2
1
Fork 0
mirror of https://github.com/yuzu-emu/yuzu.git synced 2024-07-04 23:31:19 +01:00

Merge pull request #1501 from ReinUsesLisp/half-float

gl_shader_decompiler: Implement H* instructions
This commit is contained in:
bunnei 2018-10-19 23:47:19 -04:00 committed by GitHub
commit b1f8bff7db
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
2 changed files with 458 additions and 0 deletions

View file

@ -335,6 +335,26 @@ enum class IsberdMode : u64 {
enum class IsberdShift : u64 { None = 0, U16 = 1, B32 = 2 };
enum class HalfType : u64 {
H0_H1 = 0,
F32 = 1,
H0_H0 = 2,
H1_H1 = 3,
};
enum class HalfMerge : u64 {
H0_H1 = 0,
F32 = 1,
Mrg_H0 = 2,
Mrg_H1 = 3,
};
enum class HalfPrecision : u64 {
None = 0,
FTZ = 1,
FMZ = 2,
};
enum class IpaInterpMode : u64 {
Linear = 0,
Perspective = 1,
@ -553,6 +573,70 @@ union Instruction {
BitField<49, 1, u64> negate_a;
} alu_integer;
union {
BitField<39, 1, u64> ftz;
BitField<32, 1, u64> saturate;
BitField<49, 2, HalfMerge> merge;
BitField<43, 1, u64> negate_a;
BitField<44, 1, u64> abs_a;
BitField<47, 2, HalfType> type_a;
BitField<31, 1, u64> negate_b;
BitField<30, 1, u64> abs_b;
BitField<47, 2, HalfType> type_b;
BitField<35, 2, HalfType> type_c;
} alu_half;
union {
BitField<39, 2, HalfPrecision> precision;
BitField<39, 1, u64> ftz;
BitField<52, 1, u64> saturate;
BitField<49, 2, HalfMerge> merge;
BitField<43, 1, u64> negate_a;
BitField<44, 1, u64> abs_a;
BitField<47, 2, HalfType> type_a;
} alu_half_imm;
union {
BitField<29, 1, u64> first_negate;
BitField<20, 9, u64> first;
BitField<56, 1, u64> second_negate;
BitField<30, 9, u64> second;
u32 PackImmediates() const {
// Immediates are half floats shifted.
constexpr u32 imm_shift = 6;
return static_cast<u32>((first << imm_shift) | (second << (16 + imm_shift)));
}
} half_imm;
union {
union {
BitField<37, 2, HalfPrecision> precision;
BitField<32, 1, u64> saturate;
BitField<30, 1, u64> negate_c;
BitField<35, 2, HalfType> type_c;
} rr;
BitField<57, 2, HalfPrecision> precision;
BitField<52, 1, u64> saturate;
BitField<49, 2, HalfMerge> merge;
BitField<47, 2, HalfType> type_a;
BitField<56, 1, u64> negate_b;
BitField<28, 2, HalfType> type_b;
BitField<51, 1, u64> negate_c;
BitField<53, 2, HalfType> type_reg39;
} hfma2;
union {
BitField<40, 1, u64> invert;
} popc;
@ -716,6 +800,23 @@ union Instruction {
BitField<45, 4, PredOperation> op; // op with pred39
} csetp;
union {
BitField<35, 4, PredCondition> cond;
BitField<49, 1, u64> h_and;
BitField<6, 1, u64> ftz;
BitField<45, 2, PredOperation> op;
BitField<3, 3, u64> pred3;
BitField<0, 3, u64> pred0;
BitField<43, 1, u64> negate_a;
BitField<44, 1, u64> abs_a;
BitField<47, 2, HalfType> type_a;
BitField<31, 1, u64> negate_b;
BitField<30, 1, u64> abs_b;
BitField<28, 2, HalfType> type_b;
BitField<42, 1, u64> neg_pred;
BitField<39, 3, u64> pred39;
} hsetp2;
union {
BitField<39, 3, u64> pred39;
BitField<42, 1, u64> neg_pred;
@ -730,6 +831,21 @@ union Instruction {
BitField<56, 1, u64> neg_imm;
} fset;
union {
BitField<49, 1, u64> bf;
BitField<35, 3, PredCondition> cond;
BitField<50, 1, u64> ftz;
BitField<45, 2, PredOperation> op;
BitField<43, 1, u64> negate_a;
BitField<44, 1, u64> abs_a;
BitField<47, 2, HalfType> type_a;
BitField<31, 1, u64> negate_b;
