usearch 2.21.4 → 2.25.0

package/numkong/include/numkong/trigonometry/README.md

@@ -0,0 +1,184 @@
+# Trigonometric Functions in NumKong
+
+NumKong implements element-wise trigonometric functions — sine, cosine, and arc tangent — with ~3 ulp error bounds for f32 and faithful rounding for f64.
+Each function operates on dense vectors, reading input angles (radians) and writing output values of the same length.
+The implementations derive from SLEEF (SIMD Library for Evaluating Elementary Functions), adapted for NumKong's ISA dispatch and type system.
+
+Sine:
+
+$$
+\text{sin}: \mathbb{R} \to [-1, 1]
+$$
+
+Cosine:
+
+$$
+\text{cos}: \mathbb{R} \to [-1, 1]
+$$
+
+Arc tangent:
+
+$$
+\text{atan}: \mathbb{R} \to \left(-\frac{\pi}{2}, \frac{\pi}{2}\right)
+$$
+
+Reformulating as Python pseudocode:
+
+```python
+import numpy as np
+
+def sin(a: np.ndarray) -> np.ndarray:
+    return np.sin(a)
+
+def cos(a: np.ndarray) -> np.ndarray:
+    return np.cos(a)
+
+def atan(a: np.ndarray) -> np.ndarray:
+    return np.arctan(a)
+```
+
+## Input & Output Types
+
+| Input Type | Output Type | Description                                      |
+| ---------- | ----------- | ------------------------------------------------ |
+| `f64`      | `f64`       | 64-bit IEEE 754 double precision                 |
+| `f32`      | `f32`       | 32-bit IEEE 754 single precision                 |
+| `f16`      | `f16`       | 16-bit half precision, widened to f32 internally |
+
+## Optimizations
+
+### Cody-Waite Range Reduction
+
+All trigonometric kernels reduce the input angle to $[-\pi/4, \pi/4]$ before polynomial evaluation using Cody-Waite argument reduction.
+The constant $\pi$ is split into high and low parts ($\pi_{\text{hi}} + \pi_{\text{lo}}$) to maintain precision during the subtraction $x - n\pi$: `reduced = (x - n * pi_hi) - n * pi_lo`.
+Single-part subtraction would lose ~3 bits of precision for large multiples of $\pi$; the two-part split preserves the full mantissa.
+The quadrant index $n = \text{round}(x / \pi)$ selects which trigonometric identity to apply (sin-cos swap, sign flip) via a 2-bit branch.
+
+### Minimax Polynomial Approximation
+
+`nk_each_sin_f32_serial`, `nk_each_cos_f32_serial` evaluate degree-9 minimax polynomials via Horner's method after range reduction.
+The polynomial coefficients are precomputed to minimize maximum error over $[-\pi/4, \pi/4]$ — Chebyshev-optimal, not Taylor truncation.
+Horner evaluation: `p = c9*x^2 + c7; p = p*x^2 + c5; p = p*x^2 + c3; p = p*x^2 + c1; p = p*x` — 4 FMA operations plus 1 multiply for the final odd-power term.
+`nk_each_sin_f64_serial` uses degree-19 polynomials for 52-bit mantissa coverage.
+
+### Vectorized Polynomial Evaluation
+
+`nk_each_sin_f32_haswell`, `nk_each_cos_f32_skylake` evaluate the same polynomial on 8 (AVX2) or 16 (AVX-512) elements simultaneously.
+Range reduction, quadrant selection, and polynomial evaluation all operate on packed vectors — the only scalar operation is the final sign correction via `VBLENDVPS` with the quadrant mask.
+`nk_each_sin_f32_neon` processes 4 elements per iteration using `vfmaq_f32` for the Horner chain.
+WASM v128relaxed (`nk_each_sin_f32_v128relaxed`) uses `f32x4.relaxed_madd` for the FMA steps, achieving ~2x throughput over strict `f32x4.mul` + `f32x4.add` sequences.
+
+## Performance
+
+The following performance tables are produced by manually re-running `nk_test` and `nk_bench` included internal tools to measure both accuracy and throughput at different input shapes.
+The input size is controlled by the `NK_DENSE_DIMENSIONS` environment variable and set to 256, 1024, and 4096 elements.
+The throughput is measured in GB/s as the number of input bytes per second.
+Accuracy is reported as mean ULP (units in last place) unless noted otherwise — the average number of representable floating-point values between the result and the exact answer.
+Each kernel runs for at least 20 seconds per configuration.
