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 +1,465 @@
1
+ # Element-Wise Arithmetic in NumKong
2
+
3
+ NumKong implements element-wise vector arithmetic: addition, scaling, blending, and fused multiply-add across all supported numeric types.
4
+ Each operation reads one to three input vectors and writes one output vector of the same length, with scalar coefficients $\alpha$ and $\beta$ controlling linear combinations.
5
+ Mixed-precision workflows use narrower input types (Float16, BFloat16, Float8) with Float32 intermediate computation and narrowed output.
6
+
7
+ Sum (addition):
8
+
9
+ $$
10
+ \text{result}_i = a_i + b_i
11
+ $$
12
+
13
+ Scale:
14
+
15
+ $$
16
+ \text{result}_i = \alpha \cdot a_i + \beta
17
+ $$
18
+
19
+ Blend:
20
+
21
+ $$
22
+ \text{result}_i = \alpha \cdot a_i + \beta \cdot b_i
23
+ $$
24
+
25
+ Fused multiply-add:
26
+
27
+ $$
28
+ \text{result}_i = \alpha \cdot a_i \cdot b_i + \beta \cdot c_i
29
+ $$
30
+
31
+ Reformulating as Python pseudocode:
32
+
33
+ ```python
34
+ import numpy as np
35
+
36
+ def fma(a: np.ndarray, b: np.ndarray, c: np.ndarray,
37
+ alpha: float = 1.0, beta: float = 1.0) -> np.ndarray:
38
+ return alpha * a * b + beta * c
39
+
40
+ def scale(a: np.ndarray, alpha: float = 1.0, beta: float = 0.0) -> np.ndarray:
41
+ return alpha * a + beta
42
+
43
+ def add(a: np.ndarray, b: np.ndarray) -> np.ndarray:
44
+ return a + b
45
+
46
+ def blend(a: np.ndarray, b: np.ndarray,
47
+ alpha: float = 1.0, beta: float = 1.0) -> np.ndarray:
48
+ return alpha * a + beta * b
49
+ ```
50
+
51
+ ## Input & Output Types
52
+
53
+ Real and integer element-wise operations:
54
+
55
+ | Input Type | Output Type | Description |
56
+ | ---------- | ----------- | ----------------------------------------- |
57
+ | `f64` | `f64` | 64-bit IEEE 754 double precision |
58
+ | `f32` | `f32` | 32-bit IEEE 754 single precision |
59
+ | `f16` | `f16` | 16-bit IEEE 754 half precision |
60
+ | `bf16` | `bf16` | 16-bit brain float |
61
+ | `e4m3` | `e4m3` | 8-bit Float8: 4 exponent, 3 mantissa bits |
62
+ | `e5m2` | `e5m2` | 8-bit Float8: 5 exponent, 2 mantissa bits |
63
+ | `i8` | `i8` | 8-bit signed integers, saturating |
64
+ | `u8` | `u8` | 8-bit unsigned integers, saturating |
65
+ | `i16` | `i16` | 16-bit signed integers |
66
+ | `u16` | `u16` | 16-bit unsigned integers |
67
+ | `i32` | `i32` | 32-bit signed integers |
68
+ | `u32` | `u32` | 32-bit unsigned integers |
69
+ | `i64` | `i64` | 64-bit signed integers |
70
+ | `u64` | `u64` | 64-bit unsigned integers |
71
+
72
+ Complex element-wise operations:
73
+
74
+ | Input Type | Output Type | Description |
75
+ | ---------- | ----------- | -------------------- |
76
+ | `f64c` | `f64c` | 64-bit complex pairs |
77
+ | `f32c` | `f32c` | 32-bit complex pairs |
78
+
79
+ ## Optimizations
80
+
81
+ ### Widening-Narrowing Pipeline for Sub-32-bit Types
82
+
83
+ `nk_each_fma_f16_haswell`, `nk_each_blend_bf16_neonbfdot`, `nk_each_scale_e4m3_haswell` widen inputs to Float32 before arithmetic, then narrow the result back to the original type.
84
+ The widen-compute-narrow pipeline costs 2 extra conversion instructions per element but guarantees Float32-precision intermediate results — critical for FMA where naive Float16 multiplication would lose 5+ bits of mantissa.
85
+ Haswell processes 8 Float16 elements per cycle: `VCVTPH2PS` (widen) -> `VFMADD231PS` (FMA) -> `VCVTPS2PH` (narrow), fully pipelined across 3 execution ports.
86
+
87
+ ### Saturating Integer Arithmetic
88
+
89
+ `nk_each_sum_i8_haswell`, `nk_each_sum_u8_neonhalf` use saturating addition — clamping to type bounds instead of wrapping on overflow.
90
+ Haswell uses `VPADDSB` / `VPADDUSB` for signed/unsigned 8-bit saturation in a single instruction (32 elements per cycle at YMM width).
91
+ Serial fallback implements saturation via branch-free min/max: `result = min(max(a + b, TYPE_MIN), TYPE_MAX)` with overflow detection through sign-bit comparison.
92
+
93
+ ### Complex Number Layout
94
+
95
+ `nk_each_fma_f32c_serial`, `nk_each_blend_f64c_serial` operate on interleaved real/imaginary pairs: `[re0, im0, re1, im1, ...]`.
96
+ Addition and scaling treat complex vectors as 2N-length real vectors — no special handling needed.
97
+ FMA requires cross-lane operations for the imaginary part: `re(a*b) = re(a)*re(b) - im(a)*im(b)`, implemented via `VFMADDSUB231PS` which alternates add/subtract across even/odd lanes.
98
+
99
+ ## Performance
100
+
101
+ The following performance tables are produced by manually re-running `nk_test` and `nk_bench` included internal tools to measure both accuracy and throughput at different input shapes.
102
+ The input size is controlled by the `NK_DENSE_DIMENSIONS` environment variable and set to 256, 1024, and 4096 elements.
103
+ The throughput is measured in GB/s as the number of input bytes read per second.
104
+ Accuracy is reported as mean ULP (units in last place) unless noted otherwise — the average number of representable floating-point values between the result and the exact answer.
105
+ Rows marked `🧩` use external BLAS baselines rather than NumKong kernels.
106
+ Each kernel runs for at least 20 seconds per configuration.
107
+ Benchmark threads are pinned to specific cores; on machines with heterogeneous core types (e.g., Apple P/E cores), only the fastest cores are used.
108
+ Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run in separate passes to avoid affecting throughput measurements of other kernels.
