Merge pull request #8024 from liamwhite/const-indexing

Add shader support for const buffer indirect addressing

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -123,34 +123,36 @@ 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, const Id indirect_func) {
-    if (!binding.IsImmediate()) {
+    Id buffer_offset;
-        throw NotImplementedException("Constant buffer indexing");
-    }
-    const Id cbuf{ctx.cbufs[binding.U32()].*member_ptr};
     const Id uniform_type{ctx.uniform_types.*member_ptr};
-    if (!offset.IsImmediate()) {
+    if (offset.IsImmediate()) {
-        Id index{ctx.Def(offset)};
-        if (element_size > 1) {
-            const u32 log2_element_size{static_cast<u32>(std::countr_zero(element_size))};
-            const Id shift{ctx.Const(log2_element_size)};
-            index = ctx.OpShiftRightArithmetic(ctx.U32[1], ctx.Def(offset), shift);
-        }
-        const Id access_chain{ctx.OpAccessChain(uniform_type, cbuf, ctx.u32_zero_value, index)};
-        return ctx.OpLoad(result_type, access_chain);
-    }
         // Hardware been proved to read the aligned offset (e.g. LDC.U32 at 6 will read offset 4)
         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)};
+        buffer_offset = imm_offset;
+    } else if (element_size > 1) {
+        const u32 log2_element_size{static_cast<u32>(std::countr_zero(element_size))};
+        const Id shift{ctx.Const(log2_element_size)};
+        buffer_offset = ctx.OpShiftRightArithmetic(ctx.U32[1], ctx.Def(offset), shift);
+    } else {
+        buffer_offset = ctx.Def(offset);
+    }
+    if (!binding.IsImmediate()) {
+        return ctx.OpFunctionCall(result_type, indirect_func, ctx.Def(binding), buffer_offset);
+    }
+    const Id cbuf{ctx.cbufs[binding.U32()].*member_ptr};
+    const Id access_chain{ctx.OpAccessChain(uniform_type, cbuf, ctx.u32_zero_value, buffer_offset)};
     return ctx.OpLoad(result_type, access_chain);
 }
 
 Id GetCbufU32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
-    return GetCbuf(ctx, ctx.U32[1], &UniformDefinitions::U32, sizeof(u32), binding, offset);
+    return GetCbuf(ctx, ctx.U32[1], &UniformDefinitions::U32, sizeof(u32), binding, offset,
+                   ctx.load_const_func_u32);
 }
 
 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);
+    return GetCbuf(ctx, ctx.U32[4], &UniformDefinitions::U32x4, sizeof(u32[4]), binding, offset,
+                   ctx.load_const_func_u32x4);
 }
 
 Id GetCbufElement(EmitContext& ctx, Id vector, const IR::Value& offset, u32 index_offset) {
@@ -201,7 +203,8 @@ void EmitGetIndirectBranchVariable(EmitContext&) {
 
 Id EmitGetCbufU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
     if (ctx.profile.support_descriptor_aliasing && ctx.profile.support_int8) {
-        const Id load{GetCbuf(ctx, ctx.U8, &UniformDefinitions::U8, sizeof(u8), binding, offset)};
+        const Id load{GetCbuf(ctx, ctx.U8, &UniformDefinitions::U8, sizeof(u8), binding, offset,
+                              ctx.load_const_func_u8)};
         return ctx.OpUConvert(ctx.U32[1], load);
     }
     Id element{};
@@ -217,7 +220,8 @@ Id EmitGetCbufU8(EmitContext& ctx, const IR::Value& binding, const IR::Value& of
 
 Id EmitGetCbufS8(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
     if (ctx.profile.support_descriptor_aliasing && ctx.profile.support_int8) {
-        const Id load{GetCbuf(ctx, ctx.S8, &UniformDefinitions::S8, sizeof(s8), binding, offset)};
+        const Id load{GetCbuf(ctx, ctx.S8, &UniformDefinitions::S8, sizeof(s8), binding, offset,
+                              ctx.load_const_func_u8)};
         return ctx.OpSConvert(ctx.U32[1], load);
     }
     Id element{};
@@ -233,8 +237,8 @@ Id EmitGetCbufS8(EmitContext& ctx, const IR::Value& binding, const IR::Value& of
 
