diff --git a/README.md b/README.md index 8b49b41..51947d4 100644 --- a/README.md +++ b/README.md @@ -19,7 +19,7 @@ describes**: | Source | Published under | | --- | --- | -| branch `main`, with `params.ndarray.version: v0.29.1` in `hugo.yaml` | `https://go-ndarray.github.io/docs//`, here `/docs/0.29/` | +| branch `main`, with `params.ndarray.version: v0.31.0` in `hugo.yaml` | `https://go-ndarray.github.io/docs//`, here `/docs/0.31/` | | the newest version | also `https://go-ndarray.github.io/docs/latest/` | `https://go-ndarray.github.io/docs/` redirects to `latest/`. Each push to diff --git a/content/en/dtypes.md b/content/en/dtypes.md index 84c689d..63cb85f 100644 --- a/content/en/dtypes.md +++ b/content/en/dtypes.md @@ -8,7 +8,7 @@ description: "Why the dtype is a run-time property of one Array type, how arrays tags: [dtypes, numpy] --- -> Mirrored from [`docs/design-dtypes.md`](https://github.com/go-ndarray/ndarray/blob/v0.30.0/docs/design-dtypes.md) at v0.30.0. +> Mirrored from [`docs/design-dtypes.md`](https://github.com/go-ndarray/ndarray/blob/v0.31.0/docs/design-dtypes.md) at v0.31.0. > Status: implemented (Phase 2). Supersedes decision 4 of > [plan-ndarray.md](roadmap.md) ("`float64` first"). diff --git a/content/en/linalg.md b/content/en/linalg.md index f0804f3..2619bbf 100644 --- a/content/en/linalg.md +++ b/content/en/linalg.md @@ -8,9 +8,9 @@ description: "Solve, Inv, Det, Cholesky, QR, SVD, Eig and the rest of numpy.lina tags: [linalg, numpy] --- -> Mirrored from [`docs/linalg.md`](https://github.com/go-ndarray/ndarray/blob/v0.30.0/docs/linalg.md) at v0.30.0. +> Mirrored from [`docs/linalg.md`](https://github.com/go-ndarray/ndarray/blob/v0.31.0/docs/linalg.md) at v0.31.0. -Package [`linalg`](https://pkg.go.dev/github.com/go-ndarray/ndarray@v0.30.0/linalg) gives ndarray arrays the functions of +Package [`linalg`](https://pkg.go.dev/github.com/go-ndarray/ndarray@v0.31.0/linalg) gives ndarray arrays the functions of `numpy.linalg`, written in Go without cgo or assembly of its own: the matrix products it relies on are the ndarray package's packed SIMD GEMM. diff --git a/content/en/performance.md b/content/en/performance.md index fc0cd5a..cf00d5a 100644 --- a/content/en/performance.md +++ b/content/en/performance.md @@ -1403,9 +1403,46 @@ Ruled out on that host, by A/B: nothing, and NC = 128 or 256 (a B block well inside the 1 MiB L2) moves float64 by 4–9% but float32 1024² by +4%. -The cause is not found. The quiet-host ratios above stand for a quiet host, -and this one for a contended one. The exception is float64 at 1024² on Zen 3, where we are 3% faster. -The float32 kernels for riscv64 and loong64 are still to write. +On the same host quiet again (load 2, 2026-11-01), measured back to back +with NumPy on the same pinned core, the ratio is: + +| N1, load 2, ours / NumPy | 64² | 256² | 1024² | +|---|---:|---:|---:| +| float64 | 1.09 | 1.07 | 1.16 | +| float32 | 1.08 | **1.00** | 1.07 | + +Our float64 1024² went from 192 ms busy to 137 ms quiet, and NumPy's from +124 to 118 ms. So the defect is the sensitivity to a busy host, not the +quiet-host speed. + +Hardware counters (`perf stat`, `perf record -e ll_cache_miss_rd`) locate +it. Per product, at about the same instruction count, we have 4.5× the +L1D refills, 1.6× the L2 refills and 1.7× the last-level read misses of +OpenBLAS (IPC 1.73 against 2.09). The last level is shared, so contention +there costs us more. The misses fall in the micro-kernel: 36.7M in +`gemmUA4x8D`, against 23.4M in OpenBLAS's kernel plus 2.1M in its A +packing and 1.7M in its B packing (about 10 products each). + +Six candidate causes were tested by A/B with counters, and none closes the +gap: + +- a regression; +- reading A in place; +- block sizes, alone (KC 320, MC 240; NC 128, 256, 4096); +- writing C (a fresh, zeroed C against a reused one: misses ±1%); +- software prefetch after `dgemm_kernel_8x4.S` (L2 refills 14.1M → 6.2M, + but last-level misses unchanged and the times worse); +- OpenBLAS's whole structure at once: packed A plus KC 320, MC 240, + NC 4096 (last-level misses 56M against OpenBLAS's 31M). + +What is left is how the micro-kernel itself touches memory. Closing it +would mean rewriting the N1 kernel after OpenBLAS's instruction by +instruction. + +On Zen 3, by contrast, float64 at 1024² is 3% faster than NumPy. + +The float32 kernels for riscv64 and loong64 are written (PRs #80 and #81) +but not yet measured on hardware. **Open: the allocating operations.