usearch 2.21.4 → 2.25.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (602) hide show
  1. package/README.md +24 -24
  2. package/binding.gyp +15 -37
  3. package/c/CMakeLists.txt +4 -5
  4. package/c/README.md +7 -7
  5. package/c/lib.cpp +32 -14
  6. package/c/test.c +83 -1
  7. package/c/usearch.h +24 -3
  8. package/cpp/CMakeLists.txt +4 -0
  9. package/cpp/README.md +5 -5
  10. package/cpp/bench.cpp +97 -116
  11. package/cpp/test.cpp +117 -20
  12. package/include/usearch/index.hpp +224 -48
  13. package/include/usearch/index_dense.hpp +105 -40
  14. package/include/usearch/index_plugins.hpp +1387 -204
  15. package/javascript/dist/cjs/usearch.d.ts +8 -3
  16. package/javascript/dist/cjs/usearch.js +12 -4
  17. package/javascript/dist/esm/usearch.d.ts +8 -3
  18. package/javascript/dist/esm/usearch.js +12 -4
  19. package/javascript/lib.cpp +17 -0
  20. package/javascript/usearch.ts +14 -6
  21. package/numkong/.clang-tidy +30 -0
  22. package/{simsimd → numkong}/.cmake-format.py +18 -0
  23. package/numkong/.editorconfig +12 -0
  24. package/numkong/.git-blame-ignore-revs +123 -0
  25. package/numkong/.mailmap +15 -0
  26. package/numkong/.spi.yml +11 -0
  27. package/numkong/.swift-format +14 -0
  28. package/numkong/CITATION.cff +81 -0
  29. package/numkong/CONTRIBUTING.md +500 -0
  30. package/numkong/Cargo.toml +54 -0
  31. package/numkong/MANIFEST.in +8 -0
  32. package/numkong/Package.swift +84 -0
  33. package/numkong/README.md +569 -0
  34. package/numkong/bench/README.md +217 -0
  35. package/numkong/bench/bench-browser.html +323 -0
  36. package/numkong/bench/bench.cpp +344 -0
  37. package/numkong/bench/bench.hpp +816 -0
  38. package/numkong/bench/bench.js +148 -0
  39. package/numkong/bench/bench.mjs +616 -0
  40. package/numkong/bench/bench_attention.cpp +113 -0
  41. package/numkong/bench/bench_cast.cpp +264 -0
  42. package/numkong/bench/bench_cross_amx.cpp +62 -0
  43. package/numkong/bench/bench_cross_arm.cpp +245 -0
  44. package/numkong/bench/bench_cross_blas.cpp +309 -0
  45. package/numkong/bench/bench_cross_loongarch.cpp +106 -0
  46. package/numkong/bench/bench_cross_power.cpp +84 -0
  47. package/numkong/bench/bench_cross_rvv.cpp +121 -0
  48. package/numkong/bench/bench_cross_serial.cpp +151 -0
  49. package/numkong/bench/bench_cross_sme.cpp +163 -0
  50. package/numkong/bench/bench_cross_wasm.cpp +142 -0
  51. package/numkong/bench/bench_cross_x86.cpp +455 -0
  52. package/numkong/bench/bench_curved.cpp +210 -0
  53. package/numkong/bench/bench_dot.cpp +315 -0
  54. package/numkong/bench/bench_each.cpp +503 -0
  55. package/numkong/bench/bench_geospatial.cpp +112 -0
  56. package/numkong/bench/bench_maxsim.cpp +154 -0
  57. package/numkong/bench/bench_mesh.cpp +148 -0
  58. package/numkong/bench/bench_probability.cpp +60 -0
  59. package/numkong/bench/bench_reduce.cpp +371 -0
  60. package/numkong/bench/bench_set.cpp +73 -0
  61. package/numkong/bench/bench_sparse.cpp +198 -0
  62. package/numkong/bench/bench_spatial.cpp +347 -0
  63. package/numkong/bench/bench_trigonometry.cpp +137 -0
  64. package/numkong/bench/swift/Bench.swift +506 -0
  65. package/numkong/binding.gyp +169 -0
  66. package/numkong/build.rs +484 -0
  67. package/numkong/c/dispatch.h +512 -0
  68. package/numkong/c/dispatch_bf16.c +447 -0
  69. package/numkong/c/dispatch_bf16c.c +52 -0
  70. package/numkong/c/dispatch_e2m3.c +296 -0
  71. package/numkong/c/dispatch_e3m2.c +284 -0
  72. package/numkong/c/dispatch_e4m3.c +318 -0
  73. package/numkong/c/dispatch_e5m2.c +316 -0
  74. package/numkong/c/dispatch_f16.c +452 -0
  75. package/numkong/c/dispatch_f16c.c +58 -0
  76. package/numkong/c/dispatch_f32.c +434 -0
  77. package/numkong/c/dispatch_f32c.c +99 -0
  78. package/numkong/c/dispatch_f64.c +348 -0
  79. package/numkong/c/dispatch_f64c.c +98 -0
  80. package/numkong/c/dispatch_i16.c +96 -0
  81. package/numkong/c/dispatch_i32.c +89 -0
  82. package/numkong/c/dispatch_i4.c +153 -0
  83. package/numkong/c/dispatch_i64.c +86 -0
  84. package/numkong/c/dispatch_i8.c +348 -0
  85. package/numkong/c/dispatch_other.c +348 -0
  86. package/numkong/c/dispatch_u1.c +193 -0
  87. package/numkong/c/dispatch_u16.c +124 -0
  88. package/numkong/c/dispatch_u32.c +118 -0
  89. package/numkong/c/dispatch_u4.c +153 -0
  90. package/numkong/c/dispatch_u64.c +102 -0
  91. package/numkong/c/dispatch_u8.c +364 -0
  92. package/numkong/c/numkong.c +940 -0
  93. package/numkong/cmake/nk_arm_isa_probes.cmake +61 -0
  94. package/numkong/cmake/nk_isa_probe.cmake +90 -0
  95. package/numkong/cmake/nk_loongarch_isa_probes.cmake +15 -0
  96. package/numkong/cmake/nk_math_lib.cmake +46 -0
  97. package/numkong/cmake/nk_power_isa_probes.cmake +15 -0
  98. package/numkong/cmake/nk_riscv_isa_probes.cmake +19 -0
  99. package/numkong/cmake/nk_x86_isa_probes.cmake +35 -0
  100. package/numkong/cmake/toolchain-aarch64-gnu.cmake +86 -0
  101. package/numkong/cmake/toolchain-android-arm64.cmake +25 -0
  102. package/numkong/cmake/toolchain-android-armv7.cmake +25 -0
  103. package/numkong/cmake/toolchain-loongarch64-gnu.cmake +68 -0
  104. package/numkong/cmake/toolchain-ppc64le-gnu.cmake +74 -0
  105. package/numkong/cmake/toolchain-riscv64-gnu.cmake +107 -0
  106. package/numkong/cmake/toolchain-riscv64-llvm.cmake +114 -0
  107. package/numkong/cmake/toolchain-wasi.cmake +77 -0
  108. package/numkong/cmake/toolchain-wasm.cmake +70 -0
  109. package/numkong/cmake/toolchain-wasm64.cmake +62 -0
  110. package/numkong/cmake/toolchain-x86_64-llvm.cmake +47 -0
  111. package/numkong/include/README.md +643 -0
  112. package/numkong/include/module.modulemap +129 -0
  113. package/numkong/include/numkong/attention/sapphireamx.h +1363 -0
  114. package/numkong/include/numkong/attention/sme.h +2075 -0
  115. package/numkong/include/numkong/attention.h +49 -0
  116. package/numkong/include/numkong/capabilities.h +930 -0
  117. package/numkong/include/numkong/cast/README.md +262 -0
  118. package/numkong/include/numkong/cast/diamond.h +64 -0
  119. package/numkong/include/numkong/cast/haswell.h +823 -0
  120. package/numkong/include/numkong/cast/icelake.h +475 -0
  121. package/numkong/include/numkong/cast/loongsonasx.h +252 -0
  122. package/numkong/include/numkong/cast/neon.h +1174 -0
  123. package/numkong/include/numkong/cast/powervsx.h +449 -0
  124. package/numkong/include/numkong/cast/rvv.h +970 -0
  125. package/numkong/include/numkong/cast/sapphire.h +262 -0
  126. package/numkong/include/numkong/cast/serial.h +2336 -0
  127. package/numkong/include/numkong/cast/skylake.h +915 -0
  128. package/numkong/include/numkong/cast/v128relaxed.h +537 -0
  129. package/numkong/include/numkong/cast.h +254 -0
  130. package/numkong/include/numkong/cast.hpp +44 -0
  131. package/numkong/include/numkong/curved/README.md +223 -0
  132. package/numkong/include/numkong/curved/genoa.h +182 -0
  133. package/numkong/include/numkong/curved/haswell.h +276 -0
  134. package/numkong/include/numkong/curved/neon.h +329 -0
  135. package/numkong/include/numkong/curved/neonbfdot.h +211 -0
  136. package/numkong/include/numkong/curved/rvv.h +305 -0
  137. package/numkong/include/numkong/curved/serial.h +207 -0
  138. package/numkong/include/numkong/curved/skylake.h +457 -0
  139. package/numkong/include/numkong/curved/smef64.h +530 -0
  140. package/numkong/include/numkong/curved.h +513 -0
  141. package/numkong/include/numkong/curved.hpp +144 -0
  142. package/numkong/include/numkong/dot/README.md +442 -0
  143. package/numkong/include/numkong/dot/alder.h +563 -0
  144. package/numkong/include/numkong/dot/diamond.h +158 -0
  145. package/numkong/include/numkong/dot/genoa.h +289 -0
  146. package/numkong/include/numkong/dot/haswell.h +1723 -0
  147. package/numkong/include/numkong/dot/icelake.h +997 -0
  148. package/numkong/include/numkong/dot/loongsonasx.h +671 -0
  149. package/numkong/include/numkong/dot/neon.h +869 -0
  150. package/numkong/include/numkong/dot/neonbfdot.h +243 -0
  151. package/numkong/include/numkong/dot/neonfhm.h +358 -0
  152. package/numkong/include/numkong/dot/neonfp8.h +323 -0
  153. package/numkong/include/numkong/dot/neonsdot.h +622 -0
  154. package/numkong/include/numkong/dot/powervsx.h +752 -0
  155. package/numkong/include/numkong/dot/rvv.h +722 -0
  156. package/numkong/include/numkong/dot/rvvbb.h +72 -0
  157. package/numkong/include/numkong/dot/rvvbf16.h +123 -0
  158. package/numkong/include/numkong/dot/rvvhalf.h +129 -0
  159. package/numkong/include/numkong/dot/sapphire.h +141 -0
  160. package/numkong/include/numkong/dot/serial.h +874 -0
  161. package/numkong/include/numkong/dot/sierra.h +405 -0
  162. package/numkong/include/numkong/dot/skylake.h +1116 -0
  163. package/numkong/include/numkong/dot/sve.h +420 -0
  164. package/numkong/include/numkong/dot/svebfdot.h +75 -0
  165. package/numkong/include/numkong/dot/svehalf.h +168 -0
  166. package/numkong/include/numkong/dot/svesdot.h +90 -0
  167. package/numkong/include/numkong/dot/v128relaxed.h +1293 -0
  168. package/numkong/include/numkong/dot.h +1136 -0
  169. package/numkong/include/numkong/dot.hpp +94 -0
  170. package/numkong/include/numkong/dots/README.md +496 -0
  171. package/numkong/include/numkong/dots/alder.h +117 -0
  172. package/numkong/include/numkong/dots/diamond.h +86 -0
  173. package/numkong/include/numkong/dots/genoa.h +100 -0
  174. package/numkong/include/numkong/dots/graniteamx.h +733 -0
  175. package/numkong/include/numkong/dots/haswell.h +310 -0
  176. package/numkong/include/numkong/dots/icelake.h +180 -0
  177. package/numkong/include/numkong/dots/loongsonasx.h +176 -0
  178. package/numkong/include/numkong/dots/neon.h +123 -0
  179. package/numkong/include/numkong/dots/neonbfdot.h +59 -0
  180. package/numkong/include/numkong/dots/neonfhm.h +96 -0
  181. package/numkong/include/numkong/dots/neonfp8.h +99 -0
  182. package/numkong/include/numkong/dots/neonsdot.h +144 -0
  183. package/numkong/include/numkong/dots/powervsx.h +194 -0
