fix(deps): update dependencies by tannevaled · Pull Request #106 · go-ndarray/ndarray

tannevaled · 2026-11-08T18:25:46Z

This PR contains the following updates:

Package	Change	Age	Confidence
github.com/go-asmgen/asmgen	`v0.23.0` → `v0.23.2`
github.com/go-fft/fft	`v0.28.11` → `v0.28.15`
github.com/go-ndarray/ndarray	`v0.0.0` → `v0.37.5`
github.com/go-simd/floats	`v0.11.0` → `v0.12.0`

Release Notes

go-asmgen/asmgen (github.com/go-asmgen/asmgen)

`v0.23.2`

Compare Source

What's Changed

amd64: add AVX512DQ feature probe by @tannevaled in #43

Full Changelog: go-asmgen/asmgen@v0.23.1...v0.23.2

`v0.23.1`

Compare Source

What's Changed

ci: move the Go pin from 1.27.2 to 1.27.3 by @tannevaled in #38
go.mod: raise the go directive from 1.27.1 to 1.27.3 by @tannevaled in #40
amd64: add FamilyProbe (CPUID displayed family) by @tannevaled in #42

Full Changelog: go-asmgen/asmgen@v0.23.0...v0.23.1

go-fft/fft (github.com/go-fft/fft)

`v0.28.15`

Compare Source

What's Changed

fix: require AVX512F and AVX512DQ before any AVX-512 kernel by @tannevaled in #113

Full Changelog: go-fft/fft@v0.28.14...v0.28.15

`v0.28.14`

Compare Source

What's Changed

docs(bench): Round 50, the L1D working set does not explain the Zen 4 gap by @tannevaled in #111
perf: Zen 4 runs 65,536-262,144 points on the 512-bit split layout (Round 51) by @tannevaled in #112

Full Changelog: go-fft/fft@v0.28.13...v0.28.14

`v0.28.13`

Compare Source

What's Changed

bench: refbench against scipy.fft 1.18.1 (W3), one core, one row per public function by @tannevaled in #110

Full Changelog: go-fft/fft@v0.28.12...v0.28.13

`v0.28.12`

Compare Source

What's Changed

fix(deps): update dependencies by @tannevaled in #94
docs: NEON radix 16 predicted below the M4 gap, not built (Round 45) by @tannevaled in #100
docs: NEON register-resident leaf predicted below the M4 gap, not built (Round 46) by @tannevaled in #101
docs: NEON FMA kernels predicted below the M4 gap, not built (Round 47) by @tannevaled in #102
docs: document the runtime makeslice panic above maxAlloc by @tannevaled in #103
docs(bench): Cascade Lake parity at v0.28.11 (default and FMA) by @tannevaled in #104
docs(bench): Round 48, scratch gap on Cascade Lake by @tannevaled in #105
perf: Zen 4 takes the AVX2 route up to 65,536 points (Round 49) by @tannevaled in #106

Full Changelog: go-fft/fft@v0.28.11...v0.28.12

go-ndarray/ndarray (github.com/go-ndarray/ndarray)

`v0.37.5`

Compare Source

What's Changed

perf: refbench against NumPy on one core; replace the multicore page by @tannevaled in #105

Full Changelog: v0.37.4...v0.37.5

`v0.37.4`

Compare Source

What's Changed

docs: float64 Exp on 16 cores is slowed by allocation, not the kernel by @tannevaled in #102

Full Changelog: v0.37.3...v0.37.4

`v0.37.3`

Compare Source

What's Changed

floats v0.11.0: SIMD Max/Min restored on amd64 and arm64 by @tannevaled in #101

Full Changelog: v0.37.2...v0.37.3

`v0.37.2`

Compare Source

What's Changed

floats v0.10.0: drop the max/min zero fallback and the moved sum/elementwise assembly by @tannevaled in #100

Full Changelog: v0.37.1...v0.37.2

`v0.37.1`

Compare Source

SumP runs on several cores again, with the bits of floats.SumPairwise (NumPy order) whatever the core count: it cuts the NumPy pairwise tree where NumPy cuts it and adds the subtrees back up the same tree (#99). Results are unchanged from v0.37.0.

`v0.37.0`

Compare Source

K1 of the go-scikit performance campaign (go-scikit/roadmap#27): ndarray imports go-simd/floats v0.9.0.

Changed result: Sum (and mean, var, std built on it) is now NumPy's pairwise sum (floats.SumPairwise), bit-identical to np.sum of a contiguous 1-D float64 array on every architecture. It replaces lane-parallel sums whose last bits differed from NumPy's and between architectures; it is serial at every size.

Exp, Log and the correctly rounded float32 Exp/Log, and on amd64/arm64 the float64 add/sub/mul/div/sqrt/max/min/clip kernels, are go-simd/floats' (the AVX2 and NEON exp/log kernels written here moved down: #94, #95, #96).

