numkong 7.0.0

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 (294) hide show
  1. package/LICENSE +201 -0
  2. package/README.md +495 -0
  3. package/binding.gyp +540 -0
  4. package/c/dispatch.h +512 -0
  5. package/c/dispatch_bf16.c +389 -0
  6. package/c/dispatch_bf16c.c +52 -0
  7. package/c/dispatch_e2m3.c +263 -0
  8. package/c/dispatch_e3m2.c +243 -0
  9. package/c/dispatch_e4m3.c +276 -0
  10. package/c/dispatch_e5m2.c +272 -0
  11. package/c/dispatch_f16.c +376 -0
  12. package/c/dispatch_f16c.c +58 -0
  13. package/c/dispatch_f32.c +378 -0
  14. package/c/dispatch_f32c.c +99 -0
  15. package/c/dispatch_f64.c +296 -0
  16. package/c/dispatch_f64c.c +98 -0
  17. package/c/dispatch_i16.c +96 -0
  18. package/c/dispatch_i32.c +89 -0
  19. package/c/dispatch_i4.c +150 -0
  20. package/c/dispatch_i64.c +86 -0
  21. package/c/dispatch_i8.c +289 -0
  22. package/c/dispatch_other.c +330 -0
  23. package/c/dispatch_u1.c +148 -0
  24. package/c/dispatch_u16.c +124 -0
  25. package/c/dispatch_u32.c +118 -0
  26. package/c/dispatch_u4.c +150 -0
  27. package/c/dispatch_u64.c +102 -0
  28. package/c/dispatch_u8.c +303 -0
  29. package/c/numkong.c +950 -0
  30. package/include/README.md +573 -0
  31. package/include/module.modulemap +129 -0
  32. package/include/numkong/attention/sapphireamx.h +1361 -0
  33. package/include/numkong/attention/sme.h +2066 -0
  34. package/include/numkong/attention.h +49 -0
  35. package/include/numkong/capabilities.h +748 -0
  36. package/include/numkong/cast/README.md +262 -0
  37. package/include/numkong/cast/haswell.h +975 -0
  38. package/include/numkong/cast/icelake.h +470 -0
  39. package/include/numkong/cast/neon.h +1192 -0
  40. package/include/numkong/cast/rvv.h +1021 -0
  41. package/include/numkong/cast/sapphire.h +262 -0
  42. package/include/numkong/cast/serial.h +2262 -0
  43. package/include/numkong/cast/skylake.h +856 -0
  44. package/include/numkong/cast/v128relaxed.h +180 -0
  45. package/include/numkong/cast.h +230 -0
  46. package/include/numkong/curved/README.md +223 -0
  47. package/include/numkong/curved/genoa.h +182 -0
  48. package/include/numkong/curved/haswell.h +276 -0
  49. package/include/numkong/curved/neon.h +205 -0
  50. package/include/numkong/curved/neonbfdot.h +212 -0
  51. package/include/numkong/curved/neonhalf.h +212 -0
  52. package/include/numkong/curved/rvv.h +305 -0
  53. package/include/numkong/curved/serial.h +207 -0
  54. package/include/numkong/curved/skylake.h +457 -0
  55. package/include/numkong/curved/smef64.h +506 -0
  56. package/include/numkong/curved.h +517 -0
  57. package/include/numkong/curved.hpp +144 -0
  58. package/include/numkong/dot/README.md +425 -0
  59. package/include/numkong/dot/alder.h +563 -0
  60. package/include/numkong/dot/genoa.h +315 -0
  61. package/include/numkong/dot/haswell.h +1688 -0
  62. package/include/numkong/dot/icelake.h +883 -0
  63. package/include/numkong/dot/neon.h +818 -0
  64. package/include/numkong/dot/neonbfdot.h +244 -0
  65. package/include/numkong/dot/neonfhm.h +360 -0
  66. package/include/numkong/dot/neonhalf.h +198 -0
  67. package/include/numkong/dot/neonsdot.h +508 -0
  68. package/include/numkong/dot/rvv.h +714 -0
  69. package/include/numkong/dot/rvvbb.h +72 -0
  70. package/include/numkong/dot/rvvbf16.h +123 -0
  71. package/include/numkong/dot/rvvhalf.h +129 -0
  72. package/include/numkong/dot/sapphire.h +141 -0
  73. package/include/numkong/dot/serial.h +838 -0
  74. package/include/numkong/dot/sierra.h +405 -0
  75. package/include/numkong/dot/skylake.h +1084 -0
  76. package/include/numkong/dot/sve.h +379 -0
  77. package/include/numkong/dot/svebfdot.h +74 -0
  78. package/include/numkong/dot/svehalf.h +123 -0
  79. package/include/numkong/dot/v128relaxed.h +1258 -0
  80. package/include/numkong/dot.h +1070 -0
  81. package/include/numkong/dot.hpp +94 -0
  82. package/include/numkong/dots/README.md +496 -0
  83. package/include/numkong/dots/alder.h +114 -0
  84. package/include/numkong/dots/genoa.h +94 -0
  85. package/include/numkong/dots/haswell.h +295 -0
  86. package/include/numkong/dots/icelake.h +171 -0
  87. package/include/numkong/dots/neon.h +120 -0
  88. package/include/numkong/dots/neonbfdot.h +58 -0
  89. package/include/numkong/dots/neonfhm.h +94 -0
  90. package/include/numkong/dots/neonhalf.h +57 -0
  91. package/include/numkong/dots/neonsdot.h +108 -0
  92. package/include/numkong/dots/rvv.h +2486 -0
  93. package/include/numkong/dots/sapphireamx.h +3973 -0
  94. package/include/numkong/dots/serial.h +2844 -0
  95. package/include/numkong/dots/sierra.h +97 -0
  96. package/include/numkong/dots/skylake.h +196 -0
  97. package/include/numkong/dots/sme.h +5372 -0
  98. package/include/numkong/dots/smebi32.h +461 -0
  99. package/include/numkong/dots/smef64.h +1318 -0
  100. package/include/numkong/dots/smehalf.h +47 -0
  101. package/include/numkong/dots/v128relaxed.h +294 -0
  102. package/include/numkong/dots.h +2804 -0
  103. package/include/numkong/dots.hpp +639 -0
  104. package/include/numkong/each/README.md +469 -0
  105. package/include/numkong/each/haswell.h +1658 -0
  106. package/include/numkong/each/icelake.h +272 -0
  107. package/include/numkong/each/neon.h +1104 -0
  108. package/include/numkong/each/neonbfdot.h +212 -0
  109. package/include/numkong/each/neonhalf.h +410 -0
  110. package/include/numkong/each/rvv.h +1121 -0
  111. package/include/numkong/each/sapphire.h +477 -0
  112. package/include/numkong/each/serial.h +260 -0
  113. package/include/numkong/each/skylake.h +1562 -0
  114. package/include/numkong/each.h +2146 -0
  115. package/include/numkong/each.hpp +434 -0
  116. package/include/numkong/geospatial/README.md +147 -0
  117. package/include/numkong/geospatial/haswell.h +593 -0
  118. package/include/numkong/geospatial/neon.h +571 -0
  119. package/include/numkong/geospatial/rvv.h +701 -0
  120. package/include/numkong/geospatial/serial.h +309 -0
  121. package/include/numkong/geospatial/skylake.h +577 -0
  122. package/include/numkong/geospatial/v128relaxed.h +613 -0
  123. package/include/numkong/geospatial.h +453 -0
  124. package/include/numkong/geospatial.hpp +235 -0
  125. package/include/numkong/matrix.hpp +336 -0
  126. package/include/numkong/maxsim/README.md +187 -0
  127. package/include/numkong/maxsim/alder.h +511 -0
  128. package/include/numkong/maxsim/genoa.h +115 -0
  129. package/include/numkong/maxsim/haswell.h +553 -0
  130. package/include/numkong/maxsim/icelake.h +480 -0
  131. package/include/numkong/maxsim/neonsdot.h +394 -0
  132. package/include/numkong/maxsim/sapphireamx.h +877 -0
  133. package/include/numkong/maxsim/serial.h +490 -0
  134. package/include/numkong/maxsim/sme.h +929 -0
  135. package/include/numkong/maxsim/v128relaxed.h +280 -0
  136. package/include/numkong/maxsim.h +571 -0
  137. package/include/numkong/maxsim.hpp +133 -0
  138. package/include/numkong/mesh/README.md +227 -0
  139. package/include/numkong/mesh/haswell.h +2235 -0
  140. package/include/numkong/mesh/neon.h +1329 -0
  141. package/include/numkong/mesh/neonbfdot.h +842 -0
  142. package/include/numkong/mesh/neonhalf.h +616 -0
  143. package/include/numkong/mesh/rvv.h +916 -0
  144. package/include/numkong/mesh/serial.h +742 -0
  145. package/include/numkong/mesh/skylake.h +1135 -0
  146. package/include/numkong/mesh/v128relaxed.h +1052 -0
  147. package/include/numkong/mesh.h +652 -0
  148. package/include/numkong/mesh.hpp +762 -0
  149. package/include/numkong/numkong.h +78 -0
  150. package/include/numkong/numkong.hpp +57 -0
  151. package/include/numkong/probability/README.md +173 -0
  152. package/include/numkong/probability/haswell.h +267 -0
  153. package/include/numkong/probability/neon.h +225 -0
  154. package/include/numkong/probability/rvv.h +409 -0
  155. package/include/numkong/probability/serial.h +169 -0
  156. package/include/numkong/probability/skylake.h +324 -0
  157. package/include/numkong/probability.h +383 -0
