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

emit_x64_vector_saturation: AVX2 implementation of EmitVectorUnsignedSaturatedAdd64

Browse source
This commit is contained in:
Merry 2021-05-28 14:36:58 +01:00
parent 57601f064b
commit 0a232a6fbf
1 changed files with 41 additions and 31 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

72
src/dynarmic/backend/x64/emit_x64_vector_saturation.cpp
View file

@ -272,46 +272,56 @@ void EmitX64::EmitVectorUnsignedSaturatedAdd32(EmitContext& ctx, IR::Inst* inst)
void EmitX64::EmitVectorUnsignedSaturatedAdd64(EmitContext& ctx, IR::Inst* inst) { void EmitX64::EmitVectorUnsignedSaturatedAdd64(EmitContext& ctx, IR::Inst* inst) {
auto args = ctx.reg_alloc.GetArgumentInfo(inst); auto args = ctx.reg_alloc.GetArgumentInfo(inst);
const Xbyak::Xmm result = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm addend = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
// TODO AVX2
if (code.HasHostFeature(HostFeature::AVX512_Ortho | HostFeature::AVX512DQ)) { if (code.HasHostFeature(HostFeature::AVX512_Ortho | HostFeature::AVX512DQ)) {
// Do a regular unsigned addition const Xbyak::Xmm operand1 = ctx.reg_alloc.UseXmm(args[0]);
code.vpaddq(result, result, addend); const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm result = ctx.reg_alloc.ScratchXmm();
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
// Test if an overflow happened code.vpaddq(result, operand1, operand2);
code.vpcmpuq(k1, result, addend, CmpInt::LessThan); code.vpcmpuq(k1, result, operand1, CmpInt::LessThan);
// Write 0b1111... where overflows have happened
// This is just a quick way to do this without touching memory
code.vpternlogq(result | k1, result, result, 0xFF); code.vpternlogq(result | k1, result, result, 0xFF);
// Set ZF if an overflow happened
code.ktestb(k1, k1); code.ktestb(k1, k1);
code.setnz(overflow);
code.or_(code.byte[code.r15 + code.GetJitStateInfo().offsetof_fpsr_qc], overflow);
ctx.reg_alloc.DefineValue(inst, result);
return;
}
const Xbyak::Xmm operand1 = code.HasHostFeature(HostFeature::AVX) ? ctx.reg_alloc.UseXmm(args[0]) : ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm operand2 = ctx.reg_alloc.UseXmm(args[1]);
const Xbyak::Xmm result = code.HasHostFeature(HostFeature::AVX) ? ctx.reg_alloc.ScratchXmm() : operand1;
const Xbyak::Reg8 overflow = ctx.reg_alloc.ScratchGpr().cvt8();
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
if (code.HasHostFeature(HostFeature::AVX)) {
code.vpxor(xmm0, operand1, operand2);
code.vpand(tmp, operand1, operand2);
code.vpaddq(result, operand1, operand2);
} else { } else {
const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm(); code.movaps(xmm0, operand1);
code.movaps(tmp, result); code.movaps(tmp, operand1);
code.movaps(xmm0, result);
code.pxor(xmm0, addend); code.pxor(xmm0, operand2);
code.pand(tmp, addend); code.pand(tmp, operand2);
code.paddq(result, addend); code.paddq(result, operand2);
}
code.psrlq(xmm0, 1); code.psrlq(xmm0, 1);
code.paddq(tmp, xmm0); code.paddq(tmp, xmm0);
code.psrad(tmp, 31); code.psrad(tmp, 31);
code.pshufd(tmp, tmp, 0b11110101); code.pshufd(tmp, tmp, 0b11110101);
code.por(result, tmp); code.por(result, tmp);
if (code.HasHostFeature(HostFeature::SSE41)) { if (code.HasHostFeature(HostFeature::SSE41)) {
code.ptest(tmp, tmp); code.ptest(tmp, tmp);
} else { } else {
code.movmskpd(overflow.cvt32(), tmp); code.movmskpd(overflow.cvt32(), tmp);
code.test(overflow.cvt32(), overflow.cvt32()); code.test(overflow.cvt32(), overflow.cvt32());
}
} }
code.setnz(overflow); code.setnz(overflow);