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
@@ -4,15 +4,15 @@ NumKong implements batched distance matrix computation via pre-packed dot produc
4
4
 
5
5
  Angular distance from pre-packed dot products:
6
6
 
7
- ```math
7
+ $$
8
8
  D_{ij} = 1 - \frac{C_{ij}}{\sqrt{\|A_i\|^2 \cdot \|B_j\|^2}}
9
- ```
9
+ $$
10
10
 
11
11
  Euclidean distance from pre-packed dot products:
12
12
 
13
- ```math
13
+ $$
14
14
  D_{ij} = \sqrt{\|A_i\|^2 + \|B_j\|^2 - 2 C_{ij}}
15
- ```
15
+ $$
16
16
 
17
17
  Reformulating as Python pseudocode:
18
18
 
@@ -359,222 +359,222 @@ Measured with Wasmtime v42 (Cranelift backend).
359
359
  | `nk_euclideans_packed_u4_serial` | 5.25 gso/s, 0.49 ulp | 5.64 gso/s, 0.54 ulp | 5.82 gso/s, 0.64 ulp |
360
360
  | `nk_euclideans_symmetric_u4_serial` | 2.89 gso/s, 0.48 ulp | 4.30 gso/s, 0.53 ulp | 4.86 gso/s, 0.63 ulp |
361
361
 
362
- ### Apple M4
362
+ ### Apple M5
363
363
 
364
364
  #### Native
365
365
 
366
366
  | Kernel | 256³ | 1024³ | 4096³ |
367
367
  | :--------------------------------------- | -----------------------: | -----------------------: | -----------------------: |
368
368
  | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
369
- | `nk_angulars_packed_f64_serial` | 1.96 gso/s, 0 ulp | 1.82 gso/s, 0 ulp | 2.42 gso/s, 0 ulp |
370
- | `nk_angulars_symmetric_f64_serial` | 1.41 gso/s, 0.04 ulp | 1.41 gso/s, 0.02 ulp | 1.42 gso/s, 0.009 ulp |
371
- | `nk_euclideans_packed_f64_serial` | 1.95 gso/s, 0.6 ulp | 1.82 gso/s, 0.6 ulp | 2.42 gso/s, 0.6 ulp |
372
- | `nk_euclideans_symmetric_f64_serial` | 1.41 gso/s, 0.6 ulp | 1.41 gso/s, 0.6 ulp | 1.42 gso/s, 0.6 ulp |
373
- | `nk_angulars_packed_f64_neon` | 5.44 gso/s, 7,798 ulp | 5.38 gso/s, 3,868 ulp | 5.64 gso/s, 1,720 ulp |
374
- | `nk_angulars_symmetric_f64_neon` | 5.47 gso/s, 7,660 ulp | 3.93 gso/s, 3,790 ulp | 3.77 gso/s, 1,720 ulp |
375
- | `nk_euclideans_packed_f64_neon` | 5.69 gso/s, 0.2 ulp | 5.03 gso/s, 0.2 ulp | 5.35 gso/s, 0.2 ulp |
376
- | `nk_euclideans_symmetric_f64_neon` | 5.21 gso/s, 0.2 ulp | 5.26 gso/s, 0.2 ulp | 3.99 gso/s, 0.2 ulp |
377
- | `nk_angulars_packed_f64_smef64` | 24.3 gso/s, 0.0165 ulp | 28.8 gso/s, 0.0165 ulp | 14.1 gso/s, 0.0165 ulp |
378
- | `nk_angulars_symmetric_f64_smef64` | 8.93 gso/s, 0.0182 ulp | 9.57 gso/s, 0.0182 ulp | 9.16 gso/s, 0.0182 ulp |
379
- | `nk_euclideans_packed_f64_smef64` | 32.9 gso/s, 0.236 ulp | 35.4 gso/s, 0.237 ulp | 32.4 gso/s, 0.237 ulp |
380
- | `nk_euclideans_symmetric_f64_smef64` | 9.89 gso/s, 0.281 ulp | 10.3 gso/s, 0.281 ulp | 9.83 gso/s, 0.281 ulp |
369
+ | `nk_angulars_packed_f64_serial` | 2.37 gso/s, 0 ulp | 2.35 gso/s, 0 ulp | 2.67 gso/s, 0 ulp |
370
+ | `nk_angulars_symmetric_f64_serial` | 1.36 gso/s, 0.04 ulp | 1.41 gso/s, 0.02 ulp | 1.56 gso/s, 0.01 ulp |
371
+ | `nk_euclideans_packed_f64_serial` | 2.36 gso/s, 0.6 ulp | 2.41 gso/s, 0.6 ulp | 2.67 gso/s, 0.6 ulp |
372
+ | `nk_euclideans_symmetric_f64_serial` | 1.44 gso/s, 0.6 ulp | 1.52 gso/s, 0.6 ulp | 1.56 gso/s, 0.6 ulp |
373
+ | `nk_angulars_packed_f64_neon` | 6.05 gso/s, 7,798 ulp | 6.28 gso/s, 3,868 ulp | 6.34 gso/s, 1,720 ulp |
374
+ | `nk_angulars_symmetric_f64_neon` | 5.29 gso/s, 7,660 ulp | 5.39 gso/s, 3,790 ulp | 5.44 gso/s, 1,720 ulp |
375
+ | `nk_euclideans_packed_f64_neon` | 5.97 gso/s, 0.2 ulp | 5.97 gso/s, 0.2 ulp | 6.37 gso/s, 0.2 ulp |
376
+ | `nk_euclideans_symmetric_f64_neon` | 5.25 gso/s, 0.2 ulp | 5.29 gso/s, 0.2 ulp | 5.48 gso/s, 0.2 ulp |
377
+ | `nk_angulars_packed_f64_smef64` | 40.6 gso/s, 0.02 ulp | 44.7 gso/s, 0.02 ulp | 46.0 gso/s, 0.02 ulp |
378
+ | `nk_angulars_symmetric_f64_smef64` | 19.9 gso/s, 0.02 ulp | 24.1 gso/s, 0.02 ulp | 20.8 gso/s, 0.02 ulp |
379
+ | `nk_euclideans_packed_f64_smef64` | 41.0 gso/s, 0.24 ulp | 44.9 gso/s, 0.24 ulp | 46.1 gso/s, 0.24 ulp |
380
+ | `nk_euclideans_symmetric_f64_smef64` | 20.2 gso/s, 0.28 ulp | 24.1 gso/s, 0.28 ulp | 20.9 gso/s, 0.28 ulp |
381
381
  | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
382
- | `nk_angulars_packed_f32_serial` | 10.3 gso/s, 0.1 ulp | 10.7 gso/s, 0.1 ulp | 10.5 gso/s, 0.1 ulp |
383
- | `nk_angulars_symmetric_f32_serial` | 8.26 gso/s, 0.3 ulp | 8.53 gso/s, 0.3 ulp | 8.47 gso/s, 0.1 ulp |
384
- | `nk_euclideans_packed_f32_serial` | 10.9 gso/s, 0.6 ulp | 11.0 gso/s, 0.5 ulp | 11.0 gso/s, 0.5 ulp |
385
- | `nk_euclideans_symmetric_f32_serial` | 8.47 gso/s, 3.9 ulp | 8.74 gso/s, 7.9 ulp | 8.77 gso/s, 3.4 ulp |
386
- | `nk_angulars_packed_f32_neon` | 36.5 gso/s, 0 ulp | 40.4 gso/s, 0 ulp | 41.8 gso/s, 1,740 ulp |
387
- | `nk_angulars_symmetric_f32_neon` | 9.99 gso/s, 7,690 ulp | 10.1 gso/s, 3,830 ulp | 10.4 gso/s, 1,730 ulp |
388
- | `nk_euclideans_packed_f32_neon` | 39.8 gso/s, 0.2 ulp | 41.4 gso/s, 0.2 ulp | 42.5 gso/s, 3.5 ulp |
389
- | `nk_euclideans_symmetric_f32_neon` | 10.2 gso/s, 3.8 ulp | 10.4 gso/s, 7.8 ulp | 10.7 gso/s, 3.5 ulp |
390
- | `nk_angulars_packed_f32_smef64` | 54.2 gso/s, 0.153 ulp | 69.1 gso/s, 0.153 ulp | 57.7 gso/s, 0.153 ulp |
391
- | `nk_angulars_symmetric_f32_smef64` | 33.7 gso/s, 0.128 ulp | 38.1 gso/s, 0.128 ulp | 28.2 gso/s, 0.128 ulp |
392
- | `nk_euclideans_packed_f32_smef64` | 120 gso/s, 2.16 ulp | 108 gso/s, 2.16 ulp | 131 gso/s, 2.16 ulp |
393
- | `nk_euclideans_symmetric_f32_smef64` | 37.0 gso/s, 1.50 ulp | 29.3 gso/s, 1.50 ulp | 36.5 gso/s, 1.50 ulp |
382
+ | `nk_angulars_packed_f32_serial` | 10.7 gso/s, 0.1 ulp | 11.7 gso/s, 0.1 ulp | 12.2 gso/s, 0.1 ulp |
383
