numkong 7.0.0 → 7.4.2

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 (315) hide show
  1. package/README.md +197 -124
  2. package/binding.gyp +34 -484
  3. package/c/dispatch_bf16.c +59 -1
  4. package/c/dispatch_e2m3.c +41 -8
  5. package/c/dispatch_e3m2.c +49 -8
  6. package/c/dispatch_e4m3.c +51 -9
  7. package/c/dispatch_e5m2.c +45 -1
  8. package/c/dispatch_f16.c +79 -26
  9. package/c/dispatch_f16c.c +5 -5
  10. package/c/dispatch_f32.c +56 -0
  11. package/c/dispatch_f64.c +52 -0
  12. package/c/dispatch_i4.c +3 -0
  13. package/c/dispatch_i8.c +62 -3
  14. package/c/dispatch_other.c +18 -0
  15. package/c/dispatch_u1.c +54 -9
  16. package/c/dispatch_u4.c +3 -0
  17. package/c/dispatch_u8.c +64 -3
  18. package/c/numkong.c +3 -0
  19. package/include/README.md +79 -9
  20. package/include/numkong/attention/sapphireamx.h +278 -276
  21. package/include/numkong/attention/sme.h +983 -977
  22. package/include/numkong/attention.h +1 -1
  23. package/include/numkong/capabilities.h +289 -94
  24. package/include/numkong/cast/README.md +40 -40
  25. package/include/numkong/cast/diamond.h +64 -0
  26. package/include/numkong/cast/haswell.h +42 -194
  27. package/include/numkong/cast/icelake.h +42 -37
  28. package/include/numkong/cast/loongsonasx.h +252 -0
  29. package/include/numkong/cast/neon.h +216 -249
  30. package/include/numkong/cast/powervsx.h +449 -0
  31. package/include/numkong/cast/rvv.h +223 -274
  32. package/include/numkong/cast/sapphire.h +18 -18
  33. package/include/numkong/cast/serial.h +1018 -944
  34. package/include/numkong/cast/skylake.h +82 -23
  35. package/include/numkong/cast/v128relaxed.h +462 -105
  36. package/include/numkong/cast.h +24 -0
  37. package/include/numkong/cast.hpp +44 -0
  38. package/include/numkong/curved/README.md +17 -17
  39. package/include/numkong/curved/neon.h +131 -7
  40. package/include/numkong/curved/neonbfdot.h +6 -7
  41. package/include/numkong/curved/rvv.h +26 -26
  42. package/include/numkong/curved/smef64.h +186 -182
  43. package/include/numkong/curved.h +14 -18
  44. package/include/numkong/dot/README.md +154 -137
  45. package/include/numkong/dot/alder.h +43 -43
  46. package/include/numkong/dot/diamond.h +158 -0
  47. package/include/numkong/dot/genoa.h +4 -30
  48. package/include/numkong/dot/haswell.h +215 -180
  49. package/include/numkong/dot/icelake.h +190 -76
  50. package/include/numkong/dot/loongsonasx.h +671 -0
  51. package/include/numkong/dot/neon.h +124 -73
  52. package/include/numkong/dot/neonbfdot.h +11 -12
  53. package/include/numkong/dot/neonfhm.h +44 -46
  54. package/include/numkong/dot/neonfp8.h +323 -0
  55. package/include/numkong/dot/neonsdot.h +190 -76
  56. package/include/numkong/dot/powervsx.h +752 -0
  57. package/include/numkong/dot/rvv.h +92 -84
  58. package/include/numkong/dot/rvvbf16.h +12 -12
  59. package/include/numkong/dot/rvvhalf.h +12 -12
  60. package/include/numkong/dot/sapphire.h +4 -4
  61. package/include/numkong/dot/serial.h +66 -30
  62. package/include/numkong/dot/sierra.h +31 -31
  63. package/include/numkong/dot/skylake.h +142 -110
  64. package/include/numkong/dot/sve.h +217 -177
  65. package/include/numkong/dot/svebfdot.h +10 -10
  66. package/include/numkong/dot/svehalf.h +85 -41
  67. package/include/numkong/dot/svesdot.h +89 -0
  68. package/include/numkong/dot/v128relaxed.h +124 -89
  69. package/include/numkong/dot.h +114 -48
  70. package/include/numkong/dots/README.md +203 -203
  71. package/include/numkong/dots/alder.h +12 -9
  72. package/include/numkong/dots/diamond.h +86 -0
  73. package/include/numkong/dots/genoa.h +10 -4
  74. package/include/numkong/dots/haswell.h +63 -48
  75. package/include/numkong/dots/icelake.h +27 -18
  76. package/include/numkong/dots/loongsonasx.h +176 -0
  77. package/include/numkong/dots/neon.h +14 -11
  78. package/include/numkong/dots/neonbfdot.h +4 -3
  79. package/include/numkong/dots/neonfhm.h +11 -9
  80. package/include/numkong/dots/neonfp8.h +99 -0
  81. package/include/numkong/dots/neonsdot.h +48 -12
  82. package/include/numkong/dots/powervsx.h +194 -0
  83. package/include/numkong/dots/rvv.h +451 -344
  84. package/include/numkong/dots/sapphireamx.h +1028 -984
  85. package/include/numkong/dots/serial.h +213 -197
  86. package/include/numkong/dots/sierra.h +10 -7
  87. package/include/numkong/dots/skylake.h +47 -36
  88. package/include/numkong/dots/sme.h +2001 -2364
  89. package/include/numkong/dots/smebi32.h +175 -162
  90. package/include/numkong/dots/smef64.h +328 -323
  91. package/include/numkong/dots/v128relaxed.h +64 -41
  92. package/include/numkong/dots.h +573 -293
  93. package/include/numkong/dots.hpp +45 -43
  94. package/include/numkong/each/README.md +133 -137
  95. package/include/numkong/each/haswell.h +6 -6
  96. package/include/numkong/each/icelake.h +7 -7
  97. package/include/numkong/each/neon.h +76 -42
  98. package/include/numkong/each/neonbfdot.h +11 -12
  99. package/include/numkong/each/neonhalf.h +24 -116
  100. package/include/numkong/each/rvv.h +28 -28
  101. package/include/numkong/each/sapphire.h +27 -161
  102. package/include/numkong/each/serial.h +6 -6
  103. package/include/numkong/each/skylake.h +7 -7
  104. package/include/numkong/each/v128relaxed.h +562 -0
  105. package/include/numkong/each.h +148 -62
  106. package/include/numkong/each.hpp +2 -2
  107. package/include/numkong/geospatial/README.md +18 -18
