yuzu-emu
/
yuzu
Archived
1
0
Fork 0
Code

Shader_Ir: Implement F16 Variants of F2F, F2I, I2F. 

This commit takes care of implementing the F16 Variants of the 
conversion instructions and makes sure conversions are done.
This commit is contained in:
Fernando Sahmkow 2019-07-20 17:38:25 -04:00
parent 0a67416971
commit 11f4e739bd
5 changed files with 75 additions and 18 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/video_core/engines/shader_bytecode.h
View File

@ -1018,8 +1018,6 @@ union Instruction {
} f2i;
union {
BitField<8, 2, Register::Size> src_size;
BitField<10, 2, Register::Size> dst_size;
BitField<39, 4, u64> rounding;
// H0, H1 extract for F16 missing
BitField<41, 1, u64> selector; // Guessed as some games set it, TODO: reverse this value

18
src/video_core/renderer_opengl/gl_shader_decompiler.cpp
View File

@ -1122,6 +1122,16 @@ private:
Type::Float);
}
std::string FCastHalf0(Operation operation) {
const std::string op_a = VisitOperand(operation, 0, Type::HalfFloat);
return fmt::format("({})[0]", op_a);
}
std::string FCastHalf1(Operation operation) {
const std::string op_a = VisitOperand(operation, 0, Type::HalfFloat);
return fmt::format("({})[1]", op_a);
}
template <Type type>
std::string Min(Operation operation) {
return GenerateBinaryCall(operation, "min", type, type, type);
@ -1278,6 +1288,11 @@ private:
return ApplyPrecise(operation, BitwiseCastResult(clamped, Type::HalfFloat));
}
std::string HCastFloat(Operation operation) {
const std::string op_a = VisitOperand(operation, 0, Type::Float);
return fmt::format("fromHalf2(vec2({}, 0.0f))", op_a);
}
std::string HUnpack(Operation operation) {
const std::string operand{VisitOperand(operation, 0, Type::HalfFloat)};
const auto value = [&]() -> std::string {
@ -1718,6 +1733,8 @@ private:
&GLSLDecompiler::Negate<Type::Float>,
&GLSLDecompiler::Absolute<Type::Float>,
&GLSLDecompiler::FClamp,
&GLSLDecompiler::FCastHalf0,
&GLSLDecompiler::FCastHalf1,
&GLSLDecompiler::Min<Type::Float>,
&GLSLDecompiler::Max<Type::Float>,
&GLSLDecompiler::FCos,
@ -1778,6 +1795,7 @@ private:
&GLSLDecompiler::Absolute<Type::HalfFloat>,
&GLSLDecompiler::HNegate,
&GLSLDecompiler::HClamp,
&GLSLDecompiler::HCastFloat,
&GLSLDecompiler::HUnpack,
&GLSLDecompiler::HMergeF32,
&GLSLDecompiler::HMergeH0,

18
src/video_core/renderer_vulkan/vk_shader_decompiler.cpp
View File

@ -735,6 +735,16 @@ private:
return {};
}
Id FCastHalf0(Operation operation) {
UNIMPLEMENTED();
return {};
}
Id FCastHalf1(Operation operation) {
UNIMPLEMENTED();
return {};
}
Id HNegate(Operation operation) {
UNIMPLEMENTED();
return {};
@ -745,6 +755,11 @@ private:
return {};
}
Id HCastFloat(Operation operation) {
UNIMPLEMENTED();
return {};
}
Id HUnpack(Operation operation) {
UNIMPLEMENTED();
return {};
@ -1210,6 +1225,8 @@ private:
&SPIRVDecompiler::Unary<&Module::OpFNegate, Type::Float>,
&SPIRVDecompiler::Unary<&Module::OpFAbs, Type::Float>,
&SPIRVDecompiler::Ternary<&Module::OpFClamp, Type::Float>,
&SPIRVDecompiler::FCastHalf0,
&SPIRVDecompiler::FCastHalf1,
&SPIRVDecompiler::Binary<&Module::OpFMin, Type::Float>,
&SPIRVDecompiler::Binary<&Module::OpFMax, Type::Float>,
&SPIRVDecompiler::Unary<&Module::OpCos, Type::Float>,
@ -1270,6 +1287,7 @@ private:
&SPIRVDecompiler::Unary<&Module::OpFAbs, Type::HalfFloat>,
&SPIRVDecompiler::HNegate,
&SPIRVDecompiler::HClamp,
&SPIRVDecompiler::HCastFloat,
&SPIRVDecompiler::HUnpack,
&SPIRVDecompiler::HMergeF32,
&SPIRVDecompiler::HMergeH0,