+ | `nk_angulars_symmetric_f32_serial` | 8.31 gso/s, 0.3 ulp | 8.76 gso/s, 0.3 ulp | 9.69 gso/s, 0.1 ulp |
384
+ | `nk_euclideans_packed_f32_serial` | 10.5 gso/s, 0.6 ulp | 11.4 gso/s, 0.5 ulp | 12.5 gso/s, 0.5 ulp |
385
+ | `nk_euclideans_symmetric_f32_serial` | 8.89 gso/s, 3.9 ulp | 8.91 gso/s, 7.9 ulp | 9.69 gso/s, 3.4 ulp |
386
+ | `nk_angulars_packed_f32_neon` | 37.6 gso/s, 0 ulp | 40.6 gso/s, 0 ulp | 42.2 gso/s, 1,740 ulp |
387
+ | `nk_angulars_symmetric_f32_neon` | 9.73 gso/s, 7,690 ulp | 10.5 gso/s, 3,830 ulp | 10.8 gso/s, 1,730 ulp |
388
+ | `nk_euclideans_packed_f32_neon` | 37.9 gso/s, 0.2 ulp | 39.7 gso/s, 0.2 ulp | 42.0 gso/s, 3.5 ulp |
389
+ | `nk_euclideans_symmetric_f32_neon` | 10.1 gso/s, 3.8 ulp | 10.3 gso/s, 7.8 ulp | 10.9 gso/s, 3.5 ulp |
390
+ | `nk_angulars_packed_f32_smef64` | 149 gso/s, 0.15 ulp | 230 gso/s, 0.15 ulp | 214 gso/s, 0.15 ulp |
391
+ | `nk_angulars_symmetric_f32_smef64` | 50.7 gso/s, 0.13 ulp | 85.4 gso/s, 0.13 ulp | 54.0 gso/s, 0.13 ulp |
392
+ | `nk_euclideans_packed_f32_smef64` | 151 gso/s, 2.2 ulp | 230 gso/s, 2.2 ulp | 213 gso/s, 2.2 ulp |
393
+ | `nk_euclideans_symmetric_f32_smef64` | 51.7 gso/s, 1.5 ulp | 86.1 gso/s, 1.5 ulp | 54.2 gso/s, 1.5 ulp |
394
394
  | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
395
- | `nk_angulars_packed_bf16_serial` | 14.3 gso/s, 0 ulp | 14.8 gso/s, 0.1 ulp | 14.7 gso/s, 0 ulp |
396
- | `nk_angulars_symmetric_bf16_serial` | 13.1 gso/s, 0.04 ulp | 13.5 gso/s, 0.1 ulp | 13.1 gso/s, 0.04 ulp |
397
- | `nk_euclideans_packed_bf16_serial` | 14.3 gso/s, 0.6 ulp | 14.8 gso/s, 3.1 ulp | 14.7 gso/s, 2.1 ulp |
398
- | `nk_euclideans_symmetric_bf16_serial` | 13.1 gso/s, 0.6 ulp | 13.5 gso/s, 3.1 ulp | 13.1 gso/s, 2.1 ulp |
399
- | `nk_angulars_packed_bf16_neonbfdot` | 55.8 gso/s, 0 ulp | 57.6 gso/s, 0.1 ulp | 58.5 gso/s, 0.04 ulp |
400
- | `nk_angulars_symmetric_bf16_neonbfdot` | 36.8 gso/s, 0 ulp | 38.6 gso/s, 0.1 ulp | 36.7 gso/s, 0.04 ulp |
401
- | `nk_euclideans_packed_bf16_neonbfdot` | 56.4 gso/s, 0.3 ulp | 57.7 gso/s, 2.9 ulp | 58.6 gso/s, 1.9 ulp |
402
- | `nk_euclideans_symmetric_bf16_neonbfdot` | 36.9 gso/s, 0.3 ulp | 39.2 gso/s, 2.9 ulp | 40.3 gso/s, 1.9 ulp |
403
- | `nk_angulars_packed_bf16_sme` | 272 gso/s, 0.0363 ulp | 274 gso/s, 0.0363 ulp | 377 gso/s, 0.0363 ulp |
404
- | `nk_angulars_symmetric_bf16_sme` | 103 gso/s, 0.0256 ulp | 103 gso/s, 0.0256 ulp | 120 gso/s, 0.0255 ulp |
405
- | `nk_euclideans_packed_bf16_sme` | 356 gso/s, 0.538 ulp | 335 gso/s, 0.538 ulp | 400 gso/s, 0.538 ulp |
406
- | `nk_euclideans_symmetric_bf16_sme` | 119 gso/s, 0.275 ulp | 99.8 gso/s, 0.275 ulp | 96.8 gso/s, 0.275 ulp |
395
+ | `nk_angulars_packed_bf16_serial` | 18.0 gso/s, 0 ulp | 19.9 gso/s, 0.1 ulp | 20.7 gso/s, 0 ulp |
396
+ | `nk_angulars_symmetric_bf16_serial` | 15.6 gso/s, 0.04 ulp | 16.9 gso/s, 0.1 ulp | 18.6 gso/s, 0.04 ulp |
397
+ | `nk_euclideans_packed_bf16_serial` | 19.1 gso/s, 0.6 ulp | 19.5 gso/s, 3.1 ulp | 21.8 gso/s, 2.1 ulp |
398
+ | `nk_euclideans_symmetric_bf16_serial` | 15.9 gso/s, 0.6 ulp | 16.9 gso/s, 3.1 ulp | 18.6 gso/s, 2.1 ulp |
399
+ | `nk_angulars_packed_bf16_neonbfdot` | 56.0 gso/s, 0 ulp | 57.4 gso/s, 0.1 ulp | 63.2 gso/s, 0.04 ulp |
400
+ | `nk_angulars_symmetric_bf16_neonbfdot` | 37.5 gso/s, 0 ulp | 39.6 gso/s, 0.1 ulp | 43.4 gso/s, 0.04 ulp |
401
+ | `nk_euclideans_packed_bf16_neonbfdot` | 55.7 gso/s, 0.3 ulp | 56.5 gso/s, 2.9 ulp | 62.1 gso/s, 1.9 ulp |
402
+ | `nk_euclideans_symmetric_bf16_neonbfdot` | 39.0 gso/s, 0.3 ulp | 42.1 gso/s, 2.9 ulp | 43.2 gso/s, 1.9 ulp |
403
+ | `nk_angulars_packed_bf16_sme` | 400 gso/s, 0.04 ulp | 821 gso/s, 0.04 ulp | 1,082 gso/s, 0.04 ulp |
404
+ | `nk_angulars_symmetric_bf16_sme` | 218 gso/s, 0.03 ulp | 464 gso/s, 0.03 ulp | 442 gso/s, 0.03 ulp |
405
+ | `nk_euclideans_packed_bf16_sme` | 468 gso/s, 0.54 ulp | 886 gso/s, 0.54 ulp | 1,109 gso/s, 0.54 ulp |
406
+ | `nk_euclideans_symmetric_bf16_sme` | 207 gso/s, 0.28 ulp | 473 gso/s, 0.28 ulp | 445 gso/s, 0.28 ulp |
407
407
  | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
408
- | `nk_angulars_packed_f16_serial` | 12.9 gso/s, 0.1 ulp | 13.3 gso/s, 0.1 ulp | 12.7 gso/s, 0.1 ulp |
409
- | `nk_angulars_symmetric_f16_serial` | 17.7 gso/s, 0.1 ulp | 16.9 gso/s, 0.09 ulp | 16.1 gso/s, 0.1 ulp |
410
- | `nk_euclideans_packed_f16_serial` | 12.9 gso/s, 1.1 ulp | 13.4 gso/s, 0.7 ulp | 12.7 gso/s, 5.6 ulp |
411
- | `nk_euclideans_symmetric_f16_serial` | 17.7 gso/s, 1.1 ulp | 16.9 gso/s, 0.7 ulp | 16.1 gso/s, 5.6 ulp |
412
- | `nk_angulars_packed_f16_neonhalf` | 76.7 gso/s, 0.1 ulp | 74.4 gso/s, 0.1 ulp | 83.6 gso/s, 0.1 ulp |
413
- | `nk_angulars_symmetric_f16_neonhalf` | 20.8 gso/s, 0.1 ulp | 20.2 gso/s, 0.1 ulp | 20.9 gso/s, 0.1 ulp |
414
- | `nk_euclideans_packed_f16_neonhalf` | 80.9 gso/s, 0.9 ulp | 77.3 gso/s, 0.7 ulp | 84.0 gso/s, 5.9 ulp |
415
- | `nk_euclideans_symmetric_f16_neonhalf` | 21.1 gso/s, 0.9 ulp | 20.6 gso/s, 0.6 ulp | 20.9 gso/s, 5.8 ulp |
416
- | `nk_angulars_packed_f16_neonfhm` | 101 gso/s, 0.1 ulp | 109 gso/s, 0.1 ulp | 113 gso/s, 0.1 ulp |
417
- | `nk_angulars_symmetric_f16_neonfhm` | 33.8 gso/s, 0.1 ulp | 35.9 gso/s, 0.1 ulp | 36.4 gso/s, 0.1 ulp |
418
- | `nk_euclideans_packed_f16_neonfhm` | 102 gso/s, 0.9 ulp | 106 gso/s, 0.7 ulp | 106 gso/s, 5.9 ulp |
419
- | `nk_euclideans_symmetric_f16_neonfhm` | 34.1 gso/s, 0.9 ulp | 36.0 gso/s, 0.6 ulp | 36.4 gso/s, 5.8 ulp |
420
- | `nk_angulars_packed_f16_sme` | 342 gso/s, 3,480 ulp | 344 gso/s, 3,480 ulp | 348 gso/s, 3,480 ulp |
