spirv: Add lower fp16 to fp32 pass

src/shader_recompiler/CMakeLists.txt

@@ -7,6 +7,7 @@ add_library(shader_recompiler STATIC
     backend/spirv/emit_spirv_composite.cpp
     backend/spirv/emit_spirv_context_get_set.cpp
     backend/spirv/emit_spirv_control_flow.cpp
+    backend/spirv/emit_spirv_convert.cpp
     backend/spirv/emit_spirv_floating_point.cpp
     backend/spirv/emit_spirv_integer.cpp
     backend/spirv/emit_spirv_logical.cpp
@@ -82,6 +83,7 @@ add_library(shader_recompiler STATIC
     ir_opt/dead_code_elimination_pass.cpp
     ir_opt/global_memory_to_storage_buffer_pass.cpp
     ir_opt/identity_removal_pass.cpp
+    ir_opt/lower_fp16_to_fp32.cpp
     ir_opt/passes.h
     ir_opt/ssa_rewrite_pass.cpp
     ir_opt/verification_pass.cpp

src/shader_recompiler/backend/spirv/emit_context.cpp

@@ -30,8 +30,11 @@ EmitContext::EmitContext(IR::Program& program) : Sirit::Module(0x00010000) {
     DefineCommonTypes(program.info);
     DefineCommonConstants();
     DefineSpecialVariables(program.info);
-    DefineConstantBuffers(program.info);
+
-    DefineStorageBuffers(program.info);
+    u32 binding{};
+    DefineConstantBuffers(program.info, binding);
+    DefineStorageBuffers(program.info, binding);
+
     DefineLabels(program);
 }
 
@@ -58,6 +61,12 @@ void EmitContext::DefineCommonTypes(const Info& info) {
 
     U1 = Name(TypeBool(), "u1");
 
+    // TODO: Conditionally define these
+    AddCapability(spv::Capability::Int16);
+    AddCapability(spv::Capability::Int64);
+    U16 = Name(TypeInt(16, false), "u16");
+    U64 = Name(TypeInt(64, false), "u64");
+
     F32.Define(*this, TypeFloat(32), "f32");
     U32.Define(*this, TypeInt(32, false), "u32");
 
@@ -95,12 +104,12 @@ void EmitContext::DefineSpecialVariables(const Info& info) {
     }
 }
 
-void EmitContext::DefineConstantBuffers(const Info& info) {
+void EmitContext::DefineConstantBuffers(const Info& info, u32& binding) {
     if (info.constant_buffer_descriptors.empty()) {
         return;
     }
     const Id array_type{TypeArray(U32[1], Constant(U32[1], 4096))};
-    Decorate(array_type, spv::Decoration::ArrayStride, 16U);
+    Decorate(array_type, spv::Decoration::ArrayStride, 4U);
 
     const Id struct_type{TypeStruct(array_type)};
     Name(struct_type, "cbuf_block");
@@ -111,18 +120,19 @@ void EmitContext::DefineConstantBuffers(const Info& info) {
     const Id uniform_type{TypePointer(spv::StorageClass::Uniform, struct_type)};
     uniform_u32 = TypePointer(spv::StorageClass::Uniform, U32[1]);
 
-    u32 binding{};
+    u32 index{};
     for (const Info::ConstantBufferDescriptor& desc : info.constant_buffer_descriptors) {
         const Id id{AddGlobalVariable(uniform_type, spv::StorageClass::Uniform)};
         Decorate(id, spv::Decoration::Binding, binding);
         Decorate(id, spv::Decoration::DescriptorSet, 0U);
         Name(id, fmt::format("c{}", desc.index));
         std::fill_n(cbufs.data() + desc.index, desc.count, id);
+        index += desc.count;
         binding += desc.count;
     }
 }
 
-void EmitContext::DefineStorageBuffers(const Info& info) {
+void EmitContext::DefineStorageBuffers(const Info& info, u32& binding) {
     if (info.storage_buffers_descriptors.empty()) {
         return;
     }
@@ -140,13 +150,14 @@ void EmitContext::DefineStorageBuffers(const Info& info) {
     const Id storage_type{TypePointer(spv::StorageClass::StorageBuffer, struct_type)};
     storage_u32 = TypePointer(spv::StorageClass::StorageBuffer, U32[1]);
 
-    u32 binding{};
+    u32 index{};
     for (const Info::StorageBufferDescriptor& desc : info.storage_buffers_descriptors) {
         const Id id{AddGlobalVariable(storage_type, spv::StorageClass::StorageBuffer)};
         Decorate(id, spv::Decoration::Binding, binding);
         Decorate(id, spv::Decoration::DescriptorSet, 0U);
-        Name(id, fmt::format("ssbo{}", binding));
+        Name(id, fmt::format("ssbo{}", index));
-        std::fill_n(ssbos.data() + binding, desc.count, id);
+        std::fill_n(ssbos.data() + index, desc.count, id);
+        index += desc.count;
         binding += desc.count;
     }
 }

src/shader_recompiler/backend/spirv/emit_context.h

@@ -37,6 +37,8 @@ public:
 
     Id void_id{};
     Id U1{};
+    Id U16{};
+    Id U64{};
     VectorTypes F32;
     VectorTypes U32;
     VectorTypes F16;
@@ -59,8 +61,8 @@ private:
     void DefineCommonTypes(const Info& info);
     void DefineCommonConstants();
     void DefineSpecialVariables(const Info& info);
-    void DefineConstantBuffers(const Info& info);
+    void DefineConstantBuffers(const Info& info, u32& binding);
-    void DefineStorageBuffers(const Info& info);
+    void DefineStorageBuffers(const Info& info, u32& binding);
     void DefineLabels(IR::Program& program);
 };
 

src/shader_recompiler/backend/spirv/emit_spirv.cpp

@@ -14,6 +14,8 @@
 #include "shader_recompiler/frontend/ir/microinstruction.h"
 #include "shader_recompiler/frontend/ir/program.h"
 