  184. package/numkong/include/numkong/dots/rvv.h +2593 -0
  185. package/numkong/include/numkong/dots/sapphireamx.h +4017 -0
  186. package/numkong/include/numkong/dots/serial.h +2867 -0
  187. package/numkong/include/numkong/dots/sierra.h +100 -0
  188. package/numkong/include/numkong/dots/skylake.h +207 -0
  189. package/numkong/include/numkong/dots/sme.h +5041 -0
  190. package/numkong/include/numkong/dots/smebi32.h +477 -0
  191. package/numkong/include/numkong/dots/smef64.h +1328 -0
  192. package/numkong/include/numkong/dots/smehalf.h +47 -0
  193. package/numkong/include/numkong/dots/v128relaxed.h +317 -0
  194. package/numkong/include/numkong/dots.h +3110 -0
  195. package/numkong/include/numkong/dots.hpp +641 -0
  196. package/numkong/include/numkong/each/README.md +465 -0
  197. package/numkong/include/numkong/each/haswell.h +1658 -0
  198. package/numkong/include/numkong/each/icelake.h +272 -0
  199. package/numkong/include/numkong/each/neon.h +1138 -0
  200. package/numkong/include/numkong/each/neonbfdot.h +211 -0
  201. package/numkong/include/numkong/each/neonhalf.h +318 -0
  202. package/numkong/include/numkong/each/rvv.h +1121 -0
  203. package/numkong/include/numkong/each/sapphire.h +343 -0
  204. package/numkong/include/numkong/each/serial.h +282 -0
  205. package/numkong/include/numkong/each/skylake.h +1562 -0
  206. package/numkong/include/numkong/each/v128relaxed.h +562 -0
  207. package/numkong/include/numkong/each.h +2232 -0
  208. package/numkong/include/numkong/each.hpp +434 -0
  209. package/numkong/include/numkong/geospatial/README.md +147 -0
  210. package/numkong/include/numkong/geospatial/haswell.h +633 -0
  211. package/numkong/include/numkong/geospatial/neon.h +615 -0
  212. package/numkong/include/numkong/geospatial/rvv.h +701 -0
  213. package/numkong/include/numkong/geospatial/serial.h +309 -0
  214. package/numkong/include/numkong/geospatial/skylake.h +613 -0
  215. package/numkong/include/numkong/geospatial/v128relaxed.h +652 -0
  216. package/numkong/include/numkong/geospatial.h +453 -0
  217. package/numkong/include/numkong/geospatial.hpp +235 -0
  218. package/numkong/include/numkong/matrix.hpp +336 -0
  219. package/numkong/include/numkong/maxsim/README.md +187 -0
  220. package/numkong/include/numkong/maxsim/alder.h +511 -0
  221. package/numkong/include/numkong/maxsim/genoa.h +115 -0
  222. package/numkong/include/numkong/maxsim/haswell.h +553 -0
  223. package/numkong/include/numkong/maxsim/icelake.h +480 -0
  224. package/numkong/include/numkong/maxsim/neonsdot.h +394 -0
  225. package/numkong/include/numkong/maxsim/sapphireamx.h +877 -0
  226. package/numkong/include/numkong/maxsim/serial.h +490 -0
  227. package/numkong/include/numkong/maxsim/sme.h +993 -0
  228. package/numkong/include/numkong/maxsim/v128relaxed.h +280 -0
  229. package/numkong/include/numkong/maxsim.h +571 -0
  230. package/numkong/include/numkong/maxsim.hpp +133 -0
  231. package/numkong/include/numkong/mesh/README.md +219 -0
  232. package/numkong/include/numkong/mesh/haswell.h +2013 -0
  233. package/numkong/include/numkong/mesh/neon.h +1821 -0
  234. package/numkong/include/numkong/mesh/neonbfdot.h +771 -0
  235. package/numkong/include/numkong/mesh/rvv.h +936 -0
  236. package/numkong/include/numkong/mesh/serial.h +721 -0
  237. package/numkong/include/numkong/mesh/skylake.h +2344 -0
  238. package/numkong/include/numkong/mesh/v128relaxed.h +1081 -0
  239. package/numkong/include/numkong/mesh.h +666 -0
  240. package/numkong/include/numkong/mesh.hpp +701 -0
  241. package/numkong/include/numkong/numkong.h +78 -0
  242. package/numkong/include/numkong/numkong.hpp +58 -0
  243. package/numkong/include/numkong/probability/README.md +177 -0
  244. package/numkong/include/numkong/probability/haswell.h +267 -0
  245. package/numkong/include/numkong/probability/neon.h +255 -0
  246. package/numkong/include/numkong/probability/rvv.h +414 -0
  247. package/numkong/include/numkong/probability/serial.h +181 -0
  248. package/numkong/include/numkong/probability/skylake.h +324 -0
  249. package/numkong/include/numkong/probability.h +380 -0
  250. package/numkong/include/numkong/probability.hpp +120 -0
  251. package/numkong/include/numkong/random.h +50 -0
  252. package/numkong/include/numkong/random.hpp +285 -0
  253. package/numkong/include/numkong/reduce/README.md +552 -0
  254. package/numkong/include/numkong/reduce/alder.h +636 -0
  255. package/numkong/include/numkong/reduce/genoa.h +201 -0
  256. package/numkong/include/numkong/reduce/haswell.h +3785 -0
  257. package/numkong/include/numkong/reduce/icelake.h +549 -0
  258. package/numkong/include/numkong/reduce/neon.h +3980 -0
  259. package/numkong/include/numkong/reduce/neonbfdot.h +142 -0
  260. package/numkong/include/numkong/reduce/neonfhm.h +293 -0
  261. package/numkong/include/numkong/reduce/neonsdot.h +357 -0
  262. package/numkong/include/numkong/reduce/rvv.h +3533 -0
  263. package/numkong/include/numkong/reduce/serial.h +753 -0
  264. package/numkong/include/numkong/reduce/sierra.h +338 -0
  265. package/numkong/include/numkong/reduce/skylake.h +3794 -0
  266. package/numkong/include/numkong/reduce/sve.h +52 -0
  267. package/numkong/include/numkong/reduce/v128relaxed.h +2302 -0
  268. package/numkong/include/numkong/reduce.h +1582 -0
  269. package/numkong/include/numkong/reduce.hpp +778 -0
  270. package/numkong/include/numkong/scalar/README.md +89 -0
  271. package/numkong/include/numkong/scalar/haswell.h +122 -0
  272. package/numkong/include/numkong/scalar/loongsonasx.h +74 -0
  273. package/numkong/include/numkong/scalar/neon.h +122 -0
  274. package/numkong/include/numkong/scalar/neonhalf.h +70 -0
  275. package/numkong/include/numkong/scalar/powervsx.h +96 -0
  276. package/numkong/include/numkong/scalar/rvv.h +211 -0
  277. package/numkong/include/numkong/scalar/sapphire.h +74 -0
  278. package/numkong/include/numkong/scalar/serial.h +332 -0
  279. package/numkong/include/numkong/scalar/v128relaxed.h +56 -0
  280. package/numkong/include/numkong/scalar.h +696 -0
  281. package/numkong/include/numkong/set/README.md +179 -0
  282. package/numkong/include/numkong/set/haswell.h +334 -0
  283. package/numkong/include/numkong/set/icelake.h +485 -0
  284. package/numkong/include/numkong/set/loongsonasx.h +181 -0
  285. package/numkong/include/numkong/set/neon.h +363 -0
  286. package/numkong/include/numkong/set/powervsx.h +326 -0
  287. package/numkong/include/numkong/set/rvv.h +226 -0
  288. package/numkong/include/numkong/set/rvvbb.h +117 -0
  289. package/numkong/include/numkong/set/serial.h +174 -0
  290. package/numkong/include/numkong/set/sve.h +187 -0
  291. package/numkong/include/numkong/set/v128relaxed.h +240 -0
  292. package/numkong/include/numkong/set.h +471 -0
  293. package/numkong/include/numkong/set.hpp +114 -0
  294. package/numkong/include/numkong/sets/README.md +149 -0
  295. package/numkong/include/numkong/sets/haswell.h +63 -0
  296. package/numkong/include/numkong/sets/icelake.h +66 -0
  297. package/numkong/include/numkong/sets/loongsonasx.h +52 -0
  298. package/numkong/include/numkong/sets/neon.h +61 -0
  299. package/numkong/include/numkong/sets/powervsx.h +65 -0
  300. package/numkong/include/numkong/sets/serial.h +43 -0
  301. package/numkong/include/numkong/sets/smebi32.h +1135 -0
  302. package/numkong/include/numkong/sets/v128relaxed.h +58 -0
  303. package/numkong/include/numkong/sets.h +382 -0
  304. package/numkong/include/numkong/sparse/README.md +156 -0
  305. package/numkong/include/numkong/sparse/icelake.h +463 -0
  306. package/numkong/include/numkong/sparse/neon.h +288 -0
  307. package/numkong/include/numkong/sparse/serial.h +117 -0
  308. package/numkong/include/numkong/sparse/sve2.h +504 -0
  309. package/numkong/include/numkong/sparse/turin.h +322 -0
  310. package/numkong/include/numkong/sparse.h +363 -0
  311. package/numkong/include/numkong/sparse.hpp +113 -0
  312. package/numkong/include/numkong/spatial/README.md +439 -0
  313. package/numkong/include/numkong/spatial/alder.h +604 -0
  314. package/numkong/include/numkong/spatial/diamond.h +240 -0
  315. package/numkong/include/numkong/spatial/genoa.h +222 -0
  316. package/numkong/include/numkong/spatial/haswell.h +979 -0
  317. package/numkong/include/numkong/spatial/icelake.h +1067 -0
  318. package/numkong/include/numkong/spatial/loongsonasx.h +483 -0
  319. package/numkong/include/numkong/spatial/neon.h +846 -0
  320. package/numkong/include/numkong/spatial/neonbfdot.h +164 -0
  321. package/numkong/include/numkong/spatial/neonfp8.h +258 -0
  322. package/numkong/include/numkong/spatial/neonsdot.h +429 -0
  323. package/numkong/include/numkong/spatial/powervsx.h +738 -0
  324. package/numkong/include/numkong/spatial/rvv.h +991 -0
  325. package/numkong/include/numkong/spatial/rvvbf16.h +128 -0
  326. package/numkong/include/numkong/spatial/rvvhalf.h +120 -0
  327. package/numkong/include/numkong/spatial/serial.h +347 -0
  328. package/numkong/include/numkong/spatial/sierra.h +516 -0
  329. package/numkong/include/numkong/spatial/skylake.h +605 -0
  330. package/numkong/include/numkong/spatial/sve.h +246 -0
  331. package/numkong/include/numkong/spatial/svebfdot.h +125 -0
  332. package/numkong/include/numkong/spatial/svehalf.h +136 -0
  333. package/numkong/include/numkong/spatial/svesdot.h +143 -0
  334. package/numkong/include/numkong/spatial/v128relaxed.h +969 -0
  335. package/numkong/include/numkong/spatial.h +1681 -0
  336. package/numkong/include/numkong/spatial.hpp +183 -0
  337. package/numkong/include/numkong/spatials/README.md +580 -0
  338. package/numkong/include/numkong/spatials/alder.h +94 -0
  339. package/numkong/include/numkong/spatials/diamond.h +82 -0
  340. package/numkong/include/numkong/spatials/genoa.h +94 -0
  341. package/numkong/include/numkong/spatials/graniteamx.h +173 -0
  342. package/numkong/include/numkong/spatials/haswell.h +219 -0
  343. package/numkong/include/numkong/spatials/icelake.h +113 -0