 Id EmitGetCbufU16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
     if (ctx.profile.support_descriptor_aliasing && ctx.profile.support_int16) {
-        const Id load{
+        const Id load{GetCbuf(ctx, ctx.U16, &UniformDefinitions::U16, sizeof(u16), binding, offset,
-            GetCbuf(ctx, ctx.U16, &UniformDefinitions::U16, sizeof(u16), binding, offset)};
+                              ctx.load_const_func_u16)};
         return ctx.OpUConvert(ctx.U32[1], load);
     }
     Id element{};
@@ -250,8 +254,8 @@ Id EmitGetCbufU16(EmitContext& ctx, const IR::Value& binding, const IR::Value& o
 
 Id EmitGetCbufS16(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
     if (ctx.profile.support_descriptor_aliasing && ctx.profile.support_int16) {
-        const Id load{
+        const Id load{GetCbuf(ctx, ctx.S16, &UniformDefinitions::S16, sizeof(s16), binding, offset,
-            GetCbuf(ctx, ctx.S16, &UniformDefinitions::S16, sizeof(s16), binding, offset)};
+                              ctx.load_const_func_u16)};
         return ctx.OpSConvert(ctx.U32[1], load);
     }
     Id element{};
@@ -276,7 +280,8 @@ Id EmitGetCbufU32(EmitContext& ctx, const IR::Value& binding, const IR::Value& o
 
 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);
+        return GetCbuf(ctx, ctx.F32[1], &UniformDefinitions::F32, sizeof(f32), binding, offset,
+                       ctx.load_const_func_f32);
     } else {
         const Id vector{GetCbufU32x4(ctx, binding, offset)};
         return ctx.OpBitcast(ctx.F32[1], GetCbufElement(ctx, vector, offset, 0u));
@@ -285,8 +290,8 @@ Id EmitGetCbufF32(EmitContext& ctx, const IR::Value& binding, const IR::Value& o
 
 Id EmitGetCbufU32x2(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
     if (ctx.profile.support_descriptor_aliasing) {
-        return GetCbuf(ctx, ctx.U32[2], &UniformDefinitions::U32x2, sizeof(u32[2]), binding,
+        return GetCbuf(ctx, ctx.U32[2], &UniformDefinitions::U32x2, sizeof(u32[2]), binding, offset,
-                       offset);
+                       ctx.load_const_func_u32x2);
     } else {
         const Id vector{GetCbufU32x4(ctx, binding, offset)};
         return ctx.OpCompositeConstruct(ctx.U32[2], GetCbufElement(ctx, vector, offset, 0u),

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -464,6 +464,7 @@ EmitContext::EmitContext(const Profile& profile_, const RuntimeInfo& runtime_inf
     DefineSharedMemory(program);
     DefineSharedMemoryFunctions(program);
     DefineConstantBuffers(program.info, uniform_binding);
+    DefineConstantBufferIndirectFunctions(program.info);
     DefineStorageBuffers(program.info, storage_binding);
     DefineTextureBuffers(program.info, texture_binding);
     DefineImageBuffers(program.info, image_binding);
@@ -993,7 +994,7 @@ void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) {
         }
         return;
     }
-    IR::Type types{info.used_constant_buffer_types};
+    IR::Type types{info.used_constant_buffer_types | info.used_indirect_cbuf_types};
     if (True(types & IR::Type::U8)) {
         if (profile.support_int8) {
             DefineConstBuffers(*this, info, &UniformDefinitions::U8, binding, U8, 'u', sizeof(u8));
@@ -1027,6 +1028,62 @@ void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) {
     binding += static_cast<u32>(info.constant_buffer_descriptors.size());
 }
 