BitField<30, 1, u64> abs_b;
BitField<28, 2, HalfType> type_b;
BitField<42, 1, u64> neg_pred;
BitField<39, 3, u64> pred39;
} hset2;
union {
BitField<39, 3, u64> pred39;
BitField<42, 1, u64> neg_pred;
@ -1145,6 +1261,18 @@ public:
LEA_RZ,
LEA_IMM,
LEA_HI,
HADD2_C,
HADD2_R,
HADD2_IMM,
HMUL2_C,
HMUL2_R,
HMUL2_IMM,
HFMA2_CR,
HFMA2_RC,
HFMA2_RR,
HFMA2_IMM_R,
HSETP2_R,
HSET2_R,
POPC_C,
POPC_R,
POPC_IMM,
@ -1218,9 +1346,12 @@ public:
ArithmeticImmediate,
ArithmeticInteger,
ArithmeticIntegerImmediate,
ArithmeticHalf,
ArithmeticHalfImmediate,
Bfe,
Shift,
Ffma,
Hfma2,
Flow,
Synch,
Memory,
@ -1228,6 +1359,8 @@ public:
FloatSetPredicate,
IntegerSet,
IntegerSetPredicate,
HalfSet,
HalfSetPredicate,
PredicateSetPredicate,
PredicateSetRegister,
Conversion,
@ -1389,6 +1522,18 @@ private:
INST("001101101101----", Id::LEA_IMM, Type::ArithmeticInteger, "LEA_IMM"),
INST("010010111101----", Id::LEA_RZ, Type::ArithmeticInteger, "LEA_RZ"),
INST("00011000--------", Id::LEA_HI, Type::ArithmeticInteger, "LEA_HI"),
INST("0111101-1-------", Id::HADD2_C, Type::ArithmeticHalf, "HADD2_C"),
INST("0101110100010---", Id::HADD2_R, Type::ArithmeticHalf, "HADD2_R"),
INST("0111101-0-------", Id::HADD2_IMM, Type::ArithmeticHalfImmediate, "HADD2_IMM"),
INST("0111100-1-------", Id::HMUL2_C, Type::ArithmeticHalf, "HMUL2_C"),
INST("0101110100001---", Id::HMUL2_R, Type::ArithmeticHalf, "HMUL2_R"),
INST("0111100-0-------", Id::HMUL2_IMM, Type::ArithmeticHalfImmediate, "HMUL2_IMM"),
INST("01110---1-------", Id::HFMA2_CR, Type::Hfma2, "HFMA2_CR"),
INST("01100---1-------", Id::HFMA2_RC, Type::Hfma2, "HFMA2_RC"),
INST("0101110100000---", Id::HFMA2_RR, Type::Hfma2, "HFMA2_RR"),
INST("01110---0-------", Id::HFMA2_IMM_R, Type::Hfma2, "HFMA2_R_IMM"),
INST("0101110100100---", Id::HSETP2_R, Type::HalfSetPredicate, "HSETP_R"),
INST("0101110100011---", Id::HSET2_R, Type::HalfSet, "HSET2_R"),
INST("0101000010000---", Id::MUFU, Type::Arithmetic, "MUFU"),
INST("0100110010010---", Id::RRO_C, Type::Arithmetic, "RRO_C"),
INST("0101110010010---", Id::RRO_R, Type::Arithmetic, "RRO_R"),

View file

@ -375,6 +375,49 @@ public:
}
}
/**
* Writes code that does a register assignment to a half float value operation.
* @param reg The destination register to use.
* @param elem The element to use for the operation.
* @param value The code representing the value to assign. Type has to be half float.
* @param type Half float kind of assignment.
* @param dest_num_components Number of components in the destionation.
* @param value_num_components Number of components in the value.
* @param is_saturated Optional, when True, saturates the provided value.
* @param dest_elem Optional, the destination element to use for the operation.
*/
void SetRegisterToHalfFloat(const Register& reg, u64 elem, const std::string& value,
Tegra::Shader::HalfMerge merge, u64 dest_num_components,
u64 value_num_components, bool is_saturated = false,
u64 dest_elem = 0) {
ASSERT_MSG(!is_saturated, "Unimplemented");
const std::string result = [&]() {
switch (merge) {
case Tegra::Shader::HalfMerge::H0_H1:
return "uintBitsToFloat(packHalf2x16(" + value + "))";
case Tegra::Shader::HalfMerge::F32:
// Half float instructions take the first component when doing a float cast.