  344. package/numkong/include/numkong/spatials/loongsonasx.h +153 -0
  345. package/numkong/include/numkong/spatials/neon.h +109 -0
  346. package/numkong/include/numkong/spatials/neonbfdot.h +60 -0
  347. package/numkong/include/numkong/spatials/neonfhm.h +92 -0
  348. package/numkong/include/numkong/spatials/neonfp8.h +111 -0
  349. package/numkong/include/numkong/spatials/neonsdot.h +143 -0
  350. package/numkong/include/numkong/spatials/powervsx.h +153 -0
  351. package/numkong/include/numkong/spatials/rvv.h +1976 -0
  352. package/numkong/include/numkong/spatials/sapphireamx.h +1149 -0
  353. package/numkong/include/numkong/spatials/serial.h +248 -0
  354. package/numkong/include/numkong/spatials/sierra.h +96 -0
  355. package/numkong/include/numkong/spatials/skylake.h +184 -0
  356. package/numkong/include/numkong/spatials/sme.h +1927 -0
  357. package/numkong/include/numkong/spatials/smef64.h +483 -0
  358. package/numkong/include/numkong/spatials/v128relaxed.h +240 -0
  359. package/numkong/include/numkong/spatials.h +3374 -0
  360. package/numkong/include/numkong/spatials.hpp +509 -0
  361. package/numkong/include/numkong/tensor.hpp +1981 -0
  362. package/numkong/include/numkong/trigonometry/README.md +184 -0
  363. package/numkong/include/numkong/trigonometry/haswell.h +653 -0
  364. package/numkong/include/numkong/trigonometry/neon.h +638 -0
  365. package/numkong/include/numkong/trigonometry/rvv.h +700 -0
  366. package/numkong/include/numkong/trigonometry/serial.h +703 -0
  367. package/numkong/include/numkong/trigonometry/skylake.h +725 -0
  368. package/numkong/include/numkong/trigonometry/v128relaxed.h +667 -0
  369. package/numkong/include/numkong/trigonometry.h +467 -0
  370. package/numkong/include/numkong/trigonometry.hpp +166 -0
  371. package/numkong/include/numkong/types.h +1695 -0
  372. package/numkong/include/numkong/types.hpp +6027 -0
  373. package/numkong/include/numkong/vector.hpp +766 -0
  374. package/numkong/javascript/@numkong-darwin-arm64/package.json +26 -0
  375. package/numkong/javascript/@numkong-darwin-x64/package.json +26 -0
  376. package/numkong/javascript/@numkong-linux-arm64/package.json +25 -0
  377. package/numkong/javascript/@numkong-linux-x64/package.json +25 -0
  378. package/numkong/javascript/@numkong-win32-arm64/package.json +25 -0
  379. package/numkong/javascript/@numkong-win32-x64/package.json +25 -0
  380. package/numkong/javascript/README.md +254 -0
  381. package/numkong/javascript/node-gyp-build.d.ts +4 -0
  382. package/numkong/javascript/numkong-browser.ts +40 -0
  383. package/numkong/javascript/numkong-wasm.ts +750 -0
  384. package/numkong/javascript/numkong.c +712 -0
  385. package/numkong/javascript/numkong.ts +613 -0
  386. package/{simsimd → numkong}/javascript/tsconfig-base.json +6 -3
  387. package/numkong/javascript/tsconfig-cjs.json +13 -0
  388. package/{simsimd → numkong}/javascript/tsconfig-esm.json +1 -1
  389. package/numkong/javascript/types.ts +784 -0
  390. package/numkong/numkong.gypi +104 -0
  391. package/numkong/package.json +108 -0
  392. package/numkong/probes/arm_neon.c +8 -0
  393. package/numkong/probes/arm_neon_bfdot.c +9 -0
  394. package/numkong/probes/arm_neon_fhm.c +9 -0
  395. package/numkong/probes/arm_neon_half.c +8 -0
  396. package/numkong/probes/arm_neon_sdot.c +9 -0
  397. package/numkong/probes/arm_neonfp8.c +9 -0
  398. package/numkong/probes/arm_sme.c +16 -0
  399. package/numkong/probes/arm_sme2.c +16 -0
  400. package/numkong/probes/arm_sme2p1.c +16 -0
  401. package/numkong/probes/arm_sme_bf16.c +16 -0
  402. package/numkong/probes/arm_sme_bi32.c +16 -0
  403. package/numkong/probes/arm_sme_f64.c +16 -0
  404. package/numkong/probes/arm_sme_fa64.c +14 -0
  405. package/numkong/probes/arm_sme_half.c +16 -0
  406. package/numkong/probes/arm_sme_lut2.c +15 -0
  407. package/numkong/probes/arm_sve.c +18 -0
  408. package/numkong/probes/arm_sve2.c +20 -0
  409. package/numkong/probes/arm_sve2p1.c +18 -0
  410. package/numkong/probes/arm_sve_bfdot.c +20 -0
  411. package/numkong/probes/arm_sve_half.c +18 -0
  412. package/numkong/probes/arm_sve_sdot.c +21 -0
  413. package/numkong/probes/loongarch_lasx.c +12 -0
  414. package/numkong/probes/power_vsx.c +12 -0
  415. package/numkong/probes/probe.js +127 -0
  416. package/numkong/probes/riscv_rvv.c +14 -0
  417. package/numkong/probes/riscv_rvv_bb.c +15 -0
  418. package/numkong/probes/riscv_rvv_bf16.c +17 -0
  419. package/numkong/probes/riscv_rvv_half.c +14 -0
  420. package/numkong/probes/wasm_v128relaxed.c +11 -0
  421. package/numkong/probes/x86_alder.c +17 -0
  422. package/numkong/probes/x86_diamond.c +17 -0
  423. package/numkong/probes/x86_genoa.c +17 -0
  424. package/numkong/probes/x86_graniteamx.c +19 -0
  425. package/numkong/probes/x86_haswell.c +11 -0
  426. package/numkong/probes/x86_icelake.c +17 -0
  427. package/numkong/probes/x86_sapphire.c +16 -0
  428. package/numkong/probes/x86_sapphireamx.c +18 -0
  429. package/numkong/probes/x86_sierra.c +17 -0
  430. package/numkong/probes/x86_skylake.c +15 -0
  431. package/numkong/probes/x86_turin.c +17 -0
  432. package/numkong/rust/README.md +698 -0
  433. package/numkong/rust/capabilities.rs +96 -0
  434. package/numkong/rust/cast.rs +314 -0
  435. package/numkong/rust/curved.rs +391 -0
  436. package/numkong/rust/each.rs +2721 -0
  437. package/numkong/rust/geospatial.rs +343 -0
  438. package/numkong/rust/matrix.rs +5424 -0
  439. package/numkong/rust/maxsim.rs +540 -0
  440. package/numkong/rust/mesh.rs +837 -0
  441. package/numkong/rust/numkong.rs +409 -0
  442. package/numkong/rust/probability.rs +272 -0
  443. package/numkong/rust/reduce.rs +1249 -0
  444. package/numkong/rust/set.rs +194 -0
  445. package/numkong/rust/sparse.rs +513 -0
  446. package/numkong/rust/spatial.rs +1815 -0
  447. package/numkong/rust/tensor.rs +7565 -0
  448. package/numkong/rust/types.rs +3666 -0
  449. package/numkong/rust/vector.rs +1899 -0
  450. package/numkong/rustfmt.toml +2 -0
  451. package/numkong/swift/Matrix.swift +1388 -0
  452. package/numkong/swift/MaxSim.swift +151 -0
  453. package/numkong/swift/NumKong.swift +283 -0
  454. package/numkong/swift/Numerics.swift +329 -0
  455. package/numkong/swift/README.md +433 -0
  456. package/numkong/swift/Scalar.swift +741 -0
  457. package/numkong/swift/Tensor.swift +209 -0
  458. package/numkong/test/README.md +315 -0
  459. package/numkong/test/playwright.config.ts +16 -0
  460. package/numkong/test/swift/Test.swift +534 -0
  461. package/numkong/test/test-browser-bundle-runner.mjs +36 -0
  462. package/numkong/test/test-browser-bundle.html +188 -0
  463. package/numkong/test/test-browser-runner.mjs +43 -0
  464. package/numkong/test/test-browser.html +305 -0
  465. package/numkong/test/test-dtypes.mjs +453 -0
  466. package/numkong/test/test-wasm.mjs +267 -0
  467. package/numkong/test/test.cpp +442 -0
  468. package/numkong/test/test.hpp +580 -0
  469. package/numkong/test/test.mjs +363 -0
  470. package/numkong/test/test_base.py +1041 -0
  471. package/numkong/test/test_cast.cpp +242 -0
  472. package/numkong/test/test_cross.hpp +551 -0
  473. package/numkong/test/test_cross.py +467 -0
  474. package/numkong/test/test_cross_amx.cpp +50 -0
  475. package/numkong/test/test_cross_arm.cpp +225 -0
  476. package/numkong/test/test_cross_blas.cpp +354 -0
  477. package/numkong/test/test_cross_loongarch.cpp +90 -0
  478. package/numkong/test/test_cross_power.cpp +73 -0
  479. package/numkong/test/test_cross_rvv.cpp +109 -0
  480. package/numkong/test/test_cross_serial.cpp +181 -0
  481. package/numkong/test/test_cross_sme.cpp +148 -0
  482. package/numkong/test/test_cross_wasm.cpp +141 -0
  483. package/numkong/test/test_cross_x86.cpp +442 -0
  484. package/numkong/test/test_curved.cpp +182 -0
  485. package/numkong/test/test_curved.py +207 -0
  486. package/numkong/test/test_dot.cpp +353 -0
  487. package/numkong/test/test_dot.py +296 -0
  488. package/numkong/test/test_each.cpp +437 -0
  489. package/numkong/test/test_each.py +756 -0
  490. package/numkong/test/test_geospatial.cpp +155 -0
  491. package/numkong/test/test_geospatial.py +241 -0
  492. package/numkong/test/test_maxsim.cpp +138 -0
  493. package/numkong/test/test_maxsim.py +209 -0
  494. package/numkong/test/test_mesh.cpp +203 -0
  495. package/numkong/test/test_mesh.py +394 -0
  496. package/numkong/test/test_probability.cpp +136 -0
  497. package/numkong/test/test_probability.py +166 -0
  498. package/numkong/test/test_reduce.cpp +563 -0
  499. package/numkong/test/test_reduce.py +781 -0
  500. package/numkong/test/test_set.cpp +151 -0
  501. package/numkong/test/test_set.py +120 -0
  502. package/numkong/test/test_similarities.py +579 -0
  503. package/numkong/test/test_sparse.cpp +121 -0
  504. package/numkong/test/test_sparse.py +118 -0
  505. package/numkong/test/test_spatial.cpp +518 -0
  506. package/numkong/test/test_spatial.py +269 -0
  507. package/numkong/test/test_tensor.cpp +884 -0
  508. package/numkong/test/test_tensor.py +1417 -0
  509. package/numkong/test/test_trigonometry.cpp +156 -0
  510. package/numkong/test/test_trigonometry.py +180 -0
  511. package/package.json +7 -8
  512. package/prebuilds/darwin-arm64+x64/usearch.node +0 -0
  513. package/prebuilds/linux-arm64/usearch.node +0 -0
  514. package/prebuilds/linux-x64/usearch.node +0 -0
  515. package/fp16/CMakeLists.txt +0 -176
  516. package/fp16/LICENSE +0 -11
  517. package/fp16/README.md +0 -20
  518. package/fp16/bench/alt-element.cc +0 -89
  519. package/fp16/bench/from-alt-array.cc +0 -175
  520. package/fp16/bench/from-ieee-array.cc +0 -399
  521. package/fp16/bench/ieee-element.cc +0 -272
  522. package/fp16/bench/to-alt-array.cc +0 -91
  523. package/fp16/bench/to-ieee-array.cc +0 -303
  524. package/fp16/cmake/DownloadGoogleBenchmark.cmake +0 -15
  525. package/fp16/cmake/DownloadGoogleTest.cmake +0 -15
  526. package/fp16/cmake/DownloadPSimd.cmake +0 -15
  527. package/fp16/configure.py +0 -65
