Up to [cvs.NetBSD.org] / pkgsrc / graphics / libhighway
Request diff between arbitrary revisions
Keyword substitution: kv
Default branch: MAIN
libhighway: update to 1.4.0.
Add:
Fast* math functions, sum_array example
HWY_ARCH_MAX_BYTES, HWY_MIN_BYTES, HWY_NATIVE_MASK, HWY_REGISTERS
HWY_EXPORT_AND_TEST_BEST_P
InterleaveLower/UpperBlocks, Lookup8, XorAndNot
MinMax algo, AtomicBitSet
RVV and LSX/LASX runtime dispatch, FreeBSD futex
Improvements:
MulByPow2, PopulationCount, SumsOfAdjQuadAbsDiff
ReorderWidenMulAccumulate, SumOfMulQuadAccumulate
Re-enable SVE, add i8mm for SVE/NEON_BF16
Fixes:
EVEX512 compiler change workaround, Timer::start() result truncation
BF16 dot on SVE, doc formatting, StringTable race, warnings
libhighway: update to 1.3.0.
Add:
AddLower, PairwiseAdd/Sub, MaskedAbsOr, BitsFromMask
AVX10_2 and Loongson LASX/LSX targets
AVX3_SPR F16, WASM_EMU256 F64 types
CeilInt/FloorInt, DemoteToNearestInt and F16/F64 NearestInt
Complex number operations, F16/BF16 assignment operators
emulated bf16/f16 Load/StoreInterleaved
hwy::Warn/HWY_WARN, use instead of fprintf
HWY_UNREACHABLE, HWY_VISIT_TARGETS
i16 Dot, AverageRound, RoundingShiftRight/RoundingShr
InterleaveEvenBlocks/InterleaveOddBlocks, MinMagnitude/MaxMagnitude
masked comparisons, promote, round, GetBiasedExponent
MulByPow2/MulByFloorPow2, MulRound, MulLower/MulAddLower
PositiveInfOrHighestValue/NegativeInfOrLowestValue
RVV groundwork for runtime dispatch, enable tuples
spin wait, NanoSleep, Counter2/4 barrier, Divisor64, perf_counters
Improvements:
dpbf16 WidenMulPairwiseAdd Exp2, AVX10.2 float->int, AVX3 GetExponent
header-only abort.h/cc, tests runnable with Bazel8
HWY_BROKEN_*: allow individual override
Lanes: 'optional constexpr', AllBits1
MaskedEq/Ne, NEON SumOfMulQuadAccumulate, MaskedReduceMin/Max, MulEven
Profiler: report concurrency stats, 1.36x less overhead
RVV various ops via superoptimizer
SetThreadName: support more systems
SVE2 SatWidenMulPairwiseAccumulate, SSE2/SSSE3 U16 Min/Max
TargetName: no longer returns unknown for other arch
ThreadPool autotune, avoid WakeAll
topology: add NUMA node, support Windows/Apple
Fixes:
avoid wraparound for -ftrapv, topology for offline CPUs/RVV
warnings from -Wmissing-declarations/prototypes
AdvSIMD_HPFPCvt on OSX
f32->bf16 rounding: avoid unspecified built-in cast
MSAN, PPC InvariantTicksPerSecond on QEMU, HWY_RCAST_ALIGNED, IsNaN
vqsort for ascending order, add 8-bit test
libhighway: use atomic64.mk instead of unconditional libatomic dependency
Make graphics/libhighway build on Darwin by not depending on unneeded libatomic
libhighway: update to 1.2.0.
Convert to cmake/build.mk.