+#pragma optimize("", off)
+
 namespace Shader::Backend::SPIRV {
 namespace {
 template <class Func>

src/shader_recompiler/backend/spirv/emit_spirv.h

@@ -79,26 +79,27 @@ void EmitWriteStorageU16(EmitContext& ctx);
 void EmitWriteStorageS16(EmitContext& ctx);
 void EmitWriteStorage32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
                         Id value);
-void EmitWriteStorage64(EmitContext& ctx);
+void EmitWriteStorage64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
+                        Id value);
 void EmitWriteStorage128(EmitContext& ctx);
-void EmitCompositeConstructU32x2(EmitContext& ctx);
+Id EmitCompositeConstructU32x2(EmitContext& ctx, Id e1, Id e2);
-void EmitCompositeConstructU32x3(EmitContext& ctx);
+Id EmitCompositeConstructU32x3(EmitContext& ctx, Id e1, Id e2, Id e3);
-void EmitCompositeConstructU32x4(EmitContext& ctx);
+Id EmitCompositeConstructU32x4(EmitContext& ctx, Id e1, Id e2, Id e3, Id e4);
-void EmitCompositeExtractU32x2(EmitContext& ctx);
+Id EmitCompositeExtractU32x2(EmitContext& ctx, Id composite, u32 index);
-Id EmitCompositeExtractU32x3(EmitContext& ctx, Id vector, u32 index);
+Id EmitCompositeExtractU32x3(EmitContext& ctx, Id composite, u32 index);
-void EmitCompositeExtractU32x4(EmitContext& ctx);
+Id EmitCompositeExtractU32x4(EmitContext& ctx, Id composite, u32 index);
 void EmitCompositeConstructF16x2(EmitContext& ctx);
 void EmitCompositeConstructF16x3(EmitContext& ctx);
 void EmitCompositeConstructF16x4(EmitContext& ctx);
-void EmitCompositeExtractF16x2(EmitContext& ctx);
+Id EmitCompositeExtractF16x2(EmitContext& ctx, Id composite, u32 index);
-void EmitCompositeExtractF16x3(EmitContext& ctx);
+Id EmitCompositeExtractF16x3(EmitContext& ctx, Id composite, u32 index);
-void EmitCompositeExtractF16x4(EmitContext& ctx);
+Id EmitCompositeExtractF16x4(EmitContext& ctx, Id composite, u32 index);
 void EmitCompositeConstructF32x2(EmitContext& ctx);
 void EmitCompositeConstructF32x3(EmitContext& ctx);
 void EmitCompositeConstructF32x4(EmitContext& ctx);
-void EmitCompositeExtractF32x2(EmitContext& ctx);
+Id EmitCompositeExtractF32x2(EmitContext& ctx, Id composite, u32 index);
-void EmitCompositeExtractF32x3(EmitContext& ctx);
+Id EmitCompositeExtractF32x3(EmitContext& ctx, Id composite, u32 index);
-void EmitCompositeExtractF32x4(EmitContext& ctx);
+Id EmitCompositeExtractF32x4(EmitContext& ctx, Id composite, u32 index);
 void EmitCompositeConstructF64x2(EmitContext& ctx);
 void EmitCompositeConstructF64x3(EmitContext& ctx);
 void EmitCompositeConstructF64x4(EmitContext& ctx);
@@ -116,11 +117,13 @@ void EmitBitCastF16U16(EmitContext& ctx);
 Id EmitBitCastF32U32(EmitContext& ctx, Id value);
 void EmitBitCastF64U64(EmitContext& ctx);
 void EmitPackUint2x32(EmitContext& ctx);
-void EmitUnpackUint2x32(EmitContext& ctx);
+Id EmitUnpackUint2x32(EmitContext& ctx, Id value);
-void EmitPackFloat2x16(EmitContext& ctx);
+Id EmitPackFloat2x16(EmitContext& ctx, Id value);
-void EmitUnpackFloat2x16(EmitContext& ctx);
+Id EmitUnpackFloat2x16(EmitContext& ctx, Id value);
-void EmitPackDouble2x32(EmitContext& ctx);
+Id EmitPackHalf2x16(EmitContext& ctx, Id value);
-void EmitUnpackDouble2x32(EmitContext& ctx);
+Id EmitUnpackHalf2x16(EmitContext& ctx, Id value);
+Id EmitPackDouble2x32(EmitContext& ctx, Id value);
+Id EmitUnpackDouble2x32(EmitContext& ctx, Id value);
 void EmitGetZeroFromOp(EmitContext& ctx);
 void EmitGetSignFromOp(EmitContext& ctx);
 void EmitGetCarryFromOp(EmitContext& ctx);
@@ -159,18 +162,18 @@ void EmitFPLog2(EmitContext& ctx);
 void EmitFPSaturate16(EmitContext& ctx);
 void EmitFPSaturate32(EmitContext& ctx);
 void EmitFPSaturate64(EmitContext& ctx);
-void EmitFPRoundEven16(EmitContext& ctx);
+Id EmitFPRoundEven16(EmitContext& ctx, Id value);
-void EmitFPRoundEven32(EmitContext& ctx);
+Id EmitFPRoundEven32(EmitContext& ctx, Id value);
-void EmitFPRoundEven64(EmitContext& ctx);
+Id EmitFPRoundEven64(EmitContext& ctx, Id value);
-void EmitFPFloor16(EmitContext& ctx);
+Id EmitFPFloor16(EmitContext& ctx, Id value);
-void EmitFPFloor32(EmitContext& ctx);
+Id EmitFPFloor32(EmitContext& ctx, Id value);
-void EmitFPFloor64(EmitContext& ctx);
+Id EmitFPFloor64(EmitContext& ctx, Id value);
-void EmitFPCeil16(EmitContext& ctx);
+Id EmitFPCeil16(EmitContext& ctx, Id value);
-void EmitFPCeil32(EmitContext& ctx);
+Id EmitFPCeil32(EmitContext& ctx, Id value);
-void EmitFPCeil64(EmitContext& ctx);
+Id EmitFPCeil64(EmitContext& ctx, Id value);
-void EmitFPTrunc16(EmitContext& ctx);
+Id EmitFPTrunc16(EmitContext& ctx, Id value);
-void EmitFPTrunc32(EmitContext& ctx);
+Id EmitFPTrunc32(EmitContext& ctx, Id value);
-void EmitFPTrunc64(EmitContext& ctx);
+Id EmitFPTrunc64(EmitContext& ctx, Id value);
 Id EmitIAdd32(EmitContext& ctx, IR::Inst* inst, Id a, Id b);
 void EmitIAdd64(EmitContext& ctx);
 Id EmitISub32(EmitContext& ctx, Id a, Id b);
@@ -201,25 +204,25 @@ void EmitLogicalOr(EmitContext& ctx);
 void EmitLogicalAnd(EmitContext& ctx);
 void EmitLogicalXor(EmitContext& ctx);
 void EmitLogicalNot(EmitContext& ctx);
-void EmitConvertS16F16(EmitContext& ctx);
+Id EmitConvertS16F16(EmitContext& ctx, Id value);
-void EmitConvertS16F32(EmitContext& ctx);
+Id EmitConvertS16F32(EmitContext& ctx, Id value);
-void EmitConvertS16F64(EmitContext& ctx);
+Id EmitConvertS16F64(EmitContext& ctx, Id value);
-void EmitConvertS32F16(EmitContext& ctx);
+Id EmitConvertS32F16(EmitContext& ctx, Id value);
-void EmitConvertS32F32(EmitContext& ctx);
+Id EmitConvertS32F32(EmitContext& ctx, Id value);
-void EmitConvertS32F64(EmitContext& ctx);
+Id EmitConvertS32F64(EmitContext& ctx, Id value);
-void EmitConvertS64F16(EmitContext& ctx);
+Id EmitConvertS64F16(EmitContext& ctx, Id value);
-void EmitConvertS64F32(EmitContext& ctx);
+Id EmitConvertS64F32(EmitContext& ctx, Id value);
-void EmitConvertS64F64(EmitContext& ctx);
+Id EmitConvertS64F64(EmitContext& ctx, Id value);
-void EmitConvertU16F16(EmitContext& ctx);
+Id EmitConvertU16F16(EmitContext& ctx, Id value);
-void EmitConvertU16F32(EmitContext& ctx);
+Id EmitConvertU16F32(EmitContext& ctx, Id value);
-void EmitConvertU16F64(EmitContext& ctx);
+Id EmitConvertU16F64(EmitContext& ctx, Id value);
-void EmitConvertU32F16(EmitContext& ctx);
+Id EmitConvertU32F16(EmitContext& ctx, Id value);
-void EmitConvertU32F32(EmitContext& ctx);
+Id EmitConvertU32F32(EmitContext& ctx, Id value);
-void EmitConvertU32F64(EmitContext& ctx);
+Id EmitConvertU32F64(EmitContext& ctx, Id value);
-void EmitConvertU64F16(EmitContext& ctx);
+Id EmitConvertU64F16(EmitContext& ctx, Id value);
-void EmitConvertU64F32(EmitContext& ctx);
+Id EmitConvertU64F32(EmitContext& ctx, Id value);
-void EmitConvertU64F64(EmitContext& ctx);
+Id EmitConvertU64F64(EmitContext& ctx, Id value);
-void EmitConvertU64U32(EmitContext& ctx);
+Id EmitConvertU64U32(EmitContext& ctx, Id value);
-void EmitConvertU32U64(EmitContext& ctx);
+Id EmitConvertU32U64(EmitContext& ctx, Id value);
 
