usearch 2.11.1 → 2.13.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (50) hide show
  1. package/README.md +100 -30
  2. package/binding.gyp +11 -12
  3. package/c/README.md +3 -0
  4. package/c/lib.cpp +50 -13
  5. package/c/test.c +60 -55
  6. package/c/usearch.h +16 -0
  7. package/cpp/CMakeLists.txt +6 -0
  8. package/cpp/README.md +34 -13
  9. package/cpp/bench.cpp +99 -47
  10. package/cpp/test.cpp +947 -132
  11. package/include/usearch/index.hpp +894 -272
  12. package/include/usearch/index_dense.hpp +418 -156
  13. package/include/usearch/index_plugins.hpp +99 -49
  14. package/javascript/dist/cjs/usearch.d.ts +39 -1
  15. package/javascript/dist/cjs/usearch.js +11 -1
  16. package/javascript/dist/esm/usearch.d.ts +39 -1
  17. package/javascript/dist/esm/usearch.js +11 -1
  18. package/javascript/tsconfig-esm.json +3 -2
  19. package/javascript/usearch.ts +18 -7
  20. package/package.json +2 -2
  21. package/prebuilds/darwin-arm64+x64/usearch.node +0 -0
  22. package/prebuilds/linux-arm64/usearch.node +0 -0
  23. package/prebuilds/linux-x64/usearch.node +0 -0
  24. package/simsimd/.cmake-format.py +19 -0
  25. package/simsimd/CONTRIBUTING.md +49 -8
  26. package/simsimd/Cargo.lock +1 -1
  27. package/simsimd/Cargo.toml +10 -5
  28. package/simsimd/Package.swift +16 -0
  29. package/simsimd/README.md +69 -15
  30. package/simsimd/binding.gyp +14 -19
  31. package/simsimd/build.rs +2 -0
  32. package/simsimd/c/lib.c +18 -0
  33. package/simsimd/include/module.modulemap +4 -0
  34. package/simsimd/include/simsimd/binary.h +2 -0
  35. package/simsimd/include/simsimd/dot.h +466 -32
  36. package/simsimd/include/simsimd/probability.h +20 -10
  37. package/simsimd/include/simsimd/simsimd.h +444 -98
  38. package/simsimd/include/simsimd/spatial.h +480 -44
  39. package/simsimd/include/simsimd/types.h +161 -39
  40. package/simsimd/javascript/simsimd.ts +26 -2
  41. package/simsimd/package.json +1 -1
  42. package/simsimd/rust/benches/cosine.rs +8 -2
  43. package/simsimd/rust/benches/native.rs +240 -2
  44. package/simsimd/rust/benches/sqeuclidean.rs +8 -2
  45. package/simsimd/rust/lib.rs +50 -13
  46. package/simsimd/swift/source/SimSIMD.swift +123 -0
  47. package/simsimd/swift/tests/SimSIMDTests.swift +92 -0
  48. package/prebuilds/win32-ia32/usearch.node +0 -0
  49. package/prebuilds/win32-x64/usearch.node +0 -0
  50. package/simsimd/.releaserc +0 -87
package/simsimd/include/simsimd/probability.h CHANGED
@@ -40,6 +40,8 @@ SIMSIMD_PUBLIC void simsimd_kl_f32_serial(simsimd_f32_t const* a, simsimd_f32_t
40
40
  SIMSIMD_PUBLIC void simsimd_js_f32_serial(simsimd_f32_t const* a, simsimd_f32_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
41
41
  SIMSIMD_PUBLIC void simsimd_kl_f16_serial(simsimd_f16_t const* a, simsimd_f16_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
42
42
  SIMSIMD_PUBLIC void simsimd_js_f16_serial(simsimd_f16_t const* a, simsimd_f16_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
43
+ SIMSIMD_PUBLIC void simsimd_kl_bf16_serial(simsimd_bf16_t const* a, simsimd_bf16_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
44
+ SIMSIMD_PUBLIC void simsimd_js_bf16_serial(simsimd_bf16_t const* a, simsimd_bf16_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
43
45
 