421
- | `nk_angulars_symmetric_f16_sme` | 120 gso/s, 5,050 ulp | 120 gso/s, 5,050 ulp | 94.6 gso/s, 5,080 ulp |
422
- | `nk_euclideans_packed_f16_sme` | 414 gso/s, 62,900 ulp | 419 gso/s, 62,900 ulp | 318 gso/s, 62,900 ulp |
423
- | `nk_euclideans_symmetric_f16_sme` | 104 gso/s, 91,000 ulp | 134 gso/s, 91,500 ulp | 140 gso/s, 91,500 ulp |
408
+ | `nk_angulars_packed_f16_serial` | 12.8 gso/s, 0.1 ulp | 14.4 gso/s, 0.1 ulp | 14.9 gso/s, 0.1 ulp |
409
+ | `nk_angulars_symmetric_f16_serial` | 21.7 gso/s, 0.1 ulp | 25.2 gso/s, 0.09 ulp | 28.2 gso/s, 0.1 ulp |
410
+ | `nk_euclideans_packed_f16_serial` | 13.1 gso/s, 1.1 ulp | 13.9 gso/s, 0.7 ulp | 15.7 gso/s, 5.6 ulp |
411
+ | `nk_euclideans_symmetric_f16_serial` | 23.6 gso/s, 1.1 ulp | 25.2 gso/s, 0.7 ulp | 28.4 gso/s, 5.6 ulp |
412
+ | `nk_angulars_packed_f16_neonhalf` | 72.2 gso/s, 0.1 ulp | 78.6 gso/s, 0.1 ulp | 83.8 gso/s, 0.1 ulp |
413
+ | `nk_angulars_symmetric_f16_neonhalf` | 19.3 gso/s, 0.1 ulp | 20.9 gso/s, 0.1 ulp | 21.8 gso/s, 0.1 ulp |
414
+ | `nk_euclideans_packed_f16_neonhalf` | 73.0 gso/s, 0.9 ulp | 76.2 gso/s, 0.7 ulp | 83.7 gso/s, 5.9 ulp |
415
+ | `nk_euclideans_symmetric_f16_neonhalf` | 19.2 gso/s, 0.9 ulp | 20.2 gso/s, 0.6 ulp | 21.9 gso/s, 5.8 ulp |
416
+ | `nk_angulars_packed_f16_neonfhm` | 96.2 gso/s, 0.1 ulp | 107 gso/s, 0.1 ulp | 118 gso/s, 0.1 ulp |
417
+ | `nk_angulars_symmetric_f16_neonfhm` | 35.4 gso/s, 0.1 ulp | 39.1 gso/s, 0.1 ulp | 42.5 gso/s, 0.1 ulp |
418
+ | `nk_euclideans_packed_f16_neonfhm` | 100 gso/s, 0.9 ulp | 110 gso/s, 0.7 ulp | 119 gso/s, 5.9 ulp |
419
+ | `nk_euclideans_symmetric_f16_neonfhm` | 37.2 gso/s, 0.9 ulp | 39.4 gso/s, 0.6 ulp | 42.0 gso/s, 5.8 ulp |
420
+ | `nk_angulars_packed_f16_sme` | 419 gso/s, 0.1 ulp | 839 gso/s, 0.1 ulp | 1,091 gso/s, 0.1 ulp |
421
+ | `nk_angulars_symmetric_f16_sme` | 241 gso/s, 0.1 ulp | 487 gso/s, 0.1 ulp | 450 gso/s, 0.1 ulp |
422
+ | `nk_euclideans_packed_f16_sme` | 491 gso/s, 0 ulp | 906 gso/s, 0.06 ulp | 1,118 gso/s, 2.9 ulp |
423
+ | `nk_euclideans_symmetric_f16_sme` | 227 gso/s, 0.3 ulp | 500 gso/s, 0.6 ulp | 451 gso/s, 0.3 ulp |
424
424
  | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
425
- | `nk_angulars_packed_e5m2_serial` | 8.99 gso/s, 0 ulp | 9.47 gso/s, 0 ulp | 9.55 gso/s, 0 ulp |
426
- | `nk_angulars_symmetric_e5m2_serial` | 2.53 gso/s, 0 ulp | 2.66 gso/s, 0 ulp | 5.29 gso/s, 0 ulp |
427
- | `nk_euclideans_packed_e5m2_serial` | 9.08 gso/s, 0.5 ulp | 9.35 gso/s, 0.5 ulp | 9.59 gso/s, 0.5 ulp |
428
- | `nk_euclideans_symmetric_e5m2_serial` | 2.50 gso/s, 0.5 ulp | 2.67 gso/s, 0.5 ulp | 5.32 gso/s, 0.5 ulp |
429
- | `nk_angulars_packed_e5m2_neonfhm` | 87.6 gso/s, 0 ulp | 96.1 gso/s, 0 ulp | 98.7 gso/s, 0 ulp |
430
- | `nk_angulars_symmetric_e5m2_neonfhm` | 60.0 gso/s, 0 ulp | 68.7 gso/s, 0 ulp | 63.3 gso/s, 0 ulp |
431
- | `nk_euclideans_packed_e5m2_neonfhm` | 89.4 gso/s, 0 ulp | 96.6 gso/s, 0 ulp | 98.9 gso/s, 0 ulp |
432
- | `nk_euclideans_symmetric_e5m2_neonfhm` | 57.8 gso/s, 0 ulp | 60.3 gso/s, 0 ulp | 64.4 gso/s, 0 ulp |
433
- | `nk_angulars_packed_e5m2_sme` | 287 gso/s, 0.0130 ulp | 289 gso/s, 0.0130 ulp | 250 gso/s, 0.0130 ulp |
434
- | `nk_angulars_symmetric_e5m2_sme` | 130 gso/s, 0.0130 ulp | 149 gso/s, 0.0130 ulp | 132 gso/s, 0.0130 ulp |
435
- | `nk_euclideans_packed_e5m2_sme` | 590 gso/s, 0.00483 ulp | 589 gso/s, 0.00483 ulp | 590 gso/s, 0.00482 ulp |
436
- | `nk_euclideans_symmetric_e5m2_sme` | 294 gso/s, 0.00350 ulp | 293 gso/s, 0.00350 ulp | 292 gso/s, 0.00350 ulp |
425
+ | `nk_angulars_packed_e5m2_serial` | 15.8 gso/s, 0 ulp | 16.7 gso/s, 0 ulp | 17.2 gso/s, 0 ulp |
426
+ | `nk_angulars_symmetric_e5m2_serial` | 7.78 gso/s, 0 ulp | 8.37 gso/s, 0 ulp | 8.99 gso/s, 0 ulp |
427
+ | `nk_euclideans_packed_e5m2_serial` | 15.5 gso/s, 0.5 ulp | 16.7 gso/s, 0.5 ulp | 17.2 gso/s, 0.5 ulp |
428
+ | `nk_euclideans_symmetric_e5m2_serial` | 7.93 gso/s, 0.5 ulp | 8.37 gso/s, 0.5 ulp | 8.99 gso/s, 0.5 ulp |
429
+ | `nk_angulars_packed_e5m2_neonfhm` | 84.3 gso/s, 0 ulp | 97.3 gso/s, 0 ulp | 103 gso/s, 0 ulp |
430
+ | `nk_angulars_symmetric_e5m2_neonfhm` | 58.8 gso/s, 0 ulp | 73.2 gso/s, 0 ulp | 79.3 gso/s, 0 ulp |
431
+ | `nk_euclideans_packed_e5m2_neonfhm` | 88.1 gso/s, 0 ulp | 110 gso/s, 0 ulp | 119 gso/s, 0 ulp |
432
+ | `nk_euclideans_symmetric_e5m2_neonfhm` | 66.1 gso/s, 0 ulp | 60.3 gso/s, 0 ulp | 64.4 gso/s, 0 ulp |
433
+ | `nk_angulars_packed_e5m2_sme` | 350 gso/s, 0.01 ulp | 609 gso/s, 0.01 ulp | 744 gso/s, 0.01 ulp |
434
+ | `nk_angulars_symmetric_e5m2_sme` | 138 gso/s, 0.01 ulp | 204 gso/s, 0.01 ulp | 226 gso/s, 0.01 ulp |
435
+ | `nk_euclideans_packed_e5m2_sme` | 399 gso/s, 0.005 ulp | 655 gso/s, 0.005 ulp | 762 gso/s, 0.005 ulp |
436
+ | `nk_euclideans_symmetric_e5m2_sme` | 132 gso/s, 0.004 ulp | 206 gso/s, 0.004 ulp | 227 gso/s, 0.004 ulp |
437
437
  | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
438
- | `nk_angulars_packed_e4m3_serial` | 0.837 gso/s, 0 ulp | 0.823 gso/s, 0 ulp | 0.831 gso/s, 0 ulp |
439
- | `nk_angulars_symmetric_e4m3_serial` | 0.419 gso/s, 0.03 ulp | 0.423 gso/s, 0.02 ulp | 0.852 gso/s, 0.01 ulp |
440
- | `nk_euclideans_packed_e4m3_serial` | 0.831 gso/s, 0.5 ulp | 0.840 gso/s, 0.5 ulp | 0.836 gso/s, 0.5 ulp |
441
- | `nk_euclideans_symmetric_e4m3_serial` | 0.421 gso/s, 0.5 ulp | 0.429 gso/s, 0.5 ulp | 0.857 gso/s, 0.3 ulp |
442
- | `nk_angulars_packed_e4m3_neonfhm` | 27.3 gso/s, 0 ulp | 27.8 gso/s, 0 ulp | 33.2 gso/s, 0 ulp |
443