  108. package/include/numkong/geospatial/haswell.h +365 -325
  109. package/include/numkong/geospatial/neon.h +350 -306
  110. package/include/numkong/geospatial/rvv.h +4 -4
  111. package/include/numkong/geospatial/skylake.h +376 -340
  112. package/include/numkong/geospatial/v128relaxed.h +366 -327
  113. package/include/numkong/geospatial.h +17 -17
  114. package/include/numkong/matrix.hpp +4 -4
  115. package/include/numkong/maxsim/README.md +14 -14
  116. package/include/numkong/maxsim/alder.h +6 -6
  117. package/include/numkong/maxsim/genoa.h +4 -4
  118. package/include/numkong/maxsim/haswell.h +6 -6
  119. package/include/numkong/maxsim/icelake.h +18 -18
  120. package/include/numkong/maxsim/neonsdot.h +21 -21
  121. package/include/numkong/maxsim/sapphireamx.h +14 -14
  122. package/include/numkong/maxsim/serial.h +6 -6
  123. package/include/numkong/maxsim/sme.h +221 -196
  124. package/include/numkong/maxsim/v128relaxed.h +6 -6
  125. package/include/numkong/mesh/README.md +62 -56
  126. package/include/numkong/mesh/haswell.h +339 -464
  127. package/include/numkong/mesh/neon.h +1100 -519
  128. package/include/numkong/mesh/neonbfdot.h +36 -68
  129. package/include/numkong/mesh/rvv.h +530 -435
  130. package/include/numkong/mesh/serial.h +75 -91
  131. package/include/numkong/mesh/skylake.h +1627 -302
  132. package/include/numkong/mesh/v128relaxed.h +443 -330
  133. package/include/numkong/mesh.h +63 -49
  134. package/include/numkong/mesh.hpp +4 -4
  135. package/include/numkong/numkong.h +3 -3
  136. package/include/numkong/numkong.hpp +1 -0
  137. package/include/numkong/probability/README.md +23 -19
  138. package/include/numkong/probability/neon.h +82 -52
  139. package/include/numkong/probability/rvv.h +28 -23
  140. package/include/numkong/probability/serial.h +51 -39
  141. package/include/numkong/probability.h +20 -23
  142. package/include/numkong/random.h +1 -1
  143. package/include/numkong/reduce/README.md +143 -138
  144. package/include/numkong/reduce/alder.h +81 -77
  145. package/include/numkong/reduce/haswell.h +222 -220
  146. package/include/numkong/reduce/neon.h +629 -519
  147. package/include/numkong/reduce/neonbfdot.h +7 -218
  148. package/include/numkong/reduce/neonfhm.h +9 -381
  149. package/include/numkong/reduce/neonsdot.h +9 -9
  150. package/include/numkong/reduce/rvv.h +928 -802
  151. package/include/numkong/reduce/serial.h +23 -27
  152. package/include/numkong/reduce/sierra.h +20 -20
  153. package/include/numkong/reduce/skylake.h +326 -324
  154. package/include/numkong/reduce/v128relaxed.h +52 -52
  155. package/include/numkong/reduce.h +4 -23
  156. package/include/numkong/reduce.hpp +156 -11
  157. package/include/numkong/scalar/README.md +6 -6
  158. package/include/numkong/scalar/haswell.h +26 -17
  159. package/include/numkong/scalar/loongsonasx.h +74 -0
  160. package/include/numkong/scalar/neon.h +9 -9
  161. package/include/numkong/scalar/powervsx.h +96 -0
  162. package/include/numkong/scalar/rvv.h +2 -2
  163. package/include/numkong/scalar/sapphire.h +21 -10
  164. package/include/numkong/scalar/serial.h +21 -21
  165. package/include/numkong/scalar.h +13 -0
  166. package/include/numkong/set/README.md +28 -28
  167. package/include/numkong/set/haswell.h +12 -12
  168. package/include/numkong/set/icelake.h +14 -14
  169. package/include/numkong/set/loongsonasx.h +181 -0
  170. package/include/numkong/set/neon.h +17 -18
  171. package/include/numkong/set/powervsx.h +326 -0
  172. package/include/numkong/set/rvv.h +4 -4
  173. package/include/numkong/set/serial.h +6 -6
  174. package/include/numkong/set/sve.h +60 -59
  175. package/include/numkong/set/v128relaxed.h +6 -6
  176. package/include/numkong/set.h +21 -7
  177. package/include/numkong/sets/README.md +26 -26
  178. package/include/numkong/sets/loongsonasx.h +52 -0
  179. package/include/numkong/sets/powervsx.h +65 -0
  180. package/include/numkong/sets/smebi32.h +395 -364
  181. package/include/numkong/sets.h +83 -40
  182. package/include/numkong/sparse/README.md +4 -4
  183. package/include/numkong/sparse/icelake.h +101 -101
  184. package/include/numkong/sparse/serial.h +1 -1
  185. package/include/numkong/sparse/sve2.h +137 -141
  186. package/include/numkong/sparse/turin.h +12 -12
  187. package/include/numkong/sparse.h +10 -10
  188. package/include/numkong/spatial/README.md +230 -226
  189. package/include/numkong/spatial/alder.h +113 -116
  190. package/include/numkong/spatial/diamond.h +240 -0
  191. package/include/numkong/spatial/genoa.h +0 -68
  192. package/include/numkong/spatial/haswell.h +74 -55
  193. package/include/numkong/spatial/icelake.h +539 -58
  194. package/include/numkong/spatial/loongsonasx.h +483 -0
  195. package/include/numkong/spatial/neon.h +125 -52
  196. package/include/numkong/spatial/neonbfdot.h +8 -9
  197. package/include/numkong/spatial/neonfp8.h +258 -0
  198. package/include/numkong/spatial/neonsdot.h +180 -12
  199. package/include/numkong/spatial/powervsx.h +738 -0
  200. package/include/numkong/spatial/rvv.h +146 -139
  201. package/include/numkong/spatial/rvvbf16.h +17 -12
  202. package/include/numkong/spatial/rvvhalf.h +13 -10
  203. package/include/numkong/spatial/serial.h +13 -12
  204. package/include/numkong/spatial/sierra.h +232 -39
  205. package/include/numkong/spatial/skylake.h +73 -74
  206. package/include/numkong/spatial/sve.h +93 -72