44
46
  /* Double-precision serial backends for all numeric types.
45
47
  * For single-precision computation check out the "*_serial" counterparts of those "*_accurate" functions.
@@ -48,6 +50,8 @@ SIMSIMD_PUBLIC void simsimd_kl_f32_accurate(simsimd_f32_t const* a, simsimd_f32_
48
50
  SIMSIMD_PUBLIC void simsimd_js_f32_accurate(simsimd_f32_t const* a, simsimd_f32_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
49
51
  SIMSIMD_PUBLIC void simsimd_kl_f16_accurate(simsimd_f16_t const* a, simsimd_f16_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
50
52
  SIMSIMD_PUBLIC void simsimd_js_f16_accurate(simsimd_f16_t const* a, simsimd_f16_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
53
+ SIMSIMD_PUBLIC void simsimd_kl_bf16_accurate(simsimd_bf16_t const* a, simsimd_bf16_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
54
+ SIMSIMD_PUBLIC void simsimd_js_bf16_accurate(simsimd_bf16_t const* a, simsimd_bf16_t const* b, simsimd_size_t n, simsimd_distance_t* divergence);
51
55
 
52
56
  /* SIMD-powered backends for Arm NEON, mostly using 32-bit arithmetic over 128-bit words.
53
57
  * By far the most portable backend, covering most Arm v8 devices, over a billion phones, and almost all
@@ -115,12 +119,18 @@ SIMSIMD_MAKE_JS(serial, f32, f32, SIMSIMD_IDENTIFY, SIMSIMD_F32_DIVISION_EPSILON
115
119
  SIMSIMD_MAKE_KL(serial, f16, f32, SIMSIMD_UNCOMPRESS_F16, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_kl_f16_serial
116
120
  SIMSIMD_MAKE_JS(serial, f16, f32, SIMSIMD_UNCOMPRESS_F16, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_js_f16_serial
117
121
 
122
+ SIMSIMD_MAKE_KL(serial, bf16, f32, SIMSIMD_UNCOMPRESS_BF16, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_kl_bf16_serial
123
+ SIMSIMD_MAKE_JS(serial, bf16, f32, SIMSIMD_UNCOMPRESS_BF16, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_js_bf16_serial
124
+
118
125
  SIMSIMD_MAKE_KL(accurate, f32, f64, SIMSIMD_IDENTIFY, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_kl_f32_accurate
119
126
  SIMSIMD_MAKE_JS(accurate, f32, f64, SIMSIMD_IDENTIFY, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_js_f32_accurate
120
127
 
121
128
  SIMSIMD_MAKE_KL(accurate, f16, f64, SIMSIMD_UNCOMPRESS_F16, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_kl_f16_accurate
122
129
  SIMSIMD_MAKE_JS(accurate, f16, f64, SIMSIMD_UNCOMPRESS_F16, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_js_f16_accurate
123
130
 
131
+ SIMSIMD_MAKE_KL(accurate, bf16, f64, SIMSIMD_UNCOMPRESS_BF16, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_kl_bf16_accurate
132
+ SIMSIMD_MAKE_JS(accurate, bf16, f64, SIMSIMD_UNCOMPRESS_BF16, SIMSIMD_F32_DIVISION_EPSILON) // simsimd_js_bf16_accurate
133
+
124
134
  #if SIMSIMD_TARGET_ARM
125
135
  #if SIMSIMD_TARGET_NEON
126
136
  #pragma GCC push_options
@@ -472,17 +482,17 @@ simsimd_js_f32_skylake_cycle:
472
482
 
473
483
  inline __m512h simsimd_log2_f16_sapphire(__m512h x) {
474
484
  // Extract the exponent and mantissa
475
- __m512h one = _mm512_set1_ph((_Float16)1);
485
+ __m512h one = _mm512_set1_ph((simsimd_f16_t)1);
476
486
  __m512h e = _mm512_getexp_ph(x);
477
487
  __m512h m = _mm512_getmant_ph(x, _MM_MANT_NORM_1_2, _MM_MANT_SIGN_src);
478
488
 
