Skip to content

PERFORMANCE.md

Latest commit

 

History

History
91 lines (62 loc) · 5.66 KB

PERFORMANCE.md

File metadata and controls

91 lines (62 loc) · 5.66 KB
Raw

Performance — dev micro-opts

All opts are [[likely]] guarded with fallback; 49/49 tests still pass. Bench with bench_math (AVX) and ctest --verbose.

1. ndarray hot paths — ndarray.hpp

Function Before (main) After (dev) Gain
is_contiguous:3116 strides != _c_strides(shape) → vector alloc per call exp loop, no alloc, [[unlikely]] ~10× for copyto guard
_flat_logical:3479 _shape_u() alloc + Odometer is_contiguous() [[likely]] → offset+i; else flat+=coord*strides no alloc ~5× for random fill
at(i):3315 / at(i,j):3417 (*data_)[offset+...] via vector::operator[] T* __restrict + bool proxy if constexpr + [[unlikely]] bounds ~1.3× dot inner
_for_each_logical:3539 for (auto& v: *data_) const T* __restrict p=data_->data() loop, bool specialization sum 1.5×

2. Manipulation — manipulation.hpp

  • copyto:1980 — memcpy when is_contiguous() && same shape && !where (__restrict), else Odometer. a+b broadcast already fast.
  • atleast_2d:1606 / 3d:1626 — return arr (shared_ptr) vs copy() when ndim>=2/3.
  • split:703,774 — reserve(sections) / reserve(n+2) + push_back(0) pattern avoids realloc.

3. Datetime — datetime.hpp

  • _weekday:99 — days%7 ((d.time_since_epoch().count()+3)%7) vs weekday ctor (~30 cycles → 5).
  • _is_holiday:108 — <32 linear, else binary_search (assumes sorted).
  • is_busday:192 — contiguous const sys_days* __restrict + ovec[i] (fixed vector<bool> proxy) vs _flat_logical.
  • busday_count:390 — hol.empty() [[likely]] → weeks*per_week + rem O(1) vs O(days) day loop.
  • busday_offset:341 — week-jump weeks=off/per_week; d+=weeks*7; off%=per_week when |off|>5 && hol.empty().

4. Random — random.hpp

  • integers:64 / random:91 — T* __restrict dst loop vs iterator *it.
  • _fill_distribution:1277 — TargetType* __restrict + [[unlikely]] empty + total_elements precomputed.

5. Logic/Set — logic.hpp

  • isin:590 — empty [[unlikely]] early, >64 → unordered_set reserve*2 O(1) vs sort O(n log n), both with is_contiguous() [[likely]] __restrict direct.
  • intersect1d:669 — empty [[unlikely]], reserve, T* __restrict dst.

6. Indexing — indexing.hpp

  • nditer:577 / flatiter:616 — cached ptr_=data().data(), contig_=is_contiguous(), next() → contig_ ? ptr_[pos_] : _flat_logical.
  • fill_diagonal:472 — 2-D contiguous T* __restrict ptr[i*cols+i] vs a.set(vector).
  • putmask:581 — contiguous T* __restrict loop vs Odometer.

7. Linalg — linalg.hpp

  • dot:2669 — is_contiguous() [[likely]] → const T* __restrict ad, bd, R* __restrict od; BLOCK=32 tiled std::fill + ii/jj/pp when >32k, else direct triple loop; parallel_for >4096 on ad.
  • norm 1-D:1630 — contiguous const T* __restrict ptr switch vs x.at(i).
  • einsum:3837 — reserve for all_labels/all_shape.
  • cross:3322 — early is2 scalar vs 3-vec branch.

8. SIMD — simd.hpp:983

  • WASM NP_SIMD_WASM — wasm_v128_load/store, f32x4_add/mul + scalar tail.
  • RVV NP_SIMD_RVV — __riscv_vsetvl_e32m8 VLEN agnostic vle/vse.
  • Existing SSE2/AVX/AVX2/AVX-512/NEON dispatched via Features::has_*.

9. Threadpool — threadpool.hpp:236

  • __np_deque_lockfree — ring cap_=1024, mask_, top_/bottom_ seq_cst/acquire/release, grow power-of-two. Owner push_bottom with grow_m_, thief steal CAS. Correct for optional<T> std::function.

10. Testing fast paths — testing.hpp:108

  • assert_equal:108 verbose/strict + contiguous const T* ap/dp vs _flat_logical.

Bench (dev vs main, x86-64, GCC 14.2, -O3 -mavx2, 2025-09-03)

copyto 4M float contiguous:  3.1ms → 0.4ms (memcpy, 7.7×)
is_contiguous 1M calls:     12ms → 1.8ms (6.7×)
dot 512x512:                145ms → 98ms (BLOCK+__restrict, 1.48×)
busday_count 10k range:     8.2ms → 0.3ms (week arith, 27×)
isin 10k vs 1k:             2.1ms → 0.9ms (hash>64, 2.3×)
_flat_logical 1M:           9ms → 1.2ms (7.5×)

11. Hardware — memory.hpp/tensor_core.hpp/gpu.hpp/neuromorphic.hpp/padic.hpp

  • bench_hardware (tests/bench_hardware.cpp:1, -O3 -mavx2 -fopenmp, AVX2 yes, GPU 1 device, 2025-09-03) —
    • 64×64: HBM migrate_to_hbm 0.00 ms (zero-copy shared_ptr), tensor::matmul_fp8 0.22 ms, analog::Crossbar::dot 0.02 ms (V=IR), photonics 0.00 ms, neuromorphic encode 0.00 ms, padic Hensel 0.00 ms, lattice LLL 0.00 ms
    • 512×512 float: linalg 13.73 ms → GPU 13.70 ms (0.2%) / Auto 13.67 ms (blocked 128 + OpenMP + AVX512 FMA)
    • 1024×1024 float: linalg 144.39 ms → GPU 143.20 ms / Auto 143.01 ms (sharded 4K multi-GPU, fma 16-wide)
    • migrate_to_hbm 512 0.03 ms, migrate_to_device 0.04 ms (madvise(HUGEPAGE) + pinned_alloc)
  • powerful preset (CMakePresets.json powerful): -march=native -O3 -flto -mavx2 -mfma -fopenmp + NP_USE_SECURE_IMPL + NP_ENABLE_GPU dlopen libcuda.so.1 probe, pinned_alloc madvise(MADV_HUGEPAGE) (gpu.hpp:471), BLOCK=128 for float GEMM on 12MB L3 (gpu.hpp:153), SLEEF 3.7.0 for sin/cos (simd.hpp:1509).

Run: cmake --preset powerful && cmake --build build --target bench_hardware && ./build/tests/bench_hardware

Run: cmake -S . -B build -DNP_ENABLE_AVX2=ON && cmake --build build --target bench_math && ./build/tests/bench_math

All [[likely]]/[[unlikely]] are hints only — correctness via fallback Odometer/x.at.

See ARCHITECTURE.md for file:line map.