- | `nk_angulars_symmetric_e4m3_neonfhm` | 33.0 gso/s, 0 ulp | 35.3 gso/s, 0 ulp | 35.3 gso/s, 0 ulp |
444
- | `nk_euclideans_packed_e4m3_neonfhm` | 29.5 gso/s, 0 ulp | 26.6 gso/s, 0 ulp | 27.4 gso/s, 0.2 ulp |
445
- | `nk_euclideans_symmetric_e4m3_neonfhm` | 32.1 gso/s, 0 ulp | 33.9 gso/s, 0 ulp | 33.3 gso/s, 0.2 ulp |
446
- | `nk_angulars_packed_e4m3_sme` | 90.3 gso/s, 0.0131 ulp | 89.9 gso/s, 0.0131 ulp | 96.6 gso/s, 0.0131 ulp |
447
- | `nk_angulars_symmetric_e4m3_sme` | 51.5 gso/s, 0.0131 ulp | 38.1 gso/s, 0.0131 ulp | 51.2 gso/s, 0.0131 ulp |
448
- | `nk_euclideans_packed_e4m3_sme` | 233 gso/s, 0.113 ulp | 185 gso/s, 0.113 ulp | 233 gso/s, 0.113 ulp |
449
- | `nk_euclideans_symmetric_e4m3_sme` | 64.2 gso/s, 0.113 ulp | 64.4 gso/s, 0.113 ulp | 64.2 gso/s, 0.113 ulp |
438
+ | `nk_angulars_packed_e4m3_serial` | 1.15 gso/s, 0 ulp | 1.20 gso/s, 0 ulp | 1.24 gso/s, 0 ulp |
439
+ | `nk_angulars_symmetric_e4m3_serial` | 1.22 gso/s, 0.03 ulp | 1.24 gso/s, 0.02 ulp | 1.32 gso/s, 0.01 ulp |
440
+ | `nk_euclideans_packed_e4m3_serial` | 1.23 gso/s, 0.5 ulp | 1.20 gso/s, 0.5 ulp | 1.24 gso/s, 0.5 ulp |
441
+ | `nk_euclideans_symmetric_e4m3_serial` | 1.25 gso/s, 0.5 ulp | 1.24 gso/s, 0.5 ulp | 1.32 gso/s, 0.3 ulp |
442
+ | `nk_angulars_packed_e4m3_neonfhm` | 29.1 gso/s, 0 ulp | 32.2 gso/s, 0 ulp | 34.1 gso/s, 0 ulp |
443
+ | `nk_angulars_symmetric_e4m3_neonfhm` | 32.0 gso/s, 0 ulp | 36.6 gso/s, 0 ulp | 38.9 gso/s, 0 ulp |
444
+ | `nk_euclideans_packed_e4m3_neonfhm` | 30.0 gso/s, 0 ulp | 32.2 gso/s, 0 ulp | 34.1 gso/s, 0.2 ulp |
445
+ | `nk_euclideans_symmetric_e4m3_neonfhm` | 34.1 gso/s, 0 ulp | 36.6 gso/s, 0 ulp | 38.9 gso/s, 0.2 ulp |
446
+ | `nk_angulars_packed_e4m3_sme` | 184 gso/s, 0.01 ulp | 272 gso/s, 0.01 ulp | 307 gso/s, 0.01 ulp |
447
+ | `nk_angulars_symmetric_e4m3_sme` | 56.4 gso/s, 0.01 ulp | 74.6 gso/s, 0.01 ulp | 78.5 gso/s, 0.01 ulp |
448
+ | `nk_euclideans_packed_e4m3_sme` | 200 gso/s, 0.11 ulp | 279 gso/s, 0.11 ulp | 310 gso/s, 0.11 ulp |
449
+ | `nk_euclideans_symmetric_e4m3_sme` | 55.5 gso/s, 0.11 ulp | 75.1 gso/s, 0.11 ulp | 78.4 gso/s, 0.11 ulp |
450
450
  | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
451
- | `nk_angulars_packed_e3m2_serial` | 8.34 gso/s, 0 ulp | 8.62 gso/s, 0 ulp | 9.17 gso/s, 0 ulp |
452
- | `nk_angulars_symmetric_e3m2_serial` | 2.48 gso/s, 0 ulp | 2.67 gso/s, 0 ulp | 5.36 gso/s, 0 ulp |
453
- | `nk_euclideans_packed_e3m2_serial` | 8.74 gso/s, 0.5 ulp | 9.08 gso/s, 0.5 ulp | 8.91 gso/s, 0.5 ulp |
454
- | `nk_euclideans_symmetric_e3m2_serial` | 2.50 gso/s, 0.5 ulp | 2.69 gso/s, 0.5 ulp | 5.22 gso/s, 0.5 ulp |
455
- | `nk_angulars_packed_e3m2_sme` | 183 gso/s, 0.0130 ulp | 185 gso/s, 0.0130 ulp | 177 gso/s, 0.0130 ulp |
456
- | `nk_angulars_symmetric_e3m2_sme` | 87.8 gso/s, 0.0131 ulp | 78.5 gso/s, 0.0131 ulp | 87.5 gso/s, 0.0131 ulp |
457
- | `nk_euclideans_packed_e3m2_sme` | 423 gso/s, 0 ulp | 422 gso/s, 0 ulp | 422 gso/s, 0 ulp |
458
- | `nk_euclideans_symmetric_e3m2_sme` | 121 gso/s, 0 ulp | 122 gso/s, 0 ulp | 122 gso/s, 0 ulp |
451
+ | `nk_angulars_packed_e3m2_serial` | 14.2 gso/s, 0 ulp | 14.6 gso/s, 0 ulp | 15.5 gso/s, 0 ulp |
452
+ | `nk_angulars_symmetric_e3m2_serial` | 7.77 gso/s, 0 ulp | 8.10 gso/s, 0 ulp | 9.05 gso/s, 0 ulp |
453
+ | `nk_euclideans_packed_e3m2_serial` | 13.9 gso/s, 0.5 ulp | 14.6 gso/s, 0.5 ulp | 15.5 gso/s, 0.5 ulp |
454
+ | `nk_euclideans_symmetric_e3m2_serial` | 8.08 gso/s, 0.5 ulp | 8.10 gso/s, 0.5 ulp | 9.05 gso/s, 0.5 ulp |
455
+ | `nk_angulars_packed_e3m2_sme` | 327 gso/s, 0.01 ulp | 573 gso/s, 0.01 ulp | 690 gso/s, 0.01 ulp |
456
+ | `nk_angulars_symmetric_e3m2_sme` | 124 gso/s, 0.01 ulp | 184 gso/s, 0.01 ulp | 204 gso/s, 0.01 ulp |
457
+ | `nk_euclideans_packed_e3m2_sme` | 379 gso/s, 0 ulp | 604 gso/s, 0 ulp | 702 gso/s, 0 ulp |
458
+ | `nk_euclideans_symmetric_e3m2_sme` | 119 gso/s, 0 ulp | 186 gso/s, 0 ulp | 205 gso/s, 0 ulp |
459
459
  | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
460
- | `nk_angulars_packed_e2m3_serial` | 8.60 gso/s, 0 ulp | 8.83 gso/s, 0 ulp | 9.14 gso/s, 0 ulp |
461
- | `nk_angulars_symmetric_e2m3_serial` | 2.52 gso/s, 0 ulp | 2.64 gso/s, 0 ulp | 5.43 gso/s, 0 ulp |
462
- | `nk_euclideans_packed_e2m3_serial` | 8.65 gso/s, 0.5 ulp | 9.03 gso/s, 0.5 ulp | 8.98 gso/s, 0.5 ulp |
463
- | `nk_euclideans_symmetric_e2m3_serial` | 2.53 gso/s, 0.5 ulp | 2.68 gso/s, 0.5 ulp | 5.30 gso/s, 0.5 ulp |
464
- | `nk_angulars_packed_e2m3_sme` | 141 gso/s, 0.0132 ulp | 128 gso/s, 0.0132 ulp | 132 gso/s, 0.0132 ulp |
465
- | `nk_angulars_symmetric_e2m3_sme` | 134 gso/s, 0.0130 ulp | 132 gso/s, 0.0130 ulp | 138 gso/s, 0.0130 ulp |
466
- | `nk_euclideans_packed_e2m3_sme` | 307 gso/s, 0 ulp | 307 gso/s, 0 ulp | 307 gso/s, 0 ulp |
467
- | `nk_euclideans_symmetric_e2m3_sme` | 227 gso/s, 0 ulp | 231 gso/s, 0 ulp | 231 gso/s, 0 ulp |
460
+ | `nk_angulars_packed_e2m3_serial` | 14.1 gso/s, 0 ulp | 14.8 gso/s, 0 ulp | 15.5 gso/s, 0 ulp |
461
+ | `nk_angulars_symmetric_e2m3_serial` | 7.89 gso/s, 0 ulp | 8.21 gso/s, 0 ulp | 9.09 gso/s, 0 ulp |
462
+ | `nk_euclideans_packed_e2m3_serial` | 13.6 gso/s, 0.5 ulp | 14.8 gso/s, 0.5 ulp | 15.7 gso/s, 0.5 ulp |
463
+ | `nk_euclideans_symmetric_e2m3_serial` | 7.93 gso/s, 0.5 ulp | 8.21 gso/s, 0.5 ulp | 9.09 gso/s, 0.5 ulp |
464
+ | `nk_angulars_packed_e2m3_sme` | 415 gso/s, 0.01 ulp | 926 gso/s, 0.01 ulp | 1,216 gso/s, 0.01 ulp |
465
+ | `nk_angulars_symmetric_e2m3_sme` | 170 gso/s, 0.01 ulp | 342 gso/s, 0.01 ulp | 404 gso/s, 0.01 ulp |
