steam-frame/FEX-Emu--FEX
1
0 Fork 0
mirror of https://github.com/FEX-Emu/FEX.git synced 2026-10-06 09:00:17 +02:00
Code Issues Packages Projects Releases 66 Wiki Activity

Merge pull request #3093 from Sonicadvance1/optimize_blendp

Browse Source 
OpcodeDispatcher: Optimize blendp{s,d}
This commit is contained in:
Mai authored and GitHub committed 2023-09-14 00:33:00 -04:00
2 parents 213d3c4e2b 28309a1cc5
commit 92824f5e4d
2 files changed
+152 -68
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

FEXCore/Source/Interface/Core/OpcodeDispatcher/Vector.cpp
+120 -4
View File
@@ -4026,15 +4026,131 @@ void OpDispatchBuilder::AVXVectorRound<8, true>(OpcodeArgs);
template<size_t ElementSize>
void OpDispatchBuilder::VectorBlend(OpcodeArgs) {
const auto DstSize = GetDstSize(Op);
LOGMAN_THROW_A_FMT(Op->Src[1].IsLiteral(), "Src1 needs to be literal here");
uint8_t Select = Op->Src[1].Data.Literal.Value;
OrderedNode *Dest = LoadSource(FPRClass, Op, Op->Dest, Op->Flags, -1);
OrderedNode *Src = LoadSource(FPRClass, Op, Op->Src[0], Op->Flags, -1);
for (size_t i = 0; i < (16 / ElementSize); ++i) {
if (Select & (1 << i)) {
// This could be optimized if it becomes costly
Dest = _VInsElement(16, ElementSize, i, i, Dest, Src);
if constexpr (ElementSize == 4) {
Select &= 0b1111;
switch (Select) {
case 0b0000:
// No-op
return;
case 0b0001:
// Dest[31:0] = Src[31:0]
Dest = _VInsElement(DstSize, ElementSize, 0, 0, Dest, Src);
break;
case 0b0010:
// Dest[63:32] = Src[63:32]
Dest = _VInsElement(DstSize, ElementSize, 1, 1, Dest, Src);
break;
case 0b0011:
// Dest[31:0] = Src[31:0]
// Dest[63:32] = Src[63:32]
Dest = _VInsElement(DstSize, 8, 0, 0, Dest, Src);
break;
case 0b0100:
// Dest[95:64] = Src[95:64]
Dest = _VInsElement(DstSize, ElementSize, 2, 2, Dest, Src);
break;
case 0b0101:
// Dest[31:0] = Src[31:0]
// Dest[95:64] = Src[95:64]
Dest = _VInsElement(DstSize, ElementSize, 0, 0, Dest, Src);
Dest = _VInsElement(DstSize, ElementSize, 2, 2, Dest, Src);
break;
case 0b0110:
// Dest[63:32] = Src[63:32]
// Dest[95:64] = Src[95:64]
Dest = _VInsElement(DstSize, ElementSize, 1, 1, Dest, Src);
Dest = _VInsElement(DstSize, ElementSize, 2, 2, Dest, Src);
break;
case 0b0111:
// Dest[31:0] = Src[31:0]
// Dest[63:32] = Src[63:32]
// Dest[95:64] = Src[95:64]
Dest = _VInsElement(DstSize, 8, 0, 0, Dest, Src);
Dest = _VInsElement(DstSize, ElementSize, 2, 2, Dest, Src);
break;
case 0b1000:
// Dest[127:96] = Src[127:96]
Dest = _VInsElement(DstSize, ElementSize, 3, 3, Dest, Src);
break;
case 0b1001:
// Dest[31:0] = Src[31:0]
// Dest[127:96] = Src[127:96]
Dest = _VInsElement(DstSize, ElementSize, 0, 0, Dest, Src);
Dest = _VInsElement(DstSize, ElementSize, 3, 3, Dest, Src);
break;
case 0b1010:
// Dest[63:32] = Src[63:32]