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_bitwise_conversion.cpp

@@ -34,24 +34,32 @@ void EmitPackUint2x32(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitUnpackUint2x32(EmitContext&) {
+Id EmitUnpackUint2x32(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpBitcast(ctx.U32[2], value);
 }
 
-void EmitPackFloat2x16(EmitContext&) {
+Id EmitPackFloat2x16(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpBitcast(ctx.U32[1], value);
 }
 
-void EmitUnpackFloat2x16(EmitContext&) {
+Id EmitUnpackFloat2x16(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpBitcast(ctx.F16[2], value);
 }
 
-void EmitPackDouble2x32(EmitContext&) {
+Id EmitPackHalf2x16(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpPackHalf2x16(ctx.U32[1], value);
 }
 
-void EmitUnpackDouble2x32(EmitContext&) {
+Id EmitUnpackHalf2x16(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpUnpackHalf2x16(ctx.F32[2], value);
+}
+
+Id EmitPackDouble2x32(EmitContext& ctx, Id value) {
+    return ctx.OpBitcast(ctx.F64[1], value);
+}
+
+Id EmitUnpackDouble2x32(EmitContext& ctx, Id value) {
+    return ctx.OpBitcast(ctx.U32[2], value);
 }
 
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_composite.cpp

@@ -6,28 +6,28 @@
 
 namespace Shader::Backend::SPIRV {
 
-void EmitCompositeConstructU32x2(EmitContext&) {
+Id EmitCompositeConstructU32x2(EmitContext& ctx, Id e1, Id e2) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeConstruct(ctx.U32[2], e1, e2);
 }
 
-void EmitCompositeConstructU32x3(EmitContext&) {
+Id EmitCompositeConstructU32x3(EmitContext& ctx, Id e1, Id e2, Id e3) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeConstruct(ctx.U32[3], e1, e2, e3);
 }
 
-void EmitCompositeConstructU32x4(EmitContext&) {
+Id EmitCompositeConstructU32x4(EmitContext& ctx, Id e1, Id e2, Id e3, Id e4) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeConstruct(ctx.U32[4], e1, e2, e3, e4);
 }
 
-void EmitCompositeExtractU32x2(EmitContext&) {
+Id EmitCompositeExtractU32x2(EmitContext& ctx, Id composite, u32 index) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeExtract(ctx.U32[1], composite, index);
 }
 
-Id EmitCompositeExtractU32x3(EmitContext& ctx, Id vector, u32 index) {
+Id EmitCompositeExtractU32x3(EmitContext& ctx, Id composite, u32 index) {
-    return ctx.OpCompositeExtract(ctx.U32[1], vector, index);
+    return ctx.OpCompositeExtract(ctx.U32[1], composite, index);
 }
 
-void EmitCompositeExtractU32x4(EmitContext&) {
+Id EmitCompositeExtractU32x4(EmitContext& ctx, Id composite, u32 index) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeExtract(ctx.U32[1], composite, index);
 }
 
 void EmitCompositeConstructF16x2(EmitContext&) {
@@ -42,16 +42,16 @@ void EmitCompositeConstructF16x4(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitCompositeExtractF16x2(EmitContext&) {
+Id EmitCompositeExtractF16x2(EmitContext& ctx, Id composite, u32 index) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeExtract(ctx.F16[1], composite, index);
 }
 
-void EmitCompositeExtractF16x3(EmitContext&) {
+Id EmitCompositeExtractF16x3(EmitContext& ctx, Id composite, u32 index) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeExtract(ctx.F16[1], composite, index);
 }
 
-void EmitCompositeExtractF16x4(EmitContext&) {
+Id EmitCompositeExtractF16x4(EmitContext& ctx, Id composite, u32 index) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeExtract(ctx.F16[1], composite, index);
 }
 
 void EmitCompositeConstructF32x2(EmitContext&) {
@@ -66,16 +66,16 @@ void EmitCompositeConstructF32x4(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitCompositeExtractF32x2(EmitContext&) {
+Id EmitCompositeExtractF32x2(EmitContext& ctx, Id composite, u32 index) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeExtract(ctx.F32[1], composite, index);
 }
 
-void EmitCompositeExtractF32x3(EmitContext&) {
+Id EmitCompositeExtractF32x3(EmitContext& ctx, Id composite, u32 index) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeExtract(ctx.F32[1], composite, index);
 }
 
