Aphcity/suyu
1
0
Fork 0
forked from suyu/suyu
Code Pull requests Activity Actions

decode/arithmetic_half: Fix HADD2 and HMUL2 absolute and negation bits

Browse source 
The encoding for negation and absolute value was wrong.
Extracting is now done manually. Similar instructions having different
encodings is the rule, not the exception. To keep sanity and readability
I preferred to extract the desired bit manually.

This is implemented against nxas:
8dbc389957/table.h (L68)

That is itself tested against nvdisasm (Nvidia's official disassembler).
This commit is contained in:
ReinUsesLisp 2020-04-20 03:24:00 -03:00
parent 4d7d3651f3
commit dbaebd8582
2 changed files with 37 additions and 16 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

@ -817,11 +817,9 @@ union Instruction {
BitField<32, 1, u64> saturate; BitField<32, 1, u64> saturate;
BitField<49, 2, HalfMerge> merge; BitField<49, 2, HalfMerge> merge;
BitField<43, 1, u64> negate_a;
BitField<44, 1, u64> abs_a; BitField<44, 1, u64> abs_a;
BitField<47, 2, HalfType> type_a; BitField<47, 2, HalfType> type_a;
BitField<31, 1, u64> negate_b;
BitField<30, 1, u64> abs_b; BitField<30, 1, u64> abs_b;
BitField<28, 2, HalfType> type_b; BitField<28, 2, HalfType> type_b;

51
src/video_core/shader/decode/arithmetic_half.cpp
View file

@ -19,22 +19,46 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) {
const Instruction instr = {program_code[pc]}; const Instruction instr = {program_code[pc]};
const auto opcode = OpCode::Decode(instr); const auto opcode = OpCode::Decode(instr);
if (opcode->get().GetId() == OpCode::Id::HADD2_C || bool negate_a = false;
opcode->get().GetId() == OpCode::Id::HADD2_R) { bool negate_b = false;
bool absolute_a = false;
bool absolute_b = false;
switch (opcode->get().GetId()) {
case OpCode::Id::HADD2_R:
if (instr.alu_half.ftz == 0) { if (instr.alu_half.ftz == 0) {
LOG_DEBUG(HW_GPU, "{} without FTZ is not implemented", opcode->get().GetName()); LOG_DEBUG(HW_GPU, "{} without FTZ is not implemented", opcode->get().GetName());
} }
negate_a = ((instr.value >> 43) & 1) != 0;
negate_b = ((instr.value >> 31) & 1) != 0;
absolute_a = ((instr.value >> 44) & 1) != 0;
absolute_b = ((instr.value >> 30) & 1) != 0;
break;
case OpCode::Id::HADD2_C:
if (instr.alu_half.ftz == 0) {
LOG_DEBUG(HW_GPU, "{} without FTZ is not implemented", opcode->get().GetName());
}
negate_a = ((instr.value >> 43) & 1) != 0;
negate_b = ((instr.value >> 56) & 1) != 0;
absolute_a = ((instr.value >> 44) & 1) != 0;
absolute_b = ((instr.value >> 54) & 1) != 0;
break;
case OpCode::Id::HMUL2_R:
negate_a = ((instr.value >> 43) & 1) != 0;
absolute_a = ((instr.value >> 44) & 1) != 0;
absolute_b = ((instr.value >> 30) & 1) != 0;
break;
case OpCode::Id::HMUL2_C:
negate_b = ((instr.value >> 31) & 1) != 0;
absolute_a = ((instr.value >> 44) & 1) != 0;
absolute_b = ((instr.value >> 54) & 1) != 0;
break;
} }
const bool negate_a =
opcode->get().GetId() != OpCode::Id::HMUL2_R && instr.alu_half.negate_a != 0;
const bool negate_b =
opcode->get().GetId() != OpCode::Id::HMUL2_C && instr.alu_half.negate_b != 0;
Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.alu_half.type_a); Node op_a = UnpackHalfFloat(GetRegister(instr.gpr8), instr.alu_half.type_a);
op_a = GetOperandAbsNegHalf(op_a, instr.alu_half.abs_a, negate_a); op_a = GetOperandAbsNegHalf(op_a, absolute_a, negate_a);
auto [type_b, op_b] = [&]() -> std::tuple<HalfType, Node> { auto [type_b, op_b] = [this, instr, opcode]() -> std::pair<HalfType, Node> {
switch (opcode->get().GetId()) { switch (opcode->get().GetId()) {
case OpCode::Id::HADD2_C: case OpCode::Id::HADD2_C:
case OpCode::Id::HMUL2_C: case OpCode::Id::HMUL2_C:
@ -48,17 +72,16 @@ u32 ShaderIR::DecodeArithmeticHalf(NodeBlock& bb, u32 pc) {
} }
}(); }();
op_b = UnpackHalfFloat(op_b, type_b); op_b = UnpackHalfFloat(op_b, type_b);
// redeclaration to avoid a bug in clang with reusing local bindings in lambdas op_b = GetOperandAbsNegHalf(op_b, absolute_b, negate_b);
Node op_b_alt = GetOperandAbsNegHalf(op_b, instr.alu_half.abs_b, negate_b);
Node value = [&]() { Node value = [this, opcode, op_a, op_b = op_b] {
switch (opcode->get().GetId()) { switch (opcode->get().GetId()) {
case OpCode::Id::HADD2_C: case OpCode::Id::HADD2_C:
case OpCode::Id::HADD2_R: case OpCode::Id::HADD2_R:
return Operation(OperationCode::HAdd, PRECISE, op_a, op_b_alt); return Operation(OperationCode::HAdd, PRECISE, op_a, op_b);
case OpCode::Id::HMUL2_C: case OpCode::Id::HMUL2_C:
case OpCode::Id::HMUL2_R: case OpCode::Id::HMUL2_R:
return Operation(OperationCode::HMul, PRECISE, op_a, op_b_alt); return Operation(OperationCode::HMul, PRECISE, op_a, op_b);
default: default:
UNIMPLEMENTED_MSG("Unhandled half float instruction: {}", opcode->get().GetName()); UNIMPLEMENTED_MSG("Unhandled half float instruction: {}", opcode->get().GetName());
return Immediate(0); return Immediate(0);