  207. package/include/numkong/spatial/svebfdot.h +29 -29
  208. package/include/numkong/spatial/svehalf.h +52 -26
  209. package/include/numkong/spatial/svesdot.h +142 -0
  210. package/include/numkong/spatial/v128relaxed.h +293 -41
  211. package/include/numkong/spatial.h +338 -82
  212. package/include/numkong/spatials/README.md +194 -194
  213. package/include/numkong/spatials/diamond.h +82 -0
  214. package/include/numkong/spatials/haswell.h +2 -2
  215. package/include/numkong/spatials/loongsonasx.h +153 -0
  216. package/include/numkong/spatials/neonfp8.h +111 -0
  217. package/include/numkong/spatials/neonsdot.h +34 -0
  218. package/include/numkong/spatials/powervsx.h +153 -0
  219. package/include/numkong/spatials/rvv.h +259 -243
  220. package/include/numkong/spatials/sapphireamx.h +173 -173
  221. package/include/numkong/spatials/serial.h +2 -2
  222. package/include/numkong/spatials/skylake.h +2 -2
  223. package/include/numkong/spatials/sme.h +590 -605
  224. package/include/numkong/spatials/smef64.h +139 -130
  225. package/include/numkong/spatials/v128relaxed.h +2 -2
  226. package/include/numkong/spatials.h +820 -500
  227. package/include/numkong/spatials.hpp +49 -48
  228. package/include/numkong/tensor.hpp +406 -17
  229. package/include/numkong/trigonometry/README.md +19 -19
  230. package/include/numkong/trigonometry/haswell.h +402 -401
  231. package/include/numkong/trigonometry/neon.h +386 -387
  232. package/include/numkong/trigonometry/rvv.h +52 -51
  233. package/include/numkong/trigonometry/serial.h +13 -13
  234. package/include/numkong/trigonometry/skylake.h +373 -369
  235. package/include/numkong/trigonometry/v128relaxed.h +375 -374
  236. package/include/numkong/trigonometry.h +13 -13
  237. package/include/numkong/trigonometry.hpp +2 -2
  238. package/include/numkong/types.h +287 -49
  239. package/include/numkong/types.hpp +436 -12
  240. package/include/numkong/vector.hpp +82 -14
  241. package/javascript/dist/cjs/numkong-wasm.js +6 -12
  242. package/javascript/dist/cjs/numkong.d.ts +7 -1
  243. package/javascript/dist/cjs/numkong.js +37 -11
  244. package/javascript/dist/cjs/types.d.ts +9 -0
  245. package/javascript/dist/cjs/types.js +96 -0
  246. package/javascript/dist/esm/numkong-browser.d.ts +14 -0
  247. package/javascript/dist/esm/numkong-browser.js +23 -0
  248. package/javascript/dist/esm/numkong-wasm.js +6 -12
  249. package/javascript/dist/esm/numkong.d.ts +7 -1
  250. package/javascript/dist/esm/numkong.js +37 -11
  251. package/javascript/dist/esm/types.d.ts +9 -0
  252. package/javascript/dist/esm/types.js +96 -0
  253. package/javascript/node-gyp-build.d.ts +4 -1
  254. package/javascript/numkong-browser.ts +40 -0
  255. package/javascript/numkong-wasm.ts +7 -13
  256. package/javascript/numkong.c +5 -26
  257. package/javascript/numkong.ts +36 -11
  258. package/javascript/tsconfig-base.json +1 -0
  259. package/javascript/tsconfig-cjs.json +6 -1
  260. package/javascript/types.ts +110 -0
  261. package/numkong.gypi +101 -0
  262. package/package.json +34 -13
  263. package/probes/arm_neon.c +8 -0
  264. package/probes/arm_neon_bfdot.c +9 -0
  265. package/probes/arm_neon_fhm.c +9 -0
  266. package/probes/arm_neon_half.c +8 -0
  267. package/probes/arm_neon_sdot.c +9 -0
  268. package/probes/arm_neonfp8.c +9 -0
  269. package/probes/arm_sme.c +16 -0
  270. package/probes/arm_sme2.c +16 -0
  271. package/probes/arm_sme2p1.c +16 -0
  272. package/probes/arm_sme_bf16.c +16 -0
  273. package/probes/arm_sme_bi32.c +16 -0
  274. package/probes/arm_sme_f64.c +16 -0
  275. package/probes/arm_sme_fa64.c +14 -0
  276. package/probes/arm_sme_half.c +16 -0
  277. package/probes/arm_sme_lut2.c +15 -0
  278. package/probes/arm_sve.c +18 -0
  279. package/probes/arm_sve2.c +20 -0
  280. package/probes/arm_sve2p1.c +18 -0
  281. package/probes/arm_sve_bfdot.c +20 -0
  282. package/probes/arm_sve_half.c +18 -0
  283. package/probes/arm_sve_sdot.c +21 -0
  284. package/probes/loongarch_lasx.c +12 -0
  285. package/probes/power_vsx.c +12 -0
  286. package/probes/probe.js +127 -0
  287. package/probes/riscv_rvv.c +14 -0
  288. package/probes/riscv_rvv_bb.c +15 -0
  289. package/probes/riscv_rvv_bf16.c +17 -0
  290. package/probes/riscv_rvv_half.c +14 -0
  291. package/probes/wasm_v128relaxed.c +11 -0
  292. package/probes/x86_alder.c +17 -0
  293. package/probes/x86_diamond.c +17 -0
  294. package/probes/x86_genoa.c +17 -0
  295. package/probes/x86_graniteamx.c +19 -0
  296. package/probes/x86_haswell.c +11 -0
  297. package/probes/x86_icelake.c +17 -0
  298. package/probes/x86_sapphire.c +16 -0
  299. package/probes/x86_sapphireamx.c +18 -0
  300. package/probes/x86_sierra.c +17 -0
  301. package/probes/x86_skylake.c +15 -0
  302. package/probes/x86_turin.c +17 -0
  303. package/wasm/numkong-emscripten.js +2 -0
  304. package/wasm/numkong.d.ts +14 -0
  305. package/wasm/numkong.js +1124 -0
  306. package/wasm/numkong.wasm +0 -0
  307. package/include/numkong/curved/neonhalf.h +0 -212
  308. package/include/numkong/dot/neonhalf.h +0 -198
  309. package/include/numkong/dots/neonhalf.h +0 -57
  310. package/include/numkong/mesh/neonhalf.h +0 -616
  311. package/include/numkong/reduce/neonhalf.h +0 -157
  312. package/include/numkong/spatial/neonhalf.h +0 -118
  313. package/include/numkong/spatial/sapphire.h +0 -343
  314. package/include/numkong/spatials/neonhalf.h +0 -58