-void EmitCompositeExtractF32x4(EmitContext&) {
+Id EmitCompositeExtractF32x4(EmitContext& ctx, Id composite, u32 index) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCompositeExtract(ctx.F32[1], composite, index);
 }
 
 void EmitCompositeConstructF64x2(EmitContext&) {

src/shader_recompiler/backend/spirv/emit_spirv_convert.cpp

@@ -0,0 +1,89 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "shader_recompiler/backend/spirv/emit_spirv.h"
+
+namespace Shader::Backend::SPIRV {
+
+Id EmitConvertS16F16(EmitContext& ctx, Id value) {
+    return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToS(ctx.U16, value));
+}
+
+Id EmitConvertS16F32(EmitContext& ctx, Id value) {
+    return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToS(ctx.U16, value));
+}
+
+Id EmitConvertS16F64(EmitContext& ctx, Id value) {
+    return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToS(ctx.U16, value));
+}
+
+Id EmitConvertS32F16(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToS(ctx.U32[1], value);
+}
+
+Id EmitConvertS32F32(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToS(ctx.U32[1], value);
+}
+
+Id EmitConvertS32F64(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToS(ctx.U32[1], value);
+}
+
+Id EmitConvertS64F16(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToS(ctx.U64, value);
+}
+
+Id EmitConvertS64F32(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToS(ctx.U64, value);
+}
+
+Id EmitConvertS64F64(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToS(ctx.U64, value);
+}
+
+Id EmitConvertU16F16(EmitContext& ctx, Id value) {
+    return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToU(ctx.U16, value));
+}
+
+Id EmitConvertU16F32(EmitContext& ctx, Id value) {
+    return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToU(ctx.U16, value));
+}
+
+Id EmitConvertU16F64(EmitContext& ctx, Id value) {
+    return ctx.OpUConvert(ctx.U32[1], ctx.OpConvertFToU(ctx.U16, value));
+}
+
+Id EmitConvertU32F16(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToU(ctx.U32[1], value);
+}
+
+Id EmitConvertU32F32(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToU(ctx.U32[1], value);
+}
+
+Id EmitConvertU32F64(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToU(ctx.U32[1], value);
+}
+
+Id EmitConvertU64F16(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToU(ctx.U64, value);
+}
+
+Id EmitConvertU64F32(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToU(ctx.U64, value);
+}
+
+Id EmitConvertU64F64(EmitContext& ctx, Id value) {
+    return ctx.OpConvertFToU(ctx.U64, value);
+}
+
+Id EmitConvertU64U32(EmitContext& ctx, Id value) {
+    return ctx.OpUConvert(ctx.U64, value);
+}
+
+Id EmitConvertU32U64(EmitContext& ctx, Id value) {
+    return ctx.OpUConvert(ctx.U32[1], value);
+}
+
+} // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_floating_point.cpp

@@ -169,52 +169,52 @@ void EmitFPSaturate64(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitFPRoundEven16(EmitContext&) {
+Id EmitFPRoundEven16(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpRoundEven(ctx.F16[1], value);
 }
 
-void EmitFPRoundEven32(EmitContext&) {
+Id EmitFPRoundEven32(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpRoundEven(ctx.F32[1], value);
 }
 
-void EmitFPRoundEven64(EmitContext&) {
+Id EmitFPRoundEven64(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpRoundEven(ctx.F64[1], value);
 }
 
-void EmitFPFloor16(EmitContext&) {
+Id EmitFPFloor16(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpFloor(ctx.F16[1], value);
 }
 
-void EmitFPFloor32(EmitContext&) {
+Id EmitFPFloor32(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpFloor(ctx.F32[1], value);
 }
 
-void EmitFPFloor64(EmitContext&) {
+Id EmitFPFloor64(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpFloor(ctx.F64[1], value);
 }
 
-void EmitFPCeil16(EmitContext&) {
+Id EmitFPCeil16(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCeil(ctx.F16[1], value);
 }
 
-void EmitFPCeil32(EmitContext&) {
+Id EmitFPCeil32(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCeil(ctx.F32[1], value);
 }
 
-void EmitFPCeil64(EmitContext&) {
+Id EmitFPCeil64(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpCeil(ctx.F64[1], value);
 }
 
-void EmitFPTrunc16(EmitContext&) {
+Id EmitFPTrunc16(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpTrunc(ctx.F16[1], value);
 }
 
-void EmitFPTrunc32(EmitContext&) {
+Id EmitFPTrunc32(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpTrunc(ctx.F32[1], value);
 }
 
-void EmitFPTrunc64(EmitContext&) {
+Id EmitFPTrunc64(EmitContext& ctx, Id value) {
-    throw NotImplementedException("SPIR-V Instruction");
+    return ctx.OpTrunc(ctx.F64[1], value);
 }
 
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_integer.cpp

@@ -113,20 +113,4 @@ Id EmitUGreaterThanEqual(EmitContext& ctx, Id lhs, Id rhs) {
     return ctx.OpUGreaterThanEqual(ctx.U1, lhs, rhs);
 }
 
-void EmitLogicalOr(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitLogicalAnd(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitLogicalXor(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitLogicalNot(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
 } // namespace Shader::Backend::SPIRV

src/shader_recompiler/backend/spirv/emit_spirv_logical.cpp

@@ -6,83 +6,19 @@
 
 namespace Shader::Backend::SPIRV {
 
-void EmitConvertS16F16(EmitContext&) {
+void EmitLogicalOr(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitConvertS16F32(EmitContext&) {
+void EmitLogicalAnd(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitConvertS16F64(EmitContext&) {
+void EmitLogicalXor(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitConvertS32F16(EmitContext&) {
+void EmitLogicalNot(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertS32F32(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertS32F64(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertS64F16(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertS64F32(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertS64F64(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU16F16(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU16F32(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU16F64(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU32F16(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU32F32(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU32F64(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU64F16(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU64F32(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU64F64(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU64U32(EmitContext&) {
-    throw NotImplementedException("SPIR-V Instruction");
-}
-
-void EmitConvertU32U64(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 

src/shader_recompiler/backend/spirv/emit_spirv_memory.cpp

@@ -94,8 +94,7 @@ void EmitLoadStorageS16(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-Id EmitLoadStorage32(EmitContext& ctx, const IR::Value& binding,
+Id EmitLoadStorage32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset) {
-                                const IR::Value& offset) {
     if (!binding.IsImmediate()) {
         throw NotImplementedException("Dynamic storage buffer indexing");
     }
@@ -129,8 +128,8 @@ void EmitWriteStorageS16(EmitContext&) {
     throw NotImplementedException("SPIR-V Instruction");
 }
 
-void EmitWriteStorage32(EmitContext& ctx, const IR::Value& binding,
+void EmitWriteStorage32(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
-                                   const IR::Value& offset, Id value) {
+                        Id value) {
     if (!binding.IsImmediate()) {
         throw NotImplementedException("Dynamic storage buffer indexing");
     }
@@ -140,8 +139,19 @@ void EmitWriteStorage32(EmitContext& ctx, const IR::Value& binding,
     ctx.OpStore(pointer, value);
 }
 