  528. package/fp16/confu.yaml +0 -8
  529. package/fp16/include/fp16/__init__.py +0 -0
  530. package/fp16/include/fp16/avx.py +0 -50
  531. package/fp16/include/fp16/avx2.py +0 -53
  532. package/fp16/include/fp16/bitcasts.h +0 -92
  533. package/fp16/include/fp16/fp16.h +0 -451
  534. package/fp16/include/fp16/psimd.h +0 -131
  535. package/fp16/include/fp16.h +0 -11
  536. package/fp16/jni/Android.mk +0 -6
  537. package/fp16/jni/Application.mk +0 -4
  538. package/fp16/test/alt-from-fp32-value.cc +0 -498
  539. package/fp16/test/alt-to-fp32-bits.cc +0 -263
  540. package/fp16/test/alt-to-fp32-psimd.cc +0 -145
  541. package/fp16/test/alt-to-fp32-value.cc +0 -348
  542. package/fp16/test/alt-to-fp32x2-psimd.cc +0 -245
  543. package/fp16/test/bitcasts.cc +0 -56
  544. package/fp16/test/ieee-from-fp32-value.cc +0 -523
  545. package/fp16/test/ieee-to-fp32-bits.cc +0 -333
  546. package/fp16/test/ieee-to-fp32-psimd.cc +0 -312
  547. package/fp16/test/ieee-to-fp32-value.cc +0 -420
  548. package/fp16/test/ieee-to-fp32x2-psimd.cc +0 -541
  549. package/fp16/test/peachpy/alt-xmm-to-fp32-xmm-avx.cc +0 -96
  550. package/fp16/test/peachpy/alt-xmm-to-fp32-ymm-avx2.cc +0 -96
  551. package/fp16/test/peachpy/stubs.py +0 -38
  552. package/fp16/test/tables.cc +0 -530
  553. package/fp16/test/tables.h +0 -13
  554. package/fp16/third-party/THHalf.h +0 -92
  555. package/fp16/third-party/eigen-half.h +0 -170
  556. package/fp16/third-party/float16-compressor.h +0 -80
  557. package/fp16/third-party/half.hpp +0 -2916
  558. package/fp16/third-party/npy-halffloat.h +0 -328
  559. package/simsimd/.git-blame-ignore-revs +0 -2
  560. package/simsimd/CONTRIBUTING.md +0 -321
  561. package/simsimd/Cargo.toml +0 -49
  562. package/simsimd/MANIFEST.in +0 -4
  563. package/simsimd/Package.swift +0 -28
  564. package/simsimd/README.md +0 -1493
  565. package/simsimd/binding.gyp +0 -35
  566. package/simsimd/build.rs +0 -70
  567. package/simsimd/c/lib.c +0 -374
  568. package/simsimd/include/module.modulemap +0 -4
  569. package/simsimd/include/simsimd/binary.h +0 -486
  570. package/simsimd/include/simsimd/curved.h +0 -1541
  571. package/simsimd/include/simsimd/dot.h +0 -1853
  572. package/simsimd/include/simsimd/elementwise.h +0 -2465
  573. package/simsimd/include/simsimd/geospatial.h +0 -43
  574. package/simsimd/include/simsimd/mesh.h +0 -69
  575. package/simsimd/include/simsimd/probability.h +0 -606
  576. package/simsimd/include/simsimd/simsimd.h +0 -2567
  577. package/simsimd/include/simsimd/sparse.h +0 -1384
  578. package/simsimd/include/simsimd/spatial.h +0 -2335
  579. package/simsimd/include/simsimd/types.h +0 -666
  580. package/simsimd/javascript/fallback.ts +0 -233
  581. package/simsimd/javascript/lib.c +0 -125
  582. package/simsimd/javascript/node-gyp-build.d.ts +0 -1
  583. package/simsimd/javascript/simsimd.ts +0 -179
  584. package/simsimd/javascript/tsconfig-cjs.json +0 -8
  585. package/simsimd/package-lock.json +0 -1761
  586. package/simsimd/package.json +0 -61
  587. package/simsimd/rust/lib.rs +0 -1821
  588. package/simsimd/scripts/bench.cxx +0 -1170
  589. package/simsimd/scripts/bench.ipynb +0 -187
  590. package/simsimd/scripts/bench.js +0 -127
  591. package/simsimd/scripts/bench.rs +0 -876
  592. package/simsimd/scripts/bench_vectors.py +0 -734
  593. package/simsimd/scripts/bench_vectors_live.py +0 -213
  594. package/simsimd/scripts/test.c +0 -187
  595. package/simsimd/scripts/test.mjs +0 -220
  596. package/simsimd/scripts/test.py +0 -1505
  597. package/simsimd/swift/SimSIMD.swift +0 -149
  598. package/simsimd/swift/Test.swift +0 -128
  599. /package/{simsimd → numkong}/CODE_OF_CONDUCT.md +0 -0
  600. /package/{simsimd → numkong}/LICENSE +0 -0
  601. /package/{simsimd → numkong}/javascript/dist-package-cjs.json +0 -0
  602. /package/{simsimd → numkong}/javascript/dist-package-esm.json +0 -0
@@ -0,0 +1,580 @@
1
+ # Batched Distance Matrices in NumKong
2
+
3
+ NumKong implements batched distance matrix computation via pre-packed dot products plus normalization. Angular distance and Euclidean distance are computed from the packed dot product output without materializing an intermediate C matrix.
4
+
5
+ Angular distance from pre-packed dot products:
6
+
7
+ $$
8
+ D_{ij} = 1 - \frac{C_{ij}}{\sqrt{\|A_i\|^2 \cdot \|B_j\|^2}}
9
+ $$
10
+
11
+ Euclidean distance from pre-packed dot products:
12
+
13
+ $$
14
+ D_{ij} = \sqrt{\|A_i\|^2 + \|B_j\|^2 - 2 C_{ij}}
15
+ $$
16
+
17
+ Reformulating as Python pseudocode:
18
+
19
+ ```python
20
+ import numpy as np
21
+
22
+ def angulars_packed(a: np.ndarray, b: np.ndarray) -> np.ndarray:
23
+ dots = a @ b.T
24
+ a_norms = np.sum(a ** 2, axis=1, keepdims=True)
25
+ b_norms = np.sum(b ** 2, axis=1, keepdims=True)
26
+ return 1 - dots / np.sqrt(a_norms * b_norms.T)
27
+
28
+ def euclideans_packed(a: np.ndarray, b: np.ndarray) -> np.ndarray:
29
+ dots = a @ b.T
30
+ a_norms = np.sum(a ** 2, axis=1, keepdims=True)
31
+ b_norms = np.sum(b ** 2, axis=1, keepdims=True)
32
+ return np.sqrt(np.maximum(a_norms + b_norms.T - 2 * dots, 0))
33
+ ```
34
+
35
+ ## Input & Output Types
36
+
37
+ | Input Type | Output Type | Description |
38
+ | ---------- | ----------- | ---------------------------------------------- |
39
+ | `f64` | `f64` | 64-bit IEEE 754 double precision |
40
+ | `f32` | `f32` | 32-bit IEEE 754 single precision |
41
+ | `f16` | `f32` | 16-bit IEEE 754 half precision, widened output |
42
+ | `bf16` | `f32` | 16-bit brain float, widened output |
43
+ | `e4m3` | `f32` | 8-bit Float8: 4 exponent, 3 mantissa bits |
44
+ | `e5m2` | `f32` | 8-bit Float8: 5 exponent, 2 mantissa bits |
45
+ | `e2m3` | `f32` | 8-bit MX format: 2 exponent, 3 mantissa bits |
46
+ | `e3m2` | `f32` | 8-bit MX format: 3 exponent, 2 mantissa bits |
47
+ | `i8` | `f32` | 8-bit signed integers, float output |
48
+ | `u8` | `f32` | 8-bit unsigned integers, float output |
49
+ | `i4` | `f32` | 4-bit signed integers, float output |
50
+ | `u4` | `f32` | 4-bit unsigned integers, float output |
51
+
52
+ ## Optimizations
53
+
54
+ ### Distance-from-Dot Algebraic Reduction
55
+
56
+ `nk_angulars_packed_f32_haswell`, `nk_angulars_packed_f32_skylake`, `nk_euclideans_packed_f32_haswell`, `nk_euclideans_packed_f32_skylake` derive distance matrices from pre-packed dot product output without materializing an intermediate result matrix.
57
+ Angular distance rewrites as $1 - \text{dot}(a,b) \cdot \text{rsqrt}(\|a\|^2 \cdot \|b\|^2)$, converting two separate square roots and a division into one rsqrt and one multiply.
58
+ Euclidean distance expands the identity $\|a - b\|^2 = \|a\|^2 + \|b\|^2 - 2 \cdot \text{dot}(a,b)$, requiring only one final sqrt per output element.
59
+ Both formulas decompose into: (1) a batched GEMM for all M×N dot products, (2) per-vector squared norms precomputed once during packing.
60
+ The singular `spatial/` kernels compute these three sums ($\sum a_i b_i$, $\sum a_i^2$, $\sum b_i^2$) in a single pass with three interleaved accumulators; the batched `spatials/` kernels separate them — norms are computed once per vector during packing, and dots come from the GEMM — trading register pressure for amortized cost across the full M×N output.
61
+
62
+ ### Serial vs Vectorized Sqrt and Rsqrt Cost
63
+
64
+ `nk_angular_through_f32_from_dot_serial_` uses the Quake 3 fast inverse square root (magic constant `0x5F375A86`, three Newton-Raphson iterations, ~34.9 correct bits for Float32) to compute `dot * rsqrt(query_norm * target_norm)`.
65
+ `nk_angular_through_f32_from_dot_haswell_` replaces this with hardware `_mm_rsqrt_ps` (~12-bit approximation, 5cy latency, 1/cy on port 0) plus one Newton-Raphson refinement step (~22–24 correct bits).
66
+ `nk_euclidean_through_f32_from_dot_serial_` computes `sqrt(x)` as `x * rsqrt(x)` — reusing the same rsqrt path.
67
+ `nk_euclidean_through_f32_from_dot_haswell_` uses exact `_mm_sqrt_ps` (11cy latency, 7cy throughput for XMM) instead of the rsqrt approximation — the subtraction $\|a\|^2 + \|b\|^2 - 2 \cdot \text{dot}$ can produce values near zero where rsqrt error would be amplified by the subsequent multiply.
68
+ For Float64, all backends use exact division and sqrt — no fast rsqrt approximation, since reaching 52 mantissa bits of precision would need 4+ Newton-Raphson iterations, negating the speed advantage.
69
+ The 4-wide finalizer batching amortizes these costs: one rsqrt or sqrt call processes 4 output elements simultaneously, hiding the latency behind the GEMM tile's computation.
70
+
71
+ ### Norm Precomputation in Packed Buffers
72
+
73
+ `nk_dots_pack_f32_serial`, `nk_dots_pack_f32_haswell`, `nk_dots_pack_bf16_haswell` compute per-column squared norms $\|b_j\|^2 = \sum_k b_{jk}^2 = \text{dot}(b_j, b_j)$ during the packing step via `nk_reduce_moments_*` primitives.
74
+ The squared norm is a self-dot-product — already a byproduct of touching every element for type conversion and layout transformation.
75
+ Angular and Euclidean finalizers read norms from packed buffer metadata, eliminating a separate O(N·K) norm pass over B.
76
+
77
+ ## Performance
78
+
79
+ The following performance tables are produced by manually re-running `nk_test` and `nk_bench` included internal tools to measure both accuracy and throughput at different input shapes.
80
+ The input size is controlled by `NK_MATRIX_HEIGHT`, `NK_MATRIX_WIDTH`, and `NK_MATRIX_DEPTH` environment variables, all set to the same value for batched distance computations over square matrices.
81
+ Columns show throughput for 256³, 1024³, and 4096³ configurations.
82
+ The throughput is measured in GSO/s as Giga Scalar Operations per Second, with $\text{ops} = 2 \cdot M \cdot N \cdot K$ complexity for computing $M \times N$ pairwise distances over $K$-dimensional vectors.
83
+ Accuracy is reported as mean ULP (units in last place) unless noted otherwise — the average number of representable floating-point values between the result and the exact answer.