+void EmitContext::DefineConstantBufferIndirectFunctions(const Info& info) {
+    if (!info.uses_cbuf_indirect) {
+        return;
+    }
+    const auto make_accessor{[&](Id buffer_type, Id UniformDefinitions::*member_ptr) {
+        const Id func_type{TypeFunction(buffer_type, U32[1], U32[1])};
+        const Id func{OpFunction(buffer_type, spv::FunctionControlMask::MaskNone, func_type)};
+        const Id binding{OpFunctionParameter(U32[1])};
+        const Id offset{OpFunctionParameter(U32[1])};
+
+        AddLabel();
+
+        const Id merge_label{OpLabel()};
+        const Id uniform_type{uniform_types.*member_ptr};
+
+        std::array<Id, Info::MAX_CBUFS> buf_labels;
+        std::array<Sirit::Literal, Info::MAX_CBUFS> buf_literals;
+        for (u32 i = 0; i < Info::MAX_CBUFS; i++) {
+            buf_labels[i] = OpLabel();
+            buf_literals[i] = Sirit::Literal{i};
+        }
+        OpSelectionMerge(merge_label, spv::SelectionControlMask::MaskNone);
+        OpSwitch(binding, buf_labels[0], buf_literals, buf_labels);
+        for (u32 i = 0; i < Info::MAX_CBUFS; i++) {
+            AddLabel(buf_labels[i]);
+            const Id cbuf{cbufs[i].*member_ptr};
+            const Id access_chain{OpAccessChain(uniform_type, cbuf, u32_zero_value, offset)};
+            const Id result{OpLoad(buffer_type, access_chain)};
+            OpReturnValue(result);
+        }
+        AddLabel(merge_label);
+        OpUnreachable();
+        OpFunctionEnd();
+        return func;
+    }};
+    IR::Type types{info.used_indirect_cbuf_types};
+    if (True(types & IR::Type::U8)) {
+        load_const_func_u8 = make_accessor(U8, &UniformDefinitions::U8);
+    }
+    if (True(types & IR::Type::U16)) {
+        load_const_func_u16 = make_accessor(U16, &UniformDefinitions::U16);
+    }
+    if (True(types & IR::Type::F32)) {
+        load_const_func_f32 = make_accessor(F32[1], &UniformDefinitions::F32);
+    }
+    if (True(types & IR::Type::U32)) {
+        load_const_func_u32 = make_accessor(U32[1], &UniformDefinitions::U32);
+    }
+    if (True(types & IR::Type::U32x2)) {
+        load_const_func_u32x2 = make_accessor(U32[2], &UniformDefinitions::U32x2);
+    }
+    if (True(types & IR::Type::U32x4)) {
+        load_const_func_u32x4 = make_accessor(U32[4], &UniformDefinitions::U32x4);
+    }
+}
+
 void EmitContext::DefineStorageBuffers(const Info& info, u32& binding) {
     if (info.storage_buffers_descriptors.empty()) {
         return;

src/shader_recompiler/backend/spirv/spirv_emit_context.h

@@ -294,6 +294,13 @@ public:
 
     std::vector<Id> interfaces;
 
+    Id load_const_func_u8{};
+    Id load_const_func_u16{};
+    Id load_const_func_u32{};
+    Id load_const_func_f32{};
+    Id load_const_func_u32x2{};
+    Id load_const_func_u32x4{};
+
 private:
     void DefineCommonTypes(const Info& info);
     void DefineCommonConstants();
@@ -302,6 +309,7 @@ private:
     void DefineSharedMemory(const IR::Program& program);
     void DefineSharedMemoryFunctions(const IR::Program& program);
     void DefineConstantBuffers(const Info& info, u32& binding);
+    void DefineConstantBufferIndirectFunctions(const Info& info);
     void DefineStorageBuffers(const Info& info, u32& binding);
     void DefineTextureBuffers(const Info& info, u32& binding);
     void DefineImageBuffers(const Info& info, u32& binding);

src/shader_recompiler/frontend/maxwell/translate/impl/load_constant.cpp

@@ -11,10 +11,20 @@ namespace Shader::Maxwell {
 using namespace LDC;
 namespace {
 std::pair<IR::U32, IR::U32> Slot(IR::IREmitter& ir, Mode mode, const IR::U32& imm_index,
-                                 const IR::U32& reg, const IR::U32& imm) {
+                                 const IR::U32& reg, const IR::U32& imm_offset) {
     switch (mode) {
     case Mode::Default:
-        return {imm_index, ir.IAdd(reg, imm)};
+        return {imm_index, ir.IAdd(reg, imm_offset)};
+    case Mode::IS: {
+        // Segmented addressing mode
+        // Ra+imm_offset points into a flat mapping of const buffer
+        // address space
+        const IR::U32 address{ir.IAdd(reg, imm_offset)};
+        const IR::U32 index{ir.BitFieldExtract(address, ir.Imm32(16), ir.Imm32(16))};
+        const IR::U32 offset{ir.BitFieldExtract(address, ir.Imm32(0), ir.Imm32(16))};
+
+        return {ir.IAdd(index, imm_index), offset};
+    }
     default:
         break;
     }

src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp

@@ -29,6 +29,46 @@ void AddConstantBufferDescriptor(Info& info, u32 index, u32 count) {
                  });
 }
 