466
+ | `nk_euclideans_packed_e2m3_sme` | 470 gso/s, 0 ulp | 1,011 gso/s, 0 ulp | 1,269 gso/s, 0 ulp |
467
+ | `nk_euclideans_symmetric_e2m3_sme` | 163 gso/s, 0 ulp | 348 gso/s, 0 ulp | 408 gso/s, 0 ulp |
468
468
  | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
469
- | `nk_angulars_packed_i8_serial` | 12.8 gso/s, 0 ulp | 13.6 gso/s, 0 ulp | 14.3 gso/s, 0 ulp |
470
- | `nk_angulars_symmetric_i8_serial` | 4.30 gso/s, 0 ulp | 4.57 gso/s, 0 ulp | 8.96 gso/s, 0 ulp |
471
- | `nk_euclideans_packed_i8_serial` | 13.1 gso/s, 0.4 ulp | 14.1 gso/s, 0.4 ulp | 13.4 gso/s, 0.4 ulp |
472
- | `nk_euclideans_symmetric_i8_serial` | 4.28 gso/s, 0.4 ulp | 4.43 gso/s, 0.4 ulp | 8.66 gso/s, 0.4 ulp |
473
- | `nk_angulars_packed_i8_neonsdot` | 267 gso/s, 0 ulp | 357 gso/s, 0 ulp | 369 gso/s, 0 ulp |
474
- | `nk_angulars_symmetric_i8_neonsdot` | 69.7 gso/s, 0 ulp | 78.6 gso/s, 0 ulp | 82.6 gso/s, 0 ulp |
475
- | `nk_euclideans_packed_i8_neonsdot` | 294 gso/s, 0 ulp | 365 gso/s, 0 ulp | 476 gso/s, 0 ulp |
476
- | `nk_euclideans_symmetric_i8_neonsdot` | 70.5 gso/s, 0 ulp | 83.9 gso/s, 0 ulp | 87.3 gso/s, 0 ulp |
477
- | `nk_angulars_packed_i8_sme` | 219 gso/s, 0.0133 ulp | 219 gso/s, 0.0133 ulp | 223 gso/s, 0.0133 ulp |
478
- | `nk_angulars_symmetric_i8_sme` | 177 gso/s, 0.0134 ulp | 177 gso/s, 0.0134 ulp | 177 gso/s, 0.0134 ulp |
479
- | `nk_euclideans_packed_i8_sme` | 228 gso/s, 0 ulp | 246 gso/s, 0 ulp | 243 gso/s, 0 ulp |
480
- | `nk_euclideans_symmetric_i8_sme` | 159 gso/s, 0 ulp | 167 gso/s, 0 ulp | 169 gso/s, 0 ulp |
469
+ | `nk_angulars_packed_i8_serial` | 18.3 gso/s, 0 ulp | 20.0 gso/s, 0 ulp | 20.2 gso/s, 0 ulp |
470
+ | `nk_angulars_symmetric_i8_serial` | 13.5 gso/s, 0 ulp | 13.9 gso/s, 0 ulp | 14.8 gso/s, 0 ulp |
471
+ | `nk_euclideans_packed_i8_serial` | 18.7 gso/s, 0.4 ulp | 20.0 gso/s, 0.4 ulp | 20.2 gso/s, 0.4 ulp |
472
+ | `nk_euclideans_symmetric_i8_serial` | 13.7 gso/s, 0.4 ulp | 13.9 gso/s, 0.4 ulp | 14.8 gso/s, 0.4 ulp |
473
+ | `nk_angulars_packed_i8_neonsdot` | 280 gso/s, 0 ulp | 357 gso/s, 0 ulp | 477 gso/s, 0 ulp |
474
+ | `nk_angulars_symmetric_i8_neonsdot` | 74.0 gso/s, 0 ulp | 86.9 gso/s, 0 ulp | 87.2 gso/s, 0 ulp |
475
+ | `nk_euclideans_packed_i8_neonsdot` | 305 gso/s, 0 ulp | 419 gso/s, 0 ulp | 477 gso/s, 0 ulp |
476
+ | `nk_euclideans_symmetric_i8_neonsdot` | 73.4 gso/s, 0 ulp | 87.0 gso/s, 0 ulp | 87.2 gso/s, 0 ulp |
477
+ | `nk_angulars_packed_i8_sme` | 492 gso/s, 0.01 ulp | 1,356 gso/s, 0.01 ulp | 2,166 gso/s, 0.01 ulp |
478
+ | `nk_angulars_symmetric_i8_sme` | 200 gso/s, 0.01 ulp | 873 gso/s, 0.01 ulp | 1,214 gso/s, 0.01 ulp |
479
+ | `nk_euclideans_packed_i8_sme` | 584 gso/s, 0 ulp | 1,546 gso/s, 0 ulp | 2,263 gso/s, 0 ulp |
480
+ | `nk_euclideans_symmetric_i8_sme` | 201 gso/s, 0 ulp | 917 gso/s, 0 ulp | 1,256 gso/s, 0 ulp |
481
481
  | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
482
- | `nk_angulars_packed_u8_serial` | 10.1 gso/s, 0.3 ulp | 10.3 gso/s, 0.3 ulp | 10.6 gso/s, 0.3 ulp |
483
- | `nk_angulars_symmetric_u8_serial` | 4.88 gso/s, 0.3 ulp | 5.11 gso/s, 0.3 ulp | 10.2 gso/s, 0.3 ulp |
484
- | `nk_euclideans_packed_u8_serial` | 10.1 gso/s, 0.5 ulp | 10.5 gso/s, 0.5 ulp | 10.0 gso/s, 0.6 ulp |
485
- | `nk_euclideans_symmetric_u8_serial` | 4.83 gso/s, 0.5 ulp | 5.10 gso/s, 0.5 ulp | 9.83 gso/s, 0.6 ulp |
486
- | `nk_angulars_packed_u8_neonsdot` | 264 gso/s, 0.3 ulp | 345 gso/s, 0.3 ulp | 379 gso/s, 0.3 ulp |
487
- | `nk_angulars_symmetric_u8_neonsdot` | 69.7 gso/s, 0.3 ulp | 78.6 gso/s, 0.3 ulp | 82.3 gso/s, 0.3 ulp |
488
- | `nk_euclideans_packed_u8_neonsdot` | 283 gso/s, 0 ulp | 383 gso/s, 0 ulp | 477 gso/s, 0 ulp |
489
- | `nk_euclideans_symmetric_u8_neonsdot` | 70.8 gso/s, 0 ulp | 85.3 gso/s, 0 ulp | 87.6 gso/s, 0 ulp |
490
- | `nk_angulars_packed_u8_sme` | 181 gso/s, 0.324 ulp | 181 gso/s, 0.324 ulp | 198 gso/s, 0.324 ulp |
491
- | `nk_angulars_symmetric_u8_sme` | 193 gso/s, 0.323 ulp | 192 gso/s, 0.323 ulp | 159 gso/s, 0.323 ulp |
492
- | `nk_euclideans_packed_u8_sme` | 254 gso/s, 0 ulp | 270 gso/s, 0 ulp | 225 gso/s, 0 ulp |
493
- | `nk_euclideans_symmetric_u8_sme` | 170 gso/s, 0 ulp | 215 gso/s, 0 ulp | 162 gso/s, 0 ulp |
482
+ | `nk_angulars_packed_u8_serial` | 15.5 gso/s, 0.3 ulp | 16.3 gso/s, 0.3 ulp | 17.4 gso/s, 0.3 ulp |
483
+ | `nk_angulars_symmetric_u8_serial` | 15.7 gso/s, 0.3 ulp | 16.2 gso/s, 0.3 ulp | 17.5 gso/s, 0.3 ulp |
484
+ | `nk_euclideans_packed_u8_serial` | 16.4 gso/s, 0.5 ulp | 16.3 gso/s, 0.5 ulp | 17.4 gso/s, 0.6 ulp |
485
+ | `nk_euclideans_symmetric_u8_serial` | 16.1 gso/s, 0.5 ulp | 16.2 gso/s, 0.5 ulp | 17.5 gso/s, 0.6 ulp |
486
+ | `nk_angulars_packed_u8_neonsdot` | 284 gso/s, 0.3 ulp | 369 gso/s, 0.3 ulp | 470 gso/s, 0.3 ulp |
487
+ | `nk_angulars_symmetric_u8_neonsdot` | 72.4 gso/s, 0.3 ulp | 87.4 gso/s, 0.3 ulp | 87.7 gso/s, 0.3 ulp |
488
+ | `nk_euclideans_packed_u8_neonsdot` | 302 gso/s, 0 ulp | 419 gso/s, 0 ulp | 470 gso/s, 0 ulp |
489
+ | `nk_euclideans_symmetric_u8_neonsdot` | 72.0 gso/s, 0 ulp | 87.0 gso/s, 0 ulp | 87.7 gso/s, 0 ulp |
490
+ | `nk_angulars_packed_u8_sme` | 492 gso/s, 0.32 ulp | 1,369 gso/s, 0.32 ulp | 2,169 gso/s, 0.32 ulp |
491
+ | `nk_angulars_symmetric_u8_sme` | 201 gso/s, 0.32 ulp | 874 gso/s, 0.32 ulp | 1,217 gso/s, 0.32 ulp |
492
+ | `nk_euclideans_packed_u8_sme` | 584 gso/s, 0 ulp | 1,545 gso/s, 0 ulp | 2,260 gso/s, 0 ulp |
493
+ | `nk_euclideans_symmetric_u8_sme` | 199 gso/s, 0 ulp | 905 gso/s, 0 ulp | 1,248 gso/s, 0 ulp |
