suyu/dynarmic
26
1
Fork 1
Code Issues Pull requests Projects Releases Packages Wiki Activity

A64: Implement TRN2

Browse source
This commit is contained in:
Lioncash 2018-04-13 14:45:13 -04:00 committed by MerryMage
parent 30e158f8d0
commit 3b6db59850
2 changed files with 54 additions and 17 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/frontend/A64/decoder/a64.inc
View file

@ -518,7 +518,7 @@ INST(USRA_1, "USRA", "01111
INST(TRN1, "TRN1", "0Q001110zz0mmmmm001010nnnnnddddd")
INST(ZIP1, "ZIP1", "0Q001110zz0mmmmm001110nnnnnddddd")
//INST(UZP2, "UZP2", "0Q001110zz0mmmmm010110nnnnnddddd")
//INST(TRN2, "TRN2", "0Q001110zz0mmmmm011010nnnnnddddd")
INST(TRN2, "TRN2", "0Q001110zz0mmmmm011010nnnnnddddd")
INST(ZIP2, "ZIP2", "0Q001110zz0mmmmm011110nnnnnddddd")
// Data Processing - FP and SIMD - SIMD Extract

69
src/frontend/A64/translate/impl/simd_permute.cpp
View file

@ -4,43 +4,80 @@
* General Public License version 2 or any later version.
*/
#include <tuple>
#include "frontend/A64/translate/impl/impl.h"
namespace Dynarmic::A64 {
bool TranslatorVisitor::TRN1(bool Q, Imm<2> size, Vec Vm, Vec Vn, Vec Vd) {
if (!Q && size == 0b11) {
return ReservedValue();
}
enum class Transposition {
TRN1,
TRN2,
};
static void VectorTranspose(TranslatorVisitor& v, bool Q, Imm<2> size, Vec Vm, Vec Vn, Vec Vd,
Transposition type) {
const size_t datasize = Q ? 128 : 64;
const size_t esize = 8 << size.ZeroExtend();
const u8 esize = static_cast<u8>(8 << size.ZeroExtend());
const IR::U128 m = V(datasize, Vm);
const IR::U128 n = V(datasize, Vn);
const IR::U128 m = v.V(datasize, Vm);
const IR::U128 n = v.V(datasize, Vn);
const IR::U128 result = [&] {
switch (esize) {
case 8:
case 16:
case 32: {
// Create a mask of elements we care about (e.g. for 8-bit 0x00FF00FF00FF00FF)
const u64 mask_element = Common::Ones<u64>(esize);
const u64 mask_value = Common::Replicate<u64>(mask_element, esize * 2);
// Create a mask of elements we care about (e.g. for 8-bit: 0x00FF00FF00FF00FF for TRN1
// and 0xFF00FF00FF00FF00 for TRN2)
const u64 mask_element = [&] {
const size_t shift = type == Transposition::TRN1 ? 0 : esize;
return Common::Ones<u64>(esize) << shift;
}();
const size_t doubled_esize = esize * 2;
const u64 mask_value = Common::Replicate<u64>(mask_element, doubled_esize);
const IR::U128 mask = ir.VectorBroadcast(64, I(64, mask_value));
const IR::U128 anded_m = ir.VectorAnd(m, mask);
const IR::U128 anded_n = ir.VectorAnd(n, mask);
return ir.VectorOr(ir.VectorLogicalShiftLeft(esize * 2, anded_m, esize), anded_n);
const IR::U128 mask = v.ir.VectorBroadcast(64, v.I(64, mask_value));
const IR::U128 anded_m = v.ir.VectorAnd(m, mask);
const IR::U128 anded_n = v.ir.VectorAnd(n, mask);
if (type == Transposition::TRN1) {
return v.ir.VectorOr(v.ir.VectorLogicalShiftLeft(doubled_esize, anded_m, esize), anded_n);
}
return v.ir.VectorOr(v.ir.VectorLogicalShiftRight(doubled_esize, anded_n, esize), anded_m);
}
case 64: {
default:
return ir.VectorSetElement(esize, n, 1, ir.VectorGetElement(esize, m, 0));
const auto [src, src_idx, dst, dst_idx] = [type, m, n] {
if (type == Transposition::TRN1) {
return std::make_tuple(m, 0, n, 1);
}
return std::make_tuple(n, 1, m, 0);
}();
return v.ir.VectorSetElement(esize, dst, dst_idx, v.ir.VectorGetElement(esize, src, src_idx));
}
}
}();
V(datasize, Vd, result);
v.V(datasize, Vd, result);
}
bool TranslatorVisitor::TRN1(bool Q, Imm<2> size, Vec Vm, Vec Vn, Vec Vd) {
if (!Q && size == 0b11) {
return ReservedValue();
}
VectorTranspose(*this, Q, size, Vm, Vn, Vd, Transposition::TRN1);
return true;
}
bool TranslatorVisitor::TRN2(bool Q, Imm<2> size, Vec Vm, Vec Vn, Vec Vd) {
if (!Q && size == 0b11) {
return ReservedValue();
}
VectorTranspose(*this, Q, size, Vm, Vn, Vd, Transposition::TRN2);
return true;
}