`v0.36.0`

Compare Source

What's Changed

AVX2 kernel for float64 Log: 1.55x faster than NumPy by @tannevaled in #96

Full Changelog: v0.35.0...v0.36.0

`v0.35.0`

Compare Source

What's Changed

fix(deps): update dependencies by @tannevaled in #89
docs: design of the dense linear-algebra foundation (scipy.linalg side) by @tannevaled in #92
Float32 results from the Workspace; cheaper range check in the AVX2 log by @tannevaled in #94

Full Changelog: v0.34.0...v0.35.0

`v0.34.0`

Compare Source

What's Changed

float32 Log: correctly rounded (CORE-MATH logf) with an AVX2 kernel at NumPy parity by @tannevaled in #91

Full Changelog: v0.33.0...v0.34.0

`v0.33.0`

Compare Source

What's Changed

docs: coverage inventory vs NumPy, scipy.linalg and MATLAB by @tannevaled in #88
perf: float32 Exp in AVX2, correctly rounded and as fast as NumPy by @tannevaled in #90

Full Changelog: v0.32.0...v0.33.0

`v0.32.0`

Compare Source

What's Changed

fix: float32 Exp and Log correctly rounded, proven over every float32 by @tannevaled in #87

Full Changelog: v0.31.0...v0.32.0

`v0.31.0`

Compare Source

What's Changed

docs: N1 diagnosis — fresh quiet-host ratios, counters, causes ruled out by @tannevaled in #85
perf: float32 Exp and Log on the float64 kernels (2-3.5x faster) by @tannevaled in #86

Full Changelog: v0.30.0...v0.31.0

`v0.30.0`

Compare Source

What's Changed

docs: products slow down on a busy N1 host, OpenBLAS does not (open defect) by @tannevaled in #83
chore(deps): update dependency go to v1.27.3 by @tannevaled in #60
feat(random): allocation-free scalar draws (Float64, FillUniform, IntN, FillIntN, ShuffleInt64) by @tannevaled in #84

Full Changelog: v0.29.1...v0.30.0

`v0.29.1`

Compare Source

What's Changed

perf(gemm): read a row-major B in place on the small-product path by @tannevaled in #82

Full Changelog: v0.29.0...v0.29.1

`v0.29.0`

Compare Source

What's Changed

fix(deps): update dependencies by @tannevaled in #73
perf(gemm): native float32 GEMM on s390x by @tannevaled in #79

Full Changelog: v0.28.0...v0.29.0

`v0.28.0`

Compare Source

What's Changed

perf(gemm): native float32 GEMM on ppc64le by @tannevaled in #78

Full Changelog: v0.27.0...v0.28.0

`v0.27.0`

Compare Source

What's Changed

perf(gemm): read a row-major A in place (BLIS sup kernels) by @tannevaled in #77

Full Changelog: v0.26.1...v0.27.0

`v0.26.1`

Compare Source

What's Changed

perf(gemm): pack full B panels with whole-row array moves by @tannevaled in #75

Full Changelog: v0.26.0...v0.26.1

`v0.26.0`

Compare Source

What's Changed

docs: the allocating operations' cost, measured and split by @tannevaled in #72
perf(linalg): native float32 GEMM on amd64 and arm64 by @tannevaled in #74

Full Changelog: v0.25.0...v0.26.0

`v0.25.0`

Compare Source

What's Changed

perf: loong64 Max/Min (float64, float32) in LSX integer arithmetic by @tannevaled in #71

Full Changelog: v0.24.0...v0.25.0

`v0.24.0`

Compare Source

What's Changed

perf: ppc64le Max/Min (float64, float32) with an integer NaN mask by @tannevaled in #70

Full Changelog: v0.23.0...v0.24.0

`v0.23.0`

Compare Source

What's Changed

ci: run the loong64 qemu lane with LASX, and require it by @tannevaled in #67
ci: run the s390x qemu lane with VXE, and require it by @tannevaled in #68
perf: s390x float32 Max/Min (VFMAXSB, z14+) by @tannevaled in #69

Full Changelog: v0.22.0...v0.23.0

`v0.22.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-26.

Commits since v0.21.0:

151c4c1 perf: RVV Max/Min (float64, float32); fix a small-n RVV sum; run RVV in CI (#66)

`v0.21.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-26.

Commits since v0.20.0:

543d1d0 perf: float32 MatMul on the float64 GEMM (#65)

`v0.20.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-26.

Commits since v0.19.1:

f4ab9ab perf: float32 Max/Min read float32 directly; SSE2 and NEON kernels (#64)

`v0.19.1`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-26.

Commits since v0.19.0:

5759867 ci: install qemu-user, not the virtual qemu-user-static (Ubuntu 26.04) (#63)
725ee5d fix: amd64 Max/Min over signed zeros depended on element order (#62)

`v0.19.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-25.

