Skip to content

Latest commit

 

History

History
191 lines (148 loc) · 4.88 KB

op_wishlist.md

File metadata and controls

191 lines (148 loc) · 4.88 KB
Raw

Potential new ops for Highway

[TOC]

Wishlist

F32RoundToNearestEven

NEON dot product

QuickSelect algo

numpy

NeUnordered

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

Lookup128 for 32x 32-bit and 16x 64-bit. permutex2var on AVX-512, else Gather. Also Lookup64 and Lookup32.

ReduceMin/MaxOrNaN

Document Reduce/Min NaN behavior

_mm512_getmant, _mm512_scalef, _mm512_getexp (f32/f64)

Clear lowest mask bit

Remaining math functions for hwy/contrib/math

High-precision! Consider copying from SLEEF. See #1650.

cbrt, cosh, erf, exp2, fmod, hypot, ilogb, lgamma, logb, modf, nextafter, nexttoward, pow, scalbn, tan, tgamma

Remaining STL functions for hwy/contrib/algo

Range coder

Port https://github.com/richgel999/sserangecoding to Highway (~50 instructions).

Iguana (fast LZ + ANS)

Port https://github.com/SnellerInc/sneller/tree/master/ion/zion/iguana (Go+assembly) to Highway.

AfterN

= Not(FirstN()), replaces several instances. WHILEGE on SVE.

52x52=104-bit multiply

For crypto. Native on Icelake+.

RVV codegen

  • New tuple interface for segment load/store
  • Use new mask<->vec cast instruction, possibly for OddEven, ExpandLoad
  • rgather_vx for broadcasting redsum result?

SVE codegen

  • SVE2.1: TBLQ for TableLookupBytes
  • SVE2: use XAR for RotateRight
  • 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

Add emu256 target

Reuse same wasm256 file, #if for wasm-specific parts. Use reserved avx slot.

Conflict detection

For hash tables. Use VPCONFLICT on ZEN4.

Div (integer division) and Mod

Issue 633. Consider promoting to f64 and back. Or: op to compute inverse.

AddSub and MulAddSub

Subtracts for even lanes, adds for odd. Interval arithmetic? Numpy?

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

  • Signbit
  • ConvertF64<->I32 (math-inl)
  • Copysign (math)
  • CopySignToAbs (math)
  • Neg
  • Compress
  • Mask ops (math)
  • RebindMask
  • Not
  • FP16 conversions
  • Scatter
  • Gather
  • Pause
  • Abs i64
  • FirstN
  • Compare i64
  • AESRound
  • CLMul (GCM)
  • TableLookupBytesOr0 (AES)
  • FindFirstTrue (strlen)
  • NE
  • Combine partial
  • LoadMaskBits (FirstN)
  • MaskedLoad
  • Bf16 promote2
  • ConcatOdd/Even
  • SwapAdjacentBlocks
  • OddEvenBlocks
  • CompressBlendedStore
  • RotateRight (Reverse2 i16)
  • Compare128
  • OrAnd
  • IfNegativeThenElse
  • MulFixedPoint15 (codec)
  • Insert/ExtractLane
  • IsNan
  • IsFinite
  • StoreInterleaved
  • LoadInterleaved (codec)
  • Or3/Xor3
  • NotXor (sort)
  • FindKnownFirstTrue (sort)
  • CompressStore 8-bit
  • ExpandLoad (hash)
  • Zen4 target (sort)
  • SSE2 target - by johnplatts
  • AbsDiff int - by johnplatts
  • Le integer - by johnplatts
  • LeadingZeroCount - by johnplatts in #1276
  • 8-bit Mul
  • (Neg)MulAdd for integer
  • AESRoundInv etc - by johnplatts in #1286
  • OddEven for <64bit lanes: use Set of wider constant 0_1
  • Shl for 8-bit
  • Shr for 8-bit
  • Faster Reverse2 16-bit
  • Add Reverse2 for 8-bit
  • TwoTablesLookupLanes - by johnplatts in #1303
  • Add 8/16-bit TableLookupLanes - by johnplatts in #1303
  • FindLastTrue - by johnplatts in #1308
  • Vec2, Create/Get functions
  • PromoteTo for all types (#915) - by johnplatts in #1387
  • atan2
  • Slide1Up/Down - by johnplatts in #1496
  • MaxOfLanes, MinOfLanes returning scalar
  • Add DupEven for 16-bit - by johnplatts in #1431
  • AVX3_SPR target
  • MaskedGather returns zero for mask=false.
  • GatherIndexN/ScatterIndexN
  • MaskedScatter
  • float64 support for WASM
  • LoadNOr
  • PromoteEvenTo - by johnplatts
  • Masked add/sub/div
  • ReduceMin/Max like ReduceSum, in addition to Min/MaxOfLanes
  • Reductions for 8-bit
  • RVV: Fix remaining 8-bit table lookups for large vectors