shader_recompiler: frontend: `SOPC` and `SOPK` handling separated; more cmp opcodes (#634)

This commit is contained in:
psucien 2024-08-28 22:27:47 +02:00 committed by GitHub
parent d2e4a200fb
commit 3fbb68048e
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 159 additions and 85 deletions

View File

@ -6,98 +6,150 @@
namespace Shader::Gcn { namespace Shader::Gcn {
void Translator::EmitScalarAlu(const GcnInst& inst) { void Translator::EmitScalarAlu(const GcnInst& inst) {
switch (inst.encoding) {
case InstEncoding::SOPC: {
EmitSOPC(inst);
break;
}
case InstEncoding::SOPK: {
EmitSOPK(inst);
break;
}
default:
switch (inst.opcode) {
case Opcode::S_MOV_B32:
return S_MOV(inst);
case Opcode::S_MUL_I32:
return S_MUL_I32(inst);
case Opcode::S_AND_SAVEEXEC_B64:
return S_AND_SAVEEXEC_B64(inst);
case Opcode::S_MOV_B64:
return S_MOV_B64(inst);
case Opcode::S_OR_B64:
return S_OR_B64(NegateMode::None, false, inst);
case Opcode::S_NOR_B64:
return S_OR_B64(NegateMode::Result, false, inst);
case Opcode::S_XOR_B64:
return S_OR_B64(NegateMode::None, true, inst);
case Opcode::S_ORN2_B64:
return S_OR_B64(NegateMode::Src1, false, inst);
case Opcode::S_AND_B64:
return S_AND_B64(NegateMode::None, inst);
case Opcode::S_NAND_B64:
return S_AND_B64(NegateMode::Result, inst);
case Opcode::S_ANDN2_B64:
return S_AND_B64(NegateMode::Src1, inst);
case Opcode::S_NOT_B64:
return S_NOT_B64(inst);
case Opcode::S_ADD_I32:
return S_ADD_I32(inst);
case Opcode::S_AND_B32:
return S_AND_B32(inst);
case Opcode::S_ASHR_I32:
return S_ASHR_I32(inst);
case Opcode::S_OR_B32:
return S_OR_B32(inst);
case Opcode::S_LSHL_B32:
return S_LSHL_B32(inst);
case Opcode::S_LSHR_B32:
return S_LSHR_B32(inst);
case Opcode::S_CSELECT_B32:
return S_CSELECT_B32(inst);
case Opcode::S_CSELECT_B64:
return S_CSELECT_B64(inst);
case Opcode::S_BFE_U32:
return S_BFE_U32(inst);
case Opcode::S_BFM_B32:
return S_BFM_B32(inst);
case Opcode::S_BREV_B32:
return S_BREV_B32(inst);
case Opcode::S_ADD_U32:
return S_ADD_U32(inst);
case Opcode::S_ADDC_U32:
return S_ADDC_U32(inst);
case Opcode::S_SUB_U32:
case Opcode::S_SUB_I32:
return S_SUB_U32(inst);
case Opcode::S_MIN_U32:
return S_MIN_U32(inst);
case Opcode::S_MAX_U32:
return S_MAX_U32(inst);
case Opcode::S_WQM_B64:
break;
default:
LogMissingOpcode(inst);
}
break;
}
}
void Translator::EmitSOPC(const GcnInst& inst) {
switch (inst.opcode) { switch (inst.opcode) {
case Opcode::S_MOVK_I32: case Opcode::S_CMP_EQ_I32:
return S_MOVK(inst); return S_CMP(ConditionOp::EQ, true, inst);
case Opcode::S_MOV_B32:
return S_MOV(inst);
case Opcode::S_MUL_I32:
return S_MUL_I32(inst);
case Opcode::S_AND_SAVEEXEC_B64:
return S_AND_SAVEEXEC_B64(inst);
case Opcode::S_MOV_B64:
return S_MOV_B64(inst);
case Opcode::S_CMP_LT_U32:
return S_CMP(ConditionOp::LT, false, inst);
case Opcode::S_CMP_LE_U32:
return S_CMP(ConditionOp::LE, false, inst);
case Opcode::S_CMP_LG_U32:
return S_CMP(ConditionOp::LG, false, inst);
case Opcode::S_CMP_LT_I32:
return S_CMP(ConditionOp::LT, true, inst);
case Opcode::S_CMP_LG_I32: case Opcode::S_CMP_LG_I32:
return S_CMP(ConditionOp::LG, true, inst); return S_CMP(ConditionOp::LG, true, inst);
case Opcode::S_CMP_GT_I32: case Opcode::S_CMP_GT_I32:
return S_CMP(ConditionOp::GT, true, inst); return S_CMP(ConditionOp::GT, true, inst);
case Opcode::S_CMP_LE_I32:
return S_CMP(ConditionOp::LE, true, inst);
case Opcode::S_CMP_GE_I32: case Opcode::S_CMP_GE_I32:
return S_CMP(ConditionOp::GE, true, inst); return S_CMP(ConditionOp::GE, true, inst);
case Opcode::S_CMP_EQ_I32: case Opcode::S_CMP_LT_I32:
return S_CMP(ConditionOp::EQ, true, inst); return S_CMP(ConditionOp::LT, true, inst);
case Opcode::S_CMP_LE_I32:
return S_CMP(ConditionOp::LE, true, inst);
case Opcode::S_CMP_EQ_U32: case Opcode::S_CMP_EQ_U32:
return S_CMP(ConditionOp::EQ, false, inst); return S_CMP(ConditionOp::EQ, false, inst);
case Opcode::S_CMP_GE_U32: case Opcode::S_CMP_LG_U32:
return S_CMP(ConditionOp::GE, false, inst); return S_CMP(ConditionOp::LG, false, inst);
case Opcode::S_CMP_GT_U32: case Opcode::S_CMP_GT_U32:
return S_CMP(ConditionOp::GT, false, inst); return S_CMP(ConditionOp::GT, false, inst);
case Opcode::S_OR_B64: case Opcode::S_CMP_GE_U32:
return S_OR_B64(NegateMode::None, false, inst); return S_CMP(ConditionOp::GE, false, inst);
case Opcode::S_NOR_B64: case Opcode::S_CMP_LT_U32:
return S_OR_B64(NegateMode::Result, false, inst); return S_CMP(ConditionOp::LT, false, inst);