  158. package/include/numkong/probability.hpp +120 -0
  159. package/include/numkong/random.h +50 -0
  160. package/include/numkong/random.hpp +285 -0
  161. package/include/numkong/reduce/README.md +547 -0
  162. package/include/numkong/reduce/alder.h +632 -0
  163. package/include/numkong/reduce/genoa.h +201 -0
  164. package/include/numkong/reduce/haswell.h +3783 -0
  165. package/include/numkong/reduce/icelake.h +549 -0
  166. package/include/numkong/reduce/neon.h +3841 -0
  167. package/include/numkong/reduce/neonbfdot.h +353 -0
  168. package/include/numkong/reduce/neonfhm.h +665 -0
  169. package/include/numkong/reduce/neonhalf.h +157 -0
  170. package/include/numkong/reduce/neonsdot.h +357 -0
  171. package/include/numkong/reduce/rvv.h +3407 -0
  172. package/include/numkong/reduce/serial.h +757 -0
  173. package/include/numkong/reduce/sierra.h +338 -0
  174. package/include/numkong/reduce/skylake.h +3792 -0
  175. package/include/numkong/reduce/v128relaxed.h +2302 -0
  176. package/include/numkong/reduce.h +1597 -0
  177. package/include/numkong/reduce.hpp +633 -0
  178. package/include/numkong/scalar/README.md +89 -0
  179. package/include/numkong/scalar/haswell.h +113 -0
  180. package/include/numkong/scalar/neon.h +122 -0
  181. package/include/numkong/scalar/neonhalf.h +70 -0
  182. package/include/numkong/scalar/rvv.h +211 -0
  183. package/include/numkong/scalar/sapphire.h +63 -0
  184. package/include/numkong/scalar/serial.h +332 -0
  185. package/include/numkong/scalar/v128relaxed.h +56 -0
  186. package/include/numkong/scalar.h +683 -0
  187. package/include/numkong/set/README.md +179 -0
  188. package/include/numkong/set/haswell.h +334 -0
  189. package/include/numkong/set/icelake.h +485 -0
  190. package/include/numkong/set/neon.h +364 -0
  191. package/include/numkong/set/rvv.h +226 -0
  192. package/include/numkong/set/rvvbb.h +117 -0
  193. package/include/numkong/set/serial.h +174 -0
  194. package/include/numkong/set/sve.h +185 -0
  195. package/include/numkong/set/v128relaxed.h +240 -0
  196. package/include/numkong/set.h +457 -0
  197. package/include/numkong/set.hpp +114 -0
  198. package/include/numkong/sets/README.md +149 -0
  199. package/include/numkong/sets/haswell.h +63 -0
  200. package/include/numkong/sets/icelake.h +66 -0
  201. package/include/numkong/sets/neon.h +61 -0
  202. package/include/numkong/sets/serial.h +43 -0
  203. package/include/numkong/sets/smebi32.h +1099 -0
  204. package/include/numkong/sets/v128relaxed.h +58 -0
  205. package/include/numkong/sets.h +339 -0
  206. package/include/numkong/sparse/README.md +156 -0
  207. package/include/numkong/sparse/icelake.h +463 -0
  208. package/include/numkong/sparse/neon.h +288 -0
  209. package/include/numkong/sparse/serial.h +117 -0
  210. package/include/numkong/sparse/sve2.h +507 -0
  211. package/include/numkong/sparse/turin.h +322 -0
  212. package/include/numkong/sparse.h +363 -0
  213. package/include/numkong/sparse.hpp +113 -0
  214. package/include/numkong/spatial/README.md +435 -0
  215. package/include/numkong/spatial/alder.h +607 -0
  216. package/include/numkong/spatial/genoa.h +290 -0
  217. package/include/numkong/spatial/haswell.h +960 -0
  218. package/include/numkong/spatial/icelake.h +586 -0
  219. package/include/numkong/spatial/neon.h +773 -0
  220. package/include/numkong/spatial/neonbfdot.h +165 -0
  221. package/include/numkong/spatial/neonhalf.h +118 -0
  222. package/include/numkong/spatial/neonsdot.h +261 -0
  223. package/include/numkong/spatial/rvv.h +984 -0
  224. package/include/numkong/spatial/rvvbf16.h +123 -0
  225. package/include/numkong/spatial/rvvhalf.h +117 -0
  226. package/include/numkong/spatial/sapphire.h +343 -0
  227. package/include/numkong/spatial/serial.h +346 -0
  228. package/include/numkong/spatial/sierra.h +323 -0
  229. package/include/numkong/spatial/skylake.h +606 -0
  230. package/include/numkong/spatial/sve.h +224 -0
  231. package/include/numkong/spatial/svebfdot.h +122 -0
  232. package/include/numkong/spatial/svehalf.h +109 -0
  233. package/include/numkong/spatial/v128relaxed.h +717 -0
  234. package/include/numkong/spatial.h +1425 -0
  235. package/include/numkong/spatial.hpp +183 -0
  236. package/include/numkong/spatials/README.md +580 -0
  237. package/include/numkong/spatials/alder.h +94 -0
  238. package/include/numkong/spatials/genoa.h +94 -0
  239. package/include/numkong/spatials/haswell.h +219 -0
  240. package/include/numkong/spatials/icelake.h +113 -0
  241. package/include/numkong/spatials/neon.h +109 -0
  242. package/include/numkong/spatials/neonbfdot.h +60 -0
  243. package/include/numkong/spatials/neonfhm.h +92 -0
  244. package/include/numkong/spatials/neonhalf.h +58 -0
  245. package/include/numkong/spatials/neonsdot.h +109 -0
  246. package/include/numkong/spatials/rvv.h +1960 -0
  247. package/include/numkong/spatials/sapphireamx.h +1149 -0
  248. package/include/numkong/spatials/serial.h +226 -0
  249. package/include/numkong/spatials/sierra.h +96 -0
  250. package/include/numkong/spatials/skylake.h +184 -0
  251. package/include/numkong/spatials/sme.h +1901 -0
  252. package/include/numkong/spatials/smef64.h +465 -0
  253. package/include/numkong/spatials/v128relaxed.h +240 -0
  254. package/include/numkong/spatials.h +3021 -0
  255. package/include/numkong/spatials.hpp +508 -0
  256. package/include/numkong/tensor.hpp +1592 -0
  257. package/include/numkong/trigonometry/README.md +184 -0
  258. package/include/numkong/trigonometry/haswell.h +652 -0
  259. package/include/numkong/trigonometry/neon.h +639 -0
  260. package/include/numkong/trigonometry/rvv.h +699 -0
  261. package/include/numkong/trigonometry/serial.h +703 -0
  262. package/include/numkong/trigonometry/skylake.h +721 -0
  263. package/include/numkong/trigonometry/v128relaxed.h +666 -0
  264. package/include/numkong/trigonometry.h +467 -0
  265. package/include/numkong/trigonometry.hpp +166 -0
  266. package/include/numkong/types.h +1384 -0
  267. package/include/numkong/types.hpp +5603 -0
  268. package/include/numkong/vector.hpp +698 -0
  269. package/javascript/README.md +246 -0
  270. package/javascript/dist/cjs/numkong-wasm.d.ts +166 -0
  271. package/javascript/dist/cjs/numkong-wasm.js +617 -0
  272. package/javascript/dist/cjs/numkong.d.ts +343 -0
  273. package/javascript/dist/cjs/numkong.js +523 -0
  274. package/javascript/dist/cjs/package.json +3 -0
  275. package/javascript/dist/cjs/types.d.ts +284 -0
  276. package/javascript/dist/cjs/types.js +653 -0
  277. package/javascript/dist/esm/numkong-wasm.d.ts +166 -0
  278. package/javascript/dist/esm/numkong-wasm.js +595 -0
  279. package/javascript/dist/esm/numkong.d.ts +343 -0
  280. package/javascript/dist/esm/numkong.js +452 -0
  281. package/javascript/dist/esm/package.json +3 -0
  282. package/javascript/dist/esm/types.d.ts +284 -0
  283. package/javascript/dist/esm/types.js +630 -0
  284. package/javascript/dist-package-cjs.json +3 -0
  285. package/javascript/dist-package-esm.json +3 -0
  286. package/javascript/node-gyp-build.d.ts +1 -0
  287. package/javascript/numkong-wasm.ts +756 -0
  288. package/javascript/numkong.c +689 -0
  289. package/javascript/numkong.ts +575 -0
  290. package/javascript/tsconfig-base.json +39 -0
  291. package/javascript/tsconfig-cjs.json +8 -0
  292. package/javascript/tsconfig-esm.json +8 -0
  293. package/javascript/types.ts +674 -0
  294. package/package.json +87 -0
@@ -0,0 +1,469 @@
1
+ # Element-Wise Arithmetic in NumKong
2
+
3
+ NumKong implements element-wise vector arithmetic: addition, scaling, blending, and fused multiply-add across all supported numeric types.
4
+ Each operation reads one to three input vectors and writes one output vector of the same length, with scalar coefficients $\alpha$ and $\beta$ controlling linear combinations.
5
+ Mixed-precision workflows use narrower input types (Float16, BFloat16, Float8) with Float32 intermediate computation and narrowed output.