  315. package/javascript/README.md +0 -246
@@ -7,15 +7,15 @@ Used internally by packing routines for norm precomputation and by distance kern
7
7
 
8
8
  Moments:
9
9
 
10
- ```math
10
+ $$
11
11
  \text{sum} = \sum a_i, \quad \text{sumsq} = \sum a_i^2
12
- ```
12
+ $$
13
13
 
14
14
  Min-max:
15
15
 
16
- ```math
16
+ $$
17
17
  \text{min} = \min_i a_i, \quad \text{argmin} = \arg\min_i a_i
18
- ```
18
+ $$
19
19
 
20
20
  Reformulating as Python pseudocode:
21
21
 
@@ -365,183 +365,188 @@ Measured with Wasmtime v42 (Cranelift backend).
365
365
  | `nk_reduce_minmax_u64_serial` | 7.83 gb/s | 7.59 gb/s | 7.74 gb/s |
366
366
  | `nk_reduce_minmax_u64_v128relaxed` | 2.54 gb/s | 2.59 gb/s | 2.64 gb/s |
367
367
 
368
- ### Apple M4
368
+ ### Apple M5
369
369
 
370
370
  #### Native
371
371
 
372
372
  | Kernel | 256 | 1024 | 4096 |
373
373
  | :--------------------------------- | -----------------------: | -----------------------: | -----------------------: |
374
374
  | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
375
- | `nk_reduce_moments_f64_serial` | 4.97 gb/s, 0 ulp | 4.98 gb/s, 0 ulp | 5.01 gb/s, 0 ulp |
376
- | `nk_reduce_minmax_f64_serial` | 10.3 gb/s, 0 ulp | 10.3 gb/s, 0 ulp | 10.0 gb/s, 0 ulp |
377
- | `nk_reduce_moments_f64_neon` | 11.4 gb/s, 0 ulp | 11.3 gb/s, 0 ulp | 11.1 gb/s, 0 ulp |
378
- | `nk_reduce_minmax_f64_neon` | 14.1 gb/s, 0 ulp | 13.1 gb/s, 0 ulp | 12.8 gb/s, 0 ulp |
375
+ | `nk_reduce_moments_f64_serial` | 6.41 gb/s, 0 ulp | 7.19 gb/s, 0 ulp | 6.71 gb/s, 0 ulp |
376
+ | `nk_reduce_minmax_f64_serial` | 16.4 gb/s, 0 ulp | 16.1 gb/s, 0 ulp | 16.1 gb/s, 0 ulp |
377
+ | `nk_reduce_moments_f64_neon` | 16.4 gb/s, 0 ulp | 16.3 gb/s, 0 ulp | 17.0 gb/s, 0 ulp |
378
+ | `nk_reduce_minmax_f64_neon` | 18.0 gb/s, 0 ulp | 17.0 gb/s, 0 ulp | 16.2 gb/s, 0 ulp |
379
379
  | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
380
- | `nk_reduce_moments_f32_serial` | 2.33 gb/s, 0 ulp | 2.32 gb/s, 0 ulp | 2.35 gb/s, 0 ulp |
381
- | `nk_reduce_minmax_f32_serial` | 5.39 gb/s, 0 ulp | 5.19 gb/s, 0 ulp | 5.20 gb/s, 0 ulp |
382
- | `nk_reduce_moments_f32_neon` | 11.4 gb/s, 0.4 ulp | 8.38 gb/s, 1.8 ulp | 7.27 gb/s, 1.1 ulp |
383
- | `nk_reduce_minmax_f32_neon` | 14.2 gb/s, 0 ulp | 12.6 gb/s, 0 ulp | 13.0 gb/s, 0 ulp |
380
+ | `nk_reduce_moments_f32_serial` | 3.06 gb/s, 0 ulp | 3.40 gb/s, 0 ulp | 3.15 gb/s, 0 ulp |
381
+ | `nk_reduce_minmax_f32_serial` | 8.29 gb/s, 0 ulp | 7.97 gb/s, 0 ulp | 8.00 gb/s, 0 ulp |
382
+ | `nk_reduce_moments_f32_neon` | 16.6 gb/s, 0.4 ulp | 10.8 gb/s, 1.8 ulp | 9.78 gb/s, 1.1 ulp |
383
+ | `nk_reduce_minmax_f32_neon` | 19.7 gb/s, 0 ulp | 17.3 gb/s, 0 ulp | 16.5 gb/s, 0 ulp |
384
384
  | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
385
- | `nk_reduce_moments_bf16_serial` | 1.09 gb/s, 0 ulp | 1.10 gb/s, 0 ulp | 1.11 gb/s, 0 ulp |
386
- | `nk_reduce_minmax_bf16_serial` | 1.56 gb/s, 0 ulp | 1.69 gb/s, 0 ulp | 1.75 gb/s, 0 ulp |
387
- | `nk_reduce_moments_bf16_neonbfdot` | 18.6 gb/s, 0 ulp | 20.3 gb/s, 0.4 ulp | 21.0 gb/s, 0.3 ulp |
388
- | `nk_reduce_minmax_bf16_neonbfdot` | 13.7 gb/s, 0 ulp | 14.5 gb/s, 0 ulp | 15.4 gb/s, 0 ulp |
385
+ | `nk_reduce_moments_bf16_serial` | 1.49 gb/s, 0 ulp | 1.62 gb/s, 0 ulp | 1.55 gb/s, 0 ulp |
386
+ | `nk_reduce_minmax_bf16_serial` | 2.19 gb/s, 0 ulp | 2.49 gb/s, 0 ulp | 2.69 gb/s, 0 ulp |
387
+ | `nk_reduce_moments_bf16_neonbfdot` | 26.1 gb/s, 0 ulp | 28.2 gb/s, 0.4 ulp | 29.9 gb/s, 0.3 ulp |
389
388
  | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
390
- | `nk_reduce_moments_f16_serial` | 1.10 gb/s, 0 ulp | 1.09 gb/s, 0 ulp | 1.09 gb/s, 0 ulp |
391
- | `nk_reduce_minmax_f16_serial` | 1.17 gb/s, 0 ulp | 1.26 gb/s, 0 ulp | 1.33 gb/s, 0 ulp |
392
- | `nk_reduce_moments_f16_neonhalf` | 15.5 gb/s, 0.1 ulp | 11.1 gb/s, 0.1 ulp | 8.74 gb/s, 0.8 ulp |
389
+ | `nk_reduce_moments_f16_serial` | 1.46 gb/s, 0 ulp | 1.55 gb/s, 0 ulp | 1.51 gb/s, 0 ulp |
390
+ | `nk_reduce_minmax_f16_serial` | 1.62 gb/s, 0 ulp | 1.87 gb/s, 0 ulp | 2.02 gb/s, 0 ulp |
391
+ | `nk_reduce_moments_f16_neonhalf` | 21.2 gb/s, 0.1 ulp | 15.4 gb/s, 0.1 ulp | 10.6 gb/s, 0.8 ulp |
393
392
  | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
394
- | `nk_reduce_moments_e5m2_serial` | 0.531 gb/s, 0 ulp | 0.534 gb/s, 0 ulp | 0.536 gb/s, 0 ulp |
395
- | `nk_reduce_minmax_e5m2_serial` | 0.803 gb/s, 0 ulp | 0.887 gb/s, 0 ulp | 0.880 gb/s, 0 ulp |
396
- | `nk_reduce_moments_e5m2_neonfhm` | 7.95 gb/s, 0 ulp | 5.31 gb/s, 0 ulp | 4.07 gb/s, 0 ulp |
397
- | `nk_reduce_minmax_e5m2_neonfhm` | 14.0 gb/s, 0 ulp | 16.8 gb/s, 0 ulp | 17.1 gb/s, 0 ulp |
393
+ | `nk_reduce_moments_e5m2_serial` | 0.750 gb/s, 0 ulp | 0.830 gb/s, 0 ulp | 0.758 gb/s, 0 ulp |
394
+ | `nk_reduce_minmax_e5m2_serial` | 1.10 gb/s, 0 ulp | 1.34 gb/s, 0 ulp | 1.39 gb/s, 0 ulp |
395
+ | `nk_reduce_moments_e5m2_neon` | 10.4 gb/s, ? ulp | 7.47 gb/s, ? ulp | 5.22 gb/s, ? ulp |
396
+ | `nk_reduce_moments_e5m2_neonfhm` | 12.4 gb/s, 0 ulp | 7.31 gb/s, 0 ulp | 4.78 gb/s, 0 ulp |
397
+ | `nk_reduce_minmax_e5m2_neon` | 15.5 gb/s, 0 ulp | 17.4 gb/s, 0 ulp | 18.1 gb/s, 0 ulp |
398
398
  | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
399
- | `nk_reduce_moments_e4m3_serial` | 0.382 gb/s, 0 ulp | 0.386 gb/s, 0 ulp | 0.387 gb/s, 0 ulp |
400
- | `nk_reduce_minmax_e4m3_serial` | 0.823 gb/s, 0 ulp | 0.884 gb/s, 0 ulp | 0.918 gb/s, 0 ulp |
401
- | `nk_reduce_moments_e4m3_neonfhm` | 2.90 gb/s, 0 ulp | 2.98 gb/s, 0 ulp | 2.89 gb/s, 0 ulp |
402