-void EmitWriteStorage64(EmitContext&) {
+void EmitWriteStorage64(EmitContext& ctx, const IR::Value& binding, const IR::Value& offset,
-    throw NotImplementedException("SPIR-V Instruction");
+                        Id value) {
+    if (!binding.IsImmediate()) {
+        throw NotImplementedException("Dynamic storage buffer indexing");
+    }
+    // TODO: Support reinterpreting bindings, guaranteed to be aligned
+    const Id ssbo{ctx.ssbos[binding.U32()]};
+    const Id low_index{StorageIndex(ctx, offset, sizeof(u32))};
+    const Id high_index{ctx.OpIAdd(ctx.U32[1], low_index, ctx.Constant(ctx.U32[1], 1U))};
+    const Id low_pointer{ctx.OpAccessChain(ctx.storage_u32, ssbo, ctx.u32_zero_value, low_index)};
+    const Id high_pointer{ctx.OpAccessChain(ctx.storage_u32, ssbo, ctx.u32_zero_value, high_index)};
+    ctx.OpStore(low_pointer, ctx.OpCompositeExtract(ctx.U32[1], value, 0U));
+    ctx.OpStore(high_pointer, ctx.OpCompositeExtract(ctx.U32[1], value, 1U));
 }
 
 void EmitWriteStorage128(EmitContext&) {

src/shader_recompiler/frontend/ir/condition.h

@@ -4,8 +4,8 @@
 
 #pragma once
 
-#include <string>
 #include <compare>
+#include <string>
 
 #include <fmt/format.h>
 

src/shader_recompiler/frontend/ir/ir_emitter.cpp

@@ -547,11 +547,11 @@ F32 IREmitter::FPSqrt(const F32& value) {
 
 F16F32F64 IREmitter::FPSaturate(const F16F32F64& value) {
     switch (value.Type()) {
-    case Type::U16:
+    case Type::F16:
         return Inst<F16>(Opcode::FPSaturate16, value);
-    case Type::U32:
+    case Type::F32:
         return Inst<F32>(Opcode::FPSaturate32, value);
-    case Type::U64:
+    case Type::F64:
         return Inst<F64>(Opcode::FPSaturate64, value);
     default:
         ThrowInvalidType(value.Type());
@@ -560,11 +560,11 @@ F16F32F64 IREmitter::FPSaturate(const F16F32F64& value) {
 
 F16F32F64 IREmitter::FPRoundEven(const F16F32F64& value) {
     switch (value.Type()) {
-    case Type::U16:
+    case Type::F16:
         return Inst<F16>(Opcode::FPRoundEven16, value);
-    case Type::U32:
+    case Type::F32:
         return Inst<F32>(Opcode::FPRoundEven32, value);
-    case Type::U64:
+    case Type::F64:
         return Inst<F64>(Opcode::FPRoundEven64, value);
     default:
         ThrowInvalidType(value.Type());
@@ -573,11 +573,11 @@ F16F32F64 IREmitter::FPRoundEven(const F16F32F64& value) {
 
 F16F32F64 IREmitter::FPFloor(const F16F32F64& value) {
     switch (value.Type()) {
-    case Type::U16:
+    case Type::F16:
         return Inst<F16>(Opcode::FPFloor16, value);
-    case Type::U32:
+    case Type::F32:
         return Inst<F32>(Opcode::FPFloor32, value);
-    case Type::U64:
+    case Type::F64:
         return Inst<F64>(Opcode::FPFloor64, value);
     default:
         ThrowInvalidType(value.Type());
@@ -586,11 +586,11 @@ F16F32F64 IREmitter::FPFloor(const F16F32F64& value) {
 
 F16F32F64 IREmitter::FPCeil(const F16F32F64& value) {
     switch (value.Type()) {
-    case Type::U16:
+    case Type::F16:
         return Inst<F16>(Opcode::FPCeil16, value);
-    case Type::U32:
+    case Type::F32:
         return Inst<F32>(Opcode::FPCeil32, value);
-    case Type::U64:
+    case Type::F64:
         return Inst<F64>(Opcode::FPCeil64, value);
     default:
         ThrowInvalidType(value.Type());
@@ -599,11 +599,11 @@ F16F32F64 IREmitter::FPCeil(const F16F32F64& value) {
 
 F16F32F64 IREmitter::FPTrunc(const F16F32F64& value) {
     switch (value.Type()) {
-    case Type::U16:
+    case Type::F16:
         return Inst<F16>(Opcode::FPTrunc16, value);
-    case Type::U32:
+    case Type::F32:
         return Inst<F32>(Opcode::FPTrunc32, value);
-    case Type::U64:
+    case Type::F64:
         return Inst<F64>(Opcode::FPTrunc64, value);
     default:
         ThrowInvalidType(value.Type());
@@ -729,33 +729,33 @@ U32U64 IREmitter::ConvertFToS(size_t bitsize, const F16F32F64& value) {
     switch (bitsize) {
     case 16:
         switch (value.Type()) {
-        case Type::U16:
+        case Type::F16:
             return Inst<U32>(Opcode::ConvertS16F16, value);
-        case Type::U32:
+        case Type::F32:
             return Inst<U32>(Opcode::ConvertS16F32, value);
-        case Type::U64:
+        case Type::F64:
             return Inst<U32>(Opcode::ConvertS16F64, value);
         default:
             ThrowInvalidType(value.Type());
         }
     case 32:
         switch (value.Type()) {
-        case Type::U16:
+        case Type::F16:
             return Inst<U32>(Opcode::ConvertS32F16, value);
-        case Type::U32:
+        case Type::F32:
             return Inst<U32>(Opcode::ConvertS32F32, value);
-        case Type::U64:
+        case Type::F64:
             return Inst<U32>(Opcode::ConvertS32F64, value);
         default:
             ThrowInvalidType(value.Type());
         }
     case 64:
         switch (value.Type()) {
-        case Type::U16:
+        case Type::F16:
             return Inst<U64>(Opcode::ConvertS64F16, value);
-        case Type::U32:
+        case Type::F32:
             return Inst<U64>(Opcode::ConvertS64F32, value);
-        case Type::U64:
+        case Type::F64:
             return Inst<U64>(Opcode::ConvertS64F64, value);
         default:
             ThrowInvalidType(value.Type());
@@ -769,33 +769,33 @@ U32U64 IREmitter::ConvertFToU(size_t bitsize, const F16F32F64& value) {
     switch (bitsize) {
     case 16:
         switch (value.Type()) {
-        case Type::U16:
+        case Type::F16:
             return Inst<U32>(Opcode::ConvertU16F16, value);
-        case Type::U32:
+        case Type::F32:
             return Inst<U32>(Opcode::ConvertU16F32, value);
-        case Type::U64:
+        case Type::F64:
             return Inst<U32>(Opcode::ConvertU16F64, value);
         default:
             ThrowInvalidType(value.Type());
         }
     case 32:
         switch (value.Type()) {
-        case Type::U16:
+        case Type::F16:
             return Inst<U32>(Opcode::ConvertU32F16, value);
-        case Type::U32:
+        case Type::F32:
             return Inst<U32>(Opcode::ConvertU32F32, value);
-        case Type::U64:
+        case Type::F64:
             return Inst<U32>(Opcode::ConvertU32F64, value);
         default:
             ThrowInvalidType(value.Type());
         }
     case 64:
         switch (value.Type()) {
-        case Type::U16:
+        case Type::F16:
             return Inst<U64>(Opcode::ConvertU64F16, value);
-        case Type::U32:
+        case Type::F32:
             return Inst<U64>(Opcode::ConvertU64F32, value);
-        case Type::U64:
+        case Type::F64:
             return Inst<U64>(Opcode::ConvertU64F64, value);
         default:
             ThrowInvalidType(value.Type());
@@ -829,10 +829,10 @@ U32U64 IREmitter::ConvertU(size_t result_bitsize, const U32U64& value) {
     case 64:
         switch (value.Type()) {
         case Type::U32:
+            return Inst<U64>(Opcode::ConvertU64U32, value);
+        case Type::U64:
             // Nothing to do
             return value;
-        case Type::U64:
-            return Inst<U64>(Opcode::ConvertU64U32, value);
         default:
             break;
         }