479
489
  // Compute the polynomial using Horner's method
480
- __m512h p = _mm512_set1_ph((_Float16)-3.4436006e-2f);
481
- p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((_Float16)3.1821337e-1f));
482
- p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((_Float16)-1.2315303f));
483
- p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((_Float16)2.5988452f));
484
- p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((_Float16)-3.3241990f));
485
- p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((_Float16)3.1157899f));
490
+ __m512h p = _mm512_set1_ph((simsimd_f16_t)-3.4436006e-2f);
491
+ p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((simsimd_f16_t)3.1821337e-1f));
492
+ p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((simsimd_f16_t)-1.2315303f));
493
+ p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((simsimd_f16_t)2.5988452f));
494
+ p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((simsimd_f16_t)-3.3241990f));
495
+ p = _mm512_fmadd_ph(m, p, _mm512_set1_ph((simsimd_f16_t)3.1157899f));
486
496
 
487
497
  return _mm512_add_ph(_mm512_mul_ph(p, _mm512_sub_ph(m, one)), e);
488
498
  }
@@ -490,7 +500,7 @@ inline __m512h simsimd_log2_f16_sapphire(__m512h x) {
490
500
  SIMSIMD_PUBLIC void simsimd_kl_f16_sapphire(simsimd_f16_t const* a, simsimd_f16_t const* b, simsimd_size_t n,
491
501
  simsimd_distance_t* result) {
492
502
  __m512h sum_vec = _mm512_setzero_ph();
493
- __m512h epsilon_vec = _mm512_set1_ph((_Float16)SIMSIMD_F16_DIVISION_EPSILON);
503
+ __m512h epsilon_vec = _mm512_set1_ph((simsimd_f16_t)SIMSIMD_F16_DIVISION_EPSILON);
494
504
  __m512h a_vec, b_vec;
495
505
 
496
506
  simsimd_kl_f16_sapphire_cycle:
@@ -519,7 +529,7 @@ SIMSIMD_PUBLIC void simsimd_js_f16_sapphire(simsimd_f16_t const* a, simsimd_f16_
519
529
  simsimd_distance_t* result) {
520
530
  __m512h sum_a_vec = _mm512_setzero_ph();
521
531
  __m512h sum_b_vec = _mm512_setzero_ph();
522
- __m512h epsilon_vec = _mm512_set1_ph((_Float16)SIMSIMD_F16_DIVISION_EPSILON);
532
+ __m512h epsilon_vec = _mm512_set1_ph((simsimd_f16_t)SIMSIMD_F16_DIVISION_EPSILON);
523
533
  __m512h a_vec, b_vec;
524
534
 
525
535
  simsimd_js_f16_sapphire_cycle:
@@ -533,7 +543,7 @@ simsimd_js_f16_sapphire_cycle:
533
543
  b_vec = _mm512_castsi512_ph(_mm512_loadu_epi16(b));
534
544
  a += 32, b += 32, n -= 32;
535
545
  }
536
- __m512h m_vec = _mm512_mul_ph(_mm512_add_ph(a_vec, b_vec), _mm512_set1_ph((_Float16)0.5f));
546
+ __m512h m_vec = _mm512_mul_ph(_mm512_add_ph(a_vec, b_vec), _mm512_set1_ph((simsimd_f16_t)0.5f));
537
547
  __mmask32 nonzero_mask_a = _mm512_cmp_ph_mask(a_vec, epsilon_vec, _CMP_GE_OQ);
538
548
  __mmask32 nonzero_mask_b = _mm512_cmp_ph_mask(b_vec, epsilon_vec, _CMP_GE_OQ);
539
549
  __mmask32 nonzero_mask = nonzero_mask_a & nonzero_mask_b;