fabi/Xenia-Canary
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge branch 'master' of https://github.com/xenia-project/xenia into canary_experimental

Browse Source
This commit is contained in:
Gliniak
2022-12-31 10:53:25 +01:00
parent 9f0d3d4c5b e55cb737c1
commit 26415cb8b1
38 changed files with 632 additions and 6570 deletions
Download Patch File Download Diff File Expand all files Collapse all files

111
src/xenia/cpu/backend/x64/x64_seq_vector.cc
View File

@@ -755,6 +755,19 @@ struct VECTOR_SUB
// src1/src2.
e.vpsubd(e.xmm1, src1, src2);
if (e.IsFeatureEnabled(kX64EmitAVX512Ortho)) {
// If the result is less or equal to the first operand then
// we did not underflow
Opmask not_underflow = e.k1;
// _mm_cmple_epu32_mask
e.vpcmpud(not_underflow, e.xmm1, src1, 0x2);
// Copy over values that did not underflow, write zero
// everywhere else
e.vmovdqa32(dest | not_underflow | e.T_z, e.xmm1);
return;
}
// If result is greater than either of the inputs, we've
// underflowed (only need to check one input)
// if (res > src1) then underflowed
@@ -766,6 +779,21 @@ struct VECTOR_SUB
} else {
e.vpsubd(e.xmm1, src1, src2);
if (e.IsFeatureEnabled(kX64EmitAVX512Ortho |
kX64EmitAVX512DQ)) {
e.vmovdqa32(e.xmm3, src1);
e.vpternlogd(e.xmm3, e.xmm1, src2, 0b00011000);
const Opmask saturate = e.k1;
e.vpmovd2m(saturate, e.xmm3);
e.vpsrad(e.xmm2, e.xmm1, 31);
e.vpxord(e.xmm2, e.xmm2, e.GetXmmConstPtr(XMMSignMaskI32));
e.vpblendmd(dest | saturate, e.xmm1, e.xmm2);
return;
}
// We can only overflow if the signs of the operands are
// opposite. If signs are opposite and result sign isn't the
// same as src1's sign, we've overflowed. if ((s32b)((src1 ^
@@ -2296,7 +2324,23 @@ struct PERMUTE_V128
} else {
e.vxorps(e.xmm0, i.src1, e.GetXmmConstPtr(XMMSwapWordMask));
}
if (e.IsFeatureEnabled(kX64EmitAVX512Ortho | kX64EmitAVX512VBMI)) {
Xmm table_lo = e.xmm1;
if (i.src2.is_constant) {
e.LoadConstantXmm(table_lo, i.src2.constant());
} else {
table_lo = i.src2;
}
Opmask zeroes = e.k1;
// _mm_cmple_epu8_mask
e.vpcmpub(zeroes, e.xmm0, e.GetXmmConstPtr(XMMPermuteControl15), 2);
e.vpermb(i.dest.reg() | zeroes | e.T_z, e.xmm0, table_lo);
return;
}
e.vpand(e.xmm0, e.GetXmmConstPtr(XMMPermuteByteMask));
if (i.src2.is_constant) {
e.LoadConstantXmm(i.dest, i.src2.constant());
e.vpshufb(i.dest, i.dest, e.xmm0);
@@ -2312,6 +2356,39 @@ struct PERMUTE_V128
// General permute.
// Control mask needs to be shuffled.
// TODO(benvanik): do constants here instead of in generated code.
if (e.IsFeatureEnabled(kX64EmitAVX512Ortho | kX64EmitAVX512BW |
kX64EmitAVX512VBMI)) {
Xmm table_idx = e.xmm0;
if (i.src1.is_constant) {
e.LoadConstantXmm(table_idx, i.src1.constant());
e.vxorps(table_idx, table_idx, e.GetXmmConstPtr(XMMSwapWordMask));
} else {
e.vxorps(table_idx, i.src1, e.GetXmmConstPtr(XMMSwapWordMask));
}
Xmm table_lo = e.xmm1;
if (i.src2.value->IsConstantZero()) {
e.vpxor(table_lo, table_lo);
} else if (i.src2.is_constant) {
e.LoadConstantXmm(table_lo, i.src2.constant());
} else {
table_lo = i.src2;
}
Xmm table_hi = e.xmm2;
if (i.src3.value->IsConstantZero()) {
e.vpxor(table_hi, table_hi);
} else if (i.src3.is_constant) {
e.LoadConstantXmm(table_hi, i.src3.constant());
} else {
table_hi = i.src3;
}
e.vpermi2b(table_idx, table_lo, table_hi);
e.vmovdqu8(i.dest, table_idx);
return;
}
if (i.src1.is_constant) {
e.LoadConstantXmm(e.xmm2, i.src1.constant());
e.vxorps(e.xmm2, e.xmm2, e.GetXmmConstPtr(XMMSwapWordMask));
@@ -2319,6 +2396,7 @@ struct PERMUTE_V128
e.vxorps(e.xmm2, i.src1, e.GetXmmConstPtr(XMMSwapWordMask));
}
e.vpand(e.xmm2, e.GetXmmConstPtr(XMMPermuteByteMask));
Xmm src2_shuf = e.xmm0;
if (i.src2.value->IsConstantZero()) {
e.vpxor(src2_shuf, src2_shuf);
@@ -2346,8 +2424,39 @@ struct PERMUTE_V128
static void EmitByInt16(X64Emitter& e, const EmitArgType& i) {
// src1 is an array of indices corresponding to positions within src2 and
// src3.
// src3
if (e.IsFeatureEnabled(kX64EmitAVX512Ortho | kX64EmitAVX512BW)) {
e.LoadConstantXmm(e.xmm1, vec128s(0x1));
Xmm table_idx = e.xmm0;
if (i.src1.is_constant) {
e.LoadConstantXmm(table_idx, i.src1.constant());
e.vpxord(table_idx, table_idx, e.xmm1);
} else {
e.vpxord(table_idx, i.src1, e.xmm1);
}
Xmm table_lo = e.xmm1;
if (i.src2.is_constant) {
e.LoadConstantXmm(table_lo, i.src2.constant());
} else {
table_lo = i.src2;
}
Xmm table_hi = e.xmm2;
if (i.src3.is_constant) {
e.LoadConstantXmm(table_hi, i.src3.constant());
} else {
table_hi = i.src3;
}
e.vpermi2w(table_idx, table_lo, table_hi);
e.vmovdqu8(i.dest, table_idx);
return;
}
assert_true(i.src1.is_constant);
vec128_t perm = (i.src1.constant() & vec128s(0xF)) ^ vec128s(0x1);
vec128_t perm_ctrl = vec128b(0);
for (int i = 0; i < 8; i++) {