+Benchmark threads are pinned to specific cores; on machines with heterogeneous core types (e.g., Apple P/E cores), only the fastest cores are used.
+Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run in separate passes to avoid affecting throughput measurements of other kernels.
+
+### Intel Sapphire Rapids
+
+#### Native
+
+| Kernel                     |                      256 |                     1024 |                     4096 |
+| :------------------------- | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                    | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f64_serial`   |        0.994 gb/s, 0 ulp |        0.783 gb/s, 0 ulp |        0.827 gb/s, 0 ulp |
+| `nk_each_cos_f64_serial`   |        0.906 gb/s, 0 ulp |        0.784 gb/s, 0 ulp |        0.824 gb/s, 0 ulp |
+| `nk_each_atan_f64_serial`  |        0.307 gb/s, 0 ulp |        0.291 gb/s, 0 ulp |        0.291 gb/s, 0 ulp |
+| `nk_each_sin_f64_haswell`  |         4.59 gb/s, 0 ulp |         4.19 gb/s, 0 ulp |         4.04 gb/s, 0 ulp |
+| `nk_each_cos_f64_haswell`  |         4.25 gb/s, 0 ulp |         4.14 gb/s, 0 ulp |         3.92 gb/s, 0 ulp |
+| `nk_each_atan_f64_haswell` |         3.83 gb/s, 0 ulp |         3.21 gb/s, 0 ulp |         3.49 gb/s, 0 ulp |
+| `nk_each_sin_f64_skylake`  |         7.65 gb/s, 0 ulp |         6.55 gb/s, 0 ulp |         4.70 gb/s, 0 ulp |
+| `nk_each_cos_f64_skylake`  |         7.88 gb/s, 0 ulp |         5.76 gb/s, 0 ulp |         5.01 gb/s, 0 ulp |
+| `nk_each_atan_f64_skylake` |         5.08 gb/s, 0 ulp |         4.72 gb/s, 0 ulp |         4.58 gb/s, 0 ulp |
+| __f32__                    | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f32_serial`   |         6.29 gb/s, 5 ulp |         6.07 gb/s, 5 ulp |         5.41 gb/s, 5 ulp |
+| `nk_each_cos_f32_serial`   |        7.03 gb/s, 15 ulp |        6.24 gb/s, 15 ulp |        5.16 gb/s, 15 ulp |
+| `nk_each_atan_f32_serial`  |      0.642 gb/s, 0.4 ulp |      0.541 gb/s, 0.4 ulp |      0.567 gb/s, 0.4 ulp |
+| `nk_each_sin_f32_haswell`  |         10.0 gb/s, 5 ulp |         7.36 gb/s, 5 ulp |         5.63 gb/s, 5 ulp |
+| `nk_each_cos_f32_haswell`  |        7.82 gb/s, 15 ulp |        7.11 gb/s, 15 ulp |        5.09 gb/s, 15 ulp |
+| `nk_each_atan_f32_haswell` |       7.63 gb/s, 0.4 ulp |       5.94 gb/s, 0.4 ulp |       5.38 gb/s, 0.4 ulp |
+| `nk_each_sin_f32_skylake`  |         11.9 gb/s, 5 ulp |         9.14 gb/s, 5 ulp |         5.43 gb/s, 5 ulp |
+| `nk_each_cos_f32_skylake`  |        10.4 gb/s, 15 ulp |        8.26 gb/s, 15 ulp |        5.40 gb/s, 15 ulp |
+| `nk_each_atan_f32_skylake` |       9.07 gb/s, 0.4 ulp |       7.80 gb/s, 0.4 ulp |       5.75 gb/s, 0.4 ulp |
+| __f16__                    | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f16_serial`   |      0.112 gb/s, 0.9 ulp |      0.102 gb/s, 1.1 ulp |      0.110 gb/s, 0.9 ulp |
+| `nk_each_cos_f16_serial`   |       0.105 gb/s, 12 ulp |      0.0962 gb/s, 12 ulp |      0.0976 gb/s, 12 ulp |
+| `nk_each_atan_f16_serial`  |     0.0208 gb/s, 6.4 ulp |     0.0201 gb/s, 6.7 ulp |     0.0204 gb/s, 6.6 ulp |
+| `nk_each_sin_f16_skylake`  |     6.05 gb/s, 8.41K ulp |     5.81 gb/s, 8.43K ulp |     5.24 gb/s, 8.41K ulp |
+| `nk_each_cos_f16_skylake`  |     6.05 gb/s, 8.34K ulp |     5.20 gb/s, 8.34K ulp |     5.09 gb/s, 8.35K ulp |
+| `nk_each_atan_f16_skylake` |     4.86 gb/s, 16.5K ulp |     5.25 gb/s, 16.6K ulp |     4.76 gb/s, 16.5K ulp |
+
+#### WASM
+
+Measured with Wasmtime v42 (Cranelift backend).