109
+
110
+ ### Intel Sapphire Rapids
111
+
112
+ #### Native
113
+
114
+ | Kernel | 256 | 1024 | 4096 |
115
+ | :--------------------------- | -----------------------: | -----------------------: | -----------------------: |
116
+ | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
117
+ | `sum_f64_with_blas` 🧩 | 12.1 gb/s | 8.63 gb/s | 6.83 gb/s |
118
+ | `each_blend_f64_with_blas` 🧩 | 11.3 gb/s | 8.80 gb/s | 6.71 gb/s |
119
+ | `nk_each_sum_f64_serial` | 16.7 gb/s, 0 ulp | 17.2 gb/s, 0 ulp | 10.9 gb/s, 0 ulp |
120
+ | `nk_each_sum_f64_haswell` | 14.7 gb/s, 0 ulp | 10.2 gb/s, 0 ulp | 7.99 gb/s, 0 ulp |
121
+ | `nk_each_sum_f64_skylake` | 16.4 gb/s, 0 ulp | 16.7 gb/s, 0 ulp | 8.59 gb/s, 0 ulp |
122
+ | `nk_each_scale_f64_serial` | 11.4 gb/s, 0 ulp | 12.5 gb/s, 0 ulp | 7.99 gb/s, 0 ulp |
123
+ | `nk_each_scale_f64_haswell` | 9.61 gb/s, 0 ulp | 9.22 gb/s, 0 ulp | 5.08 gb/s, 0 ulp |
124
+ | `nk_each_scale_f64_skylake` | 11.2 gb/s, 0 ulp | 11.9 gb/s, 0 ulp | 6.30 gb/s, 0 ulp |
125
+ | `nk_each_blend_f64_serial` | 16.4 gb/s, 1.4 ulp | 16.4 gb/s, 1.1 ulp | 11.8 gb/s, 1.1 ulp |
126
+ | `nk_each_blend_f64_haswell` | 13.4 gb/s, 1.5 ulp | 11.2 gb/s, 1.5 ulp | 7.85 gb/s, 1.1 ulp |
127
+ | `nk_each_blend_f64_skylake` | 16.5 gb/s, 1.7 ulp | 15.9 gb/s, 1.5 ulp | 8.55 gb/s, 1.1 ulp |
128
+ | `nk_each_fma_f64_serial` | 20.1 gb/s, 1.5 ulp | 20.8 gb/s, 1.5 ulp | 11.7 gb/s, 1.3 ulp |
129
+ | `nk_each_fma_f64_haswell` | 16.8 gb/s, 1.5 ulp | 11.1 gb/s, 1.5 ulp | 9.42 gb/s, 2.8 ulp |
130
+ | `nk_each_fma_f64_skylake` | 19.8 gb/s, 1.4 ulp | 20.4 gb/s, 1.5 ulp | 11.6 gb/s, 2.7 ulp |
131
+ | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
132
+ | `sum_f32_with_blas` 🧩 | 11.6 gb/s | 12.8 gb/s | 7.57 gb/s |
133
+ | `each_blend_f32_with_blas` 🧩 | 10.3 gb/s | 9.60 gb/s | 6.69 gb/s |
134
+ | `nk_each_sum_f32_serial` | 16.9 gb/s, 0 ulp | 18.8 gb/s, 0 ulp | 15.7 gb/s, 0 ulp |
135
+ | `nk_each_sum_f32_haswell` | 14.6 gb/s, 0 ulp | 14.3 gb/s, 0 ulp | 8.49 gb/s, 0 ulp |
136
+ | `nk_each_sum_f32_skylake` | 16.7 gb/s, 0 ulp | 17.9 gb/s, 0 ulp | 16.5 gb/s, 0 ulp |
137
+ | `nk_each_scale_f32_serial` | 10.8 gb/s, 0 ulp | 13.0 gb/s, 0 ulp | 12.1 gb/s, 0 ulp |
138
+ | `nk_each_scale_f32_haswell` | 9.26 gb/s, 0 ulp | 10.2 gb/s, 0 ulp | 6.22 gb/s, 0 ulp |
139
+ | `nk_each_scale_f32_skylake` | 12.3 gb/s, 0 ulp | 11.9 gb/s, 0 ulp | 12.9 gb/s, 0 ulp |
140
+ | `nk_each_blend_f32_serial` | 16.2 gb/s, 351 ulp | 18.7 gb/s, 2.0 ulp | 17.4 gb/s, 1.4 ulp |
141
+ | `nk_each_blend_f32_haswell` | 14.5 gb/s, 2.3 ulp | 14.2 gb/s, 2.1 ulp | 8.16 gb/s, 1.3 ulp |
142
+ | `nk_each_blend_f32_skylake` | 15.7 gb/s, 1.9 ulp | 17.2 gb/s, 1.8 ulp | 15.9 gb/s, 1.3 ulp |
143
+ | `nk_each_fma_f32_serial` | 20.1 gb/s, 1.4 ulp | 18.7 gb/s, 2.1 ulp | 19.0 gb/s, 1.6 ulp |
144
+ | `nk_each_fma_f32_haswell` | 18.4 gb/s, 1.4 ulp | 15.3 gb/s, 1.8 ulp | 9.52 gb/s, 1.5 ulp |
145
+ | `nk_each_fma_f32_skylake` | 20.8 gb/s, 1.4 ulp | 19.3 gb/s, 1.7 ulp | 16.5 gb/s, 1.5 ulp |
146
+ | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
147
+ | `nk_each_sum_bf16_serial` | 0.175 gb/s, 0 ulp | 0.178 gb/s, 0 ulp | 0.173 gb/s, 0 ulp |
148
+ | `nk_each_sum_bf16_haswell` | 7.97 gb/s, 0 ulp | 9.33 gb/s, 0 ulp | 10.3 gb/s, 0 ulp |
149
+ | `nk_each_sum_bf16_skylake` | 11.4 gb/s, 0 ulp | 11.2 gb/s, 0 ulp | 14.1 gb/s, 0 ulp |
150
+ | `nk_each_scale_bf16_serial` | 0.128 gb/s, 0 ulp | 0.119 gb/s, 0 ulp | 0.132 gb/s, 0 ulp |
151
+ | `nk_each_scale_bf16_haswell` | 6.08 gb/s, 0 ulp | 6.55 gb/s, 0 ulp | 6.92 gb/s, 0 ulp |
152
+ | `nk_each_scale_bf16_skylake` | 7.43 gb/s, 0 ulp | 8.04 gb/s, 0 ulp | 8.45 gb/s, 0 ulp |
153
+ | `nk_each_blend_bf16_serial` | 0.211 gb/s, 0 ulp | 0.204 gb/s, 0 ulp | 0.224 gb/s, 0 ulp |
154
+ | `nk_each_blend_bf16_haswell` | 8.58 gb/s, 2.2 ulp | 9.44 gb/s, 1.5 ulp | 10.2 gb/s, 1.5 ulp |
155