1.2.0
Add InterleaveEven/InterleaveOdd, BitShuffle, GatherIndexNOr
Add IsNegative, IfNegativeThenElseZero, IfNegativeThenZeroElse
Add NEON_BF16, HWY_VERSION_GE/LT, HWY_EXPORT_T/HWY_DYNAMIC_DISPATCH_T
Add PromoteInRangeTo/ConvertInRangeTo/DemoteInRangeTo
Add Rol/Ror, RotateLeft/RotateLeftSame/RotateRightSame
Add SatWidenMulPairwiseAccumulate, SatWidenMulAccumFixedPoint
Add stats.h, bit_set.h, IsEitherNaN
Add UI8/UI32/UI64 MulHigh, I64 MulEven/MulOdd/Mul128
Add WidenMulAccumulate, MulEvenAdd, MulOddAdd
contrib/bit_pack: support 32/64-bit lanes
contrib/math: Add Exp2, Hypot
contrib/matvec: Add MatVecAdd
contrib/sort: Add VQ/HeapSelect, partial sort
contrib/topology: add affinity, detect topology/cache size/CPU name
Enable runtime dispatch for NEON/RVV, bazel modules, abort handler
Remove DASSERT for negative Gather indices
Support opting out of GUnit dependency
Use SPR/ZEN4 bf16 dot product
Known GCC 13 RVV issue: parts of sort_test and bit_pack_test disabled
Known Clang RVV/QEMU issue: incorrect rounding mode in upper/lower halves
1.1.0
Add BitCastScalar, DispatchedTarget, Foreach
Add Div/Mod and MaskedDiv/ModOr, SaturatedAbs, SaturatedNeg
Add InterleaveWholeLower/Upper, Dup128VecFromValues
Add IsInteger, IsIntegerLaneType, RemoveVolatile, RemoveCvRef
Add MaskedAdd/Sub/Mul/Div/Gather/Min/Max/SatAdd/SatSubOr
Add MaskFalse, IfNegativeThenNegOrUndefIfZero, PromoteEven/OddTo
Add ReduceMin/Max, 8-bit reductions, f16 <-> f64 conversions
Add Span, AlignedArray, matrix-vector mul
Add SumsOf2/4, I8 SumsOf8, SumsOfAdjQuadAbsDiff, SumsOfShuffledQuadAbsDiff
Add ThreadPool, hierarchical profiler
Build: use bazel_platforms
Enable clang16 Arm/PPC runtime dispatch, F16 for GCC AVX3_SPR
Extend Dot to f32*bf16, FMA to integer
Fix: RVV 8-bit overflow, UB in vqsort, big-endian bugs, PPC HTM
Improved codegen in various ops, fp16/bf16 tests and conversions
New targets: HWY_Z14, HWY_Z15
Test: add foreign_arch builders, CodeQL
*: replace CMAKE_ARGS with CMAKE_CONFIGURE_ARGS
Update libhighway to version 1.0.7 The reason for updating is that the only libhighway user (wip/libjxl) needs 1.0.7 to build its latests version 0.9.0. Due to that nothing in pkgsrc-current depends on libhighway and the only thing that needs it needs 1.0.7, bump BUILDLINK_API_DEPENDS to 1.0.7 .
libhighway: Option-ize googletest dependency for ABI reasons (same as other packages, it's a dependency that might require a much newer compiler) While here, add missing libatomic dependency
*: Re-apply SunOS linker argument removals. These are currently duplicated in mk/platform/SunOS.mk but the generic removals will be removed soon in favour of per-package removals, due to flags getting leaked into installed files.
Update libhighway to version 1.0.1 Pkgsrc changes: Build shared libs. Use gcc 8 and over to fix build. Include devel/googletest in buildlink3.mk since one of the installed header files include gtest/gtest.h . Changes from changelog: * Add Eq128, i64 Mul, unsigned->float ConvertTo * Faster sort for few unique keys, more robust pivot selection * Fix: floating-point generator for sort tests, Min/MaxOfLanes for i16 * Fix: avoid always_inline in debug, link atomic * GCC warnings: string.h, maybe-uninitialized, ignored-attributes * GCC warnings: preprocessor int overflow, spurious use-after-free/overflow * Doc: <=HWY_AVX3, Full32/64/128, how to use generic-inl * ABI change: 64-bit target values, more room for expansion * Add CompressBlocksNot, CompressNot, Lt128Upper, Min/Max128Upper, TruncateTo * Add HWY_SVE2_128 target * Sort speedups especially for 128-bit * Documentation clarifications * Faster NEON CountTrue/FindFirstTrue/AllFalse/AllTrue * Improved SVE codegen * Fix u16x8 ConcatEven/Odd, SSSE3 i64 Lt * MSVC 2017 workarounds * Support for runtime dispatch on Arm/GCC/Linux
graphics/libhighway: import libhighway-0.17.0 Renamed to libhighway due to textproc/highway. Highway is a C++ library for SIMD (Single Instruction, Multiple Data), i.e. applying the same operation to 'lanes'. ## Why Highway? - more portable (same source code) than platform-specific intrinsics, - works on a wider range of compilers than compiler-specific vector extensions, - more dependable than autovectorization, - easier to write/maintain than assembly language, - supports runtime dispatch, - supports variable-length vector architectures. ## Current status Supported targets: scalar, SSE4, AVX2, AVX-512, NEON (ARMv7 and v8), WASM SIMD. Ports to RVV and SVE/SVE2 are in progress.