494
494
  | __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
495
- | `nk_angulars_packed_i4_serial` | 26.6 gso/s, 0.3 ulp | 27.9 gso/s, 0.3 ulp | 28.1 gso/s, 0.3 ulp |
496
- | `nk_angulars_symmetric_i4_serial` | 7.66 gso/s, 0.3 ulp | 7.84 gso/s, 0.3 ulp | 7.80 gso/s, 0.3 ulp |
497
- | `nk_euclideans_packed_i4_serial` | 26.6 gso/s, 0.5 ulp | 27.6 gso/s, 0.5 ulp | 27.6 gso/s, 0.6 ulp |
498
- | `nk_euclideans_symmetric_i4_serial` | 7.75 gso/s, 0.5 ulp | 7.75 gso/s, 0.5 ulp | 7.82 gso/s, 0.6 ulp |
499
- | `nk_angulars_packed_i4_neonsdot` | 207 gso/s, 0.3 ulp | 263 gso/s, 0.3 ulp | 279 gso/s, 0.3 ulp |
500
- | `nk_angulars_symmetric_i4_neonsdot` | 97.6 gso/s, 0.3 ulp | 135 gso/s, 0.3 ulp | 151 gso/s, 0.3 ulp |
501
- | `nk_euclideans_packed_i4_neonsdot` | 218 gso/s, 0.0 ulp | 266 gso/s, 0.0 ulp | 303 gso/s, 0.0 ulp |
502
- | `nk_euclideans_symmetric_i4_neonsdot` | 101 gso/s, 0.0 ulp | 138 gso/s, 0.0 ulp | 156 gso/s, 0.0 ulp |
503
- | `nk_angulars_packed_i4_sme` | 251 gso/s, 0.324 ulp | 251 gso/s, 0.324 ulp | 256 gso/s, 0.324 ulp |
504
- | `nk_angulars_symmetric_i4_sme` | 274 gso/s, 0.323 ulp | 273 gso/s, 0.323 ulp | 281 gso/s, 0.323 ulp |
505
- | `nk_euclideans_packed_i4_sme` | 260 gso/s, 0 ulp | 319 gso/s, 0 ulp | 275 gso/s, 0 ulp |
506
- | `nk_euclideans_symmetric_i4_sme` | 258 gso/s, 0 ulp | 272 gso/s, 0 ulp | 264 gso/s, 0 ulp |
495
+ | `nk_angulars_packed_i4_serial` | 17.3 gso/s, 0.3 ulp | 18.2 gso/s, 0.3 ulp | 19.6 gso/s, 0.3 ulp |
496
+ | `nk_angulars_symmetric_i4_serial` | 14.3 gso/s, 0.3 ulp | 14.9 gso/s, 0.3 ulp | 15.6 gso/s, 0.3 ulp |
497
+ | `nk_euclideans_packed_i4_serial` | 17.9 gso/s, 0.5 ulp | 18.2 gso/s, 0.5 ulp | 19.6 gso/s, 0.6 ulp |
498
+ | `nk_euclideans_symmetric_i4_serial` | 14.6 gso/s, 0.5 ulp | 14.9 gso/s, 0.5 ulp | 15.6 gso/s, 0.6 ulp |
499
+ | `nk_angulars_packed_i4_neonsdot` | 215 gso/s, 0.3 ulp | 284 gso/s, 0.3 ulp | 291 gso/s, 0.3 ulp |
500
+ | `nk_angulars_symmetric_i4_neonsdot` | 104 gso/s, 0.3 ulp | 162 gso/s, 0.3 ulp | 171 gso/s, 0.3 ulp |
501
+ | `nk_euclideans_packed_i4_neonsdot` | 225 gso/s, 0 ulp | 284 gso/s, 0 ulp | 291 gso/s, 0 ulp |
502
+ | `nk_euclideans_symmetric_i4_neonsdot` | 105 gso/s, 0 ulp | 162 gso/s, 0 ulp | 171 gso/s, 0 ulp |
503
+ | `nk_angulars_packed_i4_sme` | 486 gso/s, 0.32 ulp | 1,309 gso/s, 0.32 ulp | 2,041 gso/s, 0.32 ulp |
504
+ | `nk_angulars_symmetric_i4_sme` | 201 gso/s, 0.32 ulp | 913 gso/s, 0.32 ulp | 1,488 gso/s, 0.32 ulp |
505
+ | `nk_euclideans_packed_i4_sme` | 576 gso/s, 0 ulp | 1,453 gso/s, 0 ulp | 2,126 gso/s, 0 ulp |
506
+ | `nk_euclideans_symmetric_i4_sme` | 200 gso/s, 0 ulp | 948 gso/s, 0 ulp | 1,527 gso/s, 0 ulp |
507
507
  | __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
508
- | `nk_angulars_packed_u4_serial` | 24.1 gso/s, 0.3 ulp | 25.2 gso/s, 0.3 ulp | 25.9 gso/s, 0.3 ulp |
509
- | `nk_angulars_symmetric_u4_serial` | 8.23 gso/s, 0.3 ulp | 8.59 gso/s, 0.3 ulp | 8.53 gso/s, 0.3 ulp |
510
- | `nk_euclideans_packed_u4_serial` | 24.6 gso/s, 0.5 ulp | 25.5 gso/s, 0.5 ulp | 25.5 gso/s, 0.6 ulp |
511
- | `nk_euclideans_symmetric_u4_serial` | 8.24 gso/s, 0.5 ulp | 8.42 gso/s, 0.5 ulp | 8.54 gso/s, 0.6 ulp |
512
- | `nk_angulars_packed_u4_neonsdot` | 230 gso/s, 0.3 ulp | 302 gso/s, 0.3 ulp | 328 gso/s, 0.3 ulp |
513
- | `nk_angulars_symmetric_u4_neonsdot` | 98.1 gso/s, 0.3 ulp | 134 gso/s, 0.3 ulp | 156 gso/s, 0.3 ulp |
514
- | `nk_euclideans_packed_u4_neonsdot` | 243 gso/s, 0.0 ulp | 311 gso/s, 0.0 ulp | 353 gso/s, 0.0 ulp |
515
- | `nk_euclideans_symmetric_u4_neonsdot` | 101 gso/s, 0.0 ulp | 135 gso/s, 0.0 ulp | 158 gso/s, 0.0 ulp |
516
- | `nk_angulars_packed_u4_sme` | 243 gso/s, 0.324 ulp | 242 gso/s, 0.324 ulp | 263 gso/s, 0.324 ulp |
517
- | `nk_angulars_symmetric_u4_sme` | 300 gso/s, 0.323 ulp | 307 gso/s, 0.323 ulp | 278 gso/s, 0.323 ulp |
518
- | `nk_euclideans_packed_u4_sme` | 307 gso/s, 0 ulp | 287 gso/s, 0 ulp | 287 gso/s, 0 ulp |
519
- | `nk_euclideans_symmetric_u4_sme` | 283 gso/s, 0 ulp | 292 gso/s, 0 ulp | 271 gso/s, 0 ulp |
508
+ | `nk_angulars_packed_u4_serial` | 18.0 gso/s, 0.3 ulp | 19.4 gso/s, 0.3 ulp | 20.6 gso/s, 0.3 ulp |
509
+ | `nk_angulars_symmetric_u4_serial` | 15.5 gso/s, 0.3 ulp | 16.4 gso/s, 0.3 ulp | 17.4 gso/s, 0.3 ulp |
510
+ | `nk_euclideans_packed_u4_serial` | 19.1 gso/s, 0.5 ulp | 19.4 gso/s, 0.5 ulp | 20.6 gso/s, 0.6 ulp |
511
+ | `nk_euclideans_symmetric_u4_serial` | 15.7 gso/s, 0.5 ulp | 16.4 gso/s, 0.5 ulp | 17.4 gso/s, 0.6 ulp |
512
+ | `nk_angulars_packed_u4_neonsdot` | 241 gso/s, 0.3 ulp | 319 gso/s, 0.3 ulp | 340 gso/s, 0.3 ulp |
513
+ | `nk_angulars_symmetric_u4_neonsdot` | 107 gso/s, 0.3 ulp | 166 gso/s, 0.3 ulp | 173 gso/s, 0.3 ulp |
514
+ | `nk_euclideans_packed_u4_neonsdot` | 250 gso/s, 0 ulp | 340 gso/s, 0 ulp | 340 gso/s, 0 ulp |
515
+ | `nk_euclideans_symmetric_u4_neonsdot` | 105 gso/s, 0 ulp | 173 gso/s, 0 ulp | 173 gso/s, 0 ulp |
516
+ | `nk_angulars_packed_u4_sme` | 490 gso/s, 0.32 ulp | 1,322 gso/s, 0.32 ulp | 2,081 gso/s, 0.32 ulp |
517
+ | `nk_angulars_symmetric_u4_sme` | 205 gso/s, 0.32 ulp | 974 gso/s, 0.32 ulp | 1,682 gso/s, 0.32 ulp |
518
+ | `nk_euclideans_packed_u4_sme` | 582 gso/s, 0 ulp | 1,487 gso/s, 0 ulp | 2,162 gso/s, 0 ulp |
519
+ | `nk_euclideans_symmetric_u4_sme` | 205 gso/s, 0 ulp | 1,013 gso/s, 0 ulp | 1,734 gso/s, 0 ulp |
520
520
 
521
521
  #### WASM
522
522
 
523
- Measured with Wasmtime v42 (Cranelift backend).
523
+ Measured with Wasmtime v43 (Cranelift backend).