- | `nk_reduce_minmax_e4m3_neonfhm` | 14.0 gb/s, 0 ulp | 16.7 gb/s, 0 ulp | 17.8 gb/s, 0 ulp |
399
+ | `nk_reduce_moments_e4m3_serial` | 0.569 gb/s, 0 ulp | 0.640 gb/s, 0 ulp | 0.599 gb/s, 0 ulp |
400
+ | `nk_reduce_minmax_e4m3_serial` | 1.14 gb/s, 0 ulp | 1.39 gb/s, 0 ulp | 1.42 gb/s, 0 ulp |
401
+ | `nk_reduce_moments_e4m3_neon` | 6.52 gb/s, ? ulp | 5.52 gb/s, ? ulp | 4.88 gb/s, ? ulp |
402
+ | `nk_reduce_moments_e4m3_neonfhm` | 4.21 gb/s, 0 ulp | 4.23 gb/s, 0 ulp | 4.11 gb/s, 0 ulp |
403
+ | `nk_reduce_minmax_e4m3_neon` | 15.5 gb/s, 0 ulp | 17.4 gb/s, 0 ulp | 17.5 gb/s, 0 ulp |
403
404
  | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
404
- | `nk_reduce_moments_e3m2_serial` | 0.537 gb/s, 0 ulp | 0.532 gb/s, 0 ulp | 0.539 gb/s, 0 ulp |
405
- | `nk_reduce_minmax_e3m2_serial` | 0.506 gb/s, 0 ulp | 0.500 gb/s, 0 ulp | 0.503 gb/s, 0 ulp |
405
+ | `nk_reduce_moments_e3m2_serial` | 0.750 gb/s, 0 ulp | 0.806 gb/s, 0 ulp | 0.773 gb/s, 0 ulp |
406
+ | `nk_reduce_minmax_e3m2_serial` | 0.640 gb/s, 0 ulp | 0.641 gb/s, 0 ulp | 0.634 gb/s, 0 ulp |
407
+ | `nk_reduce_moments_e3m2_neon` | 8.64 gb/s, ? ulp | 6.96 gb/s, ? ulp | 5.70 gb/s, ? ulp |
408
+ | `nk_reduce_minmax_e3m2_neon` | 16.6 gb/s, 0 ulp | 19.1 gb/s, 0 ulp | 18.9 gb/s, 0 ulp |
406
409
  | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
407
- | `nk_reduce_moments_e2m3_serial` | 0.538 gb/s, 0 ulp | 0.539 gb/s, 0 ulp | 0.536 gb/s, 0 ulp |
408
- | `nk_reduce_minmax_e2m3_serial` | 0.503 gb/s, 0 ulp | 0.504 gb/s, 0 ulp | 0.502 gb/s, 0 ulp |
409
- | `nk_reduce_moments_e2m3_neonsdot` | 19.4 gb/s, 0 ulp | 19.5 gb/s, 0 ulp | 18.7 gb/s, 0 ulp |
410
+ | `nk_reduce_moments_e2m3_serial` | 0.753 gb/s, 0 ulp | 0.821 gb/s, 0 ulp | 0.775 gb/s, 0 ulp |
411
+ | `nk_reduce_minmax_e2m3_serial` | 0.640 gb/s, 0 ulp | 0.643 gb/s, 0 ulp | 0.618 gb/s, 0 ulp |
412
+ | `nk_reduce_moments_e2m3_neon` | 17.1 gb/s, ? ulp | 15.2 gb/s, ? ulp | 11.9 gb/s, ? ulp |
413
+ | `nk_reduce_moments_e2m3_neonsdot` | 29.4 gb/s, 0 ulp | 29.3 gb/s, 0 ulp | 24.4 gb/s, 0 ulp |
414
+ | `nk_reduce_minmax_e2m3_neon` | 17.0 gb/s, 0 ulp | 19.3 gb/s, 0 ulp | 19.6 gb/s, 0 ulp |
410
415
  | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
411
- | `nk_reduce_moments_i8_serial` | 2.44 gb/s | 2.59 gb/s | 2.59 gb/s |
412
- | `nk_reduce_minmax_i8_serial` | 1.30 gb/s | 1.34 gb/s | 1.34 gb/s |
413
- | `nk_reduce_moments_i8_neon` | 16.5 gb/s | 12.4 gb/s | 9.71 gb/s |
414
- | `nk_reduce_minmax_i8_neon` | 17.9 gb/s | 20.3 gb/s | 19.0 gb/s |
415
- | `nk_reduce_moments_i8_neonsdot` | 29.8 gb/s | 28.9 gb/s | 22.0 gb/s |
416
+ | `nk_reduce_moments_i8_serial` | 3.00 gb/s | 3.50 gb/s | 3.21 gb/s |
417
+ | `nk_reduce_minmax_i8_serial` | 1.81 gb/s | 1.94 gb/s | 1.89 gb/s |
418
+ | `nk_reduce_moments_i8_neon` | 27.9 gb/s | 17.7 gb/s | 13.3 gb/s |
419
+ | `nk_reduce_minmax_i8_neon` | 25.7 gb/s | 30.3 gb/s | 27.8 gb/s |
420
+ | `nk_reduce_moments_i8_neonsdot` | 44.4 gb/s | 46.8 gb/s | 33.4 gb/s |
416
421
  | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
417
- | `nk_reduce_moments_u8_serial` | 2.49 gb/s | 2.59 gb/s | 2.57 gb/s |
418
- | `nk_reduce_minmax_u8_serial` | 1.32 gb/s | 1.31 gb/s | 1.31 gb/s |
419
- | `nk_reduce_moments_u8_neon` | 16.7 gb/s | 12.1 gb/s | 9.20 gb/s |
420
- | `nk_reduce_minmax_u8_neon` | 18.2 gb/s | 20.4 gb/s | 21.4 gb/s |
421
- | `nk_reduce_moments_u8_neonsdot` | 29.6 gb/s | 29.2 gb/s | 22.6 gb/s |
422
+ | `nk_reduce_moments_u8_serial` | 3.08 gb/s | 3.51 gb/s | 3.23 gb/s |
423
+ | `nk_reduce_minmax_u8_serial` | 1.84 gb/s | 1.94 gb/s | 1.90 gb/s |
424
+ | `nk_reduce_moments_u8_neon` | 29.2 gb/s | 18.0 gb/s | 13.6 gb/s |
425
+ | `nk_reduce_minmax_u8_neon` | 27.0 gb/s | 28.3 gb/s | 29.2 gb/s |
426
+ | `nk_reduce_moments_u8_neonsdot` | 45.5 gb/s | 43.5 gb/s | 33.3 gb/s |
422
427
  | __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
423
- | `nk_reduce_moments_i4_serial` | 1.55 gb/s | 1.82 gb/s | 1.85 gb/s |
424
- | `nk_reduce_minmax_i4_serial` | 0.512 gb/s | 0.524 gb/s | 0.532 gb/s |
428
+ | `nk_reduce_moments_i4_serial` | 2.06 gb/s | 2.57 gb/s | 2.39 gb/s |
429
+ | `nk_reduce_minmax_i4_serial` | 0.701 gb/s | 0.794 gb/s | 0.803 gb/s |
425
430
  | __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
426
- | `nk_reduce_moments_u4_serial` | 1.63 gb/s | 1.94 gb/s | 2.00 gb/s |
427
- | `nk_reduce_minmax_u4_serial` | 0.536 gb/s | 0.544 gb/s | 0.548 gb/s |
431
+ | `nk_reduce_moments_u4_serial` | 2.20 gb/s | 2.75 gb/s | 2.62 gb/s |
432
+ | `nk_reduce_minmax_u4_serial` | 0.741 gb/s | 0.845 gb/s | 0.856 gb/s |
428
433
  | __u1__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
429
- | `nk_reduce_moments_u1_serial` | 1.24 gb/s | 1.32 gb/s | 1.35 gb/s |
430
- | `nk_reduce_minmax_u1_serial` | 7.22 gb/s | 28.9 gb/s | 111 gb/s |
434
+ | `nk_reduce_moments_u1_serial` | 1.79 gb/s | 1.99 gb/s | 2.03 gb/s |
435
+ | `nk_reduce_minmax_u1_serial` | 9.74 gb/s | 38.5 gb/s | 121 gb/s |
431
436
  | __i16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
432
- | `nk_reduce_moments_i16_serial` | 4.97 gb/s | 5.16 gb/s | 5.17 gb/s |
433
- | `nk_reduce_minmax_i16_serial` | 2.08 gb/s | 2.05 gb/s | 2.08 gb/s |
434
- | `nk_reduce_moments_i16_neon` | 16.3 gb/s | 11.9 gb/s | 9.28 gb/s |
435
- | `nk_reduce_minmax_i16_neon` | 20.4 gb/s | 20.1 gb/s | 21.1 gb/s |
437
+ | `nk_reduce_moments_i16_serial` | 6.02 gb/s | 7.03 gb/s | 6.70 gb/s |
438
+ | `nk_reduce_minmax_i16_serial` | 3.66 gb/s | 3.88 gb/s | 3.82 gb/s |
439