src/shader_recompiler/frontend/ir/microinstruction.cpp

@@ -216,6 +216,10 @@ void Inst::ReplaceUsesWith(Value replacement) {
     }
 }
 
+void Inst::ReplaceOpcode(IR::Opcode opcode) {
+    op = opcode;
+}
+
 void Inst::Use(const Value& value) {
     Inst* const inst{value.Inst()};
     ++inst->use_count;

src/shader_recompiler/frontend/ir/microinstruction.h

@@ -86,6 +86,8 @@ public:
 
     void ReplaceUsesWith(Value replacement);
 
+    void ReplaceOpcode(IR::Opcode opcode);
+
     template <typename FlagsType>
     requires(sizeof(FlagsType) <= sizeof(u32) && std::is_trivially_copyable_v<FlagsType>)
         [[nodiscard]] FlagsType Flags() const noexcept {

src/shader_recompiler/frontend/ir/opcodes.inc

@@ -119,8 +119,10 @@ OPCODE(PackUint2x32,                                        U64,            U32x
 OPCODE(UnpackUint2x32,                                      U32x2,          U64,                                                            )
 OPCODE(PackFloat2x16,                                       U32,            F16x2,                                                          )
 OPCODE(UnpackFloat2x16,                                     F16x2,          U32,                                                            )
-OPCODE(PackDouble2x32,                                      U64,            U32x2,                                                          )
+OPCODE(PackHalf2x16,                                        U32,            F32x2,                                                          )
-OPCODE(UnpackDouble2x32,                                    U32x2,          U64,                                                            )
+OPCODE(UnpackHalf2x16,                                      F32x2,          U32,                                                            )
+OPCODE(PackDouble2x32,                                      F64,            U32x2,                                                          )
+OPCODE(UnpackDouble2x32,                                    U32x2,          F64,                                                            )
 
 // Pseudo-operation, handled specially at final emit
 OPCODE(GetZeroFromOp,                                       U1,             Opaque,                                                         )

src/shader_recompiler/frontend/maxwell/program.cpp

@@ -56,6 +56,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
             .post_order_blocks{},
         });
     }
+    Optimization::LowerFp16ToFp32(program);
     for (IR::Function& function : functions) {
         function.post_order_blocks = PostOrder(function.blocks);
         Optimization::SsaRewritePass(function.post_order_blocks);
@@ -69,6 +70,7 @@ IR::Program TranslateProgram(ObjectPool<IR::Inst>& inst_pool, ObjectPool<IR::Blo
         Optimization::VerificationPass(function);
     }
     Optimization::CollectShaderInfoPass(program);
+
     fmt::print(stdout, "{}\n", IR::DumpProgram(program));
     return program;
 }

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

@@ -34,7 +34,7 @@ union F2I {
     BitField<8, 2, DestFormat> dest_format;
     BitField<10, 2, SrcFormat> src_format;
     BitField<12, 1, u64> is_signed;
-    BitField<39, 1, Rounding> rounding;
+    BitField<39, 2, Rounding> rounding;
     BitField<49, 1, u64> half;
     BitField<44, 1, u64> ftz;
     BitField<45, 1, u64> abs;
@@ -55,6 +55,28 @@ size_t BitSize(DestFormat dest_format) {
     }
 }
 
+IR::F64 UnpackCbuf(TranslatorVisitor& v, u64 insn) {
+    union {
+        u64 raw;
+        BitField<20, 14, s64> offset;
+        BitField<34, 5, u64> binding;
+    } const cbuf{insn};
+    if (cbuf.binding >= 18) {
+        throw NotImplementedException("Out of bounds constant buffer binding {}", cbuf.binding);
+    }
+    if (cbuf.offset >= 0x4'000 || cbuf.offset < 0) {
+        throw NotImplementedException("Out of bounds constant buffer offset {}", cbuf.offset * 4);
+    }
+    if (cbuf.offset % 2 != 0) {
+        throw NotImplementedException("Unaligned F64 constant buffer offset {}", cbuf.offset * 4);
+    }
+    const IR::U32 binding{v.ir.Imm32(static_cast<u32>(cbuf.binding))};
+    const IR::U32 byte_offset{v.ir.Imm32(static_cast<u32>(cbuf.offset) * 4 + 4)};
+    const IR::U32 cbuf_data{v.ir.GetCbuf(binding, byte_offset)};
+    const IR::Value vector{v.ir.CompositeConstruct(v.ir.Imm32(0U), cbuf_data)};
+    return v.ir.PackDouble2x32(vector);
+}
+
 void TranslateF2I(TranslatorVisitor& v, u64 insn, const IR::F16F32F64& src_a) {
     // F2I is used to convert from a floating point value to an integer
     const F2I f2i{insn};
@@ -82,19 +104,16 @@ void TranslateF2I(TranslatorVisitor& v, u64 insn, const IR::F16F32F64& src_a) {
     const size_t bitsize{BitSize(f2i.dest_format)};
     const IR::U16U32U64 result{v.ir.ConvertFToI(bitsize, is_signed, rounded_value)};
 
+    if (bitsize == 64) {
+        const IR::Value vector{v.ir.UnpackUint2x32(result)};
+        v.X(f2i.dest_reg + 0, IR::U32{v.ir.CompositeExtract(vector, 0)});
+        v.X(f2i.dest_reg + 1, IR::U32{v.ir.CompositeExtract(vector, 1)});
+    } else {
         v.X(f2i.dest_reg, result);
+    }
 
     if (f2i.cc != 0) {
-        v.SetZFlag(v.ir.GetZeroFromOp(result));
+        throw NotImplementedException("F2I CC");
-        if (is_signed) {
-            v.SetSFlag(v.ir.GetSignFromOp(result));
-        } else {
-            v.ResetSFlag();
-        }
-        v.ResetCFlag();
-
-        // TODO: Investigate if out of bound conversions sets the overflow flag
-        v.ResetOFlag();
     }
 }
 } // Anonymous namespace
@@ -118,12 +137,25 @@ void TranslatorVisitor::F2I_reg(u64 insn) {
                                           f2i.base.src_format.Value());
         }
     }()};
-
     TranslateF2I(*this, insn, op_a);
 }
 