// Dest[127:96] = Src[127:96]
Dest = _VInsElement(DstSize, ElementSize, 1, 1, Dest, Src);
Dest = _VInsElement(DstSize, ElementSize, 3, 3, Dest, Src);
break;
case 0b1011:
// Dest[31:0] = Src[31:0]
// Dest[63:32] = Src[63:32]
// Dest[95:64] = Src[95:64]
Dest = _VInsElement(DstSize, 8, 0, 0, Dest, Src);
Dest = _VInsElement(DstSize, ElementSize, 3, 3, Dest, Src);
break;
case 0b1100:
// Dest[95:64] = Src[95:64]
// Dest[127:96] = Src[127:96]
Dest = _VInsElement(DstSize, 8, 1, 1, Dest, Src);
break;
case 0b1101:
// Dest[31:0] = Src[31:0]
// Dest[95:64] = Src[95:64]
// Dest[127:96] = Src[127:96]
Dest = _VInsElement(DstSize, ElementSize, 0, 0, Dest, Src);
Dest = _VInsElement(DstSize, 8, 1, 1, Dest, Src);
break;
case 0b1110:
// Dest[63:32] = Src[63:32]
// Dest[95:64] = Src[95:64]
// Dest[127:96] = Src[127:96]
Dest = _VInsElement(DstSize, ElementSize, 1, 1, Dest, Src);
Dest = _VInsElement(DstSize, 8, 1, 1, Dest, Src);
break;
case 0b1111:
// Copy
Dest = Src;
break;
default: break;
}
}
else if constexpr (ElementSize == 8) {
Select &= 0b11;
switch (Select) {
case 0b00:
// No-op
return;
case 0b01:
// Dest[63:0] = Src[63:0]
Dest = _VInsElement(DstSize, ElementSize, 0, 0, Dest, Src);
break;
case 0b10:
// Dest[127:64] = Src[127:64]
Dest = _VInsElement(DstSize, ElementSize, 1, 1, Dest, Src);
break;
case 0b11:
// Copy
Dest = Src;
break;
}
}
else {
for (size_t i = 0; i < (16 / ElementSize); ++i) {
if (Select & (1 << i)) {
// This could be optimized if it becomes costly
Dest = _VInsElement(16, ElementSize, i, i, Dest, Src);
}
}
}
StoreResult(FPRClass, Op, Dest, -1);
unittests/InstructionCountCI/H0F3A.json
+32 -64
View File
@@ -280,15 +280,12 @@
]
},
"blendps xmm0, xmm1, 0000b": {
"ExpectedInstructionCount": 2,
"Optimal": "No",
"ExpectedInstructionCount": 0,
"Optimal": "Yes",
"Comment": [
"0x66 0x0f 0x3a 0x0c"
],
"ExpectedArm64ASM": [
"mov v2.16b, v16.16b",
"mov v16.16b, v2.16b"
]
"ExpectedArm64ASM": []
},
"blendps xmm0, xmm1, 0001b": {
"ExpectedInstructionCount": 1,
@@ -311,18 +308,13 @@
]
},
"blendps xmm0, xmm1, 0011b": {
"ExpectedInstructionCount": 6,
"Optimal": "No",
"ExpectedInstructionCount": 1,
"Optimal": "Yes",
"Comment": [
"0x66 0x0f 0x3a 0x0c"
],
"ExpectedArm64ASM": [
"mov v0.16b, v16.16b",
"mov v0.s[0], v17.s[0]",
"mov v2.16b, v0.16b",
"mov v0.16b, v2.16b",
"mov v0.s[1], v17.s[1]",
"mov v16.16b, v0.16b"
"mov v16.d[0], v17.d[0]"
]
},
"blendps xmm0, xmm1, 0100b": {
@@ -366,16 +358,15 @@
]
},
"blendps xmm0, xmm1, 0111b": {
"ExpectedInstructionCount": 7,
"ExpectedInstructionCount": 6,
"Optimal": "No",
"Comment": [
"0x66 0x0f 0x3a 0x0c"
],
"ExpectedArm64ASM": [
"mov v0.16b, v16.16b",
"mov v0.s[0], v17.s[0]",
"mov v0.d[0], v17.d[0]",
"mov v2.16b, v0.16b",
"mov v2.s[1], v17.s[1]",
"mov v0.16b, v2.16b",