84
+ Each kernel runs for at least 20 seconds per configuration.
85
+ Benchmark threads are pinned to specific cores; on machines with heterogeneous core types (e.g., Apple P/E cores), only the fastest cores are used.
86
+ Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run in separate passes to avoid affecting throughput measurements of other kernels.
87
+
88
+ ### Intel Sapphire Rapids
89
+
90
+ #### Native
91
+
92
+ | Kernel | 256³ | 1024³ | 4096³ |
93
+ | :----------------------------------------- | -----------------------: | -----------------------: | -----------------------: |
94
+ | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
95
+ | `nk_angulars_packed_f64_serial` | 0.578 gso/s, 0 ulp | 0.691 gso/s, 0 ulp | 0.787 gso/s, 0 ulp |
96
+ | `nk_angulars_symmetric_f64_serial` | 0.477 gso/s, 0 ulp | 0.569 gso/s, 0 ulp | 1.24 gso/s, 0 ulp |
97
+ | `nk_euclideans_packed_f64_serial` | 0.569 gso/s, 0.6 ulp | 0.692 gso/s, 0.6 ulp | 0.775 gso/s, 0.6 ulp |
98
+ | `nk_euclideans_symmetric_f64_serial` | 0.477 gso/s, 0.6 ulp | 0.562 gso/s, 0.6 ulp | 1.26 gso/s, 0.3 ulp |
99
+ | `nk_angulars_packed_f64_haswell` | 5.89 gso/s, 0 ulp | 6.04 gso/s, 0 ulp | 6.08 gso/s, 0 ulp |
100
+ | `nk_angulars_symmetric_f64_haswell` | 5.17 gso/s, 0 ulp | 5.56 gso/s, 0 ulp | 11.3 gso/s, 0 ulp |
101
+ | `nk_euclideans_packed_f64_haswell` | 5.83 gso/s, 0.2 ulp | 6.21 gso/s, 0.2 ulp | 6.24 gso/s, 0.2 ulp |
102
+ | `nk_euclideans_symmetric_f64_haswell` | 5.33 gso/s, 0.2 ulp | 5.62 gso/s, 0.2 ulp | 11.7 gso/s, 0.2 ulp |
103
+ | `nk_angulars_packed_f64_skylake` | 7.56 gso/s, 0 ulp | 8.46 gso/s, 0 ulp | 8.92 gso/s, 0 ulp |
104
+ | `nk_angulars_symmetric_f64_skylake` | 7.37 gso/s, 0 ulp | 8.66 gso/s, 0 ulp | 17.1 gso/s, 0 ulp |
105
+ | `nk_euclideans_packed_f64_skylake` | 8.06 gso/s, 0.2 ulp | 8.37 gso/s, 0.2 ulp | 8.06 gso/s, 0.2 ulp |
106
+ | `nk_euclideans_symmetric_f64_skylake` | 7.14 gso/s, 0.2 ulp | 8.43 gso/s, 0.2 ulp | 17.4 gso/s, 0.2 ulp |
107
+ | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
108
+ | `nk_angulars_packed_f32_serial` | 15.0 gso/s, 0.1 ulp | 16.3 gso/s, 0.1 ulp | 16.4 gso/s, 0.1 ulp |
109
+ | `nk_angulars_symmetric_f32_serial` | 3.86 gso/s, 0.1 ulp | 4.29 gso/s, 0.1 ulp | 8.62 gso/s, 0.1 ulp |
110
+ | `nk_euclideans_packed_f32_serial` | 15.3 gso/s, 0.6 ulp | 17.0 gso/s, 0.5 ulp | 17.0 gso/s, 0.5 ulp |
111
+ | `nk_euclideans_symmetric_f32_serial` | 3.97 gso/s, 0.6 ulp | 4.16 gso/s, 0.5 ulp | 8.38 gso/s, 0.3 ulp |
112
+ | `nk_angulars_packed_f32_haswell` | 29.3 gso/s, 0 ulp | 31.6 gso/s, 0 ulp | 31.6 gso/s, 0 ulp |
113
+ | `nk_angulars_symmetric_f32_haswell` | 21.4 gso/s, 0 ulp | 24.8 gso/s, 0 ulp | 52 gso/s, 0 ulp |
114
+ | `nk_euclideans_packed_f32_haswell` | 29.7 gso/s, 0.2 ulp | 32 gso/s, 0.2 ulp | 32.9 gso/s, 0.2 ulp |
115
+ | `nk_euclideans_symmetric_f32_haswell` | 21.8 gso/s, 0.2 ulp | 25.7 gso/s, 0.2 ulp | 53 gso/s, 0.2 ulp |
116
+ | `nk_angulars_packed_f32_skylake` | 33.3 gso/s, 0 ulp | 39.4 gso/s, 0 ulp | 37.5 gso/s, 0 ulp |
117
+ | `nk_angulars_symmetric_f32_skylake` | 24.8 gso/s, 0 ulp | 25.5 gso/s, 0 ulp | 61.4 gso/s, 0 ulp |
118
+ | `nk_euclideans_packed_f32_skylake` | 34.4 gso/s, 0.2 ulp | 40.3 gso/s, 0.2 ulp | 40.3 gso/s, 0.2 ulp |
119
+ | `nk_euclideans_symmetric_f32_skylake` | 25.1 gso/s, 0.2 ulp | 29.3 gso/s, 0.2 ulp | 65.9 gso/s, 0.2 ulp |
120
+ | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
121
+ | `nk_angulars_packed_bf16_serial` | 1.18 gso/s, 0 ulp | 1.21 gso/s, 0 ulp | 1.19 gso/s, 0.1 ulp |
122
+ | `nk_angulars_symmetric_bf16_serial` | 1.19 gso/s, 0 ulp | 1.18 gso/s, 0 ulp | 2.35 gso/s, 0 ulp |
123
+ | `nk_euclideans_packed_bf16_serial` | 1.20 gso/s, 0.6 ulp | 1.18 gso/s, 0.6 ulp | 1.16 gso/s, 6.0 ulp |
124
+ | `nk_euclideans_symmetric_bf16_serial` | 1.11 gso/s, 0.6 ulp | 1.14 gso/s, 0.6 ulp | 2.34 gso/s, 0.4 ulp |
125
+ | `nk_angulars_packed_bf16_haswell` | 54.6 gso/s, 0 ulp | 65.7 gso/s, 0 ulp | 66.1 gso/s, 0.1 ulp |
126
+ | `nk_angulars_symmetric_bf16_haswell` | 38.3 gso/s, 0 ulp | 50.1 gso/s, 0 ulp | 106 gso/s, 0 ulp |
127
+ | `nk_euclideans_packed_bf16_haswell` | 58 gso/s, 0.2 ulp | 65.7 gso/s, 0.3 ulp | 70.7 gso/s, 5.8 ulp |
128
+ | `nk_euclideans_symmetric_bf16_haswell` | 38.6 gso/s, 0.2 ulp | 49.8 gso/s, 0.3 ulp | 109 gso/s, 0.3 ulp |
129
+ | `nk_angulars_packed_bf16_skylake` | 67.8 gso/s, 0 ulp | 87.7 gso/s, 0 ulp | 86.4 gso/s, 0.1 ulp |
130
+ | `nk_angulars_symmetric_bf16_skylake` | 48.8 gso/s, 0 ulp | 58.7 gso/s, 0 ulp | 125 gso/s, 0 ulp |
131
+ | `nk_euclideans_packed_bf16_skylake` | 64 gso/s, 0.2 ulp | 87.4 gso/s, 0.3 ulp | 90.8 gso/s, 5.8 ulp |
132
+ | `nk_euclideans_symmetric_bf16_skylake` | 48.8 gso/s, 0.2 ulp | 58.9 gso/s, 0.3 ulp | 121 gso/s, 0.3 ulp |
133
+ | `nk_angulars_packed_bf16_genoa` | 59.7 gso/s, 0 ulp | 81.9 gso/s, 0 ulp | 87.2 gso/s, 0 ulp |
134
+ | `nk_angulars_symmetric_bf16_genoa` | 54.9 gso/s, 0 ulp | 61.2 gso/s, 0 ulp | 137 gso/s, 0 ulp |
135
+ | `nk_euclideans_packed_bf16_genoa` | 63 gso/s, 0.2 ulp | 79.6 gso/s, 0.3 ulp | 87.3 gso/s, 0.3 ulp |
136
+ | `nk_euclideans_symmetric_bf16_genoa` | 53.4 gso/s, 0.2 ulp | 60.2 gso/s, 0.3 ulp | 130 gso/s, 0.3 ulp |
137
+ | `nk_angulars_packed_bf16_sapphireamx` | 287 gso/s, 0 ulp | 364 gso/s, 0 ulp | 582 gso/s, 0 ulp |
138
+ | `nk_angulars_symmetric_bf16_sapphireamx` | 75.7 gso/s, 0 ulp | 114 gso/s, 0 ulp | 116 gso/s, 0 ulp |
139
+ | `nk_euclideans_packed_bf16_sapphireamx` | 328 gso/s, 0.3 ulp | 573 gso/s, 0.3 ulp | 632 gso/s, 0.3 ulp |
140
+ | `nk_euclideans_symmetric_bf16_sapphireamx` | 76.3 gso/s, 0.3 ulp | 115 gso/s, 0.3 ulp | 123 gso/s, 0.3 ulp |
141
+ | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
142
+ | `nk_angulars_packed_f16_serial` | 7.46 gso/s, 0.1 ulp | 7.97 gso/s, 0.1 ulp | 8.12 gso/s, 0.1 ulp |
143
+ | `nk_angulars_symmetric_f16_serial` | 4.04 gso/s, 0.1 ulp | 4.09 gso/s, 0.1 ulp | 8.13 gso/s, 0.1 ulp |
144
+ | `nk_euclideans_packed_f16_serial` | 7.69 gso/s, 0.7 ulp | 7.73 gso/s, 1.1 ulp | 8.34 gso/s, 0.6 ulp |
145
+ | `nk_euclideans_symmetric_f16_serial` | 4.08 gso/s, 0.7 ulp | 4.19 gso/s, 1.1 ulp | 8.23 gso/s, 0.5 ulp |
146
+ | `nk_angulars_packed_f16_haswell` | 62 gso/s, 0.1 ulp | 74.4 gso/s, 0.1 ulp | 70.6 gso/s, 0.1 ulp |
147
+ | `nk_angulars_symmetric_f16_haswell` | 38.3 gso/s, 0.1 ulp | 54.9 gso/s, 0.1 ulp | 121 gso/s, 0.1 ulp |
148
+ | `nk_euclideans_packed_f16_haswell` | 62.9 gso/s, 0.4 ulp | 75.2 gso/s, 0.9 ulp | 75.7 gso/s, 0.5 ulp |
149
+ | `nk_euclideans_symmetric_f16_haswell` | 39.6 gso/s, 0.4 ulp | 54.2 gso/s, 0.9 ulp | 123 gso/s, 0.3 ulp |
150
+ | `nk_angulars_packed_f16_skylake` | 66.6 gso/s, 0.1 ulp | 85.2 gso/s, 0.1 ulp | 88.3 gso/s, 0 ulp |
151
+ | `nk_angulars_symmetric_f16_skylake` | 50.1 gso/s, 0.1 ulp | 57.7 gso/s, 0.1 ulp | 126 gso/s, 0 ulp |
152
+ | `nk_euclideans_packed_f16_skylake` | 69.6 gso/s, 0.4 ulp | 93.3 gso/s, 0.9 ulp | 91 gso/s, 0.5 ulp |
153
+ | `nk_euclideans_symmetric_f16_skylake` | 49.4 gso/s, 0.4 ulp | 59.8 gso/s, 0.9 ulp | 134 gso/s, 0.3 ulp |
154
+ | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
155