524
524
 
525
525
  | Kernel | 256³ | 1024³ | 4096³ |
526
526
  | :----------------------------------------- | -----------------------: | -----------------------: | -----------------------: |
527
527
  | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
528
- | `nk_angulars_packed_f64_serial` | 2.04 gso/s, 0 ulp | 5.33 gso/s, 0 ulp | 5.36 gso/s, 0 ulp |
529
- | `nk_angulars_symmetric_f64_serial` | 2.10 gso/s, 0 ulp | 5.54 gso/s, 0 ulp | 11.2 gso/s, 0 ulp |
530
- | `nk_euclideans_packed_f64_serial` | 3.88 gso/s, 0.4 ulp | 5.38 gso/s, 0.4 ulp | 5.61 gso/s, 0.4 ulp |
531
- | `nk_euclideans_symmetric_f64_serial` | 5.20 gso/s, 0.4 ulp | 5.54 gso/s, 0.4 ulp | 11.2 gso/s, 0.4 ulp |
532
- | `nk_angulars_packed_f64_v128relaxed` | 30.1 gso/s, 0.1 ulp | 36.2 gso/s, 0.1 ulp | 37.1 gso/s, 0.1 ulp |
533
- | `nk_angulars_symmetric_f64_v128relaxed` | 9.59 gso/s, 0.1 ulp | 10.7 gso/s, 0.1 ulp | 20.9 gso/s, 0.1 ulp |
534
- | `nk_euclideans_packed_f64_v128relaxed` | 30.4 gso/s, 0.4 ulp | 36.3 gso/s, 0.4 ulp | 37.1 gso/s, 0.4 ulp |
535
- | `nk_euclideans_symmetric_f64_v128relaxed` | 9.58 gso/s, 0.4 ulp | 10.7 gso/s, 0.4 ulp | 20.9 gso/s, 0.4 ulp |
528
+ | `nk_angulars_packed_f64_serial` | 2.33 gso/s, 0 ulp | 2.12 gso/s, 0 ulp | 2.19 gso/s, 0 ulp |
529
+ | `nk_angulars_symmetric_f64_serial` | 1.35 gso/s, 0 ulp | 1.41 gso/s, 0 ulp | 1.57 gso/s, 0 ulp |
530
+ | `nk_euclideans_packed_f64_serial` | 2.35 gso/s, 0.4 ulp | 2.37 gso/s, 0.4 ulp | 2.50 gso/s, 0.4 ulp |
531
+ | `nk_euclideans_symmetric_f64_serial` | 1.40 gso/s, 0.4 ulp | 1.50 gso/s, 0.4 ulp | 1.59 gso/s, 0.4 ulp |
532
+ | `nk_angulars_packed_f64_v128relaxed` | 5.65 gso/s, 0.1 ulp | 5.44 gso/s, 0.1 ulp | 6.22 gso/s, 0.1 ulp |
533
+ | `nk_angulars_symmetric_f64_v128relaxed` | 5.29 gso/s, 0.1 ulp | 5.74 gso/s, 0.1 ulp | 6.01 gso/s, 0.1 ulp |
534
+ | `nk_euclideans_packed_f64_v128relaxed` | 5.69 gso/s, 0.4 ulp | 6.05 gso/s, 0.4 ulp | 6.22 gso/s, 0.4 ulp |
535
+ | `nk_euclideans_symmetric_f64_v128relaxed` | 5.29 gso/s, 0.4 ulp | 5.89 gso/s, 0.4 ulp | 6.02 gso/s, 0.4 ulp |
536
536
  | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
537
- | `nk_angulars_packed_f32_serial` | 8.31 gso/s, 0.1 ulp | 23.8 gso/s, 0.1 ulp | 26.3 gso/s, 0.1 ulp |
538
- | `nk_angulars_symmetric_f32_serial` | 6.06 gso/s, 0.1 ulp | 17.7 gso/s, 0.1 ulp | 35.0 gso/s, 0.1 ulp |
539
- | `nk_euclideans_packed_f32_serial` | 8.71 gso/s, 0.3 ulp | 25.7 gso/s, 0.3 ulp | 26.3 gso/s, 0.3 ulp |
540
- | `nk_euclideans_symmetric_f32_serial` | 6.15 gso/s, 0.3 ulp | 17.7 gso/s, 0.3 ulp | 34.9 gso/s, 0.3 ulp |
541
- | `nk_angulars_packed_f32_v128relaxed` | 56.7 gso/s, 0.1 ulp | 66.9 gso/s, 0.1 ulp | 65.8 gso/s, 0.1 ulp |
542
- | `nk_angulars_symmetric_f32_v128relaxed` | 18.5 gso/s, 0.1 ulp | 20.9 gso/s, 0.1 ulp | 39.0 gso/s, 0.1 ulp |
543
- | `nk_euclideans_packed_f32_v128relaxed` | 57.2 gso/s, 0.2 ulp | 67.1 gso/s, 0.2 ulp | 65.8 gso/s, 0.2 ulp |
544
- | `nk_euclideans_symmetric_f32_v128relaxed` | 18.5 gso/s, 0.2 ulp | 20.9 gso/s, 0.2 ulp | 39.0 gso/s, 0.2 ulp |
537
+ | `nk_angulars_packed_f32_serial` | 10.3 gso/s, 0.1 ulp | 10.1 gso/s, 0.1 ulp | 10.6 gso/s, 0.1 ulp |
538
+ | `nk_angulars_symmetric_f32_serial` | 8.18 gso/s, 0.1 ulp | 8.59 gso/s, 0.1 ulp | 9.54 gso/s, 0.1 ulp |
539
+ | `nk_euclideans_packed_f32_serial` | 10.4 gso/s, 0.3 ulp | 10.4 gso/s, 0.3 ulp | 11.0 gso/s, 0.3 ulp |
540
+ | `nk_euclideans_symmetric_f32_serial` | 8.58 gso/s, 0.3 ulp | 8.67 gso/s, 0.3 ulp | 9.52 gso/s, 0.3 ulp |
541
+ | `nk_angulars_packed_f32_v128relaxed` | 25.2 gso/s, 0.1 ulp | 30.7 gso/s, 0.1 ulp | 32.3 gso/s, 0.1 ulp |
542
+ | `nk_angulars_symmetric_f32_v128relaxed` | 9.91 gso/s, 0.1 ulp | 10.9 gso/s, 0.1 ulp | 11.1 gso/s, 0.1 ulp |
543
+ | `nk_euclideans_packed_f32_v128relaxed` | 26.6 gso/s, 0.2 ulp | 30.7 gso/s, 0.2 ulp | 32.2 gso/s, 0.2 ulp |
544
+ | `nk_euclideans_symmetric_f32_v128relaxed` | 9.98 gso/s, 0.2 ulp | 10.9 gso/s, 0.2 ulp | 11.1 gso/s, 0.2 ulp |
545
545
  | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
546
- | `nk_angulars_packed_bf16_serial` | 6.43 gso/s, 0.3 ulp | 18.6 gso/s, 0.3 ulp | 20.5 gso/s, 0.3 ulp |
547
- | `nk_angulars_symmetric_bf16_serial` | 7.55 gso/s, 0.3 ulp | 25.4 gso/s, 0.3 ulp | 52.0 gso/s, 0.3 ulp |
548
- | `nk_euclideans_packed_bf16_serial` | 6.43 gso/s, 5.3 ulp | 18.6 gso/s, 5.3 ulp | 20.6 gso/s, 5.3 ulp |
549
- | `nk_euclideans_symmetric_bf16_serial` | 7.27 gso/s, 5.3 ulp | 18.4 gso/s, 5.3 ulp | 52.0 gso/s, 5.3 ulp |
550
- | `nk_angulars_packed_bf16_v128relaxed` | 52.9 gso/s, 0.3 ulp | 57.2 gso/s, 0.3 ulp | 57.0 gso/s, 0.3 ulp |
551
- | `nk_angulars_symmetric_bf16_v128relaxed` | 16.5 gso/s, 0.3 ulp | 17.8 gso/s, 0.3 ulp | 34.8 gso/s, 0.3 ulp |
552
- | `nk_euclideans_packed_bf16_v128relaxed` | 53.3 gso/s, 5.3 ulp | 57.3 gso/s, 5.3 ulp | 56.7 gso/s, 5.3 ulp |
553
- | `nk_euclideans_symmetric_bf16_v128relaxed` | 16.6 gso/s, 5.3 ulp | 17.8 gso/s, 5.3 ulp | 34.7 gso/s, 5.3 ulp |
546
+ | `nk_angulars_packed_bf16_serial` | 21.7 gso/s, 0.3 ulp | 21.3 gso/s, 0.3 ulp | 24.3 gso/s, 0.3 ulp |
547
+ | `nk_angulars_symmetric_bf16_serial` | 22.2 gso/s, 0.3 ulp | 24.4 gso/s, 0.3 ulp | 27.8 gso/s, 0.3 ulp |
548
+ | `nk_euclideans_packed_bf16_serial` | 19.1 gso/s, 5.3 ulp | 21.4 gso/s, 5.3 ulp | 24.2 gso/s, 5.3 ulp |
549
+ | `nk_euclideans_symmetric_bf16_serial` | 22.0 gso/s, 5.3 ulp | 24.7 gso/s, 5.3 ulp | 27.7 gso/s, 5.3 ulp |
550
+ | `nk_angulars_packed_bf16_v128relaxed` | 70.2 gso/s, 0.3 ulp | 82.3 gso/s, 0.3 ulp | 89.9 gso/s, 0.3 ulp |
551
+ | `nk_angulars_symmetric_bf16_v128relaxed` | 36.9 gso/s, 0.3 ulp | 44.9 gso/s, 0.3 ulp | 47.3 gso/s, 0.3 ulp |
552