"mov v0.s[2], v17.s[2]",
"mov v16.16b, v0.16b"
@@ -422,22 +413,6 @@
]
},
"blendps xmm0, xmm1, 1011b": {
"ExpectedInstructionCount": 7,
"Optimal": "No",
"Comment": [
"0x66 0x0f 0x3a 0x0c"
],
"ExpectedArm64ASM": [
"mov v0.16b, v16.16b",
"mov v0.s[0], v17.s[0]",
"mov v2.16b, v0.16b",
"mov v2.s[1], v17.s[1]",
"mov v0.16b, v2.16b",
"mov v0.s[3], v17.s[3]",
"mov v16.16b, v0.16b"
]
},
"blendps xmm0, xmm1, 1100b": {
"ExpectedInstructionCount": 6,
"Optimal": "No",
"Comment": [
@@ -445,15 +420,25 @@
],
"ExpectedArm64ASM": [
"mov v0.16b, v16.16b",
"mov v0.s[2], v17.s[2]",
"mov v0.d[0], v17.d[0]",
"mov v2.16b, v0.16b",
"mov v0.16b, v2.16b",
"mov v0.s[3], v17.s[3]",
"mov v16.16b, v0.16b"
]
},
"blendps xmm0, xmm1, 1100b": {
"ExpectedInstructionCount": 1,
"Optimal": "Yes",
"Comment": [
"0x66 0x0f 0x3a 0x0c"
],
"ExpectedArm64ASM": [
"mov v16.d[1], v17.d[1]"
]
},
"blendps xmm0, xmm1, 1101b": {
"ExpectedInstructionCount": 7,
"ExpectedInstructionCount": 6,
"Optimal": "No",
"Comment": [
"0x66 0x0f 0x3a 0x0c"
@@ -462,14 +447,13 @@
"mov v0.16b, v16.16b",
"mov v0.s[0], v17.s[0]",
"mov v2.16b, v0.16b",
"mov v2.s[2], v17.s[2]",
"mov v0.16b, v2.16b",
"mov v0.s[3], v17.s[3]",
"mov v0.d[1], v17.d[1]",
"mov v16.16b, v0.16b"
]
},
"blendps xmm0, xmm1, 1110b": {
"ExpectedInstructionCount": 7,
"ExpectedInstructionCount": 6,
"Optimal": "No",
"Comment": [
"0x66 0x0f 0x3a 0x0c"
@@ -478,39 +462,28 @@
"mov v0.16b, v16.16b",
"mov v0.s[1], v17.s[1]",
"mov v2.16b, v0.16b",
"mov v2.s[2], v17.s[2]",
"mov v0.16b, v2.16b",
"mov v0.s[3], v17.s[3]",
"mov v0.d[1], v17.d[1]",
"mov v16.16b, v0.16b"
]
},
"blendps xmm0, xmm1, 1111b": {
"ExpectedInstructionCount": 8,
"Optimal": "No",
"ExpectedInstructionCount": 1,
"Optimal": "Yes",
"Comment": [
"0x66 0x0f 0x3a 0x0c"
],
"ExpectedArm64ASM": [
"mov v0.16b, v16.16b",
"mov v0.s[0], v17.s[0]",
"mov v2.16b, v0.16b",
"mov v2.s[1], v17.s[1]",
"mov v2.s[2], v17.s[2]",
"mov v0.16b, v2.16b",
"mov v0.s[3], v17.s[3]",
"mov v16.16b, v0.16b"
"mov v16.16b, v17.16b"
]
},
"blendpd xmm0, xmm1, 00b": {
"ExpectedInstructionCount": 2,
"Optimal": "No",
"ExpectedInstructionCount": 0,
"Optimal": "Yes",
"Comment": [
"0x66 0x0f 0x3a 0x0d"
],
"ExpectedArm64ASM": [
"mov v2.16b, v16.16b",
"mov v16.16b, v2.16b"
]
"ExpectedArm64ASM": []
},
"blendpd xmm0, xmm1, 01b": {
"ExpectedInstructionCount": 1,
@@ -533,18 +506,13 @@
]
},
"blendpd xmm0, xmm1, 11b": {
"ExpectedInstructionCount": 6,
"Optimal": "No",
"ExpectedInstructionCount": 1,
"Optimal": "Yes",
"Comment": [
"0x66 0x0f 0x3a 0x0d"
],
"ExpectedArm64ASM": [
"mov v0.16b, v16.16b",
"mov v0.d[0], v17.d[0]",
"mov v2.16b, v0.16b",
"mov v0.16b, v2.16b",
"mov v0.d[1], v17.d[1]",
"mov v16.16b, v0.16b"
"mov v16.16b, v17.16b"
]
},
"pblendw xmm0, xmm1, 00000000b": {