Commits since v0.18.0:

2432766 fix(deps): update dependencies (#57)
8388dc6 perf: float32 Sqrt kernels on six targets; float32 survey against NumPy (#61)

`v0.18.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-19.

Commits since v0.17.0:

8979001 perf: float32 Add/Sub/Mul/Div kernels on six targets; *Into writes typed results in place (#59)

`v0.17.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-18.

Commits since v0.16.0:

0906fd8 perf: float32 pairwise sum blocks on s390x (VFASB, z14+) (#58)

`v0.16.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-18.

Commits since v0.15.0:

cfbca31 perf: float32 pairwise sum blocks on VSX (ppc64le) (#56)

`v0.15.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-18.

Commits since v0.14.0:

365f1d1 perf: float32 pairwise sum blocks on RVV (riscv64) and LSX (loong64) (#55)

`v0.14.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-18.

Commits since v0.13.0:

8bd10b2 perf: narrow integer sums widen as they read, without an int64 copy (#52)
1af734c ci: an A/B benchmark workflow on the hosted amd64 and arm64 runners (#53)
c44fc7b perf: float32 sum blocks on SSE2 and NEON; integer widening folds without bounds checks (#54)

`v0.13.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-18.

Commits since v0.12.0:

7a0a00c fix(deps): update module github.com/go-fft/fft to v0.27.1 (#48)
a1f82eb test: FuzzMixedDTypes, the values of mixed-dtype operations (#50)
0659ebc perf: fold contiguous reductions directly; NumPy's pairwise sum for float32/complex (#51)

`v0.12.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-11.

Commits since v0.11.0:

575b6f3 test: FuzzUnmarshalBinary; audit entry for the decoder (#47)
b41dd2d feat: Repr prints exactly what NumPy 2.5.3's repr prints (#49)

`v0.11.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-10.

Commits since v0.10.0:

4057f6e feat(linalg): numpy.linalg in pure Go (#46)

`v0.10.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-10.

Commits since v0.9.1:

844966e docs(README): riscv64 and z17 results in the performance summary (#40)
9b4813d docs(README): docs badge says Hugo + Hextra (the site left MkDocs) (#41)
3217f28 feat: thirteen dtypes with NumPy 2.x promotion (#42)
90c2da1 feat(fft): numpy.fft on arrays, over go-fft (#43)
d032aeb feat: MarshalBinary/UnmarshalBinary, NumPy-style Repr, notebook Display (#44)
17739d2 feat(random): numpy.random with NumPy's numbers for the same seed (#45)

`v0.9.1`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-09.

Commits since v0.9.0:

dcf4f7a simd(s390x): VFMAXDB/VFMINDB max/min (Java mode) on z14+, measured on a z17 (#39)

`v0.9.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-09.

Commits since v0.8.0:

d699b21 simd(s390x): 4x8 GEMM micro-kernel; measured on a real IBM z17; Go floor 1.27.2; playground links; correct the loong64 1.27.2 claim (#38)

`v0.8.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-08.

Commits since v0.7.1:

6c851d5 simd(s390x): vector-facility kernels for sum, dot, sqrt, add/sub/mul/div; native-endian AT_HWCAP (#37)

`v0.7.1`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-08.

Commits since v0.7.0:

c71adc5 simd(riscv64): 6x8 RVV GEMM micro-kernel (LMUL=4, vl=8 on any VLEN) (#36)

`v0.7.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-08.

Commits since v0.6.1:

266c0c8 test: parallel stress of the helper pool; docs: Zen 3 re-measurement (#34)
67bc4f9 simd(riscv64): RVV kernels for sum, dot, sqrt and add/sub/mul/div (#35)

`v0.6.1`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-07.

Commits since v0.6.0:

cddfe75 docs(README): mark the Dot vs threaded-BLAS figure as pre-v0.6.0 (#32)
f3fc70b fix(pool): do not retain the last job's arrays; stop helpers before a caller panic propagates (#33)

`v0.6.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-07.

Commits since v0.5.2:

0241ff3 perf(parallel): persistent helper pool for parallelFor; finer reduction partials (#31)

`v0.5.2`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-07.

Commits since v0.5.1:

cbe4aea arm64: NEON div and max/min, Go 1.27 mnemonics for sqrt/add/sub/mul; parallelFor findings (#30)

`v0.5.1`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-07.

Commits since v0.5.0:

bea7aa9 deps: go-asmgen v0.16.0, regenerated (comments only); README Go badge 1.27.1; loong64 go1.27.1 audit entry (#29)

`v0.5.0`

Compare Source

Release backfilled on 2026-10-26; tag created 2026-10-05.

Commits since v0.4.1:

09d8cd6 build: raise the Go floor to 1.27.1 (#27)