+ | `nk_euclideans_packed_bf16_v128relaxed` | 76.4 gso/s, 5.3 ulp | 87.3 gso/s, 5.3 ulp | 89.7 gso/s, 5.3 ulp |
553
+ | `nk_euclideans_symmetric_bf16_v128relaxed` | 37.1 gso/s, 5.3 ulp | 43.1 gso/s, 5.3 ulp | 45.5 gso/s, 5.3 ulp |
554
554
  | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
555
- | `nk_angulars_packed_e2m3_serial` | 2.04 gso/s, 0 ulp | 5.36 gso/s, 0 ulp | 5.62 gso/s, 0 ulp |
556
- | `nk_angulars_symmetric_e2m3_serial` | 3.00 gso/s, 0 ulp | 7.21 gso/s, 0 ulp | 15.2 gso/s, 0 ulp |
557
- | `nk_euclideans_packed_e2m3_serial` | 3.58 gso/s, 0.3 ulp | 5.40 gso/s, 0.3 ulp | 5.64 gso/s, 0.3 ulp |
558
- | `nk_euclideans_symmetric_e2m3_serial` | 6.64 gso/s, 0.3 ulp | 7.44 gso/s, 0.3 ulp | 15.3 gso/s, 0.3 ulp |
559
- | `nk_angulars_packed_e2m3_v128relaxed` | 34.5 gso/s, 0 ulp | 37.8 gso/s, 0 ulp | 35.7 gso/s, 0 ulp |
560
- | `nk_angulars_symmetric_e2m3_v128relaxed` | 31.3 gso/s, 0 ulp | 36.8 gso/s, 0 ulp | 72.5 gso/s, 0 ulp |
561
- | `nk_euclideans_packed_e2m3_v128relaxed` | 34.3 gso/s, 0 ulp | 37.8 gso/s, 0 ulp | 35.9 gso/s, 0 ulp |
562
- | `nk_euclideans_symmetric_e2m3_v128relaxed` | 31.7 gso/s, 0 ulp | 36.8 gso/s, 0 ulp | 72.5 gso/s, 0 ulp |
555
+ | `nk_angulars_packed_e2m3_serial` | 5.78 gso/s, 0 ulp | 5.93 gso/s, 0 ulp | 6.28 gso/s, 0 ulp |
556
+ | `nk_angulars_symmetric_e2m3_serial` | 6.52 gso/s, 0 ulp | 8.09 gso/s, 0 ulp | 8.52 gso/s, 0 ulp |
557
+ | `nk_euclideans_packed_e2m3_serial` | 5.25 gso/s, 0.3 ulp | 5.94 gso/s, 0.3 ulp | 6.20 gso/s, 0.3 ulp |
558
+ | `nk_euclideans_symmetric_e2m3_serial` | 6.99 gso/s, 0.3 ulp | 8.08 gso/s, 0.3 ulp | 8.52 gso/s, 0.3 ulp |
559
+ | `nk_angulars_packed_e2m3_v128relaxed` | 36.7 gso/s, 0 ulp | 38.8 gso/s, 0 ulp | 39.9 gso/s, 0 ulp |
560
+ | `nk_angulars_symmetric_e2m3_v128relaxed` | 31.4 gso/s, 0 ulp | 37.3 gso/s, 0 ulp | 39.5 gso/s, 0 ulp |
561
+ | `nk_euclideans_packed_e2m3_v128relaxed` | 36.8 gso/s, 0 ulp | 38.9 gso/s, 0 ulp | 40.0 gso/s, 0 ulp |
562
+ | `nk_euclideans_symmetric_e2m3_v128relaxed` | 31.8 gso/s, 0 ulp | 37.5 gso/s, 0 ulp | 39.5 gso/s, 0 ulp |
563
563
  | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
564
- | `nk_angulars_packed_i8_serial` | 5.12 gso/s, 0 ulp | 13.0 gso/s, 0 ulp | 13.7 gso/s, 0 ulp |
565
- | `nk_angulars_symmetric_i8_serial` | 6.54 gso/s, 0 ulp | 17.0 gso/s, 0 ulp | 36.5 gso/s, 0 ulp |
566
- | `nk_euclideans_packed_i8_serial` | 12.2 gso/s, 0.5 ulp | 13.0 gso/s, 0.5 ulp | 13.7 gso/s, 0.5 ulp |
567
- | `nk_euclideans_symmetric_i8_serial` | 13.7 gso/s, 0.5 ulp | 17.4 gso/s, 0.5 ulp | 36.5 gso/s, 0.5 ulp |
568
- | `nk_angulars_packed_i8_v128relaxed` | 45.5 gso/s, 0 ulp | 51.1 gso/s, 0 ulp | 50.6 gso/s, 0 ulp |
569
- | `nk_angulars_symmetric_i8_v128relaxed` | 41.3 gso/s, 0 ulp | 49.9 gso/s, 0 ulp | 102 gso/s, 0 ulp |
570
- | `nk_euclideans_packed_i8_v128relaxed` | 46.0 gso/s, 0 ulp | 51.3 gso/s, 0 ulp | 50.7 gso/s, 0 ulp |
571
- | `nk_euclideans_symmetric_i8_v128relaxed` | 41.5 gso/s, 0 ulp | 50.0 gso/s, 0 ulp | 102 gso/s, 0 ulp |
564
+ | `nk_angulars_packed_i8_serial` | 13.7 gso/s, 0 ulp | 16.4 gso/s, 0 ulp | 17.3 gso/s, 0 ulp |
565
+ | `nk_angulars_symmetric_i8_serial` | 10.5 gso/s, 0 ulp | 12.8 gso/s, 0 ulp | 13.4 gso/s, 0 ulp |
566
+ | `nk_euclideans_packed_i8_serial` | 14.4 gso/s, 0.5 ulp | 16.5 gso/s, 0.5 ulp | 17.3 gso/s, 0.5 ulp |
567
+ | `nk_euclideans_symmetric_i8_serial` | 11.3 gso/s, 0.5 ulp | 12.6 gso/s, 0.5 ulp | 13.4 gso/s, 0.5 ulp |
568
+ | `nk_angulars_packed_i8_v128relaxed` | 45.2 gso/s, 0 ulp | 50.0 gso/s, 0 ulp | 52.0 gso/s, 0 ulp |
569
+ | `nk_angulars_symmetric_i8_v128relaxed` | 37.7 gso/s, 0 ulp | 47.5 gso/s, 0 ulp | 50.4 gso/s, 0 ulp |
570
+ | `nk_euclideans_packed_i8_v128relaxed` | 45.6 gso/s, 0 ulp | 50.2 gso/s, 0 ulp | 52.0 gso/s, 0 ulp |
571
+ | `nk_euclideans_symmetric_i8_v128relaxed` | 37.4 gso/s, 0 ulp | 46.8 gso/s, 0 ulp | 50.4 gso/s, 0 ulp |
572
572
  | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
573
- | `nk_angulars_packed_u8_serial` | 5.60 gso/s, 0.3 ulp | 13.4 gso/s, 0.3 ulp | 14.1 gso/s, 0.3 ulp |
574
- | `nk_angulars_symmetric_u8_serial` | 7.91 gso/s, 0.3 ulp | 17.1 gso/s, 0.3 ulp | 36.9 gso/s, 0.3 ulp |
575
- | `nk_euclideans_packed_u8_serial` | 12.5 gso/s, 0.4 ulp | 13.4 gso/s, 0.4 ulp | 14.1 gso/s, 0.4 ulp |
576
- | `nk_euclideans_symmetric_u8_serial` | 13.7 gso/s, 0.4 ulp | 17.6 gso/s, 0.4 ulp | 36.9 gso/s, 0.4 ulp |
577
- | `nk_angulars_packed_u8_v128relaxed` | 74.7 gso/s, 0.3 ulp | 86.1 gso/s, 0.3 ulp | 83.3 gso/s, 0.3 ulp |
578
- | `nk_angulars_symmetric_u8_v128relaxed` | 54.1 gso/s, 0.3 ulp | 76.1 gso/s, 0.3 ulp | 160 gso/s, 0.3 ulp |
579
- | `nk_euclideans_packed_u8_v128relaxed` | 75.6 gso/s, 0 ulp | 86.4 gso/s, 0 ulp | 84.2 gso/s, 0 ulp |
580
- | `nk_euclideans_symmetric_u8_v128relaxed` | 54.5 gso/s, 0 ulp | 76.2 gso/s, 0 ulp | 161 gso/s, 0 ulp |
573
+ | `nk_angulars_packed_u8_serial` | 14.7 gso/s, 0.3 ulp | 17.0 gso/s, 0.3 ulp | 17.8 gso/s, 0.3 ulp |
574
+ | `nk_angulars_symmetric_u8_serial` | 10.9 gso/s, 0.3 ulp | 13.2 gso/s, 0.3 ulp | 13.9 gso/s, 0.3 ulp |
575
+ | `nk_euclideans_packed_u8_serial` | 14.9 gso/s, 0.4 ulp | 17.0 gso/s, 0.4 ulp | 17.8 gso/s, 0.4 ulp |
576
+ | `nk_euclideans_symmetric_u8_serial` | 11.7 gso/s, 0.4 ulp | 13.1 gso/s, 0.4 ulp | 13.9 gso/s, 0.4 ulp |
577
+ | `nk_angulars_packed_u8_v128relaxed` | 43.7 gso/s, 0.3 ulp | 49.0 gso/s, 0.3 ulp | 50.7 gso/s, 0.3 ulp |
578
+ | `nk_angulars_symmetric_u8_v128relaxed` | 34.7 gso/s, 0.3 ulp | 45.6 gso/s, 0.3 ulp | 48.4 gso/s, 0.3 ulp |
579
+ | `nk_euclideans_packed_u8_v128relaxed` | 44.8 gso/s, 0 ulp | 49.0 gso/s, 0 ulp | 50.7 gso/s, 0 ulp |
580
+ | `nk_euclideans_symmetric_u8_v128relaxed` | 35.0 gso/s, 0 ulp | 44.2 gso/s, 0 ulp | 48.5 gso/s, 0 ulp |