+ | `nk_reduce_moments_i16_neon` | 23.0 gb/s | 16.2 gb/s | 11.9 gb/s |
440
+ | `nk_reduce_minmax_i16_neon` | 26.8 gb/s | 28.5 gb/s | 25.9 gb/s |
436
441
  | __u16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
437
- | `nk_reduce_moments_u16_serial` | 4.97 gb/s | 5.11 gb/s | 5.20 gb/s |
438
- | `nk_reduce_minmax_u16_serial` | 2.59 gb/s | 2.66 gb/s | 2.62 gb/s |
439
- | `nk_reduce_moments_u16_neon` | 16.6 gb/s | 12.0 gb/s | 9.35 gb/s |
440
- | `nk_reduce_minmax_u16_neon` | 20.2 gb/s | 21.1 gb/s | 21.4 gb/s |
442
+ | `nk_reduce_moments_u16_serial` | 6.13 gb/s | 6.79 gb/s | 6.45 gb/s |
443
+ | `nk_reduce_minmax_u16_serial` | 2.72 gb/s | 2.68 gb/s | 2.67 gb/s |
444
+ | `nk_reduce_moments_u16_neon` | 23.1 gb/s | 16.3 gb/s | 11.6 gb/s |
445
+ | `nk_reduce_minmax_u16_neon` | 26.6 gb/s | 28.5 gb/s | 25.8 gb/s |
441
446
  | __i32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
442
- | `nk_reduce_moments_i32_serial` | 5.66 gb/s | 5.62 gb/s | 5.77 gb/s |
443
- | `nk_reduce_minmax_i32_serial` | 6.17 gb/s | 6.21 gb/s | 6.24 gb/s |
444
- | `nk_reduce_moments_i32_neon` | 6.48 gb/s | 5.58 gb/s | 5.29 gb/s |
445
- | `nk_reduce_minmax_i32_neon` | 20.9 gb/s | 19.3 gb/s | 21.2 gb/s |
447
+ | `nk_reduce_moments_i32_serial` | 7.70 gb/s | 8.09 gb/s | 7.52 gb/s |
448
+ | `nk_reduce_minmax_i32_serial` | 7.35 gb/s | 7.49 gb/s | 7.65 gb/s |
449
+ | `nk_reduce_moments_i32_neon` | 8.88 gb/s | 6.74 gb/s | 6.60 gb/s |
450
+ | `nk_reduce_minmax_i32_neon` | 27.3 gb/s | 27.9 gb/s | 25.8 gb/s |
446
451
  | __u32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
447
- | `nk_reduce_moments_u32_serial` | 5.72 gb/s | 5.70 gb/s | 5.88 gb/s |
448
- | `nk_reduce_minmax_u32_serial` | 5.60 gb/s | 5.59 gb/s | 5.93 gb/s |
449
- | `nk_reduce_moments_u32_neon` | 13.7 gb/s | 10.3 gb/s | 9.07 gb/s |
450
- | `nk_reduce_minmax_u32_neon` | 20.6 gb/s | 21.4 gb/s | 21.3 gb/s |
452
+ | `nk_reduce_moments_u32_serial` | 7.51 gb/s | 8.10 gb/s | 7.73 gb/s |
453
+ | `nk_reduce_minmax_u32_serial` | 7.28 gb/s | 7.69 gb/s | 7.68 gb/s |
454
+ | `nk_reduce_moments_u32_neon` | 19.0 gb/s | 12.8 gb/s | 11.5 gb/s |
455
+ | `nk_reduce_minmax_u32_neon` | 27.6 gb/s | 28.3 gb/s | 26.4 gb/s |
451
456
  | __i64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
452
- | `nk_reduce_moments_i64_serial` | 7.53 gb/s | 7.55 gb/s | 7.59 gb/s |
453
- | `nk_reduce_minmax_i64_serial` | 10.0 gb/s | 10.6 gb/s | 10.5 gb/s |
454
- | `nk_reduce_moments_i64_neon` | 11.4 gb/s | 10.5 gb/s | 10.2 gb/s |
455
- | `nk_reduce_minmax_i64_neon` | 14.3 gb/s | 13.4 gb/s | 13.0 gb/s |
457
+ | `nk_reduce_moments_i64_serial` | 10.4 gb/s | 12.9 gb/s | 11.9 gb/s |
458
+ | `nk_reduce_minmax_i64_serial` | 13.7 gb/s | 14.1 gb/s | 14.0 gb/s |
459
+ | `nk_reduce_moments_i64_neon` | 15.6 gb/s | 13.1 gb/s | 12.9 gb/s |
460
+ | `nk_reduce_minmax_i64_neon` | 18.0 gb/s | 16.6 gb/s | 15.6 gb/s |
456
461
  | __u64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
457
- | `nk_reduce_moments_u64_serial` | 7.11 gb/s | 7.44 gb/s | 7.48 gb/s |
458
- | `nk_reduce_minmax_u64_serial` | 10.5 gb/s | 10.6 gb/s | 10.4 gb/s |
459
- | `nk_reduce_moments_u64_neon` | 21.2 gb/s | 18.7 gb/s | 17.2 gb/s |
460
- | `nk_reduce_minmax_u64_neon` | 14.3 gb/s | 13.2 gb/s | 12.9 gb/s |
462
+ | `nk_reduce_moments_u64_serial` | 10.4 gb/s | 11.9 gb/s | 11.2 gb/s |
463
+ | `nk_reduce_minmax_u64_serial` | 13.4 gb/s | 14.1 gb/s | 13.8 gb/s |
464
+ | `nk_reduce_moments_u64_neon` | 29.1 gb/s | 22.5 gb/s | 22.2 gb/s |
465
+ | `nk_reduce_minmax_u64_neon` | 18.3 gb/s | 16.7 gb/s | 16.1 gb/s |
461
466
 
462
467
  #### WASM
463
468
 
464
- Measured with Wasmtime v42 (Cranelift backend).
469
+ Measured with Wasmtime v43 (Cranelift backend).
465
470
 
466
471
  | Kernel | 256 | 1024 | 4096 |
467
472
  | :----------------------------------- | -----------------------: | -----------------------: | -----------------------: |
468
473
  | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
469
- | `nk_reduce_moments_f64_serial` | 4.19 gb/s, 0 ulp | 8.01 gb/s, 0 ulp | 8.28 gb/s, 0 ulp |
470
- | `nk_reduce_moments_f64_v128relaxed` | 6.55 gb/s, 0 ulp | 16.7 gb/s, 0 ulp | 16.4 gb/s, 0 ulp |
471
- | `nk_reduce_minmax_f64_serial` | ? | 15.6 gb/s, 0 ulp | 16.2 gb/s, 0 ulp |
472
- | `nk_reduce_minmax_f64_v128relaxed` | 5.99 gb/s, 0 ulp | 13.8 gb/s, 0 ulp | 15.6 gb/s, 0 ulp |
474
+ | `nk_reduce_moments_f64_serial` | 7.05 gb/s, 0 ulp | 6.89 gb/s, 0 ulp | 7.04 gb/s, 0 ulp |
475
+ | `nk_reduce_moments_f64_v128relaxed` | 17.9 gb/s, 0 ulp | 17.9 gb/s, 0 ulp | 17.9 gb/s, 0 ulp |
476
+ | `nk_reduce_minmax_f64_serial` | 11.8 gb/s, 0 ulp | 11.5 gb/s, 0 ulp | 11.4 gb/s, 0 ulp |
477
+ | `nk_reduce_minmax_f64_v128relaxed` | 16.4 gb/s, 0 ulp | 16.8 gb/s, 0 ulp | 16.9 gb/s, 0 ulp |
473
478
  | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
474
- | `nk_reduce_moments_f32_serial` | 1.75 gb/s, 0 ulp | 3.39 gb/s, 0 ulp | 3.50 gb/s, 0 ulp |
475
- | `nk_reduce_moments_f32_v128relaxed` | 4.63 gb/s, 0.1 ulp | 8.81 gb/s, 0.5 ulp | 7.11 gb/s, 0 ulp |
476
- | `nk_reduce_minmax_f32_serial` | ? | 5.27 gb/s, 0 ulp | 5.41 gb/s, 0 ulp |
477
- | `nk_reduce_minmax_f32_v128relaxed` | 4.89 gb/s, 0 ulp | 12.7 gb/s, 0 ulp | 15.3 gb/s, 0 ulp |
479
+ | `nk_reduce_moments_f32_serial` | 3.34 gb/s, 0 ulp | 3.32 gb/s, 0 ulp | 3.34 gb/s, 0 ulp |
480
+ | `nk_reduce_moments_f32_v128relaxed` | 16.7 gb/s, 0.1 ulp | 11.9 gb/s, 0.5 ulp | 10.3 gb/s, 0 ulp |
481
+ | `nk_reduce_minmax_f32_serial` | 4.40 gb/s, 0 ulp | 4.31 gb/s, 0 ulp | 4.28 gb/s, 0 ulp |