-void TranslatorVisitor::F2I_cbuf(u64) {
+void TranslatorVisitor::F2I_cbuf(u64 insn) {
-    throw NotImplementedException("{}", Opcode::F2I_cbuf);
+    const F2I f2i{insn};
+    const IR::F16F32F64 op_a{[&]() -> IR::F16F32F64 {
+        switch (f2i.src_format) {
+        case SrcFormat::F16:
+            return IR::F16{ir.CompositeExtract(ir.UnpackFloat2x16(GetCbuf(insn)), f2i.half)};
+        case SrcFormat::F32:
+            return GetCbufF(insn);
+        case SrcFormat::F64: {
+            return UnpackCbuf(*this, insn);
+        }
+        default:
+            throw NotImplementedException("Invalid F2I source format {}", f2i.src_format.Value());
+        }
+    }()};
+    TranslateF2I(*this, insn, op_a);
 }
 
 void TranslatorVisitor::F2I_imm(u64) {

src/shader_recompiler/frontend/maxwell/translate/impl/impl.h

@@ -11,7 +11,7 @@ namespace Shader::Maxwell {
 
 class TranslatorVisitor {
 public:
-    explicit TranslatorVisitor(Environment& env_, IR::Block& block) : env{env_} ,ir(block) {}
+    explicit TranslatorVisitor(Environment& env_, IR::Block& block) : env{env_}, ir(block) {}
 
     Environment& env;
     IR::IREmitter ir;

src/shader_recompiler/ir_opt/collect_shader_info_pass.cpp

@@ -26,6 +26,22 @@ void Visit(Info& info, IR::Inst& inst) {
     case IR::Opcode::LocalInvocationId:
         info.uses_local_invocation_id = true;
         break;
+    case IR::Opcode::CompositeConstructF16x2:
+    case IR::Opcode::CompositeConstructF16x3:
+    case IR::Opcode::CompositeConstructF16x4:
+    case IR::Opcode::CompositeExtractF16x2:
+    case IR::Opcode::CompositeExtractF16x3:
+    case IR::Opcode::CompositeExtractF16x4:
+    case IR::Opcode::BitCastU16F16:
+    case IR::Opcode::BitCastF16U16:
+    case IR::Opcode::PackFloat2x16:
+    case IR::Opcode::UnpackFloat2x16:
+    case IR::Opcode::ConvertS16F16:
+    case IR::Opcode::ConvertS32F16:
+    case IR::Opcode::ConvertS64F16:
+    case IR::Opcode::ConvertU16F16:
+    case IR::Opcode::ConvertU32F16:
+    case IR::Opcode::ConvertU64F16:
     case IR::Opcode::FPAbs16:
     case IR::Opcode::FPAdd16:
     case IR::Opcode::FPCeil16:
@@ -36,7 +52,7 @@ void Visit(Info& info, IR::Inst& inst) {
     case IR::Opcode::FPRoundEven16:
     case IR::Opcode::FPSaturate16:
     case IR::Opcode::FPTrunc16:
-        info.uses_fp16;
+        info.uses_fp16 = true;
         break;
     case IR::Opcode::FPAbs64:
     case IR::Opcode::FPAdd64:

src/shader_recompiler/ir_opt/constant_propagation_pass.cpp

@@ -104,12 +104,12 @@ void FoldGetPred(IR::Inst& inst) {
 bool FoldXmadMultiply(IR::Block& block, IR::Inst& inst) {
     /*
      * We are looking for this pattern:
-     *   %rhs_bfe = BitFieldUExtract %factor_a, #0, #16 (uses: 1)
+     *   %rhs_bfe = BitFieldUExtract %factor_a, #0, #16
-     *   %rhs_mul = IMul32 %rhs_bfe, %factor_b (uses: 1)
+     *   %rhs_mul = IMul32 %rhs_bfe, %factor_b
-     *   %lhs_bfe = BitFieldUExtract %factor_a, #16, #16 (uses: 1)
+     *   %lhs_bfe = BitFieldUExtract %factor_a, #16, #16
-     *   %rhs_mul = IMul32 %lhs_bfe, %factor_b (uses: 1)
+     *   %rhs_mul = IMul32 %lhs_bfe, %factor_b
-     *   %lhs_shl = ShiftLeftLogical32 %rhs_mul, #16 (uses: 1)
+     *   %lhs_shl = ShiftLeftLogical32 %rhs_mul, #16
-     *   %result  = IAdd32 %lhs_shl, %rhs_mul (uses: 10)
+     *   %result  = IAdd32 %lhs_shl, %rhs_mul
      *
      * And replacing it with
      *   %result  = IMul32 %factor_a, %factor_b

src/shader_recompiler/ir_opt/lower_fp16_to_fp32.cpp

@@ -0,0 +1,79 @@
+// Copyright 2021 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <algorithm>
+
+#include "shader_recompiler/frontend/ir/ir_emitter.h"
+#include "shader_recompiler/frontend/ir/microinstruction.h"
+#include "shader_recompiler/ir_opt/passes.h"
+
+namespace Shader::Optimization {
+namespace {
+IR::Opcode Replace(IR::Opcode op) {
+    switch (op) {
+    case IR::Opcode::FPAbs16:
+        return IR::Opcode::FPAbs32;
+    case IR::Opcode::FPAdd16:
+        return IR::Opcode::FPAdd32;
+    case IR::Opcode::FPCeil16:
+        return IR::Opcode::FPCeil32;
+    case IR::Opcode::FPFloor16:
+        return IR::Opcode::FPFloor32;
+    case IR::Opcode::FPFma16:
+        return IR::Opcode::FPFma32;
+    case IR::Opcode::FPMul16:
+        return IR::Opcode::FPMul32;
+    case IR::Opcode::FPNeg16:
+        return IR::Opcode::FPNeg32;
+    case IR::Opcode::FPRoundEven16:
+        return IR::Opcode::FPRoundEven32;
+    case IR::Opcode::FPSaturate16:
+        return IR::Opcode::FPSaturate32;
+    case IR::Opcode::FPTrunc16:
+        return IR::Opcode::FPTrunc32;
+    case IR::Opcode::CompositeConstructF16x2:
+        return IR::Opcode::CompositeConstructF32x2;
+    case IR::Opcode::CompositeConstructF16x3:
+        return IR::Opcode::CompositeConstructF32x3;
+    case IR::Opcode::CompositeConstructF16x4:
+        return IR::Opcode::CompositeConstructF32x4;
+    case IR::Opcode::CompositeExtractF16x2:
+        return IR::Opcode::CompositeExtractF32x2;
+    case IR::Opcode::CompositeExtractF16x3:
+        return IR::Opcode::CompositeExtractF32x3;
+    case IR::Opcode::CompositeExtractF16x4:
+        return IR::Opcode::CompositeExtractF32x4;
+    case IR::Opcode::ConvertS16F16:
+        return IR::Opcode::ConvertS16F32;
+    case IR::Opcode::ConvertS32F16:
+        return IR::Opcode::ConvertS32F32;
+    case IR::Opcode::ConvertS64F16:
+        return IR::Opcode::ConvertS64F32;
+    case IR::Opcode::ConvertU16F16:
+        return IR::Opcode::ConvertU16F32;
+    case IR::Opcode::ConvertU32F16:
+        return IR::Opcode::ConvertU32F32;
+    case IR::Opcode::ConvertU64F16:
+        return IR::Opcode::ConvertU64F32;
+    case IR::Opcode::PackFloat2x16:
+        return IR::Opcode::PackHalf2x16;
+    case IR::Opcode::UnpackFloat2x16:
+        return IR::Opcode::UnpackHalf2x16;
+    default:
+        return op;
+    }
+}
+} // Anonymous namespace
+
+void LowerFp16ToFp32(IR::Program& program) {
+    for (IR::Function& function : program.functions) {
+        for (IR::Block* const block : function.blocks) {
+            for (IR::Inst& inst : block->Instructions()) {
+                inst.ReplaceOpcode(Replace(inst.Opcode()));
+            }
+        }
+    }
+}
+
+} // namespace Shader::Optimization