** `Add` is still 1.5–3.2× slower than NumPy's `a+b`, for float64 as well. The kernel is no longer the difference, @@ -1432,6 +1469,35 @@ What does remove it: - **Tune the GC:** `GOGC` or `GOMEMLIMIT` for a program dominated by short-lived results. +### float32 `Exp` and `Log` on the float64 kernels (2026-11-01) + +float32 `Exp` and `Log` called `math.Exp` and `math.Log` per element: 6–14× +slower than NumPy's float32 SIMD on one Zen 3 core. They now run this +package's float64 kernels (the ports of Arm's optimized-routines, 0.504 ULP +worst case in float64). Elements go 256 at a time through a stack buffer: +widened (exact), computed in place, rounded once. Nothing is allocated. The +result is within 0.5 + 2⁻²⁹ float32 ULP, so correctly rounded except +next to a rounding midpoint. Overflow and underflow come from the rounding: +exp(89) is finite in float64 and rounds to +Inf. + +| float32, one Zen 3 core (`abbench`, 8 rounds) | before | after | NumPy 2.5.4 | after / NumPy | +|---|---:|---:|---:|---:| +| `Exp`, 16 Ki | 297 µs | 84 µs | 26 µs | 3.2 | +| `Log`, 16 Ki | 198 µs | 99 µs | 32 µs | 3.1 | +| `Exp`, 1 Mi | 18.8 ms | 5.2 ms | 1.73 ms | 3.0 | +| `Log`, 1 Mi | 12.4 ms | 6.1 ms | 2.02 ms | 3.0 | + +The tests hold the results to `float32(math.Exp/Log(x))`, within one ULP and +equal on at least 99.9% of 100 000 inputs. They also check the edges +(±0, ±Inf, NaN, overflow, the subnormals, log of a negative), every length +across the chunk boundaries, in-place use and the parallel split. + +**Open defect: still 3× slower than NumPy.** NumPy computes float32 exp and +log in float32 SIMD (`simd_exp_f32`/`simd_log_f32` in +`loops_exponent_log.dispatch.c.src`). Porting those kernels per architecture +is the next step. float64 `Log` is also 1.23× slower than NumPy at 16 Ki on +Zen 3 (85 µs against 69 µs), while float64 `Exp` is 16% faster. + ## SIMD coverage - **amd64 (SSE2)** ships hand-vectorized `sum` (4-accumulator `ADDPD`), `sqrt` diff --git a/content/en/random.md b/content/en/random.md index 73f4d53..f0bd869 100644 --- a/content/en/random.md +++ b/content/en/random.md @@ -8,9 +8,9 @@ description: "numpy.random with NumPy's numbers for the same seed: the bit gener tags: [random, numpy] --- -> Mirrored from [`docs/random.md`](https://github.com/go-ndarray/ndarray/blob/v0.30.0/docs/random.md) at v0.30.0. +> Mirrored from [`docs/random.md`](https://github.com/go-ndarray/ndarray/blob/v0.31.0/docs/random.md) at v0.31.0. -Package [`random`](https://pkg.go.dev/github.com/go-ndarray/ndarray@v0.30.0/random) is `numpy.random` over go-ndarray arrays. Its goal +Package [`random`](https://pkg.go.dev/github.com/go-ndarray/ndarray@v0.31.0/random) is `numpy.random` over go-ndarray arrays. Its goal is migration: a notebook moved from NumPy to Go keeps its numbers for the same seed. @@ -24,7 +24,7 @@ u, _ := random.Rand(3) // np.random.rand(3) ``` `Example`, `ExampleSeed` and `ExampleGenerator_ChoiceN` in -[`example_test.go`](https://github.com/go-ndarray/ndarray/blob/v0.30.0/random/example_test.go) print NumPy's own numbers. +[`example_test.go`](https://github.com/go-ndarray/ndarray/blob/v0.31.0/random/example_test.go) print NumPy's own numbers. ## What is implemented @@ -142,7 +142,7 @@ random arguments. Used as-is, that made 3% of `np.random.randn` values and 10% of `lognormal` values one ULP off. So `random` rounds `log`, `log1p`, `exp`, `expm1` and `pow` **correctly** -([`libm.go`](https://github.com/go-ndarray/ndarray/blob/v0.30.0/random/libm.go)): a fast evaluation to about 2⁻⁶⁸ with Ziv's +([`libm.go`](https://github.com/go-ndarray/ndarray/blob/v0.31.0/random/libm.go)): a fast evaluation to about 2⁻⁶⁸ with Ziv's rounding test, and a double-double fallback (about 2⁻¹⁰⁰) for the one argument in ~2¹⁵ that needs it. Checked against mpmath at 300 bits on 63,000 arguments (0 misrounded; the committed `libm.json` keeps 8,400), and the fast paths diff --git a/content/en/roadmap.md