+
+| Kernel                         |                      256 |                     1024 |                     4096 |
+| :----------------------------- | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                        | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f64_serial`       |       0.34 gb/s, 0.2 ulp |       0.38 gb/s, 0.2 ulp |       0.08 gb/s, 0.2 ulp |
+| `nk_each_cos_f64_serial`       |       0.36 gb/s, 0.3 ulp |       0.39 gb/s, 0.3 ulp |       0.08 gb/s, 0.3 ulp |
+| `nk_each_atan_f64_serial`      |       0.11 gb/s, 0.3 ulp |       0.12 gb/s, 0.3 ulp |       0.11 gb/s, 0.3 ulp |
+| `nk_each_sin_f64_v128relaxed`  |       0.59 gb/s, 0.2 ulp |       0.26 gb/s, 0.2 ulp |       0.05 gb/s, 0.2 ulp |
+| `nk_each_cos_f64_v128relaxed`  |       0.29 gb/s, 0.3 ulp |       0.50 gb/s, 0.3 ulp |       0.03 gb/s, 0.3 ulp |
+| `nk_each_atan_f64_v128relaxed` |       0.11 gb/s, 0.3 ulp |       0.48 gb/s, 0.3 ulp |       0.21 gb/s, 0.3 ulp |
+| __f32__                        | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f32_serial`       |       0.17 gb/s, 4.9 ulp |       0.51 gb/s, 4.9 ulp |       0.07 gb/s, 4.9 ulp |
+| `nk_each_cos_f32_serial`       |      0.05 gb/s, 14.4 ulp |      0.41 gb/s, 14.4 ulp |      0.10 gb/s, 14.4 ulp |
+| `nk_each_atan_f32_serial`      |       0.08 gb/s, 0.4 ulp |       0.08 gb/s, 0.4 ulp |       0.09 gb/s, 0.4 ulp |
+| `nk_each_sin_f32_v128relaxed`  |      0.13 gb/s, 20.7 ulp |      0.01 gb/s, 20.7 ulp |      0.10 gb/s, 20.7 ulp |
+| `nk_each_cos_f32_v128relaxed`  |      0.15 gb/s, 21.9 ulp |      0.32 gb/s, 21.9 ulp |      0.05 gb/s, 21.9 ulp |
+| `nk_each_atan_f32_v128relaxed` |       0.45 gb/s, 0.4 ulp |       0.39 gb/s, 0.4 ulp |       0.15 gb/s, 0.4 ulp |
+| __f16__                        | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f16_serial`       |       0.07 gb/s, 1.1 ulp |       0.07 gb/s, 1.1 ulp |       0.07 gb/s, 1.1 ulp |
+| `nk_each_cos_f16_serial`       |      0.07 gb/s, 11.8 ulp |      0.07 gb/s, 11.8 ulp |      0.07 gb/s, 11.8 ulp |
+| `nk_each_atan_f16_serial`      |       0.03 gb/s, 6.5 ulp |       0.03 gb/s, 6.5 ulp |       0.03 gb/s, 6.5 ulp |
+
+### Apple M4
+
+#### Native
+
+| Kernel                    |                      256 |                     1024 |                     4096 |
+| :------------------------ | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                   | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f64_serial`  |      0.627 gb/s, 0.2 ulp |      0.634 gb/s, 0.2 ulp |      0.639 gb/s, 0.2 ulp |
+| `nk_each_cos_f64_serial`  |      0.621 gb/s, 0.3 ulp |      0.632 gb/s, 0.3 ulp |      0.619 gb/s, 0.3 ulp |
+| `nk_each_atan_f64_serial` |      0.153 gb/s, 0.3 ulp |      0.154 gb/s, 0.3 ulp |      0.153 gb/s, 0.3 ulp |
+| `nk_each_sin_f64_neon`    |       5.94 gb/s, 0.2 ulp |       5.75 gb/s, 0.2 ulp |       5.82 gb/s, 0.2 ulp |
+| `nk_each_cos_f64_neon`    |       5.15 gb/s, 0.3 ulp |       5.36 gb/s, 0.3 ulp |       5.37 gb/s, 0.3 ulp |