+ | `nk_each_blend_bf16_skylake` | 10.3 gb/s, 2.3 ulp | 11.9 gb/s, 1.3 ulp | 13.4 gb/s, 1.5 ulp |
156
+ | `nk_each_fma_bf16_serial` | 0.264 gb/s, 0 ulp | 0.260 gb/s, 0 ulp | 0.256 gb/s, 0 ulp |
157
+ | `nk_each_fma_bf16_haswell` | 10.9 gb/s, 1.5 ulp | 10.3 gb/s, 0.9 ulp | 11.4 gb/s, 1.0 ulp |
158
+ | `nk_each_fma_bf16_skylake` | 14.1 gb/s, 1.2 ulp | 13.0 gb/s, 0.7 ulp | 15.8 gb/s, 1.1 ulp |
159
+ | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
160
+ | `nk_each_sum_f16_serial` | 33.7 gb/s, 0 ulp | 16.1 gb/s, 0 ulp | 18.8 gb/s, 0 ulp |
161
+ | `nk_each_sum_f16_haswell` | 14.4 gb/s, 0 ulp | 11.8 gb/s, 0 ulp | 9.84 gb/s, 0 ulp |
162
+ | `nk_each_sum_f16_sapphire` | 39.2 gb/s, 0 ulp | 17.0 gb/s, 0 ulp | 18.9 gb/s, 0 ulp |
163
+ | `nk_each_scale_f16_serial` | 0.423 gb/s, 0 ulp | 0.282 gb/s, 0 ulp | 0.409 gb/s, 0 ulp |
164
+ | `nk_each_scale_f16_haswell` | 8.92 gb/s, 0 ulp | 8.59 gb/s, 0 ulp | 8.15 gb/s, 0 ulp |
165
+ | `nk_each_scale_f16_skylake` | 17.0 gb/s, 0 ulp | 10.7 gb/s, 0 ulp | 12.1 gb/s, 0 ulp |
166
+ | `nk_each_blend_f16_serial` | 0.769 gb/s, 1.3 ulp | 0.669 gb/s, 1.6 ulp | 0.792 gb/s, 1.5 ulp |
167
+ | `nk_each_blend_f16_haswell` | 13.5 gb/s, 1.2 ulp | 11.0 gb/s, 1.4 ulp | 11.7 gb/s, 1.5 ulp |
168
+ | `nk_each_blend_f16_skylake` | 16.9 gb/s, 1.3 ulp | 13.9 gb/s, 1.4 ulp | 14.2 gb/s, 1.1 ulp |
169
+ | `nk_each_fma_f16_serial` | 0.965 gb/s, 1.0 ulp | 0.787 gb/s, 1.1 ulp | 0.952 gb/s, 1.2 ulp |
170
+ | `nk_each_fma_f16_haswell` | 15.2 gb/s, 1.4 ulp | 13.6 gb/s, 1.0 ulp | 15.7 gb/s, 1.1 ulp |
171
+ | `nk_each_fma_f16_skylake` | 16.3 gb/s, 1.3 ulp | 16.2 gb/s, 1.3 ulp | 15.3 gb/s, 1.1 ulp |
172
+ | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
173
+ | `nk_each_sum_e4m3_serial` | 0.0966 gb/s, 0 ulp | 0.0943 gb/s, 0 ulp | 0.0946 gb/s, 0 ulp |
174
+ | `nk_each_sum_e4m3_haswell` | 0.895 gb/s, 0 ulp | 0.772 gb/s, 0 ulp | 0.824 gb/s, 0 ulp |
175
+ | `nk_each_sum_e4m3_skylake` | 1.48 gb/s, 0 ulp | 1.34 gb/s, 0 ulp | 1.40 gb/s, 0 ulp |
176
+ | `nk_each_sum_e4m3_sapphire` | 2.12 gb/s, 0 ulp | 1.89 gb/s, 0 ulp | 2.13 gb/s, 0 ulp |
177
+ | `nk_each_scale_e4m3_serial` | 0.0550 gb/s, 0 ulp | 0.0543 gb/s, 0 ulp | 0.0570 gb/s, 0 ulp |
178
+ | `nk_each_scale_e4m3_haswell` | 0.495 gb/s, 0 ulp | 0.532 gb/s, 0 ulp | 0.540 gb/s, 0 ulp |
179
+ | `nk_each_scale_e4m3_skylake` | 1.05 gb/s, 0 ulp | 1.02 gb/s, 0 ulp | 1.10 gb/s, 0 ulp |
180
+ | `nk_each_blend_e4m3_serial` | 0.0889 gb/s, 0 ulp | 0.0927 gb/s, 0 ulp | 0.0876 gb/s, 0 ulp |
181
+ | `nk_each_blend_e4m3_haswell` | 0.807 gb/s, 0.6 ulp | 0.756 gb/s, 0 ulp | 0.789 gb/s, 0 ulp |
182
+ | `nk_each_blend_e4m3_skylake` | 1.50 gb/s, 0 ulp | 1.44 gb/s, 0 ulp | 1.48 gb/s, 0 ulp |
183
+ | `nk_each_fma_e4m3_serial` | 0.120 gb/s, 0 ulp | 0.118 gb/s, 0.9 ulp | 0.115 gb/s, 0 ulp |
184
+ | `nk_each_fma_e4m3_haswell` | 0.989 gb/s, 0 ulp | 0.909 gb/s, 0 ulp | 0.967 gb/s, 0.5 ulp |
185
+ | `nk_each_fma_e4m3_skylake` | 1.89 gb/s, 0 ulp | 1.74 gb/s, 0 ulp | 1.80 gb/s, 0 ulp |
186
+ | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
187
+ | `nk_each_sum_e5m2_serial` | 0.115 gb/s, 0 ulp | 0.114 gb/s, 0 ulp | 0.110 gb/s, 0 ulp |
188
+ | `nk_each_sum_e5m2_haswell` | 1.04 gb/s, 0 ulp | 0.987 gb/s, 0 ulp | 0.972 gb/s, 0 ulp |
189
+ | `nk_each_sum_e5m2_skylake` | 1.71 gb/s, 0 ulp | 1.76 gb/s, 0 ulp | 1.79 gb/s, 0 ulp |
190
+ | `nk_each_scale_e5m2_serial` | 0.0593 gb/s, 0 ulp | 0.0630 gb/s, 0 ulp | 0.0630 gb/s, 0 ulp |
191
+ | `nk_each_scale_e5m2_haswell` | 0.601 gb/s, 0 ulp | 0.611 gb/s, 0 ulp | 0.588 gb/s, 0 ulp |
192
+ | `nk_each_scale_e5m2_skylake` | 1.11 gb/s, 0 ulp | 1.11 gb/s, 0 ulp | 1.12 gb/s, 0 ulp |
193
+ | `nk_each_blend_e5m2_serial` | 0.108 gb/s, 0 ulp | 0.113 gb/s, 0 ulp | 0.114 gb/s, 50 ulp |
194
+ | `nk_each_blend_e5m2_haswell` | 0.999 gb/s, 0 ulp | 0.895 gb/s, 0 ulp | 0.951 gb/s, 0 ulp |
195