b/content/en/roadmap.md index a68a440..1389333 100644 --- a/content/en/roadmap.md +++ b/content/en/roadmap.md @@ -8,7 +8,7 @@ description: "The plan for go-ndarray and what ships today." tags: [roadmap] --- -> Mirrored from [`docs/plan-ndarray.md`](https://github.com/go-ndarray/ndarray/blob/v0.30.0/docs/plan-ndarray.md) at v0.30.0. +> Mirrored from [`docs/plan-ndarray.md`](https://github.com/go-ndarray/ndarray/blob/v0.31.0/docs/plan-ndarray.md) at v0.31.0. > Goal: a **pure-Go (CGO=0)** NumPy-style N-dimensional array library, as a > **standalone, reusable** module with correct scalar kernels now and a clean @@ -112,7 +112,7 @@ suite (plus the per-arch CI jobs already wired in `.github/workflows/ci.yml`). - **Performance — DONE (beats NumPy on its core ops; matmul at tuned-BLAS parity).** Several levers behind the kernel seam, measured honestly vs NumPy 2.2.4 (see [docs/perf.md](performance.md) and the repo's - [BENCHMARKS.md](https://github.com/go-ndarray/ndarray/blob/v0.30.0/BENCHMARKS.md)): + [BENCHMARKS.md](https://github.com/go-ndarray/ndarray/blob/v0.31.0/BENCHMARKS.md)): - **Multicore.** `Add/Sub/Mul/Div/Map`, `Sum/Max/Min`, the axis reductions, `Concatenate`/`Stack`, and the GEMM fan out across `GOMAXPROCS` above a size threshold (small arrays stay serial). Plus an allocation fix: the @@ -168,7 +168,7 @@ suite (plus the per-arch CI jobs already wired in `.github/workflows/ci.yml`). SIMD kernels on all six 64-bit targets for `Sum` (v0.14.0–v0.17.0) and for `Add`/`Sub`/`Mul`/`Div` (v0.18.0) and `Sqrt` (v0.19.0). Still pure Go for float32: nothing in `Max`/`Min` any more (v0.20.0 to v0.25.0 - vectorised them on all six 64-bit targets, for float32 and float64), and the transcendental ufuncs. `MatMul` runs on the float64 + vectorised them on all six 64-bit targets, for float32 and float64), and the transcendental ufuncs (float32 `Exp`/`Log` run the float64 kernels since v0.31.0, 2–3.5× faster but still 3× behind NumPy's float32 SIMD). `MatMul` runs on the float64 GEMM since v0.21.0, 77–90% faster than the triple loop, and since v0.26.0 on a native float32 GEMM on arm64 (NEON 8×8) and amd64 with FMA (6×16), a further 53–61% faster, and since v0.28.0 on ppc64le (VSX 8×16, 37–59% diff --git a/content/en/security.md b/content/en/security.md index 976c100..40052e7 100644 --- a/content/en/security.md +++ b/content/en/security.md @@ -8,7 +8,7 @@ description: "How to report a vulnerability, what the library guarantees, and th tags: [security] --- -> Mirrored from [`SECURITY.md`](https://github.com/go-ndarray/ndarray/blob/main/SECURITY.md) at v0.29.1. +> Mirrored from [`SECURITY.md`](https://github.com/go-ndarray/ndarray/blob/main/SECURITY.md) at v0.31.0. ## Reporting @@ -76,6 +76,12 @@ standard library's known vulnerabilities; `govulncheck ./...` reports none. ## Audit log +**2026-11-01 (v0.31.0).** float32 `Exp` and `Log` widen their input 256 +elements at a time into a fixed stack buffer and call the float64 kernels in +place. There is no allocation proportional to the input beyond the result, +and no index arithmetic beyond slices bounded by `len(dst)`. The source is +resliced to `len(dst)`, so a shorter source panics instead of being read past. + **2026-10-31 (v0.29.1).** The A-in-place GEMM kernels now take B's row stride (`ldb`) and may read a row-major B operand in place: `kc` rows `ldb` elements apart, `NR` elements each. The wrappers bounds-check diff --git a/hugo.yaml b/hugo.yaml index 3be1cf1..1926add 100644 --- a/hugo.yaml +++ b/hugo.yaml @@ -6,7 +6,7 @@ # have used since the site was built with MkDocs and mike. The CI sets it # through --baseURL, from params.ndarray.version below. This default is the # same build. -baseURL: "https://go-ndarray.github.io/docs/0.29/" +baseURL: "https://go-ndarray.github.io/docs/0.31/" title: "go-ndarray" # themes/hextra is the tannevaled/hextra fork, pinned at a release tag: # upstream Hextra plus opt-in features (partial menu items, combos, @@ -100,7 +100,7 @@ params: # The ndarray release these pages describe. Its minor version is the # documentation's version label (v0.9.1 -> 0.9): the CI reads it here. ndarray: - version: "v0.30.0" + version: "v0.31.0" navbar: displayTitle: true