return "float(" + value + ".x)";
case Tegra::Shader::HalfMerge::Mrg_H0:
// TODO(Rodrigo): I guess Mrg_H0 and Mrg_H1 take their respective component from the
// pack. I couldn't test this on hardware but it shouldn't really matter since most
// of the time when a Mrg_* flag is used both components will be mirrored. That
// being said, it deserves a test.
return "((" + GetRegisterAsInteger(reg, 0, false) +
" & 0xffff0000) | (packHalf2x16(" + value + ") & 0x0000ffff))";
case Tegra::Shader::HalfMerge::Mrg_H1:
return "((" + GetRegisterAsInteger(reg, 0, false) +
" & 0x0000ffff) | (packHalf2x16(" + value + ") & 0xffff0000))";
default:
UNREACHABLE();
return std::string("0");
}
}();
SetRegister(reg, elem, result, dest_num_components, value_num_components, dest_elem);
}
/**
* Writes code that does a register assignment to input attribute operation. Input attributes
* are stored as floats, so this may require conversion.
@ -877,6 +920,19 @@ private:
return fmt::format("uintBitsToFloat({})", instr.alu.GetImm20_32());
}
/// Generates code representing a vec2 pair unpacked from a half float immediate
static std::string UnpackHalfImmediate(const Instruction& instr, bool negate) {
const std::string immediate = GetHalfFloat(std::to_string(instr.half_imm.PackImmediates()));
if (!negate) {
return immediate;
}
const std::string negate_first = instr.half_imm.first_negate != 0 ? "-" : "";
const std::string negate_second = instr.half_imm.second_negate != 0 ? "-" : "";
const std::string negate_vec = "vec2(" + negate_first + "1, " + negate_second + "1)";
return '(' + immediate + " * " + negate_vec + ')';
}
/// Generates code representing a texture sampler.
std::string GetSampler(const Sampler& sampler, Tegra::Shader::TextureType type, bool is_array,
bool is_shadow) {
@ -1012,6 +1068,41 @@ private:
return result;
}
/*
* Transforms the input string GLSL operand into an unpacked half float pair.
* @note This function returns a float type pair instead of a half float pair. This is because
* real half floats are not standarized in GLSL but unpackHalf2x16 (which returns a vec2) is.
* @param operand Input operand. It has to be an unsigned integer.
* @param type How to unpack the unsigned integer to a half float pair.
* @param abs Get the absolute value of unpacked half floats.
* @param neg Get the negative value of unpacked half floats.
* @returns String corresponding to a half float pair.
*/
static std::string GetHalfFloat(const std::string& operand,
Tegra::Shader::HalfType type = Tegra::Shader::HalfType::H0_H1,
bool abs = false, bool neg = false) {
// "vec2" calls emitted in this function are intended to alias components.
const std::string value = [&]() {
switch (type) {
case Tegra::Shader::HalfType::H0_H1:
return "unpackHalf2x16(" + operand + ')';
case Tegra::Shader::HalfType::F32:
return "vec2(uintBitsToFloat(" + operand + "))";
case Tegra::Shader::HalfType::H0_H0:
case Tegra::Shader::HalfType::H1_H1: {
const bool high = type == Tegra::Shader::HalfType::H1_H1;
const char unpack_index = "xy"[high ? 1 : 0];
return "vec2(unpackHalf2x16(" + operand + ")." + unpack_index + ')';
}
default:
UNREACHABLE();
return std::string("vec2(0)");
}
}();
return GetOperandAbsNeg(value, abs, neg);
}
/*
* Returns whether the instruction at the specified offset is a 'sched' instruction.
* Sched instructions always appear before a sequence of 3 instructions.