case Opcode::S_XOR_B64: case Opcode::S_CMP_LE_U32:
return S_OR_B64(NegateMode::None, true, inst); return S_CMP(ConditionOp::LE, false, inst);
case Opcode::S_ORN2_B64: default:
return S_OR_B64(NegateMode::Src1, false, inst); LogMissingOpcode(inst);
case Opcode::S_AND_B64: }
return S_AND_B64(NegateMode::None, inst); }
case Opcode::S_NAND_B64:
return S_AND_B64(NegateMode::Result, inst); void Translator::EmitSOPK(const GcnInst& inst) {
case Opcode::S_ANDN2_B64: switch (inst.opcode) {
return S_AND_B64(NegateMode::Src1, inst); case Opcode::S_MOVK_I32:
case Opcode::S_NOT_B64: return S_MOVK(inst);
return S_NOT_B64(inst);
case Opcode::S_ADD_I32: case Opcode::S_CMPK_EQ_I32:
return S_ADD_I32(inst); return S_CMPK(ConditionOp::EQ, true, inst);
case Opcode::S_AND_B32: case Opcode::S_CMPK_LG_I32:
return S_AND_B32(inst); return S_CMPK(ConditionOp::LG, true, inst);
case Opcode::S_ASHR_I32: case Opcode::S_CMPK_GT_I32:
return S_ASHR_I32(inst); return S_CMPK(ConditionOp::GT, true, inst);
case Opcode::S_OR_B32: case Opcode::S_CMPK_GE_I32:
return S_OR_B32(inst); return S_CMPK(ConditionOp::GE, true, inst);
case Opcode::S_LSHL_B32: case Opcode::S_CMPK_LT_I32:
return S_LSHL_B32(inst); return S_CMPK(ConditionOp::LT, true, inst);
case Opcode::S_LSHR_B32: case Opcode::S_CMPK_LE_I32:
return S_LSHR_B32(inst); return S_CMPK(ConditionOp::LE, true, inst);
case Opcode::S_CSELECT_B32:
return S_CSELECT_B32(inst); case Opcode::S_CMPK_EQ_U32:
case Opcode::S_CSELECT_B64: return S_CMPK(ConditionOp::EQ, false, inst);
return S_CSELECT_B64(inst); case Opcode::S_CMPK_LG_U32:
case Opcode::S_BFE_U32: return S_CMPK(ConditionOp::LG, false, inst);
return S_BFE_U32(inst); case Opcode::S_CMPK_GT_U32:
case Opcode::S_BFM_B32: return S_CMPK(ConditionOp::GT, false, inst);
return S_BFM_B32(inst); case Opcode::S_CMPK_GE_U32:
case Opcode::S_BREV_B32: return S_CMPK(ConditionOp::GE, false, inst);
return S_BREV_B32(inst); case Opcode::S_CMPK_LT_U32:
case Opcode::S_ADD_U32: return S_CMPK(ConditionOp::LT, false, inst);
return S_ADD_U32(inst); case Opcode::S_CMPK_LE_U32:
case Opcode::S_ADDC_U32: return S_CMPK(ConditionOp::LE, false, inst);
return S_ADDC_U32(inst);
case Opcode::S_ADDK_I32: case Opcode::S_ADDK_I32:
return S_ADDK_I32(inst); return S_ADDK_I32(inst);
case Opcode::S_MULK_I32: case Opcode::S_MULK_I32:
return S_MULK_I32(inst); return S_MULK_I32(inst);
case Opcode::S_SUB_U32:
case Opcode::S_SUB_I32:
return S_SUB_U32(inst);
case Opcode::S_MIN_U32:
return S_MIN_U32(inst);
case Opcode::S_MAX_U32:
return S_MAX_U32(inst);
case Opcode::S_WQM_B64:
break;
case Opcode::S_CMPK_EQ_U32:
return S_CMPK_EQ_U32(inst);
default: default:
LogMissingOpcode(inst); LogMissingOpcode(inst);
} }
@ -154,6 +206,31 @@ void Translator::S_CMP(ConditionOp cond, bool is_signed, const GcnInst& inst) {
ir.SetScc(result); ir.SetScc(result);
} }
void Translator::S_CMPK(ConditionOp cond, bool is_signed, const GcnInst& inst) {
const s32 simm16 = inst.control.sopk.simm;
const IR::U32 lhs = GetSrc(inst.dst[0]);
const IR::U32 rhs = ir.Imm32(simm16);
const IR::U1 result = [&] {
switch (cond) {
case ConditionOp::EQ:
return ir.IEqual(lhs, rhs);
case ConditionOp::LG:
return ir.INotEqual(lhs, rhs);
case ConditionOp::GT:
return ir.IGreaterThan(lhs, rhs, is_signed);
case ConditionOp::GE:
return ir.IGreaterThanEqual(lhs, rhs, is_signed);
case ConditionOp::LT:
return ir.ILessThan(lhs, rhs, is_signed);
case ConditionOp::LE:
return ir.ILessThanEqual(lhs, rhs, is_signed);
default:
UNREACHABLE();
}
}();
ir.SetScc(result);
}
void Translator::S_AND_SAVEEXEC_B64(const GcnInst& inst) { void Translator::S_AND_SAVEEXEC_B64(const GcnInst& inst) {
// This instruction normally operates on 64-bit data (EXEC, VCC, SGPRs) // This instruction normally operates on 64-bit data (EXEC, VCC, SGPRs)
// However here we flatten it to 1-bit EXEC and 1-bit VCC. For the destination // However here we flatten it to 1-bit EXEC and 1-bit VCC. For the destination
@ -470,11 +547,4 @@ void Translator::S_MIN_U32(const GcnInst& inst) {
ir.SetScc(ir.IEqual(result, src0)); ir.SetScc(ir.IEqual(result, src0));
} }
void Translator::S_CMPK_EQ_U32(const GcnInst& inst) {
const s32 simm16 = inst.control.sopk.simm;
const IR::U32 src0{GetSrc(inst.dst[0])};
const IR::U32 src1{ir.Imm32(simm16)};
ir.SetScc(ir.IEqual(src0, src1));
}
} // namespace Shader::Gcn } // namespace Shader::Gcn

View File

@ -69,6 +69,10 @@ public:
void EmitScalarAlu(const GcnInst& inst); void EmitScalarAlu(const GcnInst& inst);
void EmitVectorAlu(const GcnInst& inst); void EmitVectorAlu(const GcnInst& inst);
// Instruction encodings
void EmitSOPC(const GcnInst& inst);
void EmitSOPK(const GcnInst& inst);
// Scalar ALU // Scalar ALU
void S_MOVK(const GcnInst& inst); void S_MOVK(const GcnInst& inst);
void S_MOV(const GcnInst& inst); void S_MOV(const GcnInst& inst);
@ -98,7 +102,7 @@ public:
void S_ADDK_I32(const GcnInst& inst); void S_ADDK_I32(const GcnInst& inst);
void S_MAX_U32(const GcnInst& inst); void S_MAX_U32(const GcnInst& inst);
void S_MIN_U32(const GcnInst& inst); void S_MIN_U32(const GcnInst& inst);
void S_CMPK_EQ_U32(const GcnInst& inst); void S_CMPK(ConditionOp cond, bool is_signed, const GcnInst& inst);
// Scalar Memory // Scalar Memory
void S_LOAD_DWORD(int num_dwords, const GcnInst& inst); void S_LOAD_DWORD(int num_dwords, const GcnInst& inst);