Id bit_offset{ctx.BitOffset16(offset)};
+        return ctx.OpBitFieldUExtract(ctx.U32[1], element, bit_offset, ctx.Const(16u));
+    }
 }
 
 Id EmitGetCbufS16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
-    const Id load{GetCbuf(ctx, ctx.S16, &UniformDefinitions::S16, sizeof(s16), binding, offset)};
+    if (ctx.profile.support_descriptor_aliasing) {
+        const Id load{
+            GetCbuf(ctx, ctx.S16, &UniformDefinitions::S16, sizeof(s16), binding, offset)};
         return ctx.OpSConvert(ctx.U32[1], load);
+    } else {
+        const Id vector{GetCbufU32x4(ctx, binding, offset)};
+        const Id element{GetCbufElement(ctx, vector, offset, 0u)};
+        const Id bit_offset{ctx.BitOffset16(offset)};
+        return ctx.OpBitFieldSExtract(ctx.U32[1], element, bit_offset, ctx.Const(16u));
+    }
 }
 
 Id EmitGetCbufU32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    if (ctx.profile.support_descriptor_aliasing) {
         return GetCbuf(ctx, ctx.U32[1], &UniformDefinitions::U32, sizeof(u32), binding, offset);
+    } else {
+        const Id vector{GetCbufU32x4(ctx, binding, offset)};
+        return GetCbufElement(ctx, vector, offset, 0u);
+    }
 }
 
 Id EmitGetCbufF32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
+    if (ctx.profile.support_descriptor_aliasing) {
         return GetCbuf(ctx, ctx.F32[1], &UniformDefinitions::F32, sizeof(f32), binding, offset);
+    } else {
+        const Id vector{GetCbufU32x4(ctx, binding, offset)};
+        return ctx.OpBitcast(ctx.F32[1], GetCbufElement(ctx, vector, offset, 0u));
+    }
 }
 
 Id EmitGetCbufU32x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
-    return GetCbuf(ctx, ctx.U32[2], &UniformDefinitions::U32x2, sizeof(u32[2]), binding, offset);
+    if (ctx.profile.support_descriptor_aliasing) {
+        return GetCbuf(ctx, ctx.U32[2], &UniformDefinitions::U32x2, sizeof(u32[2]), binding,
+                       offset);
+    } else {
+        const Id vector{GetCbufU32x4(ctx, binding, offset)};
+        return ctx.OpCompositeConstruct(ctx.U32[2], GetCbufElement(ctx, vector, offset, 0u),
+                                        GetCbufElement(ctx, vector, offset, 1u));
+    }
 }
 
 Id EmitGetAttribute(EmitContext& ctx, IR::Attribute attr, Id vertex) {
     const u32 element{static_cast<u32>(attr) % 4};
-    const auto element_id{[&] { return ctx.Const(element); }};
     if (IR::IsGeneric(attr)) {
         const u32 index{IR::GenericAttributeIndex(attr)};
         const std::optional<AttrInfo> type{AttrTypes(ctx, index)};
@@ -221,7 +285,7 @@ Id EmitGetAttribute(EmitContext& ctx, IR::Attribute attr, Id vertex) {
             return ctx.Const(0.0f);
         }
         const Id generic_id{ctx.input_generics.at(index)};
-        const Id pointer{AttrPointer(ctx, type->pointer, vertex, generic_id, element_id())};
+        const Id pointer{AttrPointer(ctx, type->pointer, vertex, generic_id, ctx.Const(element))};
         const Id value{ctx.OpLoad(type->id, pointer)};
         return type->needs_cast ? ctx.OpBitcast(ctx.F32[1], value) : value;
     }
@@ -232,8 +296,8 @@ Id EmitGetAttribute(EmitContext& ctx, IR::Attribute attr, Id vertex) {
     case IR::Attribute::PositionY:
     case IR::Attribute::PositionZ:
     case IR::Attribute::PositionW:
-        return ctx.OpLoad(
+        return ctx.OpLoad(ctx.F32[1], AttrPointer(ctx, ctx.input_f32, vertex, ctx.input_position,
-            ctx.F32[1], AttrPointer(ctx, ctx.input_f32, vertex, ctx.input_position, element_id()));
+                                                  ctx.Const(element)));
     case IR::Attribute::InstanceId:
         if (ctx.profile.support_vertex_instance_id) {
             return ctx.OpBitcast(ctx.F32[1], ctx.OpLoad(ctx.U32[1], ctx.instance_id));

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -127,6 +127,7 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::Engines::Maxw
     base_profile = Shader::Profile{
         .supported_spirv = device.IsKhrSpirv1_4Supported() ? 0x00010400U : 0x00010000U,
         .unified_descriptor_binding = true,
+        .support_descriptor_aliasing = true,
         .support_vertex_instance_id = false,
         .support_float_controls = true,
         .support_separate_denorm_behavior = float_control.denormBehaviorIndependence ==
@@ -149,9 +150,11 @@ PipelineCache::PipelineCache(RasterizerVulkan& rasterizer_, Tegra::Engines::Maxw
             device.IsExtShaderViewportIndexLayerSupported(),
         .support_viewport_mask = device.IsNvViewportArray2Supported(),
         .support_typeless_image_loads = device.IsFormatlessImageLoadSupported(),
+        .support_demote_to_helper_invocation = true,
         .warp_size_potentially_larger_than_guest = device.IsWarpSizePotentiallyBiggerThanGuest(),
         .support_int64_atomics = device.IsExtShaderAtomicInt64Supported(),
         .has_broken_spirv_clamp = driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS_KHR,
+        .has_broken_unsigned_image_offsets = false,
         .generic_input_types{},
         .fixed_state_point_size{},
         .alpha_test_func{},
@@ -312,7 +315,7 @@ std::unique_ptr<GraphicsPipeline> PipelineCache::CreateGraphicsPipeline(
     std::array<const Shader::Info*, Maxwell::MaxShaderStage> infos{};
     std::array<vk::ShaderModule, Maxwell::MaxShaderStage> modules;
 
-    u32 binding{0};
+    Shader::Backend::SPIRV::Bindings binding;
     for (size_t index = uses_vertex_a && uses_vertex_b ? 1 : 0; index < Maxwell::MaxShaderProgram; ++index) {
         if (key.unique_hashes[index] == 0) {
             continue;
@@ -398,7 +401,7 @@ std::unique_ptr<ComputePipeline> PipelineCache::CreateComputePipeline(
 
     Shader::Maxwell::Flow::CFG cfg{env, pools.flow_block, env.StartAddress()};
     Shader::IR::Program program{TranslateProgram(pools.inst, pools.block, env, cfg)};
-    u32 binding{0};
+    Shader::Backend::SPIRV::Bindings binding;
     const std::vector<u32> code{EmitSPIRV(base_profile, program, binding)};
     device.SaveShader(code);
     vk::ShaderModule spv_module{BuildShader(device, code)};