6
+
7
+ Sum (addition):
8
+
9
+ ```math
10
+ \text{result}_i = a_i + b_i
11
+ ```
12
+
13
+ Scale:
14
+
15
+ ```math
16
+ \text{result}_i = \alpha \cdot a_i + \beta
17
+ ```
18
+
19
+ Blend:
20
+
21
+ ```math
22
+ \text{result}_i = \alpha \cdot a_i + \beta \cdot b_i
23
+ ```
24
+
25
+ Fused multiply-add:
26
+
27
+ ```math
28
+ \text{result}_i = \alpha \cdot a_i \cdot b_i + \beta \cdot c_i
29
+ ```
30
+
31
+ Reformulating as Python pseudocode:
32
+
33
+ ```python
34
+ import numpy as np
35
+
36
+ def fma(a: np.ndarray, b: np.ndarray, c: np.ndarray,
37
+ alpha: float = 1.0, beta: float = 1.0) -> np.ndarray:
38
+ return alpha * a * b + beta * c
39
+
40
+ def scale(a: np.ndarray, alpha: float = 1.0, beta: float = 0.0) -> np.ndarray:
41
+ return alpha * a + beta
42
+
43
+ def add(a: np.ndarray, b: np.ndarray) -> np.ndarray:
44
+ return a + b
45
+
46
+ def blend(a: np.ndarray, b: np.ndarray,
47
+ alpha: float = 1.0, beta: float = 1.0) -> np.ndarray:
48
+ return alpha * a + beta * b
49
+ ```
50
+
51
+ ## Input & Output Types
52
+
53
+ Real and integer element-wise operations:
54
+
55
+ | Input Type | Output Type | Description |
56
+ | ---------- | ----------- | ----------------------------------------- |
57
+ | `f64` | `f64` | 64-bit IEEE 754 double precision |
58
+ | `f32` | `f32` | 32-bit IEEE 754 single precision |
59
+ | `f16` | `f16` | 16-bit IEEE 754 half precision |
60
+ | `bf16` | `bf16` | 16-bit brain float |
61
+ | `e4m3` | `e4m3` | 8-bit Float8: 4 exponent, 3 mantissa bits |
62
+ | `e5m2` | `e5m2` | 8-bit Float8: 5 exponent, 2 mantissa bits |
63
+ | `i8` | `i8` | 8-bit signed integers, saturating |
64
+ | `u8` | `u8` | 8-bit unsigned integers, saturating |
65
+ | `i16` | `i16` | 16-bit signed integers |
66
+ | `u16` | `u16` | 16-bit unsigned integers |
67
+ | `i32` | `i32` | 32-bit signed integers |
68
+ | `u32` | `u32` | 32-bit unsigned integers |
69
+ | `i64` | `i64` | 64-bit signed integers |
70
+ | `u64` | `u64` | 64-bit unsigned integers |
71
+
72
+ Complex element-wise operations:
73
+
74
+ | Input Type | Output Type | Description |
75
+ | ---------- | ----------- | -------------------- |
76
+ | `f64c` | `f64c` | 64-bit complex pairs |
77
+ | `f32c` | `f32c` | 32-bit complex pairs |
78
+
79
+ ## Optimizations
80
+
81
+ ### Widening-Narrowing Pipeline for Sub-32-bit Types
82
+
83
+ `nk_each_fma_f16_haswell`, `nk_each_blend_bf16_neonbfdot`, `nk_each_scale_e4m3_haswell` widen inputs to Float32 before arithmetic, then narrow the result back to the original type.
84
+ The widen-compute-narrow pipeline costs 2 extra conversion instructions per element but guarantees Float32-precision intermediate results — critical for FMA where naive Float16 multiplication would lose 5+ bits of mantissa.
85
+ Haswell processes 8 Float16 elements per cycle: `VCVTPH2PS` (widen) -> `VFMADD231PS` (FMA) -> `VCVTPS2PH` (narrow), fully pipelined across 3 execution ports.
86
+
87
+ ### Saturating Integer Arithmetic
88
+
89
+ `nk_each_sum_i8_haswell`, `nk_each_sum_u8_neonhalf` use saturating addition — clamping to type bounds instead of wrapping on overflow.
90
+ Haswell uses `VPADDSB` / `VPADDUSB` for signed/unsigned 8-bit saturation in a single instruction (32 elements per cycle at YMM width).
91
+ Serial fallback implements saturation via branch-free min/max: `result = min(max(a + b, TYPE_MIN), TYPE_MAX)` with overflow detection through sign-bit comparison.
92
+
93
+ ### Complex Number Layout
94
+
95
+ `nk_each_fma_f32c_serial`, `nk_each_blend_f64c_serial` operate on interleaved real/imaginary pairs: `[re0, im0, re1, im1, ...]`.
96
+ Addition and scaling treat complex vectors as 2N-length real vectors — no special handling needed.
97
+ FMA requires cross-lane operations for the imaginary part: `re(a*b) = re(a)*re(b) - im(a)*im(b)`, implemented via `VFMADDSUB231PS` which alternates add/subtract across even/odd lanes.
98
+
99
+ ## Performance
100
+
101
+ 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.
102
+ The input size is controlled by the `NK_DENSE_DIMENSIONS` environment variable and set to 256, 1024, and 4096 elements.
103
+ The throughput is measured in GB/s as the number of input bytes read per second.
104
+ 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.
105
+ Rows marked `🧩` use external BLAS baselines rather than NumKong kernels.
106
+ Each kernel runs for at least 20 seconds per configuration.
107
+ 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.
108
+ Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run in separate passes to avoid affecting throughput measurements of other kernels.
109
+
110
+ ### Intel Sapphire Rapids
111
+
112
+ #### Native
113
+
114
+ | Kernel | 256 | 1024 | 4096 |
115
+ | :--------------------------- | -----------------------: | -----------------------: | -----------------------: |
116
+ | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
117
+ | `sum_f64_with_blas` 🧩 | 12.1 gb/s | 8.63 gb/s | 6.83 gb/s |
118
+ | `each_blend_f64_with_blas` 🧩 | 11.3 gb/s | 8.80 gb/s | 6.71 gb/s |
119
+ | `nk_each_sum_f64_serial` | 16.7 gb/s, 0 ulp | 17.2 gb/s, 0 ulp | 10.9 gb/s, 0 ulp |
120
+ | `nk_each_sum_f64_haswell` | 14.7 gb/s, 0 ulp | 10.2 gb/s, 0 ulp | 7.99 gb/s, 0 ulp |
121
+ | `nk_each_sum_f64_skylake` | 16.4 gb/s, 0 ulp | 16.7 gb/s, 0 ulp | 8.59 gb/s, 0 ulp |
122
+ | `nk_each_scale_f64_serial` | 11.4 gb/s, 0 ulp | 12.5 gb/s, 0 ulp | 7.99 gb/s, 0 ulp |
123
+ | `nk_each_scale_f64_haswell` | 9.61 gb/s, 0 ulp | 9.22 gb/s, 0 ulp | 5.08 gb/s, 0 ulp |
124
+ | `nk_each_scale_f64_skylake` | 11.2 gb/s, 0 ulp | 11.9 gb/s, 0 ulp | 6.30 gb/s, 0 ulp |
125
+ | `nk_each_blend_f64_serial` | 16.4 gb/s, 1.4 ulp | 16.4 gb/s, 1.1 ulp | 11.8 gb/s, 1.1 ulp |