30
src/video_core/shader/decode/conversion.cpp
View File

@ -57,7 +57,7 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
case OpCode::Id::I2F_R:
case OpCode::Id::I2F_C:
case OpCode::Id::I2F_IMM: {
UNIMPLEMENTED_IF(instr.conversion.dst_size != Register::Size::Word);
UNIMPLEMENTED_IF(instr.conversion.dst_size == Register::Size::Long);
UNIMPLEMENTED_IF(instr.conversion.selector);
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in I2F is not implemented");
@ -82,14 +82,19 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
value = GetOperandAbsNegFloat(value, false, instr.conversion.negate_a);
SetInternalFlagsFromFloat(bb, value, instr.generates_cc);
if (instr.conversion.dst_size == Register::Size::Short) {
value = Operation(OperationCode::HCastFloat, PRECISE, value);
}
SetRegister(bb, instr.gpr0, value);
break;
}
case OpCode::Id::F2F_R:
case OpCode::Id::F2F_C:
case OpCode::Id::F2F_IMM: {
UNIMPLEMENTED_IF(instr.conversion.f2f.dst_size != Register::Size::Word);
UNIMPLEMENTED_IF(instr.conversion.f2f.src_size != Register::Size::Word);
UNIMPLEMENTED_IF(instr.conversion.dst_size == Register::Size::Long);
UNIMPLEMENTED_IF(instr.conversion.src_size == Register::Size::Long);
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in F2F is not implemented");
@ -107,6 +112,11 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
}
}();
if (instr.conversion.src_size == Register::Size::Short) {
// TODO: figure where extract is sey in the encoding
value = Operation(OperationCode::FCastHalf0, PRECISE, value);
}
value = GetOperandAbsNegFloat(value, instr.conversion.abs_a, instr.conversion.negate_a);
value = [&]() {
@ -124,19 +134,24 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
default:
UNIMPLEMENTED_MSG("Unimplemented F2F rounding mode {}",
static_cast<u32>(instr.conversion.f2f.rounding.Value()));
return Immediate(0);
return value;
}
}();
value = GetSaturatedFloat(value, instr.alu.saturate_d);
SetInternalFlagsFromFloat(bb, value, instr.generates_cc);
if (instr.conversion.dst_size == Register::Size::Short) {
value = Operation(OperationCode::HCastFloat, PRECISE, value);
}
SetRegister(bb, instr.gpr0, value);
break;
}
case OpCode::Id::F2I_R:
case OpCode::Id::F2I_C:
case OpCode::Id::F2I_IMM: {
UNIMPLEMENTED_IF(instr.conversion.src_size != Register::Size::Word);
UNIMPLEMENTED_IF(instr.conversion.src_size == Register::Size::Long);
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in F2I is not implemented");
Node value = [&]() {
@ -153,6 +168,11 @@ u32 ShaderIR::DecodeConversion(NodeBlock& bb, u32 pc) {
}
}();
if (instr.conversion.src_size == Register::Size::Short) {
// TODO: figure where extract is sey in the encoding
value = Operation(OperationCode::FCastHalf0, PRECISE, value);
}
value = GetOperandAbsNegFloat(value, instr.conversion.abs_a, instr.conversion.negate_a);
value = [&]() {

25
src/video_core/shader/node.h
View File

@ -30,6 +30,8 @@ enum class OperationCode {
FNegate, /// (MetaArithmetic, float a) -> float
FAbsolute, /// (MetaArithmetic, float a) -> float
FClamp, /// (MetaArithmetic, float value, float min, float max) -> float
FCastHalf0, /// (MetaArithmetic, f16vec2 a) -> float
FCastHalf1, /// (MetaArithmetic, f16vec2 a) -> float
FMin, /// (MetaArithmetic, float a, float b) -> float
FMax, /// (MetaArithmetic, float a, float b) -> float
FCos, /// (MetaArithmetic, float a) -> float
@ -83,17 +85,18 @@ enum class OperationCode {
UBitfieldExtract, /// (MetaArithmetic, uint value, int offset, int offset) -> uint
UBitCount, /// (MetaArithmetic, uint) -> uint
HAdd, /// (MetaArithmetic, f16vec2 a, f16vec2 b) -> f16vec2
HMul, /// (MetaArithmetic, f16vec2 a, f16vec2 b) -> f16vec2
HFma, /// (MetaArithmetic, f16vec2 a, f16vec2 b, f16vec2 c) -> f16vec2
HAbsolute, /// (f16vec2 a) -> f16vec2
HNegate, /// (f16vec2 a, bool first, bool second) -> f16vec2
HClamp, /// (f16vec2 src, float min, float max) -> f16vec2
HUnpack, /// (Tegra::Shader::HalfType, T value) -> f16vec2
HMergeF32, /// (f16vec2 src) -> float
HMergeH0, /// (f16vec2 dest, f16vec2 src) -> f16vec2
HMergeH1, /// (f16vec2 dest, f16vec2 src) -> f16vec2
HPack2, /// (float a, float b) -> f16vec2
HAdd, /// (MetaArithmetic, f16vec2 a, f16vec2 b) -> f16vec2
HMul, /// (MetaArithmetic, f16vec2 a, f16vec2 b) -> f16vec2
HFma, /// (MetaArithmetic, f16vec2 a, f16vec2 b, f16vec2 c) -> f16vec2
HAbsolute, /// (f16vec2 a) -> f16vec2
HNegate, /// (f16vec2 a, bool first, bool second) -> f16vec2
HClamp, /// (f16vec2 src, float min, float max) -> f16vec2
HCastFloat, /// (MetaArithmetic, float a) -> f16vec2
HUnpack, /// (Tegra::Shader::HalfType, T value) -> f16vec2
HMergeF32, /// (f16vec2 src) -> float
HMergeH0, /// (f16vec2 dest, f16vec2 src) -> f16vec2
HMergeH1, /// (f16vec2 dest, f16vec2 src) -> f16vec2
HPack2, /// (float a, float b) -> f16vec2
LogicalAssign, /// (bool& dst, bool src) -> void
LogicalAnd, /// (bool a, bool b) -> bool