Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 10 additions & 7 deletions docs/performance-campaign.md
Original file line number Diff line number Diff line change
Expand Up @@ -105,13 +105,13 @@ What the count is not:
|---|---|---|---|---|---|
| go-dsp/signal | `perf_test.go`, `docs/perf.md` | yes | yes | checked before each case (< 8) | best of 5; the model to generalise |
| go-statistics/stats | `perf_scipy_test.go`, `perf_test.go`, `perf_all_test.go`, `docs/perf.md` | yes (+ all-cores column) | yes, case by case | load in the table header only | median of 7; SciPy through `uv`, one JSON line per request; 355 cases, **236 slower than SciPy**, 51 of them because of go-numerics/special and 25 because of go-ndarray's generator (its `docs/perf.md`) |
| go-numerics/special | `nd/perf_test.go`, `bench/scipy_server.py`, `docs/perf.md` | scalar column yes | yes | load in the header (7.67 at the last run) | the Go scalar loop is **0.19× to 0.40×** SciPy's speed on every row read (Gamma … ErfInv) |
| go-numerics/special | `nd/perf_test.go`, `bench/scipy_server.py`, `docs/perf.md` | scalar column yes | yes | load in the header (7.67 at the last run) | the Go scalar loop is **0.07× to 0.70×** SciPy's speed on all 31 rows (BetaInc and Kv 0.07×, Zeta 0.70×), attributed by the page to double-double arithmetic (go-numerics/special#33) |
| go-ndarray/ndarray | `docs/perf.md`, `benchmarks/` (`bench_numpy.py`, `gen_report.sh`, `merge_report.py`) | **no**: multicore Go against single-threaded NumPy | — | — | NumPy 2.2.4, not the pinned 2.5.x; shell and Python report tooling |
| go-fft/fft | `docs/perf.md`, `BENCHMARKS.md`, `benchmarks/run.sh`, `report.py`, `cbench/` | — | — | — | FFTW native, numpy, scipy, gonum; shell and Python report tooling |
| go-simd/floats | `bench/numpycmp` (Go, cobra) | yes | yes, per row | refuses at ≥ 8, **once at start** | mean vs best-of-5 (rule 3) |
| go-geometry/predicates | `bench/` module (cgo + `predicates.c`) | yes | C and Go sub-benchmarks side by side in one benchmark | none | gap 1.4–3.6× (issue #4) |
| go-geometry/predicates | `bench/` module (cgo + `predicates.c`) | yes | C and Go sub-benchmarks side by side in one benchmark | none | 1.6–3.3× slower on random records, 1.6–6.9× in the exact stage D, with allocations there; leads: the range guard and `[3]float64` arguments on the fast path, the generic expansion determinant in stages B/D (go-geometry/predicates#4) |
| go-numerics/optimize, integrate | `benchmarks/run.sh` + `bench_scipy.py` + `summarize.py`; `bench/compare.sh` + `scipy_bench.py`; `tools/*bench` | not checked | not checked | not checked | shell and Python drivers |
| go-sparse/sparse | `cmd/perfbench`, `ldl/perf_test.go` | not opened | | | AMD-ordered refactor 11–21 % slower than QDLDL (issue #10) |
| go-sparse/sparse | `cmd/perfbench`, `ldl/perf_test.go` | not opened | | | `ldl` refactor 11–21 % and solve up to 13 % slower than QDLDL v0.1.9 (about 10 instructions per scatter entry against 7, about 40 bounds checks on data-dependent indices); `order.AMD` and the symbolic phase allocate per node; together they make optimize's ADMM setup 2.2–4.9× OSQP's (go-sparse/sparse#10, measured with optimize `tools/qpbench`) |
| go-approx/interp, poly; go-bignum/bigfloat | Go benchmarks; `bigfloat/cmd/bench` + `oracle/mpmath.go` | not opened beyond the count | | | |

Five repositories drive their comparison with shell or Python scripts. Those scripts are the
Expand Down Expand Up @@ -238,10 +238,10 @@ Not kernels, but the same rule:
`go-statistics/stats/internal/dd` have the same package comment and the same `zz_exptable.go`
and `zz_logtable.go`. stats already depends on special, so stats imports `special/dd` and
deletes its copy.
- **Scatter-add** (the `y[idx[k]] -= a·val[k]` of sparse LDL, where QDLDL leads by 11–21 %) has no
- **Scatter-add** (the `y[idx[k]] -= a·val[k]` of sparse LDL, where QDLDL's refactor leads by 11–21 %) has no
vector form on AVX2 or NEON, which have gathers and no scatter (go-simd/histogram's README
works through the same limit). The QDLDL gap is algorithmic first.
- **special's 0.19×–0.40× rows** are scalar against scalar: SciPy's ufunc loops are one-thread C
- **special's 0.07×–0.70× rows** are scalar against scalar: SciPy's ufunc loops are one-thread C
calling scalar code too. They are profiled before any kernel is written (the cost of
`math.Exp`/`math.Log` against the C library's, bounds checks, coefficient loops); K1 and K2 are
what follows if the profile points at elementary functions or polynomial evaluation.
Expand Down Expand Up @@ -309,5 +309,8 @@ Opened for this page, at these commits:
- The inventory also read integrate `d4587b4`, models `62a5776`, interp `23a47aa`, bigfloat
`6cd7f4c` through `tools/perfinventory`.

Issues go-geometry/predicates#4 and go-sparse/sparse#10 are cited from the campaign brief; the
GitHub API timed out when they were opened for this page.
- Issues go-geometry/predicates#4 and go-sparse/sparse#10 (read 2026-11-07).

Filed from this page: go-numerics/special#32 (one double-double package), go-numerics/special#33
(single-core gap), go-ndarray/ndarray#93 (perf page not on one core), go-scikit/roadmap#27 (W0,
`refbench`).
Loading