126
+ | `nk_each_blend_f64_haswell` | 13.4 gb/s, 1.5 ulp | 11.2 gb/s, 1.5 ulp | 7.85 gb/s, 1.1 ulp |
127
+ | `nk_each_blend_f64_skylake` | 16.5 gb/s, 1.7 ulp | 15.9 gb/s, 1.5 ulp | 8.55 gb/s, 1.1 ulp |
128
+ | `nk_each_fma_f64_serial` | 20.1 gb/s, 1.5 ulp | 20.8 gb/s, 1.5 ulp | 11.7 gb/s, 1.3 ulp |
129
+ | `nk_each_fma_f64_haswell` | 16.8 gb/s, 1.5 ulp | 11.1 gb/s, 1.5 ulp | 9.42 gb/s, 2.8 ulp |
130
+ | `nk_each_fma_f64_skylake` | 19.8 gb/s, 1.4 ulp | 20.4 gb/s, 1.5 ulp | 11.6 gb/s, 2.7 ulp |
131
+ | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
132
+ | `sum_f32_with_blas` 🧩 | 11.6 gb/s | 12.8 gb/s | 7.57 gb/s |
133
+ | `each_blend_f32_with_blas` 🧩 | 10.3 gb/s | 9.60 gb/s | 6.69 gb/s |
134
+ | `nk_each_sum_f32_serial` | 16.9 gb/s, 0 ulp | 18.8 gb/s, 0 ulp | 15.7 gb/s, 0 ulp |
135
+ | `nk_each_sum_f32_haswell` | 14.6 gb/s, 0 ulp | 14.3 gb/s, 0 ulp | 8.49 gb/s, 0 ulp |
136
+ | `nk_each_sum_f32_skylake` | 16.7 gb/s, 0 ulp | 17.9 gb/s, 0 ulp | 16.5 gb/s, 0 ulp |
137
+ | `nk_each_scale_f32_serial` | 10.8 gb/s, 0 ulp | 13.0 gb/s, 0 ulp | 12.1 gb/s, 0 ulp |
138
+ | `nk_each_scale_f32_haswell` | 9.26 gb/s, 0 ulp | 10.2 gb/s, 0 ulp | 6.22 gb/s, 0 ulp |
139
+ | `nk_each_scale_f32_skylake` | 12.3 gb/s, 0 ulp | 11.9 gb/s, 0 ulp | 12.9 gb/s, 0 ulp |
140
+ | `nk_each_blend_f32_serial` | 16.2 gb/s, 351 ulp | 18.7 gb/s, 2.0 ulp | 17.4 gb/s, 1.4 ulp |
141
+ | `nk_each_blend_f32_haswell` | 14.5 gb/s, 2.3 ulp | 14.2 gb/s, 2.1 ulp | 8.16 gb/s, 1.3 ulp |
142
+ | `nk_each_blend_f32_skylake` | 15.7 gb/s, 1.9 ulp | 17.2 gb/s, 1.8 ulp | 15.9 gb/s, 1.3 ulp |
143
+ | `nk_each_fma_f32_serial` | 20.1 gb/s, 1.4 ulp | 18.7 gb/s, 2.1 ulp | 19.0 gb/s, 1.6 ulp |
144
+ | `nk_each_fma_f32_haswell` | 18.4 gb/s, 1.4 ulp | 15.3 gb/s, 1.8 ulp | 9.52 gb/s, 1.5 ulp |
145
+ | `nk_each_fma_f32_skylake` | 20.8 gb/s, 1.4 ulp | 19.3 gb/s, 1.7 ulp | 16.5 gb/s, 1.5 ulp |
146
+ | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
147
+ | `nk_each_sum_bf16_serial` | 0.175 gb/s, 0 ulp | 0.178 gb/s, 0 ulp | 0.173 gb/s, 0 ulp |
148
+ | `nk_each_sum_bf16_haswell` | 7.97 gb/s, 0 ulp | 9.33 gb/s, 0 ulp | 10.3 gb/s, 0 ulp |
149
+ | `nk_each_sum_bf16_skylake` | 11.4 gb/s, 0 ulp | 11.2 gb/s, 0 ulp | 14.1 gb/s, 0 ulp |
150
+ | `nk_each_scale_bf16_serial` | 0.128 gb/s, 0 ulp | 0.119 gb/s, 0 ulp | 0.132 gb/s, 0 ulp |
151
+ | `nk_each_scale_bf16_haswell` | 6.08 gb/s, 0 ulp | 6.55 gb/s, 0 ulp | 6.92 gb/s, 0 ulp |
152
+ | `nk_each_scale_bf16_skylake` | 7.43 gb/s, 0 ulp | 8.04 gb/s, 0 ulp | 8.45 gb/s, 0 ulp |
153
+ | `nk_each_blend_bf16_serial` | 0.211 gb/s, 0 ulp | 0.204 gb/s, 0 ulp | 0.224 gb/s, 0 ulp |
154
+ | `nk_each_blend_bf16_haswell` | 8.58 gb/s, 2.2 ulp | 9.44 gb/s, 1.5 ulp | 10.2 gb/s, 1.5 ulp |
155
+ | `nk_each_blend_bf16_skylake` | 10.3 gb/s, 2.3 ulp | 11.9 gb/s, 1.3 ulp | 13.4 gb/s, 1.5 ulp |
156
+ | `nk_each_fma_bf16_serial` | 0.264 gb/s, 0 ulp | 0.260 gb/s, 0 ulp | 0.256 gb/s, 0 ulp |
157
+ | `nk_each_fma_bf16_haswell` | 10.9 gb/s, 1.5 ulp | 10.3 gb/s, 0.9 ulp | 11.4 gb/s, 1.0 ulp |
158
+ | `nk_each_fma_bf16_skylake` | 14.1 gb/s, 1.2 ulp | 13.0 gb/s, 0.7 ulp | 15.8 gb/s, 1.1 ulp |
159
+ | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
160
+ | `nk_each_sum_f16_serial` | 33.7 gb/s, 0 ulp | 16.1 gb/s, 0 ulp | 18.8 gb/s, 0 ulp |
161
+ | `nk_each_sum_f16_haswell` | 14.4 gb/s, 0 ulp | 11.8 gb/s, 0 ulp | 9.84 gb/s, 0 ulp |
162
+ | `nk_each_sum_f16_sapphire` | 39.2 gb/s, 0 ulp | 17.0 gb/s, 0 ulp | 18.9 gb/s, 0 ulp |
163
+ | `nk_each_scale_f16_serial` | 0.423 gb/s, 0 ulp | 0.282 gb/s, 0 ulp | 0.409 gb/s, 0 ulp |
164
+ | `nk_each_scale_f16_haswell` | 8.92 gb/s, 0 ulp | 8.59 gb/s, 0 ulp | 8.15 gb/s, 0 ulp |
165
+ | `nk_each_scale_f16_skylake` | 17.0 gb/s, 0 ulp | 10.7 gb/s, 0 ulp | 12.1 gb/s, 0 ulp |
166
+ | `nk_each_blend_f16_serial` | 0.769 gb/s, 1.3 ulp | 0.669 gb/s, 1.6 ulp | 0.792 gb/s, 1.5 ulp |
167
+ | `nk_each_blend_f16_haswell` | 13.5 gb/s, 1.2 ulp | 11.0 gb/s, 1.4 ulp | 11.7 gb/s, 1.5 ulp |
168
+ | `nk_each_blend_f16_skylake` | 16.9 gb/s, 1.3 ulp | 13.9 gb/s, 1.4 ulp | 14.2 gb/s, 1.1 ulp |
169
+ | `nk_each_fma_f16_serial` | 0.965 gb/s, 1.0 ulp | 0.787 gb/s, 1.1 ulp | 0.952 gb/s, 1.2 ulp |
170
+ | `nk_each_fma_f16_haswell` | 15.2 gb/s, 1.4 ulp | 13.6 gb/s, 1.0 ulp | 15.7 gb/s, 1.1 ulp |
171
+ | `nk_each_fma_f16_skylake` | 16.3 gb/s, 1.3 ulp | 16.2 gb/s, 1.3 ulp | 15.3 gb/s, 1.1 ulp |
172
+ | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
173
+ | `nk_each_sum_e4m3_serial` | 0.0966 gb/s, 0 ulp | 0.0943 gb/s, 0 ulp | 0.0946 gb/s, 0 ulp |
174
+ | `nk_each_sum_e4m3_haswell` | 0.895 gb/s, 0 ulp | 0.772 gb/s, 0 ulp | 0.824 gb/s, 0 ulp |
175
+ | `nk_each_sum_e4m3_skylake` | 1.48 gb/s, 0 ulp | 1.34 gb/s, 0 ulp | 1.40 gb/s, 0 ulp |
176
+ | `nk_each_sum_e4m3_sapphire` | 2.12 gb/s, 0 ulp | 1.89 gb/s, 0 ulp | 2.13 gb/s, 0 ulp |
177
+ | `nk_each_scale_e4m3_serial` | 0.0550 gb/s, 0 ulp | 0.0543 gb/s, 0 ulp | 0.0570 gb/s, 0 ulp |
178
+ | `nk_each_scale_e4m3_haswell` | 0.495 gb/s, 0 ulp | 0.532 gb/s, 0 ulp | 0.540 gb/s, 0 ulp |
179
+ | `nk_each_scale_e4m3_skylake` | 1.05 gb/s, 0 ulp | 1.02 gb/s, 0 ulp | 1.10 gb/s, 0 ulp |
180
+ | `nk_each_blend_e4m3_serial` | 0.0889 gb/s, 0 ulp | 0.0927 gb/s, 0 ulp | 0.0876 gb/s, 0 ulp |
181
+ | `nk_each_blend_e4m3_haswell` | 0.807 gb/s, 0.6 ulp | 0.756 gb/s, 0 ulp | 0.789 gb/s, 0 ulp |
182
+ | `nk_each_blend_e4m3_skylake` | 1.50 gb/s, 0 ulp | 1.44 gb/s, 0 ulp | 1.48 gb/s, 0 ulp |
183
+ | `nk_each_fma_e4m3_serial` | 0.120 gb/s, 0 ulp | 0.118 gb/s, 0.9 ulp | 0.115 gb/s, 0 ulp |
184
+ | `nk_each_fma_e4m3_haswell` | 0.989 gb/s, 0 ulp | 0.909 gb/s, 0 ulp | 0.967 gb/s, 0.5 ulp |
185
+ | `nk_each_fma_e4m3_skylake` | 1.89 gb/s, 0 ulp | 1.74 gb/s, 0 ulp | 1.80 gb/s, 0 ulp |