+| `nk_each_atan_f64_neon`   |       3.53 gb/s, 0.3 ulp |       3.50 gb/s, 0.3 ulp |       3.50 gb/s, 0.3 ulp |
+| __f32__                   | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f32_serial`  |       7.94 gb/s, 4.9 ulp |       7.94 gb/s, 4.9 ulp |       7.23 gb/s, 4.9 ulp |
+| `nk_each_cos_f32_serial`  |        7.26 gb/s, 14 ulp |        6.41 gb/s, 14 ulp |        6.52 gb/s, 14 ulp |
+| `nk_each_atan_f32_serial` |      0.128 gb/s, 0.4 ulp |      0.129 gb/s, 0.4 ulp |      0.126 gb/s, 0.4 ulp |
+| `nk_each_sin_f32_neon`    |       9.75 gb/s, 4.9 ulp |       9.44 gb/s, 4.9 ulp |       8.13 gb/s, 4.9 ulp |
+| `nk_each_cos_f32_neon`    |        8.68 gb/s, 18 ulp |        7.77 gb/s, 18 ulp |        7.84 gb/s, 18 ulp |
+| `nk_each_atan_f32_neon`   |       5.57 gb/s, 0.4 ulp |       5.00 gb/s, 0.4 ulp |       5.10 gb/s, 0.4 ulp |
+| __f16__                   | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f16_serial`  |       3.66 gb/s, 1.3 ulp |       3.71 gb/s, 1.3 ulp |       3.38 gb/s, 1.3 ulp |
+| `nk_each_cos_f16_serial`  |        3.28 gb/s, 12 ulp |        3.29 gb/s, 12 ulp |        3.15 gb/s, 12 ulp |
+| `nk_each_atan_f16_serial` |     0.0639 gb/s, 6.5 ulp |     0.0626 gb/s, 6.5 ulp |     0.0627 gb/s, 6.5 ulp |
+
+#### WASM
+
+Measured with Wasmtime v43 (Cranelift backend).
+
+| Kernel                         |                      256 |                     1024 |                     4096 |
+| :----------------------------- | -----------------------: | -----------------------: | -----------------------: |
+| __f64__                        | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f64_serial`       |      0.771 gb/s, 0.2 ulp |      0.733 gb/s, 0.2 ulp |      0.798 gb/s, 0.2 ulp |
+| `nk_each_sin_f64_v128relaxed`  |       8.65 gb/s, 0.2 ulp |       8.40 gb/s, 0.2 ulp |       9.45 gb/s, 0.2 ulp |
+| `nk_each_cos_f64_serial`       |      0.746 gb/s, 0.3 ulp |      0.703 gb/s, 0.3 ulp |      0.735 gb/s, 0.3 ulp |
+| `nk_each_cos_f64_v128relaxed`  |       9.07 gb/s, 0.3 ulp |       8.80 gb/s, 0.3 ulp |       9.58 gb/s, 0.3 ulp |
+| `nk_each_atan_f64_serial`      |      0.247 gb/s, 0.3 ulp |      0.234 gb/s, 0.3 ulp |      0.232 gb/s, 0.3 ulp |
+| `nk_each_atan_f64_v128relaxed` |       5.51 gb/s, 0.3 ulp |       5.21 gb/s, 0.3 ulp |       5.96 gb/s, 0.3 ulp |
+| __f32__                        | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
+| `nk_each_sin_f32_serial`       |       9.94 gb/s, 4.9 ulp |       9.44 gb/s, 4.9 ulp |       10.2 gb/s, 4.9 ulp |
+| `nk_each_sin_f32_v128relaxed`  |        17.8 gb/s, 20 ulp |        17.5 gb/s, 20 ulp |        17.9 gb/s, 20 ulp |
+| `nk_each_cos_f32_serial`       |        9.13 gb/s, 14 ulp |        8.96 gb/s, 14 ulp |        9.55 gb/s, 14 ulp |
+| `nk_each_cos_f32_v128relaxed`  |        16.7 gb/s, 21 ulp |        14.7 gb/s, 21 ulp |        16.9 gb/s, 21 ulp |
+| `nk_each_atan_f32_serial`      |      0.191 gb/s, 0.4 ulp |      0.185 gb/s, 0.4 ulp |      0.199 gb/s, 0.4 ulp |
+| `nk_each_atan_f32_v128relaxed` |       11.2 gb/s, 0.4 ulp |       10.7 gb/s, 0.4 ulp |       11.6 gb/s, 0.4 ulp |