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

[CPU] vpkd3d VPACK_NORMPACKED64

Browse Source
This commit is contained in:
Triang3l
2019-07-20 22:04:37 +03:00
parent 2350810f4e
commit 8c00aea5ff
5 changed files with 104 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
src/xenia/cpu/backend/x64/x64_seq_vector.cc
View File

@@ -1822,6 +1822,9 @@ struct PACK : Sequence<PACK, I<OPCODE_PACK, V128Op, V128Op, V128Op>> {
case PACK_TYPE_UINT_2101010:
EmitUINT_2101010(e, i);
break;
case PACK_TYPE_ULONG_4202020:
EmitULONG_4202020(e, i);
break;
case PACK_TYPE_8_IN_16:
Emit8_IN_16(e, i, i.instr->flags);
break;
@@ -2002,6 +2005,32 @@ struct PACK : Sequence<PACK, I<OPCODE_PACK, V128Op, V128Op, V128Op>> {
e.vshufps(e.xmm0, i.dest, i.dest, _MM_SHUFFLE(1, 0, 3, 2));
e.vorps(i.dest, e.xmm0);
}
static void EmitULONG_4202020(X64Emitter& e, const EmitArgType& i) {
// XYZ are 20 bits, signed and saturated.
// W is 4 bits, unsigned and saturated.
Xmm src;
if (i.src1.is_constant) {
src = i.dest;
e.LoadConstantXmm(src, i.src1.constant());
} else {
src = i.src1;
}
// Saturate.
e.vmaxps(i.dest, src, e.GetXmmConstPtr(XMMPackULONG_4202020_MinUnpacked));
e.vminps(i.dest, i.dest,
e.GetXmmConstPtr(XMMPackULONG_4202020_MaxUnpacked));
// Remove the unneeded bits of the floats (so excess nibbles will also be
// cleared).
e.vpand(i.dest, e.GetXmmConstPtr(XMMPackULONG_4202020_MaskUnpacked));
// Store Y and W shifted left by 4 so vpshufb can be used with them.
e.vpslld(e.xmm0, i.dest, 4);
// Place XZ where they're supposed to be.
e.vpshufb(i.dest, i.dest, e.GetXmmConstPtr(XMMPackULONG_4202020_PermuteXZ));
// Place YW.
e.vpshufb(e.xmm0, e.xmm0, e.GetXmmConstPtr(XMMPackULONG_4202020_PermuteYW));
// Merge XZ and YW.
e.vorps(i.dest, e.xmm0);
}
static __m128i EmulatePack8_IN_16_UN_UN_SAT(void*, __m128i src1,
__m128i src2) {
alignas(16) uint16_t a[8];
@@ -2214,6 +2243,9 @@ struct UNPACK : Sequence<UNPACK, I<OPCODE_UNPACK, V128Op, V128Op>> {
case PACK_TYPE_UINT_2101010:
EmitUINT_2101010(e, i);
break;
case PACK_TYPE_ULONG_4202020:
EmitULONG_4202020(e, i);
break;
case PACK_TYPE_8_IN_16:
Emit8_IN_16(e, i, i.instr->flags);
break;
@@ -2367,7 +2399,7 @@ struct UNPACK : Sequence<UNPACK, I<OPCODE_UNPACK, V128Op, V128Op>> {
e.vpaddd(i.dest, e.GetXmmConstPtr(XMM3301));
// Return quiet NaNs in case of negative overflow.
e.vcmpeqps(e.xmm0, i.dest, e.GetXmmConstPtr(XMMUnpackSHORT_Overflow));
e.vblendvps(i.dest, i.dest, e.GetXmmConstPtr(XMMUnpackOverflowNaN), e.xmm0);
e.vblendvps(i.dest, i.dest, e.GetXmmConstPtr(XMMQNaN), e.xmm0);
}
static void EmitSHORT_4(X64Emitter& e, const EmitArgType& i) {
// (VD.x) = 3.0 + (VB.x>>16)*2^-22
@@ -2396,7 +2428,7 @@ struct UNPACK : Sequence<UNPACK, I<OPCODE_UNPACK, V128Op, V128Op>> {
e.vpaddd(i.dest, e.GetXmmConstPtr(XMM3333));
// Return quiet NaNs in case of negative overflow.
e.vcmpeqps(e.xmm0, i.dest, e.GetXmmConstPtr(XMMUnpackSHORT_Overflow));
e.vblendvps(i.dest, i.dest, e.GetXmmConstPtr(XMMUnpackOverflowNaN), e.xmm0);
e.vblendvps(i.dest, i.dest, e.GetXmmConstPtr(XMMQNaN), e.xmm0);
}
static void EmitUINT_2101010(X64Emitter& e, const EmitArgType& i) {
Xmm src;
@@ -2437,10 +2469,41 @@ struct UNPACK : Sequence<UNPACK, I<OPCODE_UNPACK, V128Op, V128Op>> {
// Return quiet NaNs in case of negative overflow.
e.vcmpeqps(e.xmm0, i.dest,
e.GetXmmConstPtr(XMMUnpackUINT_2101010_Overflow));
e.vblendvps(i.dest, i.dest, e.GetXmmConstPtr(XMMUnpackOverflowNaN), e.xmm0);
e.vblendvps(i.dest, i.dest, e.GetXmmConstPtr(XMMQNaN), e.xmm0);
// To convert XYZ to -1 to 1, games multiply by 0x46004020 & sub 0x46C06030.
// For W to 0 to 1, they multiply by and subtract 0x4A2AAAAB.
}
static void EmitULONG_4202020(X64Emitter& e, const EmitArgType& i) {
Xmm src;
if (i.src1.is_constant) {
if (i.src1.value->IsConstantZero()) {
e.vmovdqa(i.dest, e.GetXmmConstPtr(XMM3331));
return;
}
src = i.dest;
e.LoadConstantXmm(src, i.src1.constant());
} else {
src = i.src1;
}
// Extract pairs of nibbles to XZYW. XZ will have excess 4 upper bits, YW
// will have excess 4 lower bits.
e.vpshufb(i.dest, src, e.GetXmmConstPtr(XMMUnpackULONG_4202020_Permute));
// Drop the excess nibble of YW.
e.vpsrld(e.xmm0, i.dest, 4);
// Merge XZ and YW now both starting at offset 0.
e.vshufps(i.dest, i.dest, e.xmm0, _MM_SHUFFLE(3, 2, 1, 0));
// Reorder as XYZW.
e.vshufps(i.dest, i.dest, _MM_SHUFFLE(3, 1, 2, 0));
// Drop the excess upper nibble in XZ and sign-extend XYZ.
e.vpslld(i.dest, 12);
e.vpsrad(i.dest, 12);
// Add 3,3,3,1.
e.vpaddd(i.dest, e.GetXmmConstPtr(XMM3331));
// Return quiet NaNs in case of negative overflow.
e.vcmpeqps(e.xmm0, i.dest,
e.GetXmmConstPtr(XMMUnpackULONG_4202020_Overflow));
e.vblendvps(i.dest, i.dest, e.GetXmmConstPtr(XMMQNaN), e.xmm0);
}
static void Emit8_IN_16(X64Emitter& e, const EmitArgType& i, uint32_t flags) {
assert_false(IsPackOutSaturate(flags));
Xmm src;