186
+ | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
187
+ | `nk_each_sum_e5m2_serial` | 0.115 gb/s, 0 ulp | 0.114 gb/s, 0 ulp | 0.110 gb/s, 0 ulp |
188
+ | `nk_each_sum_e5m2_haswell` | 1.04 gb/s, 0 ulp | 0.987 gb/s, 0 ulp | 0.972 gb/s, 0 ulp |
189
+ | `nk_each_sum_e5m2_skylake` | 1.71 gb/s, 0 ulp | 1.76 gb/s, 0 ulp | 1.79 gb/s, 0 ulp |
190
+ | `nk_each_scale_e5m2_serial` | 0.0593 gb/s, 0 ulp | 0.0630 gb/s, 0 ulp | 0.0630 gb/s, 0 ulp |
191
+ | `nk_each_scale_e5m2_haswell` | 0.601 gb/s, 0 ulp | 0.611 gb/s, 0 ulp | 0.588 gb/s, 0 ulp |
192
+ | `nk_each_scale_e5m2_skylake` | 1.11 gb/s, 0 ulp | 1.11 gb/s, 0 ulp | 1.12 gb/s, 0 ulp |
193
+ | `nk_each_blend_e5m2_serial` | 0.108 gb/s, 0 ulp | 0.113 gb/s, 0 ulp | 0.114 gb/s, 50 ulp |
194
+ | `nk_each_blend_e5m2_haswell` | 0.999 gb/s, 0 ulp | 0.895 gb/s, 0 ulp | 0.951 gb/s, 0 ulp |
195
+ | `nk_each_blend_e5m2_skylake` | 1.77 gb/s, 0 ulp | 1.65 gb/s, 0 ulp | 1.72 gb/s, 0 ulp |
196
+ | `nk_each_fma_e5m2_serial` | 0.155 gb/s, 5.1 ulp | 0.146 gb/s, 0 ulp | 0.149 gb/s, 0 ulp |
197
+ | `nk_each_fma_e5m2_haswell` | 1.24 gb/s, 0 ulp | 1.19 gb/s, 0 ulp | 1.25 gb/s, 0 ulp |
198
+ | `nk_each_fma_e5m2_skylake` | 2.36 gb/s, 0 ulp | 2.01 gb/s, 0 ulp | 2.11 gb/s, 0 ulp |
199
+ | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
200
+ | `nk_each_sum_e2m3_serial` | 0.109 gb/s, 0 ulp | 0.105 gb/s, 0 ulp | 0.110 gb/s, 0 ulp |
201
+ | `nk_each_scale_e2m3_serial` | 0.0500 gb/s, 0 ulp | 0.0474 gb/s, 0 ulp | 0.0495 gb/s, 0 ulp |
202
+ | `nk_each_blend_e2m3_serial` | 0.0864 gb/s, 0 ulp | 0.0888 gb/s, 0 ulp | 0.0908 gb/s, 0 ulp |
203
+ | `nk_each_fma_e2m3_serial` | 0.133 gb/s, 0 ulp | 0.128 gb/s, 0 ulp | 0.128 gb/s, 0 ulp |
204
+ | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
205
+ | `nk_each_sum_e3m2_serial` | 0.118 gb/s, 0 ulp | 0.115 gb/s, 0 ulp | 0.106 gb/s, 0 ulp |
206
+ | `nk_each_scale_e3m2_serial` | 0.0574 gb/s, 0 ulp | 0.0555 gb/s, 0 ulp | 0.0548 gb/s, 0 ulp |
207
+ | `nk_each_blend_e3m2_serial` | 0.110 gb/s, 0 ulp | 0.105 gb/s, 0 ulp | 0.0956 gb/s, 0 ulp |
208
+ | `nk_each_fma_e3m2_serial` | 0.151 gb/s, 0 ulp | 0.149 gb/s, 0 ulp | 0.138 gb/s, 0 ulp |
209
+ | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
210
+ | `nk_each_sum_i8_serial` | 21.7 gb/s | 15.9 gb/s | 12.5 gb/s |
211
+ | `nk_each_sum_i8_haswell` | 37.7 gb/s | 15.7 gb/s | 16.5 gb/s |
212
+ | `nk_each_sum_i8_icelake` | 47.2 gb/s | 17.7 gb/s | 17.7 gb/s |
213
+ | `nk_each_scale_i8_serial` | 2.34 gb/s | 1.50 gb/s | 1.70 gb/s |
214
+ | `nk_each_scale_i8_haswell` | 3.91 gb/s | 3.93 gb/s | 3.64 gb/s |
215
+ | `nk_each_scale_i8_skylake` | 6.74 gb/s | 6.70 gb/s | 6.93 gb/s |
216
+ | `nk_each_scale_i8_sapphire` | 23.0 gb/s | 11.4 gb/s | 10.8 gb/s |
217
+ | `nk_each_blend_i8_serial` | 3.66 gb/s | 2.23 gb/s | 2.60 gb/s |
218
+ | `nk_each_blend_i8_haswell` | 5.95 gb/s | 5.37 gb/s | 6.37 gb/s |
219
+ | `nk_each_blend_i8_sapphire` | 32.4 gb/s | 17.7 gb/s | 15.2 gb/s |
220
+ | `nk_each_fma_i8_serial` | 4.49 gb/s | 2.63 gb/s | 2.98 gb/s |
221
+ | `nk_each_fma_i8_haswell` | 7.36 gb/s | 6.84 gb/s | 7.15 gb/s |
222
+ | `nk_each_fma_i8_skylake` | 11.2 gb/s | 9.45 gb/s | 10.1 gb/s |
223
+ | `nk_each_fma_i8_sapphire` | 32.2 gb/s | 21.2 gb/s | 18.7 gb/s |
224
+ | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
225
+ | `nk_each_sum_u8_serial` | 17.0 gb/s | 13.7 gb/s | 12.5 gb/s |
226
+ | `nk_each_sum_u8_haswell` | 42.6 gb/s | 15.7 gb/s | 15.4 gb/s |
227
+ | `nk_each_sum_u8_icelake` | 45.9 gb/s | 17.3 gb/s | 17.4 gb/s |
228
+ | `nk_each_scale_u8_serial` | 2.11 gb/s | 2.03 gb/s | 1.86 gb/s |
229
+ | `nk_each_scale_u8_haswell` | 3.91 gb/s | 3.89 gb/s | 4.27 gb/s |
230
+ | `nk_each_scale_u8_skylake` | 6.93 gb/s | 5.99 gb/s | 6.70 gb/s |
231
+ | `nk_each_scale_u8_sapphire` | 24.7 gb/s | 12.1 gb/s | 11.9 gb/s |
232
+ | `nk_each_blend_u8_serial` | 3.23 gb/s | 2.62 gb/s | 3.43 gb/s |
233
+ | `nk_each_blend_u8_haswell` | 4.87 gb/s | 5.10 gb/s | 5.61 gb/s |
234
+ | `nk_each_blend_u8_sapphire` | 39.8 gb/s | 18.1 gb/s | 16.5 gb/s |
235
+ | `nk_each_fma_u8_serial` | 3.19 gb/s | 3.92 gb/s | 4.54 gb/s |
236
+ | `nk_each_fma_u8_haswell` | 6.98 gb/s | 6.29 gb/s | 7.62 gb/s |
237
+ | `nk_each_fma_u8_skylake` | 9.66 gb/s | 9.21 gb/s | 10.3 gb/s |
238
+ | `nk_each_fma_u8_sapphire` | 25.3 gb/s | 21.2 gb/s | 19.0 gb/s |
239
+ | __i16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
240
+ | `nk_each_sum_i16_serial` | 12.0 gb/s | 11.6 gb/s | 14.2 gb/s |
241
+ | `nk_each_sum_i16_haswell` | 26.1 gb/s | 16.2 gb/s | 16.8 gb/s |
242
+ | `nk_each_sum_i16_icelake` | 39.3 gb/s | 18.2 gb/s | 18.0 gb/s |
243
+ | `nk_each_scale_i16_serial` | 3.19 gb/s | 4.25 gb/s | 3.77 gb/s |
244
+ | `nk_each_scale_i16_haswell` | 7.07 gb/s | 7.42 gb/s | 7.69 gb/s |
245
+ | `nk_each_scale_i16_skylake` | 12.8 gb/s | 8.94 gb/s | 10.5 gb/s |
246
+ | `nk_each_blend_i16_serial` | 6.00 gb/s | 6.04 gb/s | 6.37 gb/s |
247
+ | `nk_each_fma_i16_serial` | 8.37 gb/s | 7.64 gb/s | 7.07 gb/s |
248
+ | `nk_each_fma_i16_haswell` | 10.7 gb/s | 12.5 gb/s | 13.4 gb/s |
249
+ | `nk_each_fma_i16_skylake` | 18.3 gb/s | 15.7 gb/s | 19.0 gb/s |
250
+ | __u16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
251
+ | `nk_each_sum_u16_serial` | 12.2 gb/s | 11.8 gb/s | 13.7 gb/s |
252
+ | `nk_each_sum_u16_haswell` | 24.1 gb/s | 14.4 gb/s | 17.7 gb/s |
253
+ | `nk_each_sum_u16_icelake` | 39.7 gb/s | 18.2 gb/s | 18.4 gb/s |
254
+ | `nk_each_scale_u16_serial` | 4.44 gb/s | 5.28 gb/s | 5.48 gb/s |
255
+ | `nk_each_scale_u16_haswell` | 7.82 gb/s | 7.21 gb/s | 7.66 gb/s |
256
+ | `nk_each_scale_u16_skylake` | 15.5 gb/s | 11.6 gb/s | 9.04 gb/s |
257
+ | `nk_each_blend_u16_serial` | 7.84 gb/s | 7.08 gb/s | 9.15 gb/s |
258
