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

EmitPackedHalvingSub{U,S}16: SSE2 implementation

Browse source
This commit is contained in:
MerryMage 2017-12-14 15:04:20 +00:00
parent a98821da41
commit dc357c780d
1 changed files with 15 additions and 34 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

49
src/backend_x64/emit_x64.cpp
View file

@ -1993,28 +1993,22 @@ void EmitX64::EmitPackedHalvingSubS8(RegAlloc& reg_alloc, IR::Block&, IR::Inst*
void EmitX64::EmitPackedHalvingSubU16(RegAlloc& reg_alloc, IR::Block&, IR::Inst* inst) { void EmitX64::EmitPackedHalvingSubU16(RegAlloc& reg_alloc, IR::Block&, IR::Inst* inst) {
auto args = reg_alloc.GetArgumentInfo(inst); auto args = reg_alloc.GetArgumentInfo(inst);
Xbyak::Reg32 minuend = reg_alloc.UseScratchGpr(args[0]).cvt32(); Xbyak::Xmm minuend = reg_alloc.UseScratchXmm(args[0]);
Xbyak::Reg32 subtrahend = reg_alloc.UseScratchGpr(args[1]).cvt32(); Xbyak::Xmm subtrahend = reg_alloc.UseScratchXmm(args[1]);
// This relies on the equality x-y == (x^y) - (((x^y)&y) << 1). // This relies on the equality x-y == (x^y) - (((x^y)&y) << 1).
// Note that x^y always contains the LSB of the result. // Note that x^y always contains the LSB of the result.
// Since we want to calculate (x+y)/2, we can instead calculate ((x^y)>>1) - ((x^y)&y). // Since we want to calculate (x-y)/2, we can instead calculate ((x^y)>>1) - ((x^y)&y).
code->xor_(minuend, subtrahend); code->pxor(minuend, subtrahend);
code->and_(subtrahend, minuend); code->pand(subtrahend, minuend);
code->shr(minuend, 1); code->psrlw(minuend, 1);
// At this point, // At this point,
// minuend := (a^b) >> 1 // minuend := (a^b) >> 1
// subtrahend := (a^b) & b // subtrahend := (a^b) & b
// We must now perform a partitioned subtraction. code->psubw(minuend, subtrahend);
// We can do this because minuend contains 15 bit fields.
// We use the extra bit in minuend as a bit to borrow from; we set this bit.
// We invert this bit at the end as this tells us if that bit was borrowed from.
code->or_(minuend, 0x80008000);
code->sub(minuend, subtrahend);
code->xor_(minuend, 0x80008000);
reg_alloc.DefineValue(inst, minuend); reg_alloc.DefineValue(inst, minuend);
} }
@ -2022,35 +2016,22 @@ void EmitX64::EmitPackedHalvingSubU16(RegAlloc& reg_alloc, IR::Block&, IR::Inst*
void EmitX64::EmitPackedHalvingSubS16(RegAlloc& reg_alloc, IR::Block&, IR::Inst* inst) { void EmitX64::EmitPackedHalvingSubS16(RegAlloc& reg_alloc, IR::Block&, IR::Inst* inst) {
auto args = reg_alloc.GetArgumentInfo(inst); auto args = reg_alloc.GetArgumentInfo(inst);
Xbyak::Reg32 minuend = reg_alloc.UseScratchGpr(args[0]).cvt32(); Xbyak::Xmm minuend = reg_alloc.UseScratchXmm(args[0]);
Xbyak::Reg32 subtrahend = reg_alloc.UseScratchGpr(args[1]).cvt32(); Xbyak::Xmm subtrahend = reg_alloc.UseScratchXmm(args[1]);
Xbyak::Reg32 carry = reg_alloc.ScratchGpr().cvt32();
// This relies on the equality x-y == (x^y) - (((x^y)&y) << 1). // This relies on the equality x-y == (x^y) - (((x^y)&y) << 1).
// Note that x^y always contains the LSB of the result. // Note that x^y always contains the LSB of the result.
// Since we want to calculate (x-y)/2, we can instead calculate ((x^y)>>1) - ((x^y)&y). // Since we want to calculate (x-y)/2, we can instead calculate ((x^y)>>>1) - ((x^y)&y).
code->xor_(minuend, subtrahend); code->pxor(minuend, subtrahend);
code->and_(subtrahend, minuend); code->pand(subtrahend, minuend);
code->mov(carry, minuend); code->psraw(minuend, 1);
code->and_(carry, 0x80008000);
code->shr(minuend, 1);
// At this point, // At this point,
// minuend := (a^b) >> 1 // minuend := (a^b) >>> 1
// subtrahend := (a^b) & b // subtrahend := (a^b) & b
// carry := (a^b) & 0x80008000
// We must now perform a partitioned subtraction. code->psubw(minuend, subtrahend);
// We can do this because minuend contains 7 bit fields.
// We use the extra bit in minuend as a bit to borrow from; we set this bit.
// We invert this bit at the end as this tells us if that bit was borrowed from.
// We then sign extend the result into this bit.
code->or_(minuend, 0x80008000);
code->sub(minuend, subtrahend);
code->xor_(minuend, 0x80008000);
code->xor_(minuend, carry);
reg_alloc.DefineValue(inst, minuend); reg_alloc.DefineValue(inst, minuend);
} }