@ -1748,6 +1839,86 @@ private:
break;
}
case OpCode::Type::ArithmeticHalf: {
if (opcode->GetId() == OpCode::Id::HADD2_C || opcode->GetId() == OpCode::Id::HADD2_R) {
ASSERT_MSG(instr.alu_half.ftz == 0, "Unimplemented");
}
const bool negate_a =
opcode->GetId() != OpCode::Id::HMUL2_R && instr.alu_half.negate_a != 0;
const bool negate_b =
opcode->GetId() != OpCode::Id::HMUL2_C && instr.alu_half.negate_b != 0;
const std::string op_a =
GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr8, 0, false), instr.alu_half.type_a,
instr.alu_half.abs_a != 0, negate_a);
std::string op_b;
switch (opcode->GetId()) {
case OpCode::Id::HADD2_C:
case OpCode::Id::HMUL2_C:
op_b = regs.GetUniform(instr.cbuf34.index, instr.cbuf34.offset,
GLSLRegister::Type::UnsignedInteger);
break;
case OpCode::Id::HADD2_R:
case OpCode::Id::HMUL2_R:
op_b = regs.GetRegisterAsInteger(instr.gpr20, 0, false);
break;
default:
UNREACHABLE();
op_b = "0";
break;
}
op_b = GetHalfFloat(op_b, instr.alu_half.type_b, instr.alu_half.abs_b != 0, negate_b);
const std::string result = [&]() {
switch (opcode->GetId()) {
case OpCode::Id::HADD2_C:
case OpCode::Id::HADD2_R:
return '(' + op_a + " + " + op_b + ')';
case OpCode::Id::HMUL2_C:
case OpCode::Id::HMUL2_R:
return '(' + op_a + " * " + op_b + ')';
default:
LOG_CRITICAL(HW_GPU, "Unhandled half float instruction: {}", opcode->GetName());
UNREACHABLE();
return std::string("0");
}
}();
regs.SetRegisterToHalfFloat(instr.gpr0, 0, result, instr.alu_half.merge, 1, 1,
instr.alu_half.saturate != 0);
break;
}
case OpCode::Type::ArithmeticHalfImmediate: {
if (opcode->GetId() == OpCode::Id::HADD2_IMM) {
ASSERT_MSG(instr.alu_half_imm.ftz == 0, "Unimplemented");
} else {
ASSERT_MSG(instr.alu_half_imm.precision == Tegra::Shader::HalfPrecision::None,
"Unimplemented");
}
const std::string op_a = GetHalfFloat(
regs.GetRegisterAsInteger(instr.gpr8, 0, false), instr.alu_half_imm.type_a,
instr.alu_half_imm.abs_a != 0, instr.alu_half_imm.negate_a != 0);
const std::string op_b = UnpackHalfImmediate(instr, true);
const std::string result = [&]() {
switch (opcode->GetId()) {
case OpCode::Id::HADD2_IMM:
return op_a + " + " + op_b;
case OpCode::Id::HMUL2_IMM:
return op_a + " * " + op_b;
default:
UNREACHABLE();
return std::string("0");
}
}();
regs.SetRegisterToHalfFloat(instr.gpr0, 0, result, instr.alu_half_imm.merge, 1, 1,
instr.alu_half_imm.saturate != 0);
break;
}
case OpCode::Type::Ffma: {
const std::string op_a = regs.GetRegisterAsFloat(instr.gpr8);
std::string op_b = instr.ffma.negate_b ? "-" : "";
@ -1792,6 +1963,59 @@ private:
instr.alu.saturate_d);
break;
}
case OpCode::Type::Hfma2: {
if (opcode->GetId() == OpCode::Id::HFMA2_RR) {
ASSERT_MSG(instr.hfma2.rr.precision == Tegra::Shader::HalfPrecision::None,
"Unimplemented");
} else {
ASSERT_MSG(instr.hfma2.precision == Tegra::Shader::HalfPrecision::None,
"Unimplemented");
}
const bool saturate = opcode->GetId() == OpCode::Id::HFMA2_RR
? instr.hfma2.rr.saturate != 0
: instr.hfma2.saturate != 0;
const std::string op_a =
GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr8, 0, false), instr.hfma2.type_a);
std::string op_b, op_c;
switch (opcode->GetId()) {
case OpCode::Id::HFMA2_CR:
op_b = GetHalfFloat(regs.GetUniform(instr.cbuf34.index, instr.cbuf34.offset,
GLSLRegister::Type::UnsignedInteger),
instr.hfma2.type_b, false, instr.hfma2.negate_b);
op_c = GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr39, 0, false),
instr.hfma2.type_reg39, false, instr.hfma2.negate_c);
break;
case OpCode::Id::HFMA2_RC:
op_b = GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr39, 0, false),
instr.hfma2.type_reg39, false, instr.hfma2.negate_b);
op_c = GetHalfFloat(regs.GetUniform(instr.cbuf34.index, instr.cbuf34.offset,
GLSLRegister::Type::UnsignedInteger),
instr.hfma2.type_b, false, instr.hfma2.negate_c);
break;
case OpCode::Id::HFMA2_RR:
op_b = GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr20, 0, false),
instr.hfma2.type_b, false, instr.hfma2.negate_b);
op_c = GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr39, 0, false),
instr.hfma2.rr.type_c, false, instr.hfma2.rr.negate_c);
break;
case OpCode::Id::HFMA2_IMM_R:
op_b = UnpackHalfImmediate(instr, true);
op_c = GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr39, 0, false),
instr.hfma2.type_reg39, false, instr.hfma2.negate_c);
break;
default:
UNREACHABLE();
op_c = op_b = "vec2(0)";
break;
}
const std::string result = '(' + op_a + " * " + op_b + " + " + op_c + ')';
regs.SetRegisterToHalfFloat(instr.gpr0, 0, result, instr.hfma2.merge, 1, 1, saturate);
break;
}
case OpCode::Type::Conversion: {
switch (opcode->GetId()) {
case OpCode::Id::I2I_R: {
@ -2611,6 +2835,51 @@ private:
}
break;
}
case OpCode::Type::HalfSetPredicate: {
ASSERT_MSG(instr.hsetp2.ftz == 0, "Unimplemented");
const std::string op_a =
GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr8, 0, false), instr.hsetp2.type_a,
instr.hsetp2.abs_a, instr.hsetp2.negate_a);
const std::string op_b = [&]() {
switch (opcode->GetId()) {
case OpCode::Id::HSETP2_R:
return GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr20, 0, false),
instr.hsetp2.type_b, instr.hsetp2.abs_a,
instr.hsetp2.negate_b);
default:
UNREACHABLE();
return std::string("vec2(0)");
}
}();
// We can't use the constant predicate as destination.