+ | `nk_each_fma_u16_serial` | 9.01 gb/s | 8.19 gb/s | 9.37 gb/s |
259
+ | `nk_each_fma_u16_haswell` | 11.2 gb/s | 12.2 gb/s | 13.9 gb/s |
260
+ | `nk_each_fma_u16_skylake` | 22.6 gb/s | 20.8 gb/s | 17.5 gb/s |
261
+ | __i32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
262
+ | `nk_each_sum_i32_serial` | 12.6 gb/s | 12.7 gb/s | 10.1 gb/s |
263
+ | `nk_each_sum_i32_haswell` | 14.6 gb/s | 15.3 gb/s | 13.9 gb/s |
264
+ | `nk_each_sum_i32_icelake` | 17.1 gb/s | 18.5 gb/s | 15.4 gb/s |
265
+ | `nk_each_scale_i32_serial` | 3.04 gb/s | 3.28 gb/s | 3.44 gb/s |
266
+ | `nk_each_scale_i32_haswell` | 8.94 gb/s | 8.72 gb/s | 8.09 gb/s |
267
+ | `nk_each_scale_i32_skylake` | 11.1 gb/s | 12.6 gb/s | 11.3 gb/s |
268
+ | `nk_each_blend_i32_serial` | 4.99 gb/s | 5.99 gb/s | 5.06 gb/s |
269
+ | `nk_each_fma_i32_serial` | 6.33 gb/s | 6.38 gb/s | 6.40 gb/s |
270
+ | `nk_each_fma_i32_haswell` | 13.5 gb/s | 16.4 gb/s | 11.2 gb/s |
271
+ | `nk_each_fma_i32_skylake` | 20.1 gb/s | 21.3 gb/s | 16.0 gb/s |
272
+ | __u32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
273
+ | `nk_each_sum_u32_serial` | 13.6 gb/s | 11.6 gb/s | 9.94 gb/s |
274
+ | `nk_each_sum_u32_haswell` | 16.0 gb/s | 17.7 gb/s | 12.7 gb/s |
275
+ | `nk_each_sum_u32_icelake` | 17.4 gb/s | 19.0 gb/s | 14.8 gb/s |
276
+ | `nk_each_scale_u32_serial` | 2.13 gb/s | 3.22 gb/s | 2.74 gb/s |
277
+ | `nk_each_scale_u32_haswell` | 8.20 gb/s | 9.40 gb/s | 9.18 gb/s |
278
+ | `nk_each_scale_u32_skylake` | 10.6 gb/s | 12.9 gb/s | 11.1 gb/s |
279
+ | `nk_each_blend_u32_serial` | 3.76 gb/s | 5.22 gb/s | 5.78 gb/s |
280
+ | `nk_each_fma_u32_serial` | 4.78 gb/s | 5.63 gb/s | 8.43 gb/s |
281
+ | `nk_each_fma_u32_haswell` | 13.7 gb/s | 16.1 gb/s | 12.1 gb/s |
282
+ | `nk_each_fma_u32_skylake` | 20.2 gb/s | 21.1 gb/s | 15.2 gb/s |
283
+ | __i64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
284
+ | `nk_each_sum_i64_serial` | 15.9 gb/s | 15.8 gb/s | 10.3 gb/s |
285
+ | `nk_each_sum_i64_icelake` | 17.7 gb/s | 19.2 gb/s | 11.1 gb/s |
286
+ | `nk_each_scale_i64_serial` | 7.44 gb/s | 9.17 gb/s | 8.72 gb/s |
287
+ | `nk_each_scale_i64_skylake` | 11.8 gb/s | 13.8 gb/s | 9.10 gb/s |
288
+ | `nk_each_blend_i64_serial` | 10.9 gb/s | 14.3 gb/s | 10.5 gb/s |
289
+ | `nk_each_fma_i64_serial` | 13.5 gb/s | 19.0 gb/s | 11.8 gb/s |
290
+ | `nk_each_fma_i64_skylake` | 21.7 gb/s | 22.2 gb/s | 11.8 gb/s |
291
+ | __u64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
292
+ | `nk_each_sum_u64_serial` | 14.7 gb/s | 16.9 gb/s | 10.5 gb/s |
293
+ | `nk_each_sum_u64_icelake` | 18.1 gb/s | 19.2 gb/s | 12.4 gb/s |
294
+ | `nk_each_scale_u64_serial` | 7.99 gb/s | 9.79 gb/s | 8.82 gb/s |
295
+ | `nk_each_scale_u64_skylake` | 11.6 gb/s | 13.9 gb/s | 7.24 gb/s |
296
+ | `nk_each_blend_u64_serial` | 11.9 gb/s | 16.5 gb/s | 13.8 gb/s |
297
+ | `nk_each_fma_u64_serial` | 15.3 gb/s | 21.6 gb/s | 14.0 gb/s |
298
+ | `nk_each_fma_u64_skylake` | 21.6 gb/s | 21.8 gb/s | 11.4 gb/s |
299
+ | __f64c__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
300
+ | `nk_each_scale_f64c_serial` | 10.6 gb/s, 3.4 ulp | 7.04 gb/s, 2.6 ulp | 5.63 gb/s, 2.0 ulp |
301
+ | `nk_each_scale_f64c_haswell` | 9.89 gb/s, 3.5 ulp | 6.45 gb/s, 2.0 ulp | 5.90 gb/s, 2.1 ulp |
302
+ | `nk_each_scale_f64c_skylake` | 9.29 gb/s, 3.5 ulp | 6.01 gb/s, 2.0 ulp | 5.67 gb/s, 2.0 ulp |
303
+ | `nk_each_blend_f64c_serial` | 16.2 gb/s, 2.5 ulp | 8.80 gb/s, 2.4 ulp | 8.23 gb/s, 2.6 ulp |
304
+ | `nk_each_blend_f64c_haswell` | 14.0 gb/s, 2.5 ulp | 8.26 gb/s, 2.4 ulp | 8.59 gb/s, 2.7 ulp |
305
+ | `nk_each_blend_f64c_skylake` | 14.3 gb/s, 2.5 ulp | 8.82 gb/s, 2.7 ulp | 7.33 gb/s, 2.7 ulp |
306
+ | `nk_each_fma_f64c_serial` | 17.4 gb/s, 4.5 ulp | 10.1 gb/s, 3.2 ulp | 9.45 gb/s, 2.8 ulp |
307
+ | `nk_each_fma_f64c_haswell` | 14.5 gb/s, 4.9 ulp | 8.86 gb/s, 3.1 ulp | 9.31 gb/s, 2.7 ulp |
308
+ | `nk_each_fma_f64c_skylake` | 16.1 gb/s, 4.0 ulp | 8.93 gb/s, 3.4 ulp | 10.4 gb/s, 2.8 ulp |
309
+ | __f32c__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
310
+ | `nk_each_scale_f32c_serial` | 10.1 gb/s, 2.1 ulp | 9.28 gb/s, 1.8 ulp | 6.20 gb/s, 2.1 ulp |
311
+ | `nk_each_scale_f32c_haswell` | 8.81 gb/s, 2.1 ulp | 8.87 gb/s, 1.8 ulp | 7.73 gb/s, 2.1 ulp |
312
+ | `nk_each_scale_f32c_skylake` | 9.67 gb/s, 2.2 ulp | 7.61 gb/s, 1.8 ulp | 7.09 gb/s, 2.1 ulp |
313
+ | `nk_each_blend_f32c_serial` | 14.6 gb/s, 6.9 ulp | 10.8 gb/s, 2.8 ulp | 8.43 gb/s, 8.7 ulp |
314
+ | `nk_each_blend_f32c_haswell` | 13.0 gb/s, 8.7 ulp | 12.1 gb/s, 2.9 ulp | 10.5 gb/s, 9.7 ulp |
315
+ | `nk_each_blend_f32c_skylake` | 14.0 gb/s, 7.7 ulp | 9.28 gb/s, 2.7 ulp | 9.47 gb/s, 10.2 ulp |
316
+ | `nk_each_fma_f32c_serial` | 15.7 gb/s, 8.8 ulp | 11.96 gb/s, 3.9 ulp | 9.33 gb/s, 8.5 ulp |
317
+ | `nk_each_fma_f32c_haswell` | 14.1 gb/s, 6.6 ulp | 10.3 gb/s, 2.9 ulp | 9.77 gb/s, 7.2 ulp |
318
+ | `nk_each_fma_f32c_skylake` | 15.5 gb/s, 9.2 ulp | 10.3 gb/s, 3.8 ulp | 9.95 gb/s, 8.3 ulp |
319
+
320
+ ### Apple M4
321
+
322
+ #### Native
323
+
324
+ | Kernel | 256 | 1024 | 4096 |
325
+ | :----------------------------- | -----------------------: | -----------------------: | -----------------------: |
326
+ | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
327
+ | `nk_each_sum_f64_serial` | 19.0 gb/s, 0 ulp | 14.0 gb/s, 0 ulp | 14.8 gb/s, 0 ulp |
328
+ | `nk_each_sum_f64_neon` | 14.4 gb/s, 0 ulp | 16.6 gb/s, 0 ulp | 23.0 gb/s, 0 ulp |
329
+ | `nk_each_scale_f64_serial` | 35.8 gb/s, 0 ulp | 16.3 gb/s, 0 ulp | 17.4 gb/s, 0 ulp |
330
+ | `nk_each_scale_f64_neon` | 46.2 gb/s, 0 ulp | 25.8 gb/s, 0 ulp | 33.0 gb/s, 0 ulp |
331
+ | `nk_each_blend_f64_serial` | 18.3 gb/s, 0 ulp | 12.4 gb/s, 0 ulp | 13.0 gb/s, 0 ulp |
332
+ | `nk_each_blend_f64_neon` | 34.0 gb/s, 1.9 ulp | 17.8 gb/s, 2.7 ulp | 24.0 gb/s, 1.8 ulp |