482
+ | `nk_reduce_minmax_f32_v128relaxed` | 14.5 gb/s, 0 ulp | 15.8 gb/s, 0 ulp | 16.7 gb/s, 0 ulp |
478
483
  | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
479
- | `nk_reduce_moments_bf16_serial` | 2.21 gb/s, 0 ulp | 2.08 gb/s, 0 ulp | 2.07 gb/s, 0 ulp |
480
- | `nk_reduce_moments_bf16_v128relaxed` | 5.49 gb/s, 0 ulp | 9.10 gb/s, 0.2 ulp | 8.28 gb/s, 1.6 ulp |
481
- | `nk_reduce_minmax_bf16_serial` | ? | 1.87 gb/s, 0 ulp | 1.99 gb/s, 0 ulp |
482
- | `nk_reduce_minmax_bf16_v128relaxed` | 4.05 gb/s, 0 ulp | 8.80 gb/s, 0 ulp | 10.4 gb/s, 0 ulp |
484
+ | `nk_reduce_moments_bf16_serial` | 1.76 gb/s, 0 ulp | 1.72 gb/s, 0 ulp | 1.72 gb/s, 0 ulp |
485
+ | `nk_reduce_moments_bf16_v128relaxed` | 20.8 gb/s, 0 ulp | 14.8 gb/s, 0.2 ulp | 11.1 gb/s, 1.6 ulp |
486
+ | `nk_reduce_minmax_bf16_serial` | 1.80 gb/s, 0 ulp | 2.01 gb/s, 0 ulp | 2.10 gb/s, 0 ulp |
487
+ | `nk_reduce_minmax_bf16_v128relaxed` | 9.22 gb/s, 0 ulp | 9.53 gb/s, 0 ulp | 10.5 gb/s, 0 ulp |
483
488
  | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
484
- | `nk_reduce_moments_f16_serial` | 2.05 gb/s, 0 ulp | 1.98 gb/s, 0 ulp | 2.00 gb/s, 0 ulp |
485
- | `nk_reduce_moments_f16_v128relaxed` | 1.78 gb/s, 0 ulp | 3.94 gb/s, 0 ulp | 4.43 gb/s, 0.3 ulp |
486
- | `nk_reduce_minmax_f16_serial` | ? | 1.86 gb/s, 0 ulp | 1.98 gb/s, 0 ulp |
487
- | `nk_reduce_minmax_f16_v128relaxed` | 1.66 gb/s, 0 ulp | 3.70 gb/s, 0 ulp | 3.90 gb/s, 0 ulp |
489
+ | `nk_reduce_moments_f16_serial` | 1.77 gb/s, 0 ulp | 1.73 gb/s, 0 ulp | 1.73 gb/s, 0 ulp |
490
+ | `nk_reduce_moments_f16_v128relaxed` | 11.8 gb/s, 0 ulp | 10.8 gb/s, 0 ulp | 9.84 gb/s, 0.3 ulp |
491
+ | `nk_reduce_minmax_f16_serial` | 1.79 gb/s, 0 ulp | 1.99 gb/s, 0 ulp | 2.09 gb/s, 0 ulp |
492
+ | `nk_reduce_minmax_f16_v128relaxed` | 6.04 gb/s, 0 ulp | 7.62 gb/s, 0 ulp | 8.11 gb/s, 0 ulp |
488
493
  | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
489
- | `nk_reduce_moments_e5m2_serial` | 1.10 gb/s, 0 ulp | 1.05 gb/s, 0 ulp | 1.06 gb/s, 0 ulp |
490
- | `nk_reduce_moments_e5m2_v128relaxed` | 1.16 gb/s, 0 ulp | 2.56 gb/s, 0 ulp | 2.83 gb/s, 0 ulp |
491
- | `nk_reduce_minmax_e5m2_serial` | ? | 1.09 gb/s, 0 ulp | 1.14 gb/s, 0 ulp |
492
- | `nk_reduce_minmax_e5m2_v128relaxed` | 3.09 gb/s, 0 ulp | 10.3 gb/s, 0 ulp | 14.6 gb/s, 0 ulp |
494
+ | `nk_reduce_moments_e5m2_serial` | 0.886 gb/s, 0 ulp | 0.865 gb/s, 0 ulp | 0.870 gb/s, 0 ulp |
495
+ | `nk_reduce_moments_e5m2_v128relaxed` | 3.53 gb/s, 0 ulp | 3.55 gb/s, 0 ulp | 3.56 gb/s, 0 ulp |
496
+ | `nk_reduce_minmax_e5m2_serial` | 1.12 gb/s, 0 ulp | 1.18 gb/s, 0 ulp | 1.22 gb/s, 0 ulp |
497
+ | `nk_reduce_minmax_e5m2_v128relaxed` | 5.50 gb/s, 0 ulp | 10.8 gb/s, 0 ulp | 14.6 gb/s, 0 ulp |
493
498
  | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
494
- | `nk_reduce_moments_e4m3_serial` | 0.731 gb/s, 0 ulp | 0.669 gb/s, 0 ulp | 0.668 gb/s, 0 ulp |
495
- | `nk_reduce_moments_e4m3_v128relaxed` | 0.945 gb/s, 0 ulp | 2.10 gb/s, 0 ulp | 2.37 gb/s, 0 ulp |
496
- | `nk_reduce_minmax_e4m3_serial` | ? | 1.09 gb/s, 0 ulp | 1.14 gb/s, 0 ulp |
497
- | `nk_reduce_minmax_e4m3_v128relaxed` | 3.47 gb/s, 0 ulp | 9.70 gb/s, 0 ulp | 14.6 gb/s, 0 ulp |
499
+ | `nk_reduce_moments_e4m3_serial` | 0.617 gb/s, 0 ulp | 0.606 gb/s, 0 ulp | 0.610 gb/s, 0 ulp |
500
+ | `nk_reduce_moments_e4m3_v128relaxed` | 3.01 gb/s, 0 ulp | 3.04 gb/s, 0 ulp | 3.05 gb/s, 0 ulp |
501
+ | `nk_reduce_minmax_e4m3_serial` | 1.09 gb/s, 0 ulp | 1.21 gb/s, 0 ulp | 1.27 gb/s, 0 ulp |
502
+ | `nk_reduce_minmax_e4m3_v128relaxed` | 6.46 gb/s, 0 ulp | 10.5 gb/s, 0 ulp | 15.0 gb/s, 0 ulp |
498
503
  | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
499
- | `nk_reduce_moments_e3m2_serial` | 1.10 gb/s, 0 ulp | 1.00 gb/s, 0 ulp | 1.06 gb/s, 0 ulp |
500
- | `nk_reduce_moments_e3m2_v128relaxed` | 2.03 gb/s, 0 ulp | 3.92 gb/s, 0 ulp | 4.01 gb/s, 0 ulp |
501
- | `nk_reduce_minmax_e3m2_serial` | ? | 0.618 gb/s, 0 ulp | 0.633 gb/s, 0 ulp |
502
- | `nk_reduce_minmax_e3m2_v128relaxed` | 3.34 gb/s, 0 ulp | 12.2 gb/s, 0 ulp | 15.8 gb/s, 0 ulp |
504
+ | `nk_reduce_moments_e3m2_serial` | 0.881 gb/s, 0 ulp | 0.873 gb/s, 0 ulp | 0.868 gb/s, 0 ulp |
505
+ | `nk_reduce_moments_e3m2_v128relaxed` | 5.41 gb/s, 0 ulp | 4.67 gb/s, 0 ulp | 4.34 gb/s, 0 ulp |
506
+ | `nk_reduce_minmax_e3m2_serial` | 0.671 gb/s, 0 ulp | 0.667 gb/s, 0 ulp | 0.666 gb/s, 0 ulp |
507
+ | `nk_reduce_minmax_e3m2_v128relaxed` | 7.95 gb/s, 0 ulp | 13.0 gb/s, 0 ulp | 16.3 gb/s, 0 ulp |
503
508
  | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
504
- | `nk_reduce_moments_e2m3_serial` | 1.10 gb/s, 0 ulp | 1.05 gb/s, 0 ulp | 1.06 gb/s, 0 ulp |
505
- | `nk_reduce_moments_e2m3_v128relaxed` | 3.41 gb/s, 0 ulp | 7.66 gb/s, 0 ulp | 8.01 gb/s, 0 ulp |
506
- | `nk_reduce_minmax_e2m3_serial` | ? | 0.618 gb/s, 0 ulp | 0.629 gb/s, 0 ulp |
507
- | `nk_reduce_minmax_e2m3_v128relaxed` | 3.90 gb/s, 0 ulp | 11.0 gb/s, 0 ulp | 15.9 gb/s, 0 ulp |
509
+ | `nk_reduce_moments_e2m3_serial` | 0.884 gb/s, 0 ulp | 0.877 gb/s, 0 ulp | 0.857 gb/s, 0 ulp |
510
+ | `nk_reduce_moments_e2m3_v128relaxed` | 9.54 gb/s, 0 ulp | 9.49 gb/s, 0 ulp | 8.72 gb/s, 0 ulp |
511
+ | `nk_reduce_minmax_e2m3_serial` | 0.670 gb/s, 0 ulp | 0.667 gb/s, 0 ulp | 0.666 gb/s, 0 ulp |
512
+ | `nk_reduce_minmax_e2m3_v128relaxed` | 9.35 gb/s, 0 ulp | 12.9 gb/s, 0 ulp | 16.4 gb/s, 0 ulp |
508
513
  | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
509
- | `nk_reduce_moments_i8_serial` | 1.86 gb/s | 3.46 gb/s | 3.53 gb/s |