+ | `nk_angulars_packed_e5m2_serial` | 0.587 gso/s, 0 ulp | 0.553 gso/s, 0 ulp | 0.563 gso/s, 0 ulp |
156
+ | `nk_angulars_symmetric_e5m2_serial` | 0.446 gso/s, 0 ulp | 0.427 gso/s, 0 ulp | 0.847 gso/s, 0 ulp |
157
+ | `nk_euclideans_packed_e5m2_serial` | 0.576 gso/s, 0.5 ulp | 0.571 gso/s, 0.5 ulp | 0.557 gso/s, 0.2 ulp |
158
+ | `nk_euclideans_symmetric_e5m2_serial` | 0.424 gso/s, 0.5 ulp | 0.437 gso/s, 0.5 ulp | 0.836 gso/s, 0.2 ulp |
159
+ | `nk_angulars_packed_e5m2_haswell` | 27.4 gso/s, 0 ulp | 30.4 gso/s, 0 ulp | 31 gso/s, 0 ulp |
160
+ | `nk_angulars_symmetric_e5m2_haswell` | 15.3 gso/s, 0 ulp | 15.7 gso/s, 0 ulp | 32.3 gso/s, 0 ulp |
161
+ | `nk_euclideans_packed_e5m2_haswell` | 28 gso/s, 0 ulp | 30.8 gso/s, 0 ulp | 30.6 gso/s, 0 ulp |
162
+ | `nk_euclideans_symmetric_e5m2_haswell` | 15.4 gso/s, 0 ulp | 15.9 gso/s, 0 ulp | 32 gso/s, 0 ulp |
163
+ | `nk_angulars_packed_e5m2_skylake` | 32.9 gso/s, 0 ulp | 36.7 gso/s, 0 ulp | 40.1 gso/s, 0 ulp |
164
+ | `nk_angulars_symmetric_e5m2_skylake` | 19 gso/s, 0 ulp | 21 gso/s, 0 ulp | 42.7 gso/s, 0 ulp |
165
+ | `nk_euclideans_packed_e5m2_skylake` | 34.1 gso/s, 0 ulp | 37.9 gso/s, 0 ulp | 39.6 gso/s, 0 ulp |
166
+ | `nk_euclideans_symmetric_e5m2_skylake` | 20 gso/s, 0 ulp | 18.4 gso/s, 0 ulp | 41.6 gso/s, 0 ulp |
167
+ | `nk_angulars_packed_e5m2_genoa` | 39.6 gso/s, 0 ulp | 46.8 gso/s, 0 ulp | 47.5 gso/s, 0 ulp |
168
+ | `nk_angulars_symmetric_e5m2_genoa` | 30 gso/s, 0 ulp | 32.5 gso/s, 0 ulp | 66.3 gso/s, 0 ulp |
169
+ | `nk_euclideans_packed_e5m2_genoa` | 42.3 gso/s, 0 ulp | 49.1 gso/s, 0 ulp | 51.3 gso/s, 0 ulp |
170
+ | `nk_euclideans_symmetric_e5m2_genoa` | 30.1 gso/s, 0 ulp | 32.8 gso/s, 0 ulp | 64.9 gso/s, 0 ulp |
171
+ | `nk_angulars_packed_e5m2_sapphireamx` | 216 gso/s, 0 ulp | 355 gso/s, 0 ulp | 427 gso/s, 0 ulp |
172
+ | `nk_angulars_symmetric_e5m2_sapphireamx` | 48.7 gso/s, 0 ulp | 73.3 gso/s, 0 ulp | 72.3 gso/s, 0 ulp |
173
+ | `nk_euclideans_packed_e5m2_sapphireamx` | 220 gso/s, 0 ulp | 375 gso/s, 0 ulp | 408 gso/s, 0 ulp |
174
+ | `nk_euclideans_symmetric_e5m2_sapphireamx` | 48.3 gso/s, 0 ulp | 73.3 gso/s, 0 ulp | 74 gso/s, 0 ulp |
175
+ | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
176
+ | `nk_angulars_packed_e4m3_serial` | 0.479 gso/s, 0 ulp | 0.473 gso/s, 0 ulp | 0.485 gso/s, 0 ulp |
177
+ | `nk_angulars_symmetric_e4m3_serial` | 0.395 gso/s, 0 ulp | 0.390 gso/s, 0 ulp | 0.795 gso/s, 0 ulp |
178
+ | `nk_euclideans_packed_e4m3_serial` | 0.467 gso/s, 0.5 ulp | 0.484 gso/s, 0.5 ulp | 0.480 gso/s, 0.5 ulp |
179
+ | `nk_euclideans_symmetric_e4m3_serial` | 0.395 gso/s, 0.5 ulp | 0.395 gso/s, 0.5 ulp | 0.781 gso/s, 0.3 ulp |
180
+ | `nk_angulars_packed_e4m3_haswell` | 20.6 gso/s, 0 ulp | 22.5 gso/s, 0 ulp | 21.8 gso/s, 0 ulp |
181
+ | `nk_angulars_symmetric_e4m3_haswell` | 12.2 gso/s, 0 ulp | 12.1 gso/s, 0 ulp | 24.7 gso/s, 0 ulp |
182
+ | `nk_euclideans_packed_e4m3_haswell` | 20.7 gso/s, 0 ulp | 22.4 gso/s, 0 ulp | 23.4 gso/s, 0.2 ulp |
183
+ | `nk_euclideans_symmetric_e4m3_haswell` | 11.2 gso/s, 0 ulp | 11.9 gso/s, 0 ulp | 24.6 gso/s, 0.1 ulp |
184
+ | `nk_angulars_packed_e4m3_skylake` | 28.8 gso/s, 0 ulp | 32.8 gso/s, 0 ulp | 31.3 gso/s, 0 ulp |
185
+ | `nk_angulars_symmetric_e4m3_skylake` | 16.4 gso/s, 0 ulp | 17.4 gso/s, 0 ulp | 35.1 gso/s, 0 ulp |
186
+ | `nk_euclideans_packed_e4m3_skylake` | 27.8 gso/s, 0 ulp | 31.2 gso/s, 0 ulp | 31.7 gso/s, 0.2 ulp |
187
+ | `nk_euclideans_symmetric_e4m3_skylake` | 16.1 gso/s, 0 ulp | 16.8 gso/s, 0 ulp | 34.4 gso/s, 0.1 ulp |
188
+ | `nk_angulars_packed_e4m3_genoa` | 40.8 gso/s, 0 ulp | 48.4 gso/s, 0 ulp | 52.1 gso/s, 0 ulp |
189
+ | `nk_angulars_symmetric_e4m3_genoa` | 30.3 gso/s, 0 ulp | 31.5 gso/s, 0 ulp | 69.2 gso/s, 0 ulp |
190
+ | `nk_euclideans_packed_e4m3_genoa` | 43.3 gso/s, 0 ulp | 50.9 gso/s, 0 ulp | 48.8 gso/s, 0.1 ulp |
191
+ | `nk_euclideans_symmetric_e4m3_genoa` | 29.9 gso/s, 0 ulp | 31.9 gso/s, 0 ulp | 64.6 gso/s, 0.1 ulp |
192
+ | `nk_angulars_packed_e4m3_sapphireamx` | 212 gso/s, 0 ulp | 325 gso/s, 0 ulp | 418 gso/s, 0 ulp |
193
+ | `nk_angulars_symmetric_e4m3_sapphireamx` | 50.5 gso/s, 0 ulp | 73.4 gso/s, 0 ulp | 72 gso/s, 0 ulp |
194
+ | `nk_euclideans_packed_e4m3_sapphireamx` | 216 gso/s, 0.1 ulp | 372 gso/s, 0.1 ulp | 394 gso/s, 0.1 ulp |
195
+ | `nk_euclideans_symmetric_e4m3_sapphireamx` | 49.3 gso/s, 0.1 ulp | 70.1 gso/s, 0.1 ulp | 73.1 gso/s, 0.1 ulp |
196
+ | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
197
+ | `nk_angulars_packed_e3m2_serial` | 0.554 gso/s, 0 ulp | 0.524 gso/s, 0 ulp | 0.534 gso/s, 0 ulp |
198
+ | `nk_angulars_symmetric_e3m2_serial` | 0.439 gso/s, 0 ulp | 0.427 gso/s, 0 ulp | 0.839 gso/s, 0 ulp |
199
+ | `nk_euclideans_packed_e3m2_serial` | 0.556 gso/s, 0.5 ulp | 0.549 gso/s, 0.5 ulp | 0.509 gso/s, 0.2 ulp |
200
+ | `nk_euclideans_symmetric_e3m2_serial` | 0.413 gso/s, 0.5 ulp | 0.427 gso/s, 0.5 ulp | 0.829 gso/s, 0.2 ulp |
201
+ | `nk_angulars_packed_e3m2_haswell` | 30.3 gso/s, 0 ulp | 32.2 gso/s, 0 ulp | 32.8 gso/s, 0 ulp |
202
+ | `nk_angulars_symmetric_e3m2_haswell` | 27.1 gso/s, 0 ulp | 32.8 gso/s, 0 ulp | 65.7 gso/s, 0 ulp |
203
+ | `nk_euclideans_packed_e3m2_haswell` | 30.1 gso/s, 0 ulp | 32.3 gso/s, 0 ulp | 33.5 gso/s, 0 ulp |
204
+ | `nk_euclideans_symmetric_e3m2_haswell` | 28.5 gso/s, 0 ulp | 32.6 gso/s, 0 ulp | 66.1 gso/s, 0 ulp |
205
+ | `nk_angulars_packed_e3m2_skylake` | 37.4 gso/s, 0 ulp | 41.4 gso/s, 0 ulp | 44.1 gso/s, 0 ulp |
206
+ | `nk_angulars_symmetric_e3m2_skylake` | 39 gso/s, 0 ulp | 41.9 gso/s, 0 ulp | 87.3 gso/s, 0 ulp |
207
+ | `nk_euclideans_packed_e3m2_skylake` | 35.7 gso/s, 0 ulp | 41.3 gso/s, 0 ulp | 43 gso/s, 0 ulp |
208
+ | `nk_euclideans_symmetric_e3m2_skylake` | 36.2 gso/s, 0 ulp | 36.4 gso/s, 0 ulp | 87.8 gso/s, 0 ulp |
209
+ | `nk_angulars_packed_e3m2_genoa` | 48 gso/s, 0 ulp | 56 gso/s, 0 ulp | 59.3 gso/s, 0 ulp |
210
+ | `nk_angulars_symmetric_e3m2_genoa` | 40 gso/s, 0 ulp | 40.8 gso/s, 0 ulp | 87.4 gso/s, 0 ulp |
211
+ | `nk_euclideans_packed_e3m2_genoa` | 49.8 gso/s, 0 ulp | 58.4 gso/s, 0 ulp | 61 gso/s, 0 ulp |
212
+ | `nk_euclideans_symmetric_e3m2_genoa` | 38.4 gso/s, 0 ulp | 41.6 gso/s, 0 ulp | 87.7 gso/s, 0 ulp |
213
+ | `nk_angulars_packed_e3m2_sapphireamx` | 238 gso/s, 0 ulp | 420 gso/s, 0 ulp | 431 gso/s, 0 ulp |
214
+ | `nk_angulars_symmetric_e3m2_sapphireamx` | 60.7 gso/s, 0 ulp | 96.5 gso/s, 0 ulp | 90.9 gso/s, 0 ulp |
215
+ | `nk_euclideans_packed_e3m2_sapphireamx` | 224 gso/s, 0 ulp | 426 gso/s, 0 ulp | 443 gso/s, 0 ulp |
216
+ | `nk_euclideans_symmetric_e3m2_sapphireamx` | 60.8 gso/s, 0 ulp | 99.2 gso/s, 0 ulp | 92.6 gso/s, 0 ulp |
217
+ | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
218
+ | `nk_angulars_packed_e2m3_serial` | 0.332 gso/s, 0 ulp | 0.325 gso/s, 0 ulp | 0.320 gso/s, 0 ulp |
219
+ | `nk_angulars_symmetric_e2m3_serial` | 0.298 gso/s, 0 ulp | 0.305 gso/s, 0 ulp | 0.568 gso/s, 0 ulp |
220
+ | `nk_euclideans_packed_e2m3_serial` | 0.324 gso/s, 0.5 ulp | 0.310 gso/s, 0.5 ulp | 0.313 gso/s, 0.2 ulp |
221
+ | `nk_euclideans_symmetric_e2m3_serial` | 0.293 gso/s, 0.5 ulp | 0.295 gso/s, 0.5 ulp | 0.586 gso/s, 0.2 ulp |
222
