OpcodeDispatcher: Implement shufps with VTBL2 in worst case

In the case that source registers are sequential then this turns in to a
load of the vector constant (2 instructions) and the single tbl
instruction.

If the registers aren't sequential then the tbl turns in to 2 moves and
then the single tbl, which with zero-cycle rename isn't too bad.

Since this is a worst case option this is significantly better than the
previous implementation doing a bunch of inserts which was always 9
instructions.
We should still strive to implement faster versions without the use of
TBL2 if possible but this makes it less of a concern.
This commit is contained in:
Ryan Houdek committed 2023-09-13 11:31:20 -07:00
1 parent e9d96ce538
commit db5056f275
3 files changed
+56 -3

No files matched your search

@@ -126,6 +126,55 @@ constexpr static auto PSHUFD_LUT {
}()
};
constexpr static auto SHUFPS_LUT {
[]() consteval {
struct LUTType {
uint64_t Val[2];
};
// 32-bit words in [127:96], [95:64], [63:32], [31:0] are selected using the 8-bit Index.
// Expectation for this LUT is to simulate SHUFPS with ARM's TBL (two register) instruction.
// SHUFPS behaviour:
// Two 32-bits words from each source are selected from each source in the lower and upper halves of the 128-bit destination.
// Dest[31:0] = Src1[<Word0>]
// Dest[63:32] = Src1[<Word1>]
// Dest[95:64] = Src2[<Word2>]
// Dest[127:96] = Src2[<Word3>]
std::array<LUTType, 256> TotalLUT{};
const uint64_t WordSelectionSrc1[4] = {
0x03'02'01'00,
0x07'06'05'04,
0x0b'0a'09'08,
0x0f'0e'0d'0c,
};
// Src2 needs to offset each byte index by 16-bytes to pull from the second source.
const uint64_t WordSelectionSrc2[4] = {
0x03'02'01'00 + (0x10101010),
0x07'06'05'04 + (0x10101010),
0x0b'0a'09'08 + (0x10101010),
0x0f'0e'0d'0c + (0x10101010),
};
for (size_t i = 0; i < 256; ++i) {
auto &LUT = TotalLUT[i];
const auto Word0 = (i >> 0) & 0b11;
const auto Word1 = (i >> 2) & 0b11;
const auto Word2 = (i >> 4) & 0b11;
const auto Word3 = (i >> 6) & 0b11;
LUT.Val[0] =
(WordSelectionSrc1[Word0] << 0) |
(WordSelectionSrc1[Word1] << 32);
LUT.Val[1] =
(WordSelectionSrc2[Word2] << 0) |
(WordSelectionSrc2[Word3] << 32);
}
return TotalLUT;
}()
};
CPUBackend::CPUBackend(FEXCore::Core::InternalThreadState *ThreadState, size_t InitialCodeSize, size_t MaxCodeSize)
: ThreadState(ThreadState), InitialCodeSize(InitialCodeSize), MaxCodeSize(MaxCodeSize) {
@@ -140,6 +189,7 @@ CPUBackend::CPUBackend(FEXCore::Core::InternalThreadState *ThreadState, size_t I
Common.IndexedNamedVectorConstantPointers[FEXCore::IR::IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_PSHUFLW] = reinterpret_cast<uint64_t>(PSHUFLW_LUT.data());
Common.IndexedNamedVectorConstantPointers[FEXCore::IR::IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_PSHUFHW] = reinterpret_cast<uint64_t>(PSHUFHW_LUT.data());
Common.IndexedNamedVectorConstantPointers[FEXCore::IR::IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_PSHUFD] = reinterpret_cast<uint64_t>(PSHUFD_LUT.data());
Common.IndexedNamedVectorConstantPointers[FEXCore::IR::IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_SHUFPS] = reinterpret_cast<uint64_t>(SHUFPS_LUT.data());
#ifndef FEX_DISABLE_TELEMETRY
// Fill in telemetry values
@@ -1384,9 +1384,11 @@ OrderedNode* OpDispatchBuilder::SHUFOpImpl(OpcodeArgs, size_t ElementSize,
return _VZip(DstSize, 8, DupSrc1, DupSrc2);
}
default:
// Intentional doing nothing.
// Uncomment if you want to see remaining fallthrough in the output logs.
//LogMan::Msg::DFmt("shufps: 0x{:x}", Shuffle);
// Use a TBL2 operation to handle this implementation. If the backend supports it.
if (CTX->BackendFeatures.SupportsVTBL2) {
auto LookupIndexes = LoadAndCacheIndexedNamedVectorConstant(DstSize, FEXCore::IR::IndexNamedVectorConstant::INDEXED_NAMED_VECTOR_SHUFPS, Shuffle * 16);
return _VTBL2(DstSize, Src1Node, Src2Node, LookupIndexes);
}
break;
}
}
+1
View File
@@ -539,6 +539,7 @@ enum IndexNamedVectorConstant : uint8_t {
INDEXED_NAMED_VECTOR_PSHUFLW = 0,
INDEXED_NAMED_VECTOR_PSHUFHW,
INDEXED_NAMED_VECTOR_PSHUFD,
INDEXED_NAMED_VECTOR_SHUFPS,
INDEXED_NAMED_VECTOR_MAX,
};