Potential new ops for Highway
Wishlist
Widening add: u16 + half u8 = u16
F32RoundToNearestEven
F16 WidenMulAccumulate on non-NEON
numpy
Loadn: Gather*, but for stride 2..4 use ld2..4.
LoadnPair: Gather with optimizations in particular for 2x64-bit, which use 128-bit loads plus Combine. Also StorePair
Clear lowest mask bit
Remaining math functions for hwy/contrib/math
High-precision! Consider copying from SLEEF. See #1650.
fmod, nexttoward
Remaining STL functions for hwy/contrib/algo
ShuffleSpan (see #3392)
Reduce (see #3374)
In-place Remove / RemoveIf, like CopyIf, but in-place.
FindLast / FindLastIf (in find-inl.h) - can use FindLastTrue.
index-returning Mismatch(d, a, b, count) (in find-inl.h) - like EqualSpan, but returns the first mismatched index.
AfterN
= Not(FirstN()), replaces several instances. WHILEGE on SVE.
RVV codegen
- Use new mask<->vec cast instruction, possibly for OddEven, ExpandLoad
rgather_vx for broadcasting redsum result?- use new vcreate intrinsics
- Use clipu for ConcatEven, ConcatOdd
x86 codegen
- SumOfLanes 8-bit also use SumsOf8+Broadcast
SVE codegen
- SVE2.1: TBLQ for
TableLookupBytes CombineShiftRightBytes use TableLookupLanes instead?Shuffle*: use TableLookupLanes instead?- Use SME once available: DUP predicate, REVD (rotate 128-bit elements by 64), SCLAMP/UCLAMP, 128-bit TRN/UZP/ZIP (also in F64MM)
emu128 codegen
#pragma unroll(1) in all loops to enable autovectorization
Dup128TableLookupBytes
Avoids having to add offset on RVV. Table must come from LoadDup128.
LoadPromoteTo
For SVE (svld1sb_u32)+WASM? Compiler can probably already fuse.
Done
IsSorted (algo)SignbitConvertF64<->I32 (math-inl)Copysign (math)CopySignToAbs (math)NegCompressMask ops (math)RebindMaskNotFP16 conversionsScatterGatherPauseAbs i64FirstNCompare i64AESRoundCLMul (GCM)TableLookupBytesOr0 (AES)FindFirstTrue (strlen)NECombine partialLoadMaskBits (FirstN)MaskedLoadBf16 promote2ConcatOdd/EvenSwapAdjacentBlocksOddEvenBlocksCompressBlendedStoreRotateRight (Reverse2 i16)Compare128OrAndIfNegativeThenElseMulFixedPoint15 (codec)Insert/ExtractLaneIsNanIsFiniteStoreInterleavedLoadInterleaved (codec)Or3/Xor3NotXor (sort)FindKnownFirstTrue (sort)CompressStore 8-bitExpandLoad (hash)Zen4 target (sort)SSE2 target - by johnplattsAbsDiff int - by johnplattsLe integer - by johnplattsLeadingZeroCount - by johnplatts in #12768-bit Mul(Neg)MulAdd for integerAESRoundInv etc - by johnplatts in #1286OddEven for <64bit lanes: use Set of wider constant 0_1Shl for 8-bitShr for 8-bitFaster Reverse2 16-bitAdd Reverse2 for 8-bitTwoTablesLookupLanes - by johnplatts in #1303Add 8/16-bit TableLookupLanes - by johnplatts in #1303FindLastTrue - by johnplatts in #1308Vec2, Create/Get functionsPromoteTo for all types (#915) - by johnplatts in #1387atan2Slide1Up/Down - by johnplatts in #1496MaxOfLanes, MinOfLanes returning scalarAdd DupEven for 16-bit - by johnplatts in #1431AVX3_SPR targetMaskedGather returns zero for mask=false.GatherIndexN/ScatterIndexNMaskedScatterfloat64 support for WASMLoadNOrPromoteEvenTo - by johnplattsMasked add/sub/divReduceMin/Max like ReduceSum, in addition to Min/MaxOfLanesReductions for 8-bitRVV: Fix remaining 8-bit table lookups for large vectorsQuickSelect algo - by enum-classNew tuple interface for segment load/storeDiv (integer division) and Mod - by johnplattsAddSub and MulAddSub - by johnplattshypot - by johnplattsexp2 - by johnplattsMin/MaxValueCount(If) (https://en.algorithmica.org/hpc/simd/masking/)cbrterfcoshtanpowLookup32tgammalgammaReduceMin/MaxOrNaNDocument Reduce/Min NaN behaviorIndexOfMin/Max (algo)AllOf / AnyOf / NoneOf (algo)EqualSpan (algo)ReverseSpan (algo)NEON dot productilogb, logb, modf, nextafterIguanaMul52NeUnorderedMinMaxValue, IndexOfMinMax (algo)_mm512_getmant (f32/f64)