+ | `nk_each_blend_e5m2_skylake` | 1.77 gb/s, 0 ulp | 1.65 gb/s, 0 ulp | 1.72 gb/s, 0 ulp |
196
+ | `nk_each_fma_e5m2_serial` | 0.155 gb/s, 5.1 ulp | 0.146 gb/s, 0 ulp | 0.149 gb/s, 0 ulp |
197
+ | `nk_each_fma_e5m2_haswell` | 1.24 gb/s, 0 ulp | 1.19 gb/s, 0 ulp | 1.25 gb/s, 0 ulp |
198
+ | `nk_each_fma_e5m2_skylake` | 2.36 gb/s, 0 ulp | 2.01 gb/s, 0 ulp | 2.11 gb/s, 0 ulp |
199
+ | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
200
+ | `nk_each_sum_e2m3_serial` | 0.109 gb/s, 0 ulp | 0.105 gb/s, 0 ulp | 0.110 gb/s, 0 ulp |
201
+ | `nk_each_scale_e2m3_serial` | 0.0500 gb/s, 0 ulp | 0.0474 gb/s, 0 ulp | 0.0495 gb/s, 0 ulp |
202
+ | `nk_each_blend_e2m3_serial` | 0.0864 gb/s, 0 ulp | 0.0888 gb/s, 0 ulp | 0.0908 gb/s, 0 ulp |
203
+ | `nk_each_fma_e2m3_serial` | 0.133 gb/s, 0 ulp | 0.128 gb/s, 0 ulp | 0.128 gb/s, 0 ulp |
204
+ | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
205
+ | `nk_each_sum_e3m2_serial` | 0.118 gb/s, 0 ulp | 0.115 gb/s, 0 ulp | 0.106 gb/s, 0 ulp |
206
+ | `nk_each_scale_e3m2_serial` | 0.0574 gb/s, 0 ulp | 0.0555 gb/s, 0 ulp | 0.0548 gb/s, 0 ulp |
207
+ | `nk_each_blend_e3m2_serial` | 0.110 gb/s, 0 ulp | 0.105 gb/s, 0 ulp | 0.0956 gb/s, 0 ulp |
208
+ | `nk_each_fma_e3m2_serial` | 0.151 gb/s, 0 ulp | 0.149 gb/s, 0 ulp | 0.138 gb/s, 0 ulp |
209
+ | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
210
+ | `nk_each_sum_i8_serial` | 21.7 gb/s | 15.9 gb/s | 12.5 gb/s |
211
+ | `nk_each_sum_i8_haswell` | 37.7 gb/s | 15.7 gb/s | 16.5 gb/s |
212
+ | `nk_each_sum_i8_icelake` | 47.2 gb/s | 17.7 gb/s | 17.7 gb/s |
213
+ | `nk_each_scale_i8_serial` | 2.34 gb/s | 1.50 gb/s | 1.70 gb/s |
214
+ | `nk_each_scale_i8_haswell` | 3.91 gb/s | 3.93 gb/s | 3.64 gb/s |
215
+ | `nk_each_scale_i8_skylake` | 6.74 gb/s | 6.70 gb/s | 6.93 gb/s |
216
+ | `nk_each_scale_i8_sapphire` | 23.0 gb/s | 11.4 gb/s | 10.8 gb/s |
217
+ | `nk_each_blend_i8_serial` | 3.66 gb/s | 2.23 gb/s | 2.60 gb/s |
218
+ | `nk_each_blend_i8_haswell` | 5.95 gb/s | 5.37 gb/s | 6.37 gb/s |
219
+ | `nk_each_blend_i8_sapphire` | 32.4 gb/s | 17.7 gb/s | 15.2 gb/s |
220
+ | `nk_each_fma_i8_serial` | 4.49 gb/s | 2.63 gb/s | 2.98 gb/s |
221
+ | `nk_each_fma_i8_haswell` | 7.36 gb/s | 6.84 gb/s | 7.15 gb/s |
222
+ | `nk_each_fma_i8_skylake` | 11.2 gb/s | 9.45 gb/s | 10.1 gb/s |
223
+ | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
224
+ | `nk_each_sum_u8_serial` | 17.0 gb/s | 13.7 gb/s | 12.5 gb/s |
225
+ | `nk_each_sum_u8_haswell` | 42.6 gb/s | 15.7 gb/s | 15.4 gb/s |
226
+ | `nk_each_sum_u8_icelake` | 45.9 gb/s | 17.3 gb/s | 17.4 gb/s |
227
+ | `nk_each_scale_u8_serial` | 2.11 gb/s | 2.03 gb/s | 1.86 gb/s |
228
+ | `nk_each_scale_u8_haswell` | 3.91 gb/s | 3.89 gb/s | 4.27 gb/s |
229
+ | `nk_each_scale_u8_skylake` | 6.93 gb/s | 5.99 gb/s | 6.70 gb/s |
230
+ | `nk_each_scale_u8_sapphire` | 24.7 gb/s | 12.1 gb/s | 11.9 gb/s |
231
+ | `nk_each_blend_u8_serial` | 3.23 gb/s | 2.62 gb/s | 3.43 gb/s |
232
+ | `nk_each_blend_u8_haswell` | 4.87 gb/s | 5.10 gb/s | 5.61 gb/s |
233
+ | `nk_each_blend_u8_sapphire` | 39.8 gb/s | 18.1 gb/s | 16.5 gb/s |
234
+ | `nk_each_fma_u8_serial` | 3.19 gb/s | 3.92 gb/s | 4.54 gb/s |
235
+ | `nk_each_fma_u8_haswell` | 6.98 gb/s | 6.29 gb/s | 7.62 gb/s |
236
+ | `nk_each_fma_u8_skylake` | 9.66 gb/s | 9.21 gb/s | 10.3 gb/s |
237
+ | __i16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
238
+ | `nk_each_sum_i16_serial` | 12.0 gb/s | 11.6 gb/s | 14.2 gb/s |
239
+ | `nk_each_sum_i16_haswell` | 26.1 gb/s | 16.2 gb/s | 16.8 gb/s |
240
+ | `nk_each_sum_i16_icelake` | 39.3 gb/s | 18.2 gb/s | 18.0 gb/s |
241
+ | `nk_each_scale_i16_serial` | 3.19 gb/s | 4.25 gb/s | 3.77 gb/s |
242
+ | `nk_each_scale_i16_haswell` | 7.07 gb/s | 7.42 gb/s | 7.69 gb/s |
243
+ | `nk_each_scale_i16_skylake` | 12.8 gb/s | 8.94 gb/s | 10.5 gb/s |
244
+ | `nk_each_blend_i16_serial` | 6.00 gb/s | 6.04 gb/s | 6.37 gb/s |
245
+ | `nk_each_fma_i16_serial` | 8.37 gb/s | 7.64 gb/s | 7.07 gb/s |
246
+ | `nk_each_fma_i16_haswell` | 10.7 gb/s | 12.5 gb/s | 13.4 gb/s |
247
+ | `nk_each_fma_i16_skylake` | 18.3 gb/s | 15.7 gb/s | 19.0 gb/s |
248
+ | __u16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
249
+ | `nk_each_sum_u16_serial` | 12.2 gb/s | 11.8 gb/s | 13.7 gb/s |
250
+ | `nk_each_sum_u16_haswell` | 24.1 gb/s | 14.4 gb/s | 17.7 gb/s |
251
+ | `nk_each_sum_u16_icelake` | 39.7 gb/s | 18.2 gb/s | 18.4 gb/s |
252
+ | `nk_each_scale_u16_serial` | 4.44 gb/s | 5.28 gb/s | 5.48 gb/s |
253
+ | `nk_each_scale_u16_haswell` | 7.82 gb/s | 7.21 gb/s | 7.66 gb/s |
254
+ | `nk_each_scale_u16_skylake` | 15.5 gb/s | 11.6 gb/s | 9.04 gb/s |
255
+ | `nk_each_blend_u16_serial` | 7.84 gb/s | 7.08 gb/s | 9.15 gb/s |
256
+ | `nk_each_fma_u16_serial` | 9.01 gb/s | 8.19 gb/s | 9.37 gb/s |
257
+ | `nk_each_fma_u16_haswell` | 11.2 gb/s | 12.2 gb/s | 13.9 gb/s |
258
+ | `nk_each_fma_u16_skylake` | 22.6 gb/s | 20.8 gb/s | 17.5 gb/s |
259
+ | __i32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
260
+ | `nk_each_sum_i32_serial` | 12.6 gb/s | 12.7 gb/s | 10.1 gb/s |
261
+ | `nk_each_sum_i32_haswell` | 14.6 gb/s | 15.3 gb/s | 13.9 gb/s |
262
+ | `nk_each_sum_i32_icelake` | 17.1 gb/s | 18.5 gb/s | 15.4 gb/s |
263
+ | `nk_each_scale_i32_serial` | 3.04 gb/s | 3.28 gb/s | 3.44 gb/s |
264
+ | `nk_each_scale_i32_haswell` | 8.94 gb/s | 8.72 gb/s | 8.09 gb/s |
265
+ | `nk_each_scale_i32_skylake` | 11.1 gb/s | 12.6 gb/s | 11.3 gb/s |
266
+ | `nk_each_blend_i32_serial` | 4.99 gb/s | 5.99 gb/s | 5.06 gb/s |
267
+ | `nk_each_fma_i32_serial` | 6.33 gb/s | 6.38 gb/s | 6.40 gb/s |
268
+ | `nk_each_fma_i32_haswell` | 13.5 gb/s | 16.4 gb/s | 11.2 gb/s |
269
+ | `nk_each_fma_i32_skylake` | 20.1 gb/s | 21.3 gb/s | 16.0 gb/s |
270
+ | __u32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
271
+ | `nk_each_sum_u32_serial` | 13.6 gb/s | 11.6 gb/s | 9.94 gb/s |
272
+ | `nk_each_sum_u32_haswell` | 16.0 gb/s | 17.7 gb/s | 12.7 gb/s |
273
+ | `nk_each_sum_u32_icelake` | 17.4 gb/s | 19.0 gb/s | 14.8 gb/s |
274
+ | `nk_each_scale_u32_serial` | 2.13 gb/s | 3.22 gb/s | 2.74 gb/s |
275
+ | `nk_each_scale_u32_haswell` | 8.20 gb/s | 9.40 gb/s | 9.18 gb/s |
276
+ | `nk_each_scale_u32_skylake` | 10.6 gb/s | 12.9 gb/s | 11.1 gb/s |
277
+ | `nk_each_blend_u32_serial` | 3.76 gb/s | 5.22 gb/s | 5.78 gb/s |
278
+ | `nk_each_fma_u32_serial` | 4.78 gb/s | 5.63 gb/s | 8.43 gb/s |
279
+ | `nk_each_fma_u32_haswell` | 13.7 gb/s | 16.1 gb/s | 12.1 gb/s |
280
+ | `nk_each_fma_u32_skylake` | 20.2 gb/s | 21.1 gb/s | 15.2 gb/s |
281
+ | __i64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
282
+ | `nk_each_sum_i64_serial` | 15.9 gb/s | 15.8 gb/s | 10.3 gb/s |
283
+ | `nk_each_sum_i64_icelake` | 17.7 gb/s | 19.2 gb/s | 11.1 gb/s |
284
+ | `nk_each_scale_i64_serial` | 7.44 gb/s | 9.17 gb/s | 8.72 gb/s |
285
+ | `nk_each_scale_i64_skylake` | 11.8 gb/s | 13.8 gb/s | 9.10 gb/s |
286
+ | `nk_each_blend_i64_serial` | 10.9 gb/s | 14.3 gb/s | 10.5 gb/s |
287
+ | `nk_each_fma_i64_serial` | 13.5 gb/s | 19.0 gb/s | 11.8 gb/s |
288
+ | `nk_each_fma_i64_skylake` | 21.7 gb/s | 22.2 gb/s | 11.8 gb/s |
289
+ | __u64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
290
+ | `nk_each_sum_u64_serial` | 14.7 gb/s | 16.9 gb/s | 10.5 gb/s |
291
+ | `nk_each_sum_u64_icelake` | 18.1 gb/s | 19.2 gb/s | 12.4 gb/s |
292
+ | `nk_each_scale_u64_serial` | 7.99 gb/s | 9.79 gb/s | 8.82 gb/s |
293
+ | `nk_each_scale_u64_skylake` | 11.6 gb/s | 13.9 gb/s | 7.24 gb/s |
294
+ | `nk_each_blend_u64_serial` | 11.9 gb/s | 16.5 gb/s | 13.8 gb/s |
295
+ | `nk_each_fma_u64_serial` | 15.3 gb/s | 21.6 gb/s | 14.0 gb/s |
296
+ | `nk_each_fma_u64_skylake` | 21.6 gb/s | 21.8 gb/s | 11.4 gb/s |
297
+ | __f64c__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
298
+ | `nk_each_scale_f64c_serial` | 10.6 gb/s, 3.4 ulp | 7.04 gb/s, 2.6 ulp | 5.63 gb/s, 2.0 ulp |
299
+ | `nk_each_scale_f64c_haswell` | 9.89 gb/s, 3.5 ulp | 6.45 gb/s, 2.0 ulp | 5.90 gb/s, 2.1 ulp |
300
+ | `nk_each_scale_f64c_skylake` | 9.29 gb/s, 3.5 ulp | 6.01 gb/s, 2.0 ulp | 5.67 gb/s, 2.0 ulp |
301
+ | `nk_each_blend_f64c_serial` | 16.2 gb/s, 2.5 ulp | 8.80 gb/s, 2.4 ulp | 8.23 gb/s, 2.6 ulp |
302
+ | `nk_each_blend_f64c_haswell` | 14.0 gb/s, 2.5 ulp | 8.26 gb/s, 2.4 ulp | 8.59 gb/s, 2.7 ulp |
303
+ | `nk_each_blend_f64c_skylake` | 14.3 gb/s, 2.5 ulp | 8.82 gb/s, 2.7 ulp | 7.33 gb/s, 2.7 ulp |
304
+ | `nk_each_fma_f64c_serial` | 17.4 gb/s, 4.5 ulp | 10.1 gb/s, 3.2 ulp | 9.45 gb/s, 2.8 ulp |
305
+ | `nk_each_fma_f64c_haswell` | 14.5 gb/s, 4.9 ulp | 8.86 gb/s, 3.1 ulp | 9.31 gb/s, 2.7 ulp |
306
+ | `nk_each_fma_f64c_skylake` | 16.1 gb/s, 4.0 ulp | 8.93 gb/s, 3.4 ulp | 10.4 gb/s, 2.8 ulp |
307
+ | __f32c__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
308
+ | `nk_each_scale_f32c_serial` | 10.1 gb/s, 2.1 ulp | 9.28 gb/s, 1.8 ulp | 6.20 gb/s, 2.1 ulp |
309
+ | `nk_each_scale_f32c_haswell` | 8.81 gb/s, 2.1 ulp | 8.87 gb/s, 1.8 ulp | 7.73 gb/s, 2.1 ulp |
310
+ | `nk_each_scale_f32c_skylake` | 9.67 gb/s, 2.2 ulp | 7.61 gb/s, 1.8 ulp | 7.09 gb/s, 2.1 ulp |
311
+ | `nk_each_blend_f32c_serial` | 14.6 gb/s, 6.9 ulp | 10.8 gb/s, 2.8 ulp | 8.43 gb/s, 8.7 ulp |
312
+ | `nk_each_blend_f32c_haswell` | 13.0 gb/s, 8.7 ulp | 12.1 gb/s, 2.9 ulp | 10.5 gb/s, 9.7 ulp |
313
+ | `nk_each_blend_f32c_skylake` | 14.0 gb/s, 7.7 ulp | 9.28 gb/s, 2.7 ulp | 9.47 gb/s, 10.2 ulp |
314
+ | `nk_each_fma_f32c_serial` | 15.7 gb/s, 8.8 ulp | 11.96 gb/s, 3.9 ulp | 9.33 gb/s, 8.5 ulp |
315
+ | `nk_each_fma_f32c_haswell` | 14.1 gb/s, 6.6 ulp | 10.3 gb/s, 2.9 ulp | 9.77 gb/s, 7.2 ulp |
316
+ | `nk_each_fma_f32c_skylake` | 15.5 gb/s, 9.2 ulp | 10.3 gb/s, 3.8 ulp | 9.95 gb/s, 8.3 ulp |
317
+
318
+ ### Apple M5
319
+
320
+ #### Native
321
+
322
+ | Kernel | 256 | 1024 | 4096 |
323
+ | :----------------------------- | -----------------------: | -----------------------: | -----------------------: |
324
+ | __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
325
+ | `nk_each_sum_f64_serial` | 54.7 gb/s, 0 ulp | 66.7 gb/s, 0 ulp | 51.8 gb/s, 0 ulp |
326
+ | `nk_each_sum_f64_neon` | 53.8 gb/s, 0 ulp | 71.3 gb/s, 0 ulp | 56.0 gb/s, 0 ulp |
327
+ | `nk_each_scale_f64_serial` | 46.4 gb/s, 0 ulp | 58.2 gb/s, 0 ulp | 48.2 gb/s, 0 ulp |
328
+ | `nk_each_scale_f64_neon` | 42.1 gb/s, 0 ulp | 50.8 gb/s, 0 ulp | 52.0 gb/s, 0 ulp |
329
+ | `nk_each_blend_f64_serial` | 75.1 gb/s, 0 ulp | 71.8 gb/s, 0 ulp | 60.5 gb/s, 0 ulp |
330
+ | `nk_each_blend_f64_neon` | 64.9 gb/s, 1.9 ulp | 69.8 gb/s, 2.7 ulp | 64.5 gb/s, 1.8 ulp |
331
+ | `nk_each_fma_f64_serial` | 83.7 gb/s, 0 ulp | 76.3 gb/s, 0 ulp | 66.0 gb/s, 0 ulp |
332
+ | `nk_each_fma_f64_neon` | 71.5 gb/s, 1.4 ulp | 71.6 gb/s, 1.6 ulp | 69.6 gb/s, 1.6 ulp |
333
+ | __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
334
+ | `nk_each_sum_f32_serial` | 89.9 gb/s, 0 ulp | 67.8 gb/s, 0 ulp | 59.6 gb/s, 0 ulp |
335
+ | `nk_each_sum_f32_neon` | 75.5 gb/s, 0 ulp | 66.2 gb/s, 0 ulp | 55.2 gb/s, 0 ulp |
336
+ | `nk_each_scale_f32_serial` | 66.0 gb/s, 0 ulp | 61.5 gb/s, 0 ulp | 54.4 gb/s, 0 ulp |
337
+ | `nk_each_scale_f32_neon` | 49.7 gb/s, 0 ulp | 44.7 gb/s, 0 ulp | 49.2 gb/s, 0 ulp |
338
+ | `nk_each_blend_f32_serial` | 87.1 gb/s, 26 ulp | 66.9 gb/s, 26 ulp | 52.1 gb/s, 2.0 ulp |
339
+ | `nk_each_blend_f32_neon` | 74.4 gb/s, 1.7 ulp | 66.7 gb/s, 1.6 ulp | 50.1 gb/s, 1.6 ulp |
340
+ | `nk_each_fma_f32_serial` | 89.4 gb/s, 2.1 ulp | 51.3 gb/s, 2.5 ulp | 56.2 gb/s, 2.2 ulp |
341
+ | `nk_each_fma_f32_neon` | 82.4 gb/s, 2.1 ulp | 55.2 gb/s, 21 ulp | 56.1 gb/s, 1.8 ulp |
342
+ | __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
343
+ | `nk_each_sum_bf16_serial` | 32.8 gb/s, 0 ulp | 31.6 gb/s, 0 ulp | 33.8 gb/s, 0 ulp |
344
+ | `nk_each_sum_bf16_neonbfdot` | 42.0 gb/s, 0 ulp | 41.7 gb/s, 0 ulp | 43.3 gb/s, 0 ulp |
345
+ | `nk_each_scale_bf16_serial` | 19.3 gb/s, 0 ulp | 17.7 gb/s, 0 ulp | 19.8 gb/s, 0 ulp |
346
+ | `nk_each_scale_bf16_neonbfdot` | 26.9 gb/s, 0 ulp | 23.8 gb/s, 0 ulp | 26.7 gb/s, 0 ulp |
347
+ | `nk_each_blend_bf16_serial` | 25.3 gb/s, 28 ulp | 24.0 gb/s, 26 ulp | 26.4 gb/s, 2.2 ulp |
348
+ | `nk_each_blend_bf16_neonbfdot` | 34.8 gb/s, 29 ulp | 36.1 gb/s, 29 ulp | 34.7 gb/s, 2.2 ulp |
349
+ | `nk_each_fma_bf16_serial` | 31.7 gb/s, 2.1 ulp | 29.3 gb/s, 2.0 ulp | 33.1 gb/s, 33 ulp |
350
+ | `nk_each_fma_bf16_neonbfdot` | 42.8 gb/s, 1.2 ulp | 39.8 gb/s, 1.2 ulp | 39.3 gb/s, 1.5 ulp |
351
+ | __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
352
+ | `nk_each_sum_f16_serial` | 91.5 gb/s, 0 ulp | 68.2 gb/s, 0 ulp | 66.1 gb/s, 0 ulp |
353
+ | `nk_each_sum_f16_neonhalf` | 83.2 gb/s, 0 ulp | 67.3 gb/s, 0 ulp | 61.2 gb/s, 0 ulp |
354
+ | `nk_each_scale_f16_serial` | 37.6 gb/s, 0 ulp | 35.9 gb/s, 0 ulp | 36.2 gb/s, 0 ulp |
355
+ | `nk_each_scale_f16_neonhalf` | 50.0 gb/s, 87.3K ulp | 48.2 gb/s, 84.8K ulp | 41.2 gb/s, 87.3K ulp |
356
+ | `nk_each_blend_f16_serial` | 38.3 gb/s, 2.0 ulp | 35.8 gb/s, 2.0 ulp | 39.1 gb/s, 2.3 ulp |
357
+ | `nk_each_blend_f16_neonhalf` | 78.4 gb/s, 91.6K ulp | 59.2 gb/s, 92.6K ulp | 66.2 gb/s, 91.9K ulp |
358
+ | `nk_each_fma_f16_serial` | 43.5 gb/s, 2.1 ulp | 37.2 gb/s, 1.8 ulp | 43.4 gb/s, 2.2 ulp |
359
+ | `nk_each_fma_f16_neonhalf` | 86.5 gb/s, 97.1K ulp | 75.4 gb/s, 96.7K ulp | 71.2 gb/s, 99.2K ulp |
360
+ | __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
361
+ | `nk_each_sum_e4m3_serial` | 0.358 gb/s, 0 ulp | 0.332 gb/s, 0 ulp | 0.369 gb/s, 0 ulp |
362
+ | `nk_each_sum_e4m3_neon` | 1.59 gb/s, 0 ulp | 1.75 gb/s, 0 ulp | 1.72 gb/s, 0 ulp |
363
+ | `nk_each_scale_e4m3_serial` | 0.148 gb/s, 0 ulp | 0.137 gb/s, 0 ulp | 0.153 gb/s, 0 ulp |
364
+ | `nk_each_scale_e4m3_neon` | 0.954 gb/s, 0 ulp | 1.07 gb/s, 0 ulp | 1.03 gb/s, 0 ulp |
365
+ | `nk_each_blend_e4m3_serial` | 0.253 gb/s, 0.4 ulp | 0.239 gb/s, 1.1 ulp | 0.259 gb/s, 2.8 ulp |
366
+ | `nk_each_blend_e4m3_neon` | 1.56 gb/s, 0.1 ulp | 1.74 gb/s, 0 ulp | 1.68 gb/s, 0 ulp |
367
+ | `nk_each_fma_e4m3_serial` | 0.334 gb/s, 0.7 ulp | 0.317 gb/s, 0.6 ulp | 0.343 gb/s, 2.2 ulp |
368
+ | `nk_each_fma_e4m3_neon` | 1.92 gb/s, 0.1 ulp | 2.20 gb/s, 0.6 ulp | 2.08 gb/s, 1.0 ulp |
369
+ | __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
370
+ | `nk_each_sum_e5m2_serial` | 0.437 gb/s, 0 ulp | 0.413 gb/s, 0 ulp | 0.445 gb/s, 0 ulp |
371
+ | `nk_each_sum_e5m2_neon` | 3.11 gb/s, 0 ulp | 3.48 gb/s, 0 ulp | 3.36 gb/s, 0 ulp |
372
+ | `nk_each_scale_e5m2_serial` | 0.180 gb/s, 0 ulp | 0.173 gb/s, 0 ulp | 0.178 gb/s, 0 ulp |
373
+ | `nk_each_scale_e5m2_neon` | 1.64 gb/s, 0 ulp | 1.82 gb/s, 0 ulp | 1.75 gb/s, 0 ulp |
374
+ | `nk_each_blend_e5m2_serial` | 0.317 gb/s, 0 ulp | 0.322 gb/s, 0 ulp | 0.316 gb/s, 4.9 ulp |
375
+ | `nk_each_blend_e5m2_neon` | 3.02 gb/s, 0.7 ulp | 3.29 gb/s, 0 ulp | 3.20 gb/s, 0 ulp |
376
+ | `nk_each_fma_e5m2_serial` | 0.476 gb/s, 1.9 ulp | 0.455 gb/s, 0.9 ulp | 0.453 gb/s, 4.0 ulp |
377
+ | `nk_each_fma_e5m2_neon` | 4.14 gb/s, 0 ulp | 4.42 gb/s, 1.3 ulp | 4.31 gb/s, 0 ulp |
378
+ | __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
379
+ | `nk_each_sum_e2m3_serial` | 0.337 gb/s, ? ulp | 0.297 gb/s, ? ulp | 0.290 gb/s, ? ulp |
380
+ | `nk_each_scale_e2m3_serial` | 0.102 gb/s, ? ulp | 0.0925 gb/s, ? ulp | 0.0939 gb/s, ? ulp |
381
+ | `nk_each_blend_e2m3_serial` | 0.226 gb/s, ? ulp | 0.201 gb/s, ? ulp | 0.201 gb/s, ? ulp |
382
+ | `nk_each_fma_e2m3_serial` | 0.368 gb/s, ? ulp | 0.331 gb/s, ? ulp | 0.331 gb/s, ? ulp |