+ | `nk_angulars_packed_e2m3_haswell` | 54.2 gso/s, 0 ulp | 61 gso/s, 0 ulp | 66.2 gso/s, 0 ulp |
223
+ | `nk_angulars_symmetric_e2m3_haswell` | 48.2 gso/s, 0 ulp | 60 gso/s, 0 ulp | 128 gso/s, 0 ulp |
224
+ | `nk_euclideans_packed_e2m3_haswell` | 55.9 gso/s, 0 ulp | 63.4 gso/s, 0 ulp | 64.8 gso/s, 0 ulp |
225
+ | `nk_euclideans_symmetric_e2m3_haswell` | 48.6 gso/s, 0 ulp | 62.1 gso/s, 0 ulp | 128 gso/s, 0 ulp |
226
+ | `nk_angulars_packed_e2m3_skylake` | 65.1 gso/s, 0 ulp | 79.4 gso/s, 0 ulp | 85.4 gso/s, 0 ulp |
227
+ | `nk_angulars_symmetric_e2m3_skylake` | 61.7 gso/s, 0 ulp | 81.1 gso/s, 0 ulp | 163 gso/s, 0 ulp |
228
+ | `nk_euclideans_packed_e2m3_skylake` | 65.1 gso/s, 0 ulp | 80.4 gso/s, 0 ulp | 80.8 gso/s, 0 ulp |
229
+ | `nk_euclideans_symmetric_e2m3_skylake` | 60.8 gso/s, 0 ulp | 62.3 gso/s, 0 ulp | 167 gso/s, 0 ulp |
230
+ | `nk_angulars_packed_e2m3_genoa` | 47.7 gso/s, 0 ulp | 55.4 gso/s, 0 ulp | 60 gso/s, 0 ulp |
231
+ | `nk_angulars_symmetric_e2m3_genoa` | 36.4 gso/s, 0 ulp | 41.5 gso/s, 0 ulp | 86.7 gso/s, 0 ulp |
232
+ | `nk_euclideans_packed_e2m3_genoa` | 50 gso/s, 0 ulp | 59.1 gso/s, 0 ulp | 58.3 gso/s, 0 ulp |
233
+ | `nk_euclideans_symmetric_e2m3_genoa` | 38 gso/s, 0 ulp | 42.3 gso/s, 0 ulp | 85.1 gso/s, 0 ulp |
234
+ | `nk_angulars_packed_e2m3_sapphireamx` | 350 gso/s, 0 ulp | 956 gso/s, 0 ulp | 1,020 gso/s, 0 ulp |
235
+ | `nk_angulars_symmetric_e2m3_sapphireamx` | 88.4 gso/s, 0 ulp | 203 gso/s, 0 ulp | 188 gso/s, 0 ulp |
236
+ | `nk_euclideans_packed_e2m3_sapphireamx` | 337 gso/s, 0 ulp | 990 gso/s, 0 ulp | 992 gso/s, 0 ulp |
237
+ | `nk_euclideans_symmetric_e2m3_sapphireamx` | 88.7 gso/s, 0 ulp | 193 gso/s, 0 ulp | 201 gso/s, 0 ulp |
238
+ | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
239
+ | `nk_angulars_packed_i8_serial` | 8.84 gso/s, 0 ulp | 9.49 gso/s, 0 ulp | 10.1 gso/s, 0 ulp |
240
+ | `nk_angulars_symmetric_i8_serial` | 4.40 gso/s, 0 ulp | 4.45 gso/s, 0 ulp | 9.58 gso/s, ? ulp |
241
+ | `nk_euclideans_packed_i8_serial` | 8.64 gso/s, 0.4 ulp | 9.84 gso/s, 0.4 ulp | 9.94 gso/s, 0.4 ulp |
242
+ | `nk_euclideans_symmetric_i8_serial` | 4.47 gso/s, 0.4 ulp | 4.64 gso/s, 0.4 ulp | 9.15 gso/s, ? ulp |
243
+ | `nk_angulars_packed_i8_haswell` | 79.5 gso/s, 0 ulp | 102 gso/s, 0 ulp | 109 gso/s, 0 ulp |
244
+ | `nk_angulars_symmetric_i8_haswell` | 60.6 gso/s, 0 ulp | 77.4 gso/s, 0 ulp | 168 gso/s, ? ulp |
245
+ | `nk_euclideans_packed_i8_haswell` | 82.5 gso/s, 0 ulp | 102 gso/s, 0 ulp | 109 gso/s, 0 ulp |
246
+ | `nk_euclideans_symmetric_i8_haswell` | 62 gso/s, 0 ulp | 76.5 gso/s, 0 ulp | 166 gso/s, ? ulp |
247
+ | `nk_angulars_packed_i8_icelake` | 155 gso/s, 0 ulp | 206 gso/s, 0 ulp | 402 gso/s, 0 ulp |
248
+ | `nk_angulars_symmetric_i8_icelake` | 103 gso/s, 0 ulp | 263 gso/s, 0 ulp | 690 gso/s, ? ulp |
249
+ | `nk_euclideans_packed_i8_icelake` | 169 gso/s, 0 ulp | 313 gso/s, 0 ulp | 393 gso/s, 0 ulp |
250
+ | `nk_euclideans_symmetric_i8_icelake` | 108 gso/s, 0 ulp | 268 gso/s, 0 ulp | 695 gso/s, ? ulp |
251
+ | `nk_angulars_packed_i8_sapphireamx` | 427 gso/s, 0 ulp | 1,020 gso/s, 0 ulp | 1,170 gso/s, 0 ulp |
252
+ | `nk_angulars_symmetric_i8_sapphireamx` | 106 gso/s, 0 ulp | 261 gso/s, 0 ulp | 210 gso/s, 0 ulp |
253
+ | `nk_euclideans_packed_i8_sapphireamx` | 428 gso/s, 0 ulp | 1,240 gso/s, 0 ulp | 1,170 gso/s, 0 ulp |
254
+ | `nk_euclideans_symmetric_i8_sapphireamx` | 104 gso/s, 0 ulp | 243 gso/s, 0 ulp | 219 gso/s, 0 ulp |
255
+ | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
256
+ | `nk_angulars_packed_u8_serial` | 12.2 gso/s, 0.3 ulp | 12.8 gso/s, 0.3 ulp | 13.0 gso/s, 0.3 ulp |
257
+ | `nk_angulars_symmetric_u8_serial` | 4.48 gso/s, 0.3 ulp | 4.73 gso/s, 0.3 ulp | 9.50 gso/s, ? ulp |
258
+ | `nk_euclideans_packed_u8_serial` | 12.0 gso/s, 0.5 ulp | 13.1 gso/s, 0.5 ulp | 13.4 gso/s, 0.6 ulp |
259
+ | `nk_euclideans_symmetric_u8_serial` | 4.52 gso/s, 0.5 ulp | 4.69 gso/s, 0.5 ulp | 9.65 gso/s, ? ulp |
260
+ | `nk_angulars_packed_u8_haswell` | 54.6 gso/s, 0.3 ulp | 87.8 gso/s, 0.3 ulp | 104 gso/s, 0.3 ulp |
261
+ | `nk_angulars_symmetric_u8_haswell` | 44.6 gso/s, 0.3 ulp | 70.2 gso/s, 0.3 ulp | 161 gso/s, ? ulp |
262
+ | `nk_euclideans_packed_u8_haswell` | 55.5 gso/s, 0.5 ulp | 87.7 gso/s, 0.5 ulp | 105 gso/s, 0.6 ulp |
263
+ | `nk_euclideans_symmetric_u8_haswell` | 45.3 gso/s, 0.5 ulp | 68.4 gso/s, 0.5 ulp | 159 gso/s, ? ulp |
264
+ | `nk_angulars_packed_u8_icelake` | 154 gso/s, 0.3 ulp | 301 gso/s, 0.3 ulp | 404 gso/s, 0.3 ulp |
265
+ | `nk_angulars_symmetric_u8_icelake` | 108 gso/s, 0.3 ulp | 267 gso/s, 0.3 ulp | 699 gso/s, ? ulp |
266
+ | `nk_euclideans_packed_u8_icelake` | 168 gso/s, 0 ulp | 300 gso/s, 0 ulp | 402 gso/s, 0 ulp |
267
+ | `nk_euclideans_symmetric_u8_icelake` | 109 gso/s, 0 ulp | 253 gso/s, 0 ulp | 695 gso/s, ? ulp |
268
+ | `nk_angulars_packed_u8_sapphireamx` | 444 gso/s, 0.2 ulp | 1,210 gso/s, 0.2 ulp | 1,220 gso/s, 0.2 ulp |
269
+ | `nk_angulars_symmetric_u8_sapphireamx` | 103 gso/s, 0.2 ulp | 257 gso/s, 0.2 ulp | 227 gso/s, 0.2 ulp |
270
+ | `nk_euclideans_packed_u8_sapphireamx` | 432 gso/s, 0 ulp | 1,240 gso/s, 0 ulp | 1,200 gso/s, 0 ulp |
271
+ | `nk_euclideans_symmetric_u8_sapphireamx` | 102 gso/s, 0 ulp | 256 gso/s, 0 ulp | 220 gso/s, 0 ulp |
272
+ | __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
273
+ | `nk_angulars_packed_i4_serial` | 3.79 gso/s, ? ulp | 3.83 gso/s, ? ulp | 4.06 gso/s, ? ulp |
274
+ | `nk_angulars_symmetric_i4_serial` | 3.52 gso/s, ? ulp | 3.58 gso/s, ? ulp | 7.08 gso/s, ? ulp |
275
+ | `nk_euclideans_packed_i4_serial` | 3.69 gso/s, ? ulp | 3.91 gso/s, ? ulp | 3.76 gso/s, ? ulp |
276
+ | `nk_euclideans_symmetric_i4_serial` | 3.45 gso/s, ? ulp | 3.64 gso/s, ? ulp | 6.99 gso/s, ? ulp |
277
+ | `nk_angulars_packed_i4_icelake` | 117 gso/s, ? ulp | 208 gso/s, ? ulp | 249 gso/s, ? ulp |
278
+ | `nk_angulars_symmetric_i4_icelake` | 103 gso/s, ? ulp | 233 gso/s, ? ulp | 561 gso/s, ? ulp |
279
+ | `nk_euclideans_packed_i4_icelake` | 121 gso/s, ? ulp | 173 gso/s, ? ulp | 246 gso/s, ? ulp |
280
+ | `nk_euclideans_symmetric_i4_icelake` | 101 gso/s, ? ulp | 228 gso/s, ? ulp | 572 gso/s, ? ulp |
281
+ | __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
282
+ | `nk_angulars_packed_u4_serial` | 5.49 gso/s, ? ulp | 5.60 gso/s, ? ulp | 5.78 gso/s, ? ulp |
283
+ | `nk_angulars_symmetric_u4_serial` | 5.18 gso/s, ? ulp | 5.57 gso/s, ? ulp | 11.5 gso/s, ? ulp |
284
+ | `nk_euclideans_packed_u4_serial` | 5.23 gso/s, ? ulp | 5.50 gso/s, ? ulp | 5.64 gso/s, ? ulp |
285
+ | `nk_euclideans_symmetric_u4_serial` | 5.22 gso/s, ? ulp | 5.47 gso/s, ? ulp | 11.1 gso/s, ? ulp |
286
+ | `nk_angulars_packed_u4_icelake` | 153 gso/s, ? ulp | 270 gso/s, ? ulp | 381 gso/s, ? ulp |
287
+ | `nk_angulars_symmetric_u4_icelake` | 122 gso/s, ? ulp | 264 gso/s, ? ulp | 658 gso/s, ? ulp |
288
+ | `nk_euclideans_packed_u4_icelake` | 158 gso/s, ? ulp | 285 gso/s, ? ulp | 385 gso/s, ? ulp |
289
+ | `nk_euclideans_symmetric_u4_icelake` | 120 gso/s, ? ulp | 279 gso/s, ? ulp | 624 gso/s, ? ulp |
290
+
291
+ #### WASM
292
+
293
+ Measured with Wasmtime v42 (Cranelift backend).
