lux/yuzu
1
0
Fork 0
mirror of https://git.h3cjp.net/H3cJP/yuzu.git synced 2024-12-15 23:35:38 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #3502 from namkazt/patch-3

Browse source 
shader_decode: Reimplement BFE instructions
This commit is contained in:
Rodrigo Locatti 2020-03-15 21:23:04 -03:00 committed by GitHub
parent d64edf21bb 3287b1247d
commit ddafc99776
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 53 additions and 29 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

11
src/video_core/engines/shader_bytecode.h
View file

@ -911,14 +911,9 @@ union Instruction {
} fadd32i;
union {
BitField<20, 8, u64> shift_position;
BitField<28, 8, u64> shift_length;
BitField<48, 1, u64> negate_b;
BitField<49, 1, u64> negate_a;
u64 GetLeftShiftValue() const {
return 32 - (shift_position + shift_length);
}
BitField<40, 1, u64> brev;
BitField<47, 1, u64> rd_cc;
BitField<48, 1, u64> is_signed;
} bfe;
union {

69
src/video_core/shader/decode/bfe.cpp
View file

@ -17,33 +17,60 @@ u32 ShaderIR::DecodeBfe(NodeBlock& bb, u32 pc) {
const Instruction instr = {program_code[pc]};
const auto opcode = OpCode::Decode(instr);
UNIMPLEMENTED_IF(instr.bfe.negate_b);
Node op_a = GetRegister(instr.gpr8);
op_a = GetOperandAbsNegInteger(op_a, false, instr.bfe.negate_a, false);
Node op_b = [&] {
switch (opcode->get().GetId()) {
case OpCode::Id::BFE_R:
return GetRegister(instr.gpr20);
case OpCode::Id::BFE_C:
return GetConstBuffer(instr.cbuf34.index, instr.cbuf34.GetOffset());
case OpCode::Id::BFE_IMM:
return Immediate(instr.alu.GetSignedImm20_20());
default:
UNREACHABLE();
return Immediate(0);
}
}();
switch (opcode->get().GetId()) {
case OpCode::Id::BFE_IMM: {
UNIMPLEMENTED_IF_MSG(instr.generates_cc,
"Condition codes generation in BFE is not implemented");
UNIMPLEMENTED_IF_MSG(instr.bfe.rd_cc, "Condition codes in BFE is not implemented");
const Node inner_shift_imm = Immediate(static_cast<u32>(instr.bfe.GetLeftShiftValue()));
const Node outer_shift_imm =
Immediate(static_cast<u32>(instr.bfe.GetLeftShiftValue() + instr.bfe.shift_position));
const bool is_signed = instr.bfe.is_signed;
const Node inner_shift =
Operation(OperationCode::ILogicalShiftLeft, NO_PRECISE, op_a, inner_shift_imm);
const Node outer_shift =
Operation(OperationCode::ILogicalShiftRight, NO_PRECISE, inner_shift, outer_shift_imm);
SetInternalFlagsFromInteger(bb, outer_shift, instr.generates_cc);
SetRegister(bb, instr.gpr0, outer_shift);
break;
}
default:
UNIMPLEMENTED_MSG("Unhandled BFE instruction: {}", opcode->get().GetName());
// using reverse parallel method in
// https://graphics.stanford.edu/~seander/bithacks.html#ReverseParallel
// note for later if possible to implement faster method.
if (instr.bfe.brev) {
const auto swap = [&](u32 s, u32 mask) {
Node v1 =
SignedOperation(OperationCode::ILogicalShiftRight, is_signed, op_a, Immediate(s));
if (mask != 0) {
v1 = SignedOperation(OperationCode::IBitwiseAnd, is_signed, std::move(v1),
Immediate(mask));
}
Node v2 = op_a;
if (mask != 0) {
v2 = SignedOperation(OperationCode::IBitwiseAnd, is_signed, std::move(v2),
Immediate(mask));
}
v2 = SignedOperation(OperationCode::ILogicalShiftLeft, is_signed, std::move(v2),
Immediate(s));
return SignedOperation(OperationCode::IBitwiseOr, is_signed, std::move(v1),
std::move(v2));
};
op_a = swap(1, 0x55555555U);
op_a = swap(2, 0x33333333U);
op_a = swap(4, 0x0F0F0F0FU);
op_a = swap(8, 0x00FF00FFU);
op_a = swap(16, 0);
}
const auto offset = SignedOperation(OperationCode::IBitfieldExtract, is_signed, op_b,
Immediate(0), Immediate(8));
const auto bits = SignedOperation(OperationCode::IBitfieldExtract, is_signed, op_b,
Immediate(8), Immediate(8));
auto result = SignedOperation(OperationCode::IBitfieldExtract, is_signed, op_a, offset, bits);
SetRegister(bb, instr.gpr0, std::move(result));
return pc;
}

2
src/video_core/shader/node_helper.cpp
View file

@ -68,6 +68,8 @@ OperationCode SignedToUnsignedCode(OperationCode operation_code, bool is_signed)
return OperationCode::UBitwiseXor;
case OperationCode::IBitwiseNot:
return OperationCode::UBitwiseNot;
case OperationCode::IBitfieldExtract:
return OperationCode::UBitfieldExtract;
case OperationCode::IBitfieldInsert:
return OperationCode::UBitfieldInsert;
case OperationCode::IBitCount: