mirrors/dynarmic
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

A64: Implement FMLA/FMLS' half-precision vector indexed variants

Browse source
This commit is contained in:
Lioncash 2019-04-13 02:19:06 -04:00 committed by MerryMage
parent 1fdd3ef8a0
commit 9a58c3f1c7
2 changed files with 43 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
src/frontend/A64/decoder/a64.inc
View file

@ -876,9 +876,9 @@ INST(SQDMULL_elt_2, "SQDMULL, SQDMULL2 (by element)", "0Q001
INST(SQDMULH_elt_2, "SQDMULH (by element)", "0Q001111zzLMmmmm1100H0nnnnnddddd")
INST(SQRDMULH_elt_2, "SQRDMULH (by element)", "0Q001111zzLMmmmm1101H0nnnnnddddd")
INST(SDOT_elt, "SDOT (by element)", "0Q001111zzLMmmmm1110H0nnnnnddddd")
//INST(FMLA_elt_3, "FMLA (by element)", "0Q00111100LMmmmm0001H0nnnnnddddd")
INST(FMLA_elt_3, "FMLA (by element)", "0Q00111100LMmmmm0001H0nnnnnddddd")
INST(FMLA_elt_4, "FMLA (by element)", "0Q0011111zLMmmmm0001H0nnnnnddddd")
//INST(FMLS_elt_3, "FMLS (by element)", "0Q00111100LMmmmm0101H0nnnnnddddd")
INST(FMLS_elt_3, "FMLS (by element)", "0Q00111100LMmmmm0101H0nnnnnddddd")
INST(FMLS_elt_4, "FMLS (by element)", "0Q0011111zLMmmmm0101H0nnnnnddddd")
//INST(FMUL_elt_3, "FMUL (by element)", "0Q00111100LMmmmm1001H0nnnnnddddd")
INST(FMUL_elt_4, "FMUL (by element)", "0Q0011111zLMmmmm1001H0nnnnnddddd")

41
src/frontend/A64/translate/impl/simd_vector_x_indexed_element.cpp
View file

@ -89,6 +89,39 @@ bool FPMultiplyByElement(TranslatorVisitor& v, bool Q, bool sz, Imm<1> L, Imm<1>
return true;
}
bool FPMultiplyByElementHalfPrecision(TranslatorVisitor& v, bool Q, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H,
Vec Vn, Vec Vd, ExtraBehavior extra_behavior) {
const size_t idxdsize = H == 1 ? 128 : 64;
const size_t index = concatenate(H, L, M).ZeroExtend();
const Vec Vm = Vmlo.ZeroExtend<Vec>();
const size_t esize = 16;
const size_t datasize = Q ? 128 : 64;
const IR::UAny element2 = v.ir.VectorGetElement(esize, v.V(idxdsize, Vm), index);
const IR::U128 operand1 = v.V(datasize, Vn);
const IR::U128 operand2 = Q ? v.ir.VectorBroadcast(esize, element2) : v.ir.VectorBroadcastLower(esize, element2);
const IR::U128 operand3 = v.V(datasize, Vd);
// TODO: We currently don't implement half-precision paths for
// regular multiplies and extended multiplies.
const IR::U128 result = [&]{
switch (extra_behavior) {
case ExtraBehavior::None:
break;
case ExtraBehavior::Extended:
break;
case ExtraBehavior::Accumulate:
return v.ir.FPVectorMulAdd(esize, operand3, operand1, operand2);
case ExtraBehavior::Subtract:
return v.ir.FPVectorMulAdd(esize, operand3, v.ir.FPVectorNeg(esize, operand1), operand2);
}
UNREACHABLE();
return IR::U128{};
}();
v.V(datasize, Vd, result);
return true;
}
using ExtensionFunction = IR::U32 (IREmitter::*)(const IR::UAny&);
bool DotProduct(TranslatorVisitor& v, bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H,
@ -276,10 +309,18 @@ bool TranslatorVisitor::FCMLA_elt(bool Q, Imm<2> size, Imm<1> L, Imm<1> M, Imm<4
return true;
}
bool TranslatorVisitor::FMLA_elt_3(bool Q, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) {
return FPMultiplyByElementHalfPrecision(*this, Q, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Accumulate);
}
bool TranslatorVisitor::FMLA_elt_4(bool Q, bool sz, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) {
return FPMultiplyByElement(*this, Q, sz, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Accumulate);
}
bool TranslatorVisitor::FMLS_elt_3(bool Q, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) {
return FPMultiplyByElementHalfPrecision(*this, Q, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Subtract);
}
bool TranslatorVisitor::FMLS_elt_4(bool Q, bool sz, Imm<1> L, Imm<1> M, Imm<4> Vmlo, Imm<1> H, Vec Vn, Vec Vd) {
return FPMultiplyByElement(*this, Q, sz, L, M, Vmlo, H, Vn, Vd, ExtraBehavior::Subtract);
}