294
+
295
+ | Kernel | 256³ | 1024³ | 4096³ |
296
+ | :----------------------------------------- | -----------------------: | -----------------------: | -----------------------: |
297
+ | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
298
+ | `nk_angulars_packed_f64_serial` | 1.38 gso/s, 0 ulp | 1.37 gso/s, 0 ulp | 1.36 gso/s, 0 ulp |
299
+ | `nk_angulars_symmetric_f64_serial` | 0.267 gso/s, 0 ulp | 0.268 gso/s, 0 ulp | 0.258 gso/s, 0 ulp |
300
+ | `nk_euclideans_packed_f64_serial` | 1.41 gso/s, 0.6 ulp | 1.37 gso/s, 0.6 ulp | 1.36 gso/s, 0.6 ulp |
301
+ | `nk_euclideans_symmetric_f64_serial` | 0.272 gso/s, 0.6 ulp | 0.271 gso/s, 0.5 ulp | 0.161 gso/s, 0.5 ulp |
302
+ | `nk_angulars_packed_f64_v128relaxed` | 10.9 gso/s, 0.1 ulp | 10.9 gso/s, 0.1 ulp | 10.9 gso/s, 0.1 ulp |
303
+ | `nk_angulars_symmetric_f64_v128relaxed` | 0.238 gso/s, 0.1 ulp | 0.240 gso/s, 0.1 ulp | 0.271 gso/s, 0.1 ulp |
304
+ | `nk_euclideans_packed_f64_v128relaxed` | 11.0 gso/s, 0.6 ulp | 11.2 gso/s, 0.6 ulp | 11.2 gso/s, 0.6 ulp |
305
+ | `nk_euclideans_symmetric_f64_v128relaxed` | 0.0463 gso/s, 0.6 ulp | 0.0465 gso/s, 0.5 ulp | 0.00806 gso/s, 0.5 ulp |
306
+ | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
307
+ | `nk_angulars_packed_f32_serial` | 4.16 gso/s, 0.1 ulp | 4.26 gso/s, 0.1 ulp | 4.39 gso/s, 0.1 ulp |
308
+ | `nk_angulars_symmetric_f32_serial` | 3.08 gso/s, 0.1 ulp | 4.88 gso/s, 0.1 ulp | 5.69 gso/s, 0.1 ulp |
309
+ | `nk_euclideans_packed_f32_serial` | 4.19 gso/s, 0.6 ulp | 4.32 gso/s, 0.6 ulp | 4.33 gso/s, 0.5 ulp |
310
+ | `nk_euclideans_symmetric_f32_serial` | 3.05 gso/s, 0.5 ulp | 4.97 gso/s, 0.5 ulp | 5.64 gso/s, 0.5 ulp |
311
+ | `nk_angulars_packed_f32_v128relaxed` | 9.41 gso/s, 0.1 ulp | 10.6 gso/s, 0.1 ulp | 10.7 gso/s, 0.1 ulp |
312
+ | `nk_angulars_symmetric_f32_v128relaxed` | 3.64 gso/s, 0.1 ulp | 6.14 gso/s, 0.1 ulp | 7.33 gso/s, 0.1 ulp |
313
+ | `nk_euclideans_packed_f32_v128relaxed` | 9.55 gso/s, 0.2 ulp | 10.6 gso/s, 0.2 ulp | 10.6 gso/s, 0.2 ulp |
314
+ | `nk_euclideans_symmetric_f32_v128relaxed` | 3.55 gso/s, 0.2 ulp | 6.15 gso/s, 0.2 ulp | 7.27 gso/s, 0.2 ulp |
315
+ | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
316
+ | `nk_angulars_packed_bf16_serial` | 4.10 gso/s, 0 ulp | 4.33 gso/s, 0.2 ulp | 4.45 gso/s, 0.6 ulp |
317
+ | `nk_angulars_symmetric_bf16_serial` | 3.74 gso/s, 0 ulp | 6.15 gso/s, 0.2 ulp | 7.39 gso/s, 0.6 ulp |
318
+ | `nk_euclideans_packed_bf16_serial` | 4.26 gso/s, 0.7 ulp | 4.35 gso/s, 6.1 ulp | 4.40 gso/s, 32 ulp |
319
+ | `nk_euclideans_symmetric_bf16_serial` | 3.80 gso/s, 0.6 ulp | 6.16 gso/s, 5.3 ulp | 7.40 gso/s, 28 ulp |
320
+ | `nk_angulars_packed_bf16_v128relaxed` | 22.0 gso/s, 0 ulp | 24.8 gso/s, 0.2 ulp | 24.7 gso/s, 0.6 ulp |
321
+ | `nk_angulars_symmetric_bf16_v128relaxed` | 4.78 gso/s, 0 ulp | 9.61 gso/s, 0.2 ulp | 12.5 gso/s, 0.6 ulp |
322
+ | `nk_euclideans_packed_bf16_v128relaxed` | 22.2 gso/s, 0.7 ulp | 24.1 gso/s, 6.1 ulp | 24.8 gso/s, 32 ulp |
323
+ | `nk_euclideans_symmetric_bf16_v128relaxed` | 4.72 gso/s, 0.3 ulp | 9.53 gso/s, 5.1 ulp | 12.4 gso/s, 28 ulp |
324
+ | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
325
+ | `nk_angulars_packed_e2m3_serial` | 2.66 gso/s, 0 ulp | 2.71 gso/s, 0 ulp | 2.63 gso/s, 0 ulp |
326
+ | `nk_angulars_symmetric_e2m3_serial` | 0.0400 gso/s, 0 ulp | 0.0413 gso/s, 0 ulp | 0.238 gso/s, 0 ulp |
327
+ | `nk_euclideans_packed_e2m3_serial` | 2.74 gso/s, 0.5 ulp | 2.70 gso/s, 0.5 ulp | 2.67 gso/s, 0.5 ulp |
328
+ | `nk_euclideans_symmetric_e2m3_serial` | 0.0403 gso/s, 0.5 ulp | 0.0411 gso/s, 0.4 ulp | 0.0401 gso/s, 0.4 ulp |
329
+ | `nk_angulars_packed_e2m3_v128relaxed` | 18.4 gso/s, 0 ulp | 18.6 gso/s, 0 ulp | 18.5 gso/s, 0 ulp |
330
+ | `nk_angulars_symmetric_e2m3_v128relaxed` | 0.0559 gso/s, 0 ulp | 0.0180 gso/s, 0 ulp | 0.131 gso/s, 0 ulp |
331
+ | `nk_euclideans_packed_e2m3_v128relaxed` | 18.5 gso/s, 0 ulp | 18.7 gso/s, 0 ulp | 18.1 gso/s, 0 ulp |
332
+ | `nk_euclideans_symmetric_e2m3_v128relaxed` | 0.206 gso/s, 0 ulp | 0.0170 gso/s, 0 ulp | 0.0554 gso/s, 0 ulp |
333
+ | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
334
+ | `nk_angulars_packed_i8_serial` | 4.73 gso/s, 0 ulp | 4.81 gso/s, 0 ulp | 4.59 gso/s, 0 ulp |
335
+ | `nk_angulars_symmetric_i8_serial` | 0.00447 gso/s, 0 ulp | 0.198 gso/s, 0 ulp | 0.190 gso/s, 0 ulp |
336
+ | `nk_euclideans_packed_i8_serial` | 4.77 gso/s, 0.5 ulp | 4.80 gso/s, 0.4 ulp | 4.65 gso/s, 0.4 ulp |
337
+ | `nk_euclideans_symmetric_i8_serial` | 0.201 gso/s, 0.5 ulp | 0.0819 gso/s, 0.4 ulp | 0.0823 gso/s, 0.4 ulp |
338
+ | `nk_angulars_packed_i8_v128relaxed` | 31.6 gso/s, 0 ulp | 31.7 gso/s, 0 ulp | 31.1 gso/s, 0 ulp |
339
+ | `nk_angulars_symmetric_i8_v128relaxed` | 0.0304 gso/s, 0 ulp | 0.0680 gso/s, 0 ulp | 0.298 gso/s, 0 ulp |
340
+ | `nk_euclideans_packed_i8_v128relaxed` | 31.5 gso/s, 0 ulp | 32.3 gso/s, 0 ulp | 30.8 gso/s, 0 ulp |
341
+ | `nk_euclideans_symmetric_i8_v128relaxed` | 0.224 gso/s, 0 ulp | 0.222 gso/s, 0 ulp | 0.143 gso/s, 0 ulp |
342
+ | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
343
+ | `nk_angulars_packed_u8_serial` | 4.26 gso/s, 0.4 ulp | 5.07 gso/s, 0.3 ulp | 5.11 gso/s, 0.3 ulp |
344
+ | `nk_angulars_symmetric_u8_serial` | 2.64 gso/s, 0.4 ulp | 4.02 gso/s, 0.3 ulp | 4.34 gso/s, 0.3 ulp |
345
+ | `nk_euclideans_packed_u8_serial` | 4.35 gso/s, 0.5 ulp | 4.67 gso/s, 0.5 ulp | 5.09 gso/s, 0.5 ulp |
346
+ | `nk_euclideans_symmetric_u8_serial` | 2.64 gso/s, 0.5 ulp | 3.97 gso/s, 0.5 ulp | 4.38 gso/s, 0.5 ulp |
347
+ | `nk_angulars_packed_u8_v128relaxed` | 23.7 gso/s, 0.3 ulp | 25.1 gso/s, 0.3 ulp | 25.8 gso/s, 0.3 ulp |
348
+ | `nk_angulars_symmetric_u8_v128relaxed` | 19.6 gso/s, 0.3 ulp | 23.2 gso/s, 0.3 ulp | 24.1 gso/s, 0.3 ulp |
349
+ | `nk_euclideans_packed_u8_v128relaxed` | 23.8 gso/s, 0 ulp | 25.3 gso/s, 0 ulp | 25.8 gso/s, 0 ulp |
350
+ | `nk_euclideans_symmetric_u8_v128relaxed` | 19.5 gso/s, 0 ulp | 23.0 gso/s, 0 ulp | 24.6 gso/s, 0 ulp |
351
+ | __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
352
+ | `nk_angulars_packed_i4_serial` | 6.22 gso/s, 0.35 ulp | 6.41 gso/s, 0.34 ulp | 6.55 gso/s, 0.35 ulp |
353
+ | `nk_angulars_symmetric_i4_serial` | 2.64 gso/s, 0.34 ulp | 3.69 gso/s, 0.34 ulp | 4.18 gso/s, 0.34 ulp |
354
+ | `nk_euclideans_packed_i4_serial` | 6.00 gso/s, 0.49 ulp | 6.43 gso/s, 0.54 ulp | 6.56 gso/s, 0.64 ulp |
355
+ | `nk_euclideans_symmetric_i4_serial` | 2.61 gso/s, 0.48 ulp | 3.68 gso/s, 0.53 ulp | 4.14 gso/s, 0.63 ulp |
356
+ | __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
357
+ | `nk_angulars_packed_u4_serial` | 5.38 gso/s, 0.35 ulp | 5.60 gso/s, 0.34 ulp | 5.81 gso/s, 0.35 ulp |
358
+ | `nk_angulars_symmetric_u4_serial` | 2.90 gso/s, 0.34 ulp | 4.28 gso/s, 0.34 ulp | 4.90 gso/s, 0.34 ulp |
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
+ | `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
+
362
+ ### Apple M5
363
+
364
+ #### Native
365
+
366
+ | Kernel | 256³ | 1024³ | 4096³ |
367
+ | :--------------------------------------- | -----------------------: | -----------------------: | -----------------------: |
368
+ | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+
521
+ #### WASM
522
+
523
+ Measured with Wasmtime v43 (Cranelift backend).
524
+
525
+ | Kernel | 256³ | 1024³ | 4096³ |
526
+ | :----------------------------------------- | -----------------------: | -----------------------: | -----------------------: |
527
+ | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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
+ | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
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 |