EmitPackedHalvingAddU8: Add SSE2 implementation

This commit is contained in:
MerryMage 2017-12-12 16:11:22 +00:00
parent 4682211729
commit fc885ac80f

View file

@ -1737,6 +1737,24 @@ void EmitX64::EmitPackedSubS16(RegAlloc& reg_alloc, IR::Block& block, IR::Inst*
void EmitX64::EmitPackedHalvingAddU8(RegAlloc& reg_alloc, IR::Block&, IR::Inst* inst) {
auto args = reg_alloc.GetArgumentInfo(inst);
if (args[0].IsInXmm() || args[1].IsInXmm()) {
Xbyak::Xmm xmm_a = reg_alloc.UseScratchXmm(args[0]);
Xbyak::Xmm xmm_b = reg_alloc.UseScratchXmm(args[1]);
Xbyak::Xmm ones = reg_alloc.ScratchXmm();
// Since,
// pavg(a, b) == (a + b + 1) >> 1
// Therefore,
// ~pavg(~a, ~b) == (a + b) >> 1
code->pcmpeqb(ones, ones);
code->pxor(xmm_a, ones);
code->pxor(xmm_b, ones);
code->pavgb(xmm_a, xmm_b);
code->pxor(xmm_a, ones);
reg_alloc.DefineValue(inst, xmm_a);
} else {
Xbyak::Reg32 reg_a = reg_alloc.UseScratchGpr(args[0]).cvt32();
Xbyak::Reg32 reg_b = reg_alloc.UseGpr(args[1]).cvt32();
Xbyak::Reg32 xor_a_b = reg_alloc.ScratchGpr().cvt32();
@ -1756,6 +1774,7 @@ void EmitX64::EmitPackedHalvingAddU8(RegAlloc& reg_alloc, IR::Block&, IR::Inst*
code->add(result, xor_a_b);
reg_alloc.DefineValue(inst, result);
}
}
void EmitX64::EmitPackedHalvingAddU16(RegAlloc& reg_alloc, IR::Block&, IR::Inst* inst) {