333
+ | `nk_each_fma_f64_serial` | 12.3 gb/s, 0 ulp | 9.51 gb/s, 0 ulp | 10.1 gb/s, 0 ulp |
334
+ | `nk_each_fma_f64_neon` | 23.5 gb/s, 1.4 ulp | 12.9 gb/s, 1.6 ulp | 19.5 gb/s, 1.6 ulp |
335
+ | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
336
+ | `nk_each_sum_f32_serial` | 21.1 gb/s, 0 ulp | 10.4 gb/s, 0 ulp | 12.9 gb/s, 0 ulp |
337
+ | `nk_each_sum_f32_neon` | 32.7 gb/s, 0 ulp | 34.7 gb/s, 0 ulp | 28.1 gb/s, 0 ulp |
338
+ | `nk_each_scale_f32_serial` | 24.9 gb/s, 0 ulp | 26.8 gb/s, 0 ulp | 17.7 gb/s, 0 ulp |
339
+ | `nk_each_scale_f32_neon` | 50.4 gb/s, 0 ulp | 42.2 gb/s, 0 ulp | 56.5 gb/s, 0 ulp |
340
+ | `nk_each_blend_f32_serial` | 16.8 gb/s, 26 ulp | 9.57 gb/s, 26 ulp | 13.7 gb/s, 2.0 ulp |
341
+ | `nk_each_blend_f32_neon` | 16.0 gb/s, 1.7 ulp | 9.46 gb/s, 1.6 ulp | 11.0 gb/s, 1.6 ulp |
342
+ | `nk_each_fma_f32_serial` | 13.6 gb/s, 2.1 ulp | 7.80 gb/s, 2.5 ulp | 9.92 gb/s, 2.2 ulp |
343
+ | `nk_each_fma_f32_neon` | 13.2 gb/s, 2.1 ulp | 8.85 gb/s, 21 ulp | 8.78 gb/s, 1.8 ulp |
344
+ | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
345
+ | `nk_each_sum_bf16_serial` | 10.6 gb/s, 0 ulp | 9.80 gb/s, 0 ulp | 10.6 gb/s, 0 ulp |
346
+ | `nk_each_sum_bf16_neonbfdot` | 24.7 gb/s, 0 ulp | 24.4 gb/s, 0 ulp | 23.4 gb/s, 0 ulp |
347
+ | `nk_each_scale_bf16_serial` | 13.1 gb/s, 0 ulp | 11.0 gb/s, 0 ulp | 12.9 gb/s, 0 ulp |
348
+ | `nk_each_scale_bf16_neonbfdot` | 28.3 gb/s, 0 ulp | 27.2 gb/s, 0 ulp | 19.5 gb/s, 0 ulp |
349
+ | `nk_each_blend_bf16_serial` | 9.13 gb/s, 28 ulp | 8.36 gb/s, 26 ulp | 8.42 gb/s, 2.2 ulp |
350
+ | `nk_each_blend_bf16_neonbfdot` | 11.6 gb/s, 29 ulp | 9.57 gb/s, 29 ulp | 8.71 gb/s, 2.2 ulp |
351
+ | `nk_each_fma_bf16_serial` | 7.68 gb/s, 2.1 ulp | 6.56 gb/s, 2.0 ulp | 6.37 gb/s, 33 ulp |
352
+ | `nk_each_fma_bf16_neonbfdot` | 8.74 gb/s, 1.2 ulp | 7.53 gb/s, 1.2 ulp | 6.22 gb/s, 1.5 ulp |
353
+ | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
354
+ | `nk_each_sum_f16_serial` | 19.0 gb/s, 0 ulp | 17.1 gb/s, 0 ulp | 11.0 gb/s, 0 ulp |
355
+ | `nk_each_sum_f16_neonhalf` | 28.9 gb/s, 0 ulp | 33.8 gb/s, 0 ulp | 20.2 gb/s, 0 ulp |
356
+ | `nk_each_scale_f16_serial` | 19.4 gb/s, 0 ulp | 18.1 gb/s, 0 ulp | 14.2 gb/s, 0 ulp |
357
+ | `nk_each_scale_f16_neonhalf` | 51.4 gb/s, 87.3K ulp | 49.7 gb/s, 84.8K ulp | 36.1 gb/s, 87.3K ulp |
358
+ | `nk_each_blend_f16_serial` | 12.0 gb/s, 2.0 ulp | 10.3 gb/s, 2.0 ulp | 9.46 gb/s, 2.3 ulp |
359
+ | `nk_each_blend_f16_neonhalf` | 18.0 gb/s, 91.6K ulp | 15.5 gb/s, 92.6K ulp | 10.1 gb/s, 91.9K ulp |
360
+ | `nk_each_fma_f16_serial` | 9.94 gb/s, 2.1 ulp | 8.29 gb/s, 1.8 ulp | 8.52 gb/s, 2.2 ulp |
361
+ | `nk_each_fma_f16_neonhalf` | 14.7 gb/s, 97.1K ulp | 11.1 gb/s, 96.7K ulp | 10.7 gb/s, 99.2K ulp |
362
+ | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
363
+ | `nk_each_sum_e4m3_serial` | 0.122 gb/s, 0 ulp | 0.123 gb/s, 0 ulp | 0.125 gb/s, 0 ulp |
364
+ | `nk_each_sum_e4m3_neon` | 0.856 gb/s, 0 ulp | 0.896 gb/s, 0 ulp | 0.909 gb/s, 0 ulp |
365
+ | `nk_each_scale_e4m3_serial` | 0.108 gb/s, 0 ulp | 0.109 gb/s, 0 ulp | 0.109 gb/s, 0 ulp |
366
+ | `nk_each_scale_e4m3_neon` | 1.03 gb/s, 0 ulp | 1.11 gb/s, 0 ulp | 1.12 gb/s, 0 ulp |
367
+ | `nk_each_blend_e4m3_serial` | 0.0907 gb/s, 0.4 ulp | 0.0907 gb/s, 1.1 ulp | 0.0885 gb/s, 2.8 ulp |
368
+ | `nk_each_blend_e4m3_neon` | 0.845 gb/s, 0.1 ulp | 0.905 gb/s, 0 ulp | 0.910 gb/s, 0 ulp |
369
+ | `nk_each_fma_e4m3_serial` | 0.0808 gb/s, 0.7 ulp | 0.0795 gb/s, 0.6 ulp | 0.0790 gb/s, 2.2 ulp |
370
+ | `nk_each_fma_e4m3_neon` | 0.707 gb/s, 0.1 ulp | 0.766 gb/s, 0.6 ulp | 0.765 gb/s, 1.0 ulp |
371
+ | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
372
+ | `nk_each_sum_e5m2_serial` | 0.163 gb/s, 0 ulp | 0.160 gb/s, 0 ulp | 0.161 gb/s, 0 ulp |
373
+ | `nk_each_sum_e5m2_neon` | 1.70 gb/s, 0 ulp | 1.87 gb/s, 0 ulp | 1.88 gb/s, 0 ulp |
374
+ | `nk_each_scale_e5m2_serial` | 0.137 gb/s, 0 ulp | 0.134 gb/s, 0 ulp | 0.134 gb/s, 0 ulp |
375
+ | `nk_each_scale_e5m2_neon` | 1.78 gb/s, 0 ulp | 1.94 gb/s, 0 ulp | 1.96 gb/s, 0 ulp |
376
+ | `nk_each_blend_e5m2_serial` | 0.126 gb/s, 0 ulp | 0.123 gb/s, 0 ulp | 0.123 gb/s, 4.9 ulp |
377
+ | `nk_each_blend_e5m2_neon` | 1.62 gb/s, 0.7 ulp | 1.79 gb/s, 0 ulp | 1.79 gb/s, 0 ulp |
378
+ | `nk_each_fma_e5m2_serial` | 0.117 gb/s, 1.9 ulp | 0.114 gb/s, 0.9 ulp | 0.114 gb/s, 4.0 ulp |
379
+ | `nk_each_fma_e5m2_neon` | 1.48 gb/s, 0 ulp | 1.62 gb/s, 1.3 ulp | 1.62 gb/s, 0 ulp |
380
+ | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
381
+ | `nk_each_sum_e2m3_serial` | 0.124 gb/s, ? ulp | 0.117 gb/s, ? ulp | 0.115 gb/s, ? ulp |
382
+ | `nk_each_scale_e2m3_serial` | 0.0793 gb/s, ? ulp | 0.0769 gb/s, ? ulp | 0.0769 gb/s, ? ulp |
383
+ | `nk_each_blend_e2m3_serial` | 0.0856 gb/s, ? ulp | 0.0828 gb/s, ? ulp | 0.0828 gb/s, ? ulp |
384
+ | `nk_each_fma_e2m3_serial` | 0.0884 gb/s, ? ulp | 0.0842 gb/s, ? ulp | 0.0852 gb/s, ? ulp |
385
+ | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
386
+ | `nk_each_sum_e3m2_serial` | 0.181 gb/s, ? ulp | 0.175 gb/s, ? ulp | 0.173 gb/s, ? ulp |
387
+ | `nk_each_scale_e3m2_serial` | 0.124 gb/s, ? ulp | 0.117 gb/s, ? ulp | 0.117 gb/s, ? ulp |
388
+ | `nk_each_blend_e3m2_serial` | 0.123 gb/s, ? ulp | 0.116 gb/s, ? ulp | 0.115 gb/s, ? ulp |
389
+ | `nk_each_fma_e3m2_serial` | 0.118 gb/s, ? ulp | 0.110 gb/s, ? ulp | 0.109 gb/s, ? ulp |
390
+ | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
391
+ | `nk_each_sum_i8_serial` | 20.6 gb/s | 24.2 gb/s | 12.6 gb/s |
392
+ | `nk_each_sum_i8_neonhalf` | 41.5 gb/s | 40.2 gb/s | 36.6 gb/s |
393
+ | `nk_each_scale_i8_serial` | 2.88 gb/s | 2.85 gb/s | 2.84 gb/s |
394
+ | `nk_each_scale_i8_neonhalf` | 24.4 gb/s | 22.7 gb/s | 12.0 gb/s |