+void AddRegisterIndexedLdc(Info& info) {
+    info.uses_cbuf_indirect = true;
+
+    // The shader can use any possible constant buffer
+    info.constant_buffer_mask = (1 << Info::MAX_CBUFS) - 1;
+
+    auto& cbufs{info.constant_buffer_descriptors};
+    cbufs.clear();
+    for (u32 i = 0; i < Info::MAX_CBUFS; i++) {
+        cbufs.push_back(ConstantBufferDescriptor{.index = i, .count = 1});
+
+        // The shader can use any possible access size
+        info.constant_buffer_used_sizes[i] = 0x10'000;
+    }
+}
+
+u32 GetElementSize(IR::Type& used_type, Shader::IR::Opcode opcode) {
+    switch (opcode) {
+    case IR::Opcode::GetCbufU8:
+    case IR::Opcode::GetCbufS8:
+        used_type |= IR::Type::U8;
+        return 1;
+    case IR::Opcode::GetCbufU16:
+    case IR::Opcode::GetCbufS16:
+        used_type |= IR::Type::U16;
+        return 2;
+    case IR::Opcode::GetCbufU32:
+        used_type |= IR::Type::U32;
+        return 4;
+    case IR::Opcode::GetCbufF32:
+        used_type |= IR::Type::F32;
+        return 4;
+    case IR::Opcode::GetCbufU32x2:
+        used_type |= IR::Type::U32x2;
+        return 8;
+    default:
+        throw InvalidArgument("Invalid opcode {}", opcode);
+    }
+}
+
 void GetPatch(Info& info, IR::Patch patch) {
     if (!IR::IsGeneric(patch)) {
         throw NotImplementedException("Reading non-generic patch {}", patch);
@@ -463,43 +503,19 @@ void VisitUsages(Info& info, IR::Inst& inst) {
     case IR::Opcode::GetCbufU32x2: {
         const IR::Value index{inst.Arg(0)};
         const IR::Value offset{inst.Arg(1)};
-        if (!index.IsImmediate()) {
+        if (index.IsImmediate()) {
-            throw NotImplementedException("Constant buffer with non-immediate index");
-        }
             AddConstantBufferDescriptor(info, index.U32(), 1);
-        u32 element_size{};
+            u32 element_size = GetElementSize(info.used_constant_buffer_types, inst.GetOpcode());
-        switch (inst.GetOpcode()) {
-        case IR::Opcode::GetCbufU8:
-        case IR::Opcode::GetCbufS8:
-            info.used_constant_buffer_types |= IR::Type::U8;
-            element_size = 1;
-            break;
-        case IR::Opcode::GetCbufU16:
-        case IR::Opcode::GetCbufS16:
-            info.used_constant_buffer_types |= IR::Type::U16;
-            element_size = 2;
-            break;
-        case IR::Opcode::GetCbufU32:
-            info.used_constant_buffer_types |= IR::Type::U32;
-            element_size = 4;
-            break;
-        case IR::Opcode::GetCbufF32:
-            info.used_constant_buffer_types |= IR::Type::F32;
-            element_size = 4;
-            break;
-        case IR::Opcode::GetCbufU32x2:
-            info.used_constant_buffer_types |= IR::Type::U32x2;
-            element_size = 8;
-            break;
-        default:
-            break;
-        }
             u32& size{info.constant_buffer_used_sizes[index.U32()]};
             if (offset.IsImmediate()) {
                 size = Common::AlignUp(std::max(size, offset.U32() + element_size), 16u);
             } else {
                 size = 0x10'000;
             }
+        } else {
+            AddRegisterIndexedLdc(info);
+            GetElementSize(info.used_indirect_cbuf_types, inst.GetOpcode());
+        }
         break;
     }
     case IR::Opcode::BindlessImageSampleImplicitLod:

src/shader_recompiler/shader_info.h

@@ -173,9 +173,11 @@ struct Info {
     bool uses_atomic_image_u32{};
     bool uses_shadow_lod{};
     bool uses_rescaling_uniform{};
+    bool uses_cbuf_indirect{};
 
     IR::Type used_constant_buffer_types{};
     IR::Type used_storage_buffer_types{};
+    IR::Type used_indirect_cbuf_types{};
 
     u32 constant_buffer_mask{};
     std::array<u32, MAX_CBUFS> constant_buffer_used_sizes{};