ASSERT(instr.hsetp2.pred3 != static_cast<u64>(Pred::UnusedIndex));
const std::string second_pred =
GetPredicateCondition(instr.hsetp2.pred39, instr.hsetp2.neg_pred != 0);
const std::string combiner = GetPredicateCombiner(instr.hsetp2.op);
const std::string component_combiner = instr.hsetp2.h_and ? "&&" : "||";
const std::string predicate =
'(' + GetPredicateComparison(instr.hsetp2.cond, op_a + ".x", op_b + ".x") + ' ' +
component_combiner + ' ' +
GetPredicateComparison(instr.hsetp2.cond, op_a + ".y", op_b + ".y") + ')';
// Set the primary predicate to the result of Predicate OP SecondPredicate
SetPredicate(instr.hsetp2.pred3,
'(' + predicate + ") " + combiner + " (" + second_pred + ')');
if (instr.hsetp2.pred0 != static_cast<u64>(Pred::UnusedIndex)) {
// Set the secondary predicate to the result of !Predicate OP SecondPredicate,
// if enabled
SetPredicate(instr.hsetp2.pred0,
"!(" + predicate + ") " + combiner + " (" + second_pred + ')');
}
break;
}
case OpCode::Type::PredicateSetRegister: {
const std::string op_a =
GetPredicateCondition(instr.pset.pred12, instr.pset.neg_pred12 != 0);
@ -2771,6 +3040,50 @@ private:
}
break;
}
case OpCode::Type::HalfSet: {
ASSERT_MSG(instr.hset2.ftz == 0, "Unimplemented");
const std::string op_a =
GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr8, 0, false), instr.hset2.type_a,
instr.hset2.abs_a != 0, instr.hset2.negate_a != 0);
const std::string op_b = [&]() {
switch (opcode->GetId()) {
case OpCode::Id::HSET2_R:
return GetHalfFloat(regs.GetRegisterAsInteger(instr.gpr20, 0, false),
instr.hset2.type_b, instr.hset2.abs_b != 0,
instr.hset2.negate_b != 0);
default:
UNREACHABLE();
return std::string("vec2(0)");
}
}();
const std::string second_pred =
GetPredicateCondition(instr.hset2.pred39, instr.hset2.neg_pred != 0);
const std::string combiner = GetPredicateCombiner(instr.hset2.op);
// HSET2 operates on each half float in the pack.
std::string result;
for (int i = 0; i < 2; ++i) {
const std::string float_value = i == 0 ? "0x00003c00" : "0x3c000000";
const std::string integer_value = i == 0 ? "0x0000ffff" : "0xffff0000";
const std::string value = instr.hset2.bf == 1 ? float_value : integer_value;
const std::string comp = std::string(".") + "xy"[i];
const std::string predicate =
"((" + GetPredicateComparison(instr.hset2.cond, op_a + comp, op_b + comp) +
") " + combiner + " (" + second_pred + "))";
result += '(' + predicate + " ? " + value + " : 0)";
if (i == 0) {
result += " | ";
}
}
regs.SetRegisterToInteger(instr.gpr0, false, 0, '(' + result + ')', 1, 1);
break;
}
case OpCode::Type::Xmad: {
ASSERT_MSG(!instr.xmad.sign_a, "Unimplemented");
ASSERT_MSG(!instr.xmad.sign_b, "Unimplemented");