395
+ | `nk_each_blend_i8_serial` | 2.29 gb/s | 2.32 gb/s | 2.24 gb/s |
396
+ | `nk_each_blend_i8_neonhalf` | 8.83 gb/s | 9.48 gb/s | 8.68 gb/s |
397
+ | `nk_each_fma_i8_serial` | 2.02 gb/s | 1.97 gb/s | 1.77 gb/s |
398
+ | `nk_each_fma_i8_neonhalf` | 7.27 gb/s | 6.97 gb/s | 4.79 gb/s |
399
+ | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
400
+ | `nk_each_sum_u8_serial` | 4.46 gb/s | 4.49 gb/s | 4.09 gb/s |
401
+ | `nk_each_sum_u8_neonhalf` | 35.0 gb/s | 38.1 gb/s | 20.2 gb/s |
402
+ | `nk_each_scale_u8_serial` | 3.38 gb/s | 3.41 gb/s | 3.34 gb/s |
403
+ | `nk_each_scale_u8_neonhalf` | 24.4 gb/s | 22.7 gb/s | 21.0 gb/s |
404
+ | `nk_each_blend_u8_serial` | 2.59 gb/s | 2.53 gb/s | 2.48 gb/s |
405
+ | `nk_each_blend_u8_neonhalf` | 9.21 gb/s | 8.53 gb/s | 8.00 gb/s |
406
+ | `nk_each_fma_u8_serial` | 2.17 gb/s | 2.16 gb/s | 1.86 gb/s |
407
+ | `nk_each_fma_u8_neonhalf` | 7.29 gb/s | 7.11 gb/s | 4.72 gb/s |
408
+ | __i16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
409
+ | `nk_each_sum_i16_serial` | 20.2 gb/s | 24.3 gb/s | 11.1 gb/s |
410
+ | `nk_each_sum_i16_neon` | 31.4 gb/s | 36.8 gb/s | 36.7 gb/s |
411
+ | `nk_each_scale_i16_serial` | 5.36 gb/s | 5.35 gb/s | 5.49 gb/s |
412
+ | `nk_each_scale_i16_neon` | 24.9 gb/s | 24.5 gb/s | 19.9 gb/s |
413
+ | `nk_each_blend_i16_serial` | 4.23 gb/s | 3.87 gb/s | 4.08 gb/s |
414
+ | `nk_each_fma_i16_serial` | 3.74 gb/s | 3.31 gb/s | 3.16 gb/s |
415
+ | `nk_each_fma_i16_neon` | 11.5 gb/s | 11.7 gb/s | 11.7 gb/s |
416
+ | __u16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
417
+ | `nk_each_sum_u16_serial` | 7.95 gb/s | 7.73 gb/s | 6.73 gb/s |
418
+ | `nk_each_sum_u16_neon` | 32.0 gb/s | 39.3 gb/s | 35.9 gb/s |
419
+ | `nk_each_scale_u16_serial` | 6.43 gb/s | 6.36 gb/s | 6.07 gb/s |
420
+ | `nk_each_scale_u16_neon` | 25.0 gb/s | 24.5 gb/s | 25.7 gb/s |
421
+ | `nk_each_blend_u16_serial` | 5.07 gb/s | 4.95 gb/s | 4.73 gb/s |
422
+ | `nk_each_fma_u16_serial` | 4.41 gb/s | 4.03 gb/s | 3.73 gb/s |
423
+ | `nk_each_fma_u16_neon` | 11.5 gb/s | 11.8 gb/s | 10.7 gb/s |
424
+ | __i32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
425
+ | `nk_each_sum_i32_serial` | 20.9 gb/s | 14.3 gb/s | 12.6 gb/s |
426
+ | `nk_each_sum_i32_neon` | 38.8 gb/s | 39.0 gb/s | 36.2 gb/s |
427
+ | `nk_each_scale_i32_serial` | 4.34 gb/s | 4.14 gb/s | 4.24 gb/s |
428
+ | `nk_each_scale_i32_neon` | 22.6 gb/s | 25.6 gb/s | 25.6 gb/s |
429
+ | `nk_each_blend_i32_serial` | 4.02 gb/s | 3.57 gb/s | 3.84 gb/s |
430
+ | `nk_each_fma_i32_serial` | 3.47 gb/s | 3.01 gb/s | 3.35 gb/s |
431
+ | `nk_each_fma_i32_neon` | 11.5 gb/s | 11.7 gb/s | 11.7 gb/s |
432
+ | __u32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
433
+ | `nk_each_sum_u32_serial` | 13.1 gb/s | 13.4 gb/s | 9.76 gb/s |
434
+ | `nk_each_sum_u32_neon` | 33.2 gb/s | 28.1 gb/s | 27.0 gb/s |
435
+ | `nk_each_scale_u32_serial` | 4.72 gb/s | 4.76 gb/s | 5.56 gb/s |
436
+ | `nk_each_scale_u32_neon` | 22.5 gb/s | 25.8 gb/s | 25.6 gb/s |
437
+ | `nk_each_blend_u32_serial` | 4.70 gb/s | 4.38 gb/s | 4.70 gb/s |
438
+ | `nk_each_fma_u32_serial` | 4.03 gb/s | 4.04 gb/s | 3.98 gb/s |
439
+ | `nk_each_fma_u32_neon` | 11.4 gb/s | 11.8 gb/s | 10.7 gb/s |
440
+ | __i64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
441
+ | `nk_each_sum_i64_serial` | 16.7 gb/s | 15.5 gb/s | 11.2 gb/s |
442
+ | `nk_each_sum_i64_neon` | 38.5 gb/s | 39.4 gb/s | 28.6 gb/s |
443
+ | `nk_each_scale_i64_serial` | 6.94 gb/s | 7.26 gb/s | 8.77 gb/s |
444
+ | `nk_each_scale_i64_neon` | 50.0 gb/s | 50.6 gb/s | 44.5 gb/s |
445
+ | `nk_each_blend_i64_serial` | 9.02 gb/s | 7.98 gb/s | 8.72 gb/s |
446
+ | `nk_each_fma_i64_serial` | 7.19 gb/s | 7.29 gb/s | 7.32 gb/s |
447
+ | `nk_each_fma_i64_neon` | 28.4 gb/s | 28.2 gb/s | 20.6 gb/s |
448
+ | __u64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
449
+ | `nk_each_sum_u64_serial` | 17.8 gb/s | 11.8 gb/s | 11.7 gb/s |
450
+ | `nk_each_sum_u64_neon` | 39.2 gb/s | 37.7 gb/s | 28.6 gb/s |
451
+ | `nk_each_scale_u64_serial` | 9.80 gb/s | 8.70 gb/s | 11.8 gb/s |
452
+ | `nk_each_scale_u64_neon` | 38.6 gb/s | 50.9 gb/s | 43.7 gb/s |
453
+ | `nk_each_blend_u64_serial` | 9.31 gb/s | 8.52 gb/s | 8.76 gb/s |
454
+ | `nk_each_fma_u64_serial` | 7.59 gb/s | 7.91 gb/s | 6.65 gb/s |
455
+ | `nk_each_fma_u64_neon` | 19.1 gb/s | 27.7 gb/s | 20.6 gb/s |
456
+ | __f64c__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
457
+ | `nk_each_scale_f64c_serial` | 15.6 gb/s, 2.2 ulp | 15.8 gb/s, 1.9 ulp | 19.7 gb/s, 2.2 ulp |
458
+ | `nk_each_scale_f64c_neon` | 24.1 gb/s, 1.5 ulp | 51.2 gb/s, 1.5 ulp | 35.5 gb/s, 1.3 ulp |
459
+ | `nk_each_blend_f64c_serial` | 12.2 gb/s, 4.2 ulp | 10.3 gb/s, 3.0 ulp | 17.4 gb/s, 2.6 ulp |
460
+ | `nk_each_blend_f64c_neon` | 17.2 gb/s, 3.2 ulp | 24.7 gb/s, 3.0 ulp | 23.1 gb/s, 2.2 ulp |
461
+ | `nk_each_fma_f64c_serial` | 6.73 gb/s, 3.4 ulp | 8.11 gb/s, 5.2 ulp | 13.0 gb/s, 2.5 ulp |
462
+ | `nk_each_fma_f64c_neon` | 12.5 gb/s, 3.2 ulp | 18.8 gb/s, 3.0 ulp | 17.8 gb/s, 2.4 ulp |
463
+ | __f32c__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
464
+ | `nk_each_scale_f32c_serial` | 23.1 gb/s, 17 ulp | 13.3 gb/s, 2.4 ulp | 13.6 gb/s, 17 ulp |
465
+ | `nk_each_scale_f32c_neon` | 18.6 gb/s, 1.8 ulp | 38.3 gb/s, 1.6 ulp | 32.9 gb/s, 1.6 ulp |
466
+ | `nk_each_blend_f32c_serial` | 13.7 gb/s, 2.4 ulp | 8.74 gb/s, 2.6 ulp | 12.4 gb/s, 28 ulp |
467
+ | `nk_each_blend_f32c_neon` | 26.9 gb/s, 2.2 ulp | 21.8 gb/s, 2.2 ulp | 22.4 gb/s, 3.2 ulp |
468
+ | `nk_each_fma_f32c_serial` | 10.2 gb/s, 3.1 ulp | 8.04 gb/s, 78 ulp | 9.05 gb/s, 3.5 ulp |
469
+ | `nk_each_fma_f32c_neon` | 22.4 gb/s, 2.9 ulp | 17.1 gb/s, 4.1 ulp | 17.8 gb/s, 81 ulp |