src/shader_recompiler/ir_opt/passes.h

@@ -24,6 +24,7 @@ void ConstantPropagationPass(IR::Block& block);
 void DeadCodeEliminationPass(IR::Block& block);
 void GlobalMemoryToStorageBufferPass(IR::Program& program);
 void IdentityRemovalPass(IR::Function& function);
+void LowerFp16ToFp32(IR::Program& program);
 void SsaRewritePass(std::span<IR::Block* const> post_order_blocks);
 void VerificationPass(const IR::Function& function);
 

src/shader_recompiler/main.cpp

@@ -67,8 +67,8 @@ int main() {
     ObjectPool<IR::Inst> inst_pool;
     ObjectPool<IR::Block> block_pool;
 
-    FileEnvironment env{"D:\\Shaders\\Database\\Oninaki\\CS8F146B41DB6BD826.bin"};
+    // FileEnvironment env{"D:\\Shaders\\Database\\Oninaki\\CS8F146B41DB6BD826.bin"};
-    // FileEnvironment env{"D:\\Shaders\\shader.bin"};
+    FileEnvironment env{"D:\\Shaders\\shader.bin"};
     block_pool.ReleaseContents();
     inst_pool.ReleaseContents();
     flow_block_pool.ReleaseContents();
@@ -76,5 +76,9 @@ int main() {
     fmt::print(stdout, "{}\n", cfg.Dot());
     IR::Program program{TranslateProgram(inst_pool, block_pool, env, cfg)};
     fmt::print(stdout, "{}\n", IR::DumpProgram(program));
-    void(Backend::SPIRV::EmitSPIRV(env, program));
+    const std::vector<u32> spirv{Backend::SPIRV::EmitSPIRV(env, program)};
+    std::FILE* const file{std::fopen("D:\\shader.spv", "wb")};
+    std::fwrite(spirv.data(), spirv.size(), sizeof(u32), file);
+    std::fclose(file);
+    std::system("spirv-dis D:\\shader.spv");
 }

src/shader_recompiler/object_pool.h

@@ -18,7 +18,7 @@ public:
     }
 
     template <typename... Args>
-    requires std::is_constructible_v<T, Args...> [[nodiscard]] T* Create(Args&&... args) {
+    requires std::is_constructible_v<T, Args...>[[nodiscard]] T* Create(Args&&... args) {
         return std::construct_at(Memory(), std::forward<Args>(args)...);
     }
 

src/video_core/renderer_vulkan/vk_compute_pass.cpp

@@ -206,6 +206,8 @@ VKComputePass::VKComputePass(const Device& device, VKDescriptorPool& descriptor_
         .codeSize = static_cast<u32>(code.size_bytes()),
         .pCode = code.data(),
     });
+    /*
+    FIXME
     pipeline = device.GetLogical().CreateComputePipeline({
         .sType = VK_STRUCTURE_TYPE_COMPUTE_PIPELINE_CREATE_INFO,
         .pNext = nullptr,
@@ -224,6 +226,7 @@ VKComputePass::VKComputePass(const Device& device, VKDescriptorPool& descriptor_
         .basePipelineHandle = nullptr,
         .basePipelineIndex = 0,
     });
+    */
 }
 
 VKComputePass::~VKComputePass() = default;

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -31,8 +31,6 @@
 #include "video_core/vulkan_common/vulkan_device.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
-#pragma optimize("", off)
-
 namespace Vulkan {
 MICROPROFILE_DECLARE(Vulkan_PipelineCache);
 
@@ -180,6 +178,12 @@ ComputePipeline PipelineCache::CreateComputePipeline(ShaderInfo* shader_info) {
         // TODO: Load from cache
     }
     const auto [info, code]{Shader::RecompileSPIRV(env, qmd.program_start)};
+
+    FILE* file = fopen("D:\\shader.spv", "wb");
+    fwrite(code.data(), 4, code.size(), file);
+    fclose(file);
+    std::system("spirv-dis D:\\shader.spv");
+
     shader_info->unique_hash = env.ComputeHash();
     shader_info->size_bytes = env.ShaderSize();
     return ComputePipeline{device, descriptor_pool, update_descriptor_queue, info,

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -36,8 +36,6 @@
 #include "video_core/vulkan_common/vulkan_device.h"
 #include "video_core/vulkan_common/vulkan_wrapper.h"
 
-#pragma optimize("", off)
-
 namespace Vulkan {
 
 using Maxwell = Tegra::Engines::Maxwell3D::Regs;

src/video_core/vulkan_common/vulkan_device.cpp

@@ -247,9 +247,9 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
             .shaderStorageImageArrayDynamicIndexing = false,
             .shaderClipDistance = false,
             .shaderCullDistance = false,
-            .shaderFloat64 = false,
+            .shaderFloat64 = true,
-            .shaderInt64 = false,
+            .shaderInt64 = true,
-            .shaderInt16 = false,
+            .shaderInt16 = true,
             .shaderResourceResidency = false,
             .shaderResourceMinLod = false,
             .sparseBinding = false,
@@ -420,8 +420,8 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
     }
     if (is_float16_supported && driver_id == VK_DRIVER_ID_INTEL_PROPRIETARY_WINDOWS) {
         // Intel's compiler crashes when using fp16 on Astral Chain, disable it for the time being.
-        LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
+        // LOG_WARNING(Render_Vulkan, "Blacklisting Intel proprietary from float16 math");
-        is_float16_supported = false;
+        // is_float16_supported = false;
     }
 
     graphics_queue = logical.GetQueue(graphics_family);