510
- | `nk_reduce_moments_i8_v128relaxed` | 6.64 gb/s | 16.8 gb/s | 13.2 gb/s |
511
- | `nk_reduce_minmax_i8_serial` | ? | 1.38 gb/s | 1.43 gb/s |
512
- | `nk_reduce_minmax_i8_v128relaxed` | 4.67 gb/s | 16.0 gb/s | 21.7 gb/s |
514
+ | `nk_reduce_moments_i8_serial` | 3.78 gb/s | 3.68 gb/s | 3.62 gb/s |
515
+ | `nk_reduce_moments_i8_v128relaxed` | 21.6 gb/s | 22.3 gb/s | 18.7 gb/s |
516
+ | `nk_reduce_minmax_i8_serial` | 1.47 gb/s | 1.49 gb/s | 1.49 gb/s |
517
+ | `nk_reduce_minmax_i8_v128relaxed` | 13.0 gb/s | 19.1 gb/s | 23.1 gb/s |
513
518
  | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
514
- | `nk_reduce_moments_u8_serial` | 2.19 gb/s | 3.45 gb/s | 3.51 gb/s |
515
- | `nk_reduce_moments_u8_v128relaxed` | 6.58 gb/s | 16.7 gb/s | 13.2 gb/s |
516
- | `nk_reduce_minmax_u8_serial` | ? | 1.44 gb/s | 1.52 gb/s |
517
- | `nk_reduce_minmax_u8_v128relaxed` | 4.52 gb/s | 15.3 gb/s | 21.6 gb/s |
519
+ | `nk_reduce_moments_u8_serial` | 3.71 gb/s | 3.68 gb/s | 3.62 gb/s |
520
+ | `nk_reduce_moments_u8_v128relaxed` | 21.6 gb/s | 22.3 gb/s | 18.8 gb/s |
521
+ | `nk_reduce_minmax_u8_serial` | 1.51 gb/s | 1.55 gb/s | 1.55 gb/s |
522
+ | `nk_reduce_minmax_u8_v128relaxed` | 12.6 gb/s | 16.6 gb/s | 22.5 gb/s |
518
523
  | __i16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
519
- | `nk_reduce_moments_i16_serial` | 4.83 gb/s | 6.94 gb/s | 7.03 gb/s |
520
- | `nk_reduce_moments_i16_v128relaxed` | 4.64 gb/s | 8.63 gb/s | 7.74 gb/s |
521
- | `nk_reduce_minmax_i16_serial` | ? | 2.75 gb/s | 2.88 gb/s |
522
- | `nk_reduce_minmax_i16_v128relaxed` | 7.26 gb/s | 12.3 gb/s | 14.9 gb/s |
524
+ | `nk_reduce_moments_i16_serial` | 7.52 gb/s | 7.35 gb/s | 7.33 gb/s |
525
+ | `nk_reduce_moments_i16_v128relaxed` | 13.3 gb/s | 9.97 gb/s | 8.83 gb/s |
526
+ | `nk_reduce_minmax_i16_serial` | 2.96 gb/s | 2.99 gb/s | 2.99 gb/s |
527
+ | `nk_reduce_minmax_i16_v128relaxed` | 18.8 gb/s | 14.5 gb/s | 16.1 gb/s |
523
528
  | __u16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
524
- | `nk_reduce_moments_u16_serial` | 5.49 gb/s | 6.94 gb/s | 6.92 gb/s |
525
- | `nk_reduce_moments_u16_v128relaxed` | 4.71 gb/s | 8.68 gb/s | 7.86 gb/s |
526
- | `nk_reduce_minmax_u16_serial` | ? | 2.88 gb/s | 3.01 gb/s |
527
- | `nk_reduce_minmax_u16_v128relaxed` | 7.79 gb/s | 12.2 gb/s | 14.9 gb/s |
529
+ | `nk_reduce_moments_u16_serial` | 7.51 gb/s | 7.28 gb/s | 7.23 gb/s |
530
+ | `nk_reduce_moments_u16_v128relaxed` | 13.2 gb/s | 9.98 gb/s | 8.84 gb/s |
531
+ | `nk_reduce_minmax_u16_serial` | 3.04 gb/s | 3.11 gb/s | 3.11 gb/s |
532
+ | `nk_reduce_minmax_u16_v128relaxed` | 16.5 gb/s | 14.5 gb/s | 16.2 gb/s |
528
533
  | __i32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
529
- | `nk_reduce_moments_i32_serial` | 4.10 gb/s | 5.38 gb/s | 5.31 gb/s |
530
- | `nk_reduce_moments_i32_v128relaxed` | 2.23 gb/s | 5.74 gb/s | 5.75 gb/s |
531
- | `nk_reduce_minmax_i32_serial` | ? | 6.38 gb/s | 6.73 gb/s |
532
- | `nk_reduce_minmax_i32_v128relaxed` | 5.36 gb/s | 12.9 gb/s | 15.4 gb/s |
534
+ | `nk_reduce_moments_i32_serial` | 5.73 gb/s | 5.66 gb/s | 5.58 gb/s |
535
+ | `nk_reduce_moments_i32_v128relaxed` | 6.40 gb/s | 6.35 gb/s | 6.39 gb/s |
536
+ | `nk_reduce_minmax_i32_serial` | 6.85 gb/s | 6.88 gb/s | 6.90 gb/s |
537
+ | `nk_reduce_minmax_i32_v128relaxed` | 14.0 gb/s | 15.9 gb/s | 16.7 gb/s |
533
538
  | __u32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
534
- | `nk_reduce_moments_u32_serial` | 3.97 gb/s | 5.35 gb/s | 5.29 gb/s |
535
- | `nk_reduce_moments_u32_v128relaxed` | 1.16 gb/s | 2.32 gb/s | 2.30 gb/s |
536
- | `nk_reduce_minmax_u32_serial` | ? | 6.38 gb/s | 6.66 gb/s |
537
- | `nk_reduce_minmax_u32_v128relaxed` | 5.45 gb/s | 12.9 gb/s | 15.7 gb/s |
539
+ | `nk_reduce_moments_u32_serial` | 6.01 gb/s | 5.70 gb/s | 5.54 gb/s |
540
+ | `nk_reduce_moments_u32_v128relaxed` | 1.65 gb/s | 1.58 gb/s | 1.56 gb/s |
541
+ | `nk_reduce_minmax_u32_serial` | 6.80 gb/s | 6.87 gb/s | 6.90 gb/s |
542
+ | `nk_reduce_minmax_u32_v128relaxed` | 14.4 gb/s | 15.9 gb/s | 16.7 gb/s |
538
543
  | __i64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
539
- | `nk_reduce_moments_i64_serial` | 5.57 gb/s | 7.93 gb/s | 7.99 gb/s |
540
- | `nk_reduce_moments_i64_v128relaxed` | 2.61 gb/s | 6.93 gb/s | 6.92 gb/s |
541
- | `nk_reduce_minmax_i64_serial` | ? | 12.7 gb/s | 13.2 gb/s |
542
- | `nk_reduce_minmax_i64_v128relaxed` | 6.85 gb/s | 13.8 gb/s | 15.8 gb/s |
544
+ | `nk_reduce_moments_i64_serial` | 9.52 gb/s | 9.73 gb/s | 9.62 gb/s |
545
+ | `nk_reduce_moments_i64_v128relaxed` | 7.42 gb/s | 7.40 gb/s | 7.50 gb/s |
546
+ | `nk_reduce_minmax_i64_serial` | 13.6 gb/s | 13.7 gb/s | 13.7 gb/s |
547
+ | `nk_reduce_minmax_i64_v128relaxed` | 16.4 gb/s | 16.9 gb/s | 17.0 gb/s |
543
548
  | __u64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
544
- | `nk_reduce_moments_u64_serial` | 7.06 gb/s | 6.46 gb/s | 6.45 gb/s |
545
- | `nk_reduce_moments_u64_v128relaxed` | 1.36 gb/s | 2.55 gb/s | 2.67 gb/s |
546
- | `nk_reduce_minmax_u64_serial` | ? | 12.7 gb/s | 13.1 gb/s |
547
- | `nk_reduce_minmax_u64_v128relaxed` | 1.72 gb/s | 2.79 gb/s | 3.07 gb/s |
549
+ | `nk_reduce_moments_u64_serial` | 10.2 gb/s | 10.3 gb/s | 10.5 gb/s |
550
+ | `nk_reduce_moments_u64_v128relaxed` | 3.04 gb/s | 2.98 gb/s | 2.94 gb/s |
551
+ | `nk_reduce_minmax_u64_serial` | 13.6 gb/s | 13.7 gb/s | 13.7 gb/s |
552
+ | `nk_reduce_minmax_u64_v128relaxed` | 3.31 gb/s | 3.33 gb/s | 3.35 gb/s |