383
+ | __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
384
+ | `nk_each_sum_e3m2_serial` | 0.544 gb/s, ? ulp | 0.472 gb/s, ? ulp | 0.458 gb/s, ? ulp |
385
+ | `nk_each_scale_e3m2_serial` | 0.176 gb/s, ? ulp | 0.151 gb/s, ? ulp | 0.143 gb/s, ? ulp |
386
+ | `nk_each_blend_e3m2_serial` | 0.355 gb/s, ? ulp | 0.302 gb/s, ? ulp | 0.292 gb/s, ? ulp |
387
+ | `nk_each_fma_e3m2_serial` | 0.506 gb/s, ? ulp | 0.433 gb/s, ? ulp | 0.436 gb/s, ? ulp |
388
+ | __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
389
+ | `nk_each_sum_i8_serial` | 94.6 gb/s | 55.5 gb/s | 42.7 gb/s |
390
+ | `nk_each_sum_i8_neonhalf` | 85.5 gb/s | 60.5 gb/s | 39.6 gb/s |
391
+ | `nk_each_scale_i8_serial` | 0.171 gb/s | 0.145 gb/s | 0.148 gb/s |
392
+ | `nk_each_scale_i8_neonhalf` | 22.6 gb/s | 22.4 gb/s | 20.8 gb/s |
393
+ | `nk_each_blend_i8_serial` | 0.302 gb/s | 0.254 gb/s | 0.272 gb/s |
394
+ | `nk_each_blend_i8_neonhalf` | 26.5 gb/s | 27.7 gb/s | 27.3 gb/s |
395
+ | `nk_each_fma_i8_serial` | 0.436 gb/s | 0.359 gb/s | 0.376 gb/s |
396
+ | __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
397
+ | `nk_each_sum_u8_serial` | 13.2 gb/s | 12.3 gb/s | 12.3 gb/s |
398
+ | `nk_each_sum_u8_neonhalf` | 85.3 gb/s | 84.8 gb/s | 67.3 gb/s |
399
+ | `nk_each_scale_u8_serial` | 0.135 gb/s | 0.120 gb/s | 0.119 gb/s |
400
+ | `nk_each_scale_u8_neonhalf` | 23.0 gb/s | 21.9 gb/s | 21.2 gb/s |
401
+ | `nk_each_blend_u8_serial` | 0.256 gb/s | 0.218 gb/s | 0.220 gb/s |
402
+ | `nk_each_blend_u8_neonhalf` | 27.4 gb/s | 27.0 gb/s | 27.4 gb/s |
403
+ | `nk_each_fma_u8_serial` | 0.356 gb/s | 0.333 gb/s | 0.323 gb/s |
404
+ | __i16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
405
+ | `nk_each_sum_i16_serial` | 100 gb/s | 69.6 gb/s | 79.2 gb/s |
406
+ | `nk_each_sum_i16_neon` | 75.3 gb/s | 65.6 gb/s | 66.9 gb/s |
407
+ | `nk_each_scale_i16_serial` | 0.358 gb/s | 0.297 gb/s | 0.284 gb/s |
408
+ | `nk_each_scale_i16_neon` | 17.6 gb/s | 18.5 gb/s | 15.5 gb/s |
409
+ | `nk_each_blend_i16_serial` | 0.644 gb/s | 0.523 gb/s | 0.492 gb/s |
410
+ | `nk_each_fma_i16_serial` | 0.889 gb/s | 0.731 gb/s | 0.712 gb/s |
411
+ | `nk_each_fma_i16_neon` | 27.9 gb/s | 27.7 gb/s | 28.1 gb/s |
412
+ | __u16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
413
+ | `nk_each_sum_u16_serial` | 25.0 gb/s | 22.4 gb/s | 23.2 gb/s |
414
+ | `nk_each_sum_u16_neon` | 79.2 gb/s | 66.8 gb/s | 68.1 gb/s |
415
+ | `nk_each_scale_u16_serial` | 0.267 gb/s | 0.245 gb/s | 0.241 gb/s |
416
+ | `nk_each_scale_u16_neon` | 17.0 gb/s | 18.9 gb/s | 18.6 gb/s |
417
+ | `nk_each_blend_u16_serial` | 0.520 gb/s | 0.432 gb/s | 0.444 gb/s |
418
+ | `nk_each_fma_u16_serial` | 0.744 gb/s | 0.664 gb/s | 0.638 gb/s |
419
+ | `nk_each_fma_u16_neon` | 28.1 gb/s | 27.9 gb/s | 28.9 gb/s |
420
+ | __i32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
421
+ | `nk_each_sum_i32_serial` | 68.6 gb/s | 62.0 gb/s | 71.4 gb/s |
422
+ | `nk_each_sum_i32_neon` | 66.2 gb/s | 66.9 gb/s | 62.4 gb/s |
423
+ | `nk_each_scale_i32_serial` | 0.672 gb/s | 0.613 gb/s | 0.589 gb/s |
424
+ | `nk_each_scale_i32_neon` | 18.2 gb/s | 19.3 gb/s | 18.6 gb/s |
425
+ | `nk_each_blend_i32_serial` | 1.27 gb/s | 1.11 gb/s | 0.993 gb/s |
426
+ | `nk_each_fma_i32_serial` | 1.71 gb/s | 1.54 gb/s | 1.47 gb/s |
427
+ | `nk_each_fma_i32_neon` | 28.4 gb/s | 28.8 gb/s | 27.6 gb/s |
428
+ | __u32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
429
+ | `nk_each_sum_u32_serial` | 56.2 gb/s | 46.0 gb/s | 50.0 gb/s |
430
+ | `nk_each_sum_u32_neon` | 84.9 gb/s | 69.7 gb/s | 55.1 gb/s |
431
+ | `nk_each_scale_u32_serial` | 0.598 gb/s | 0.555 gb/s | 0.552 gb/s |
432
+ | `nk_each_scale_u32_neon` | 18.4 gb/s | 18.8 gb/s | 18.5 gb/s |
433
+ | `nk_each_blend_u32_serial` | 1.02 gb/s | 0.982 gb/s | 0.953 gb/s |
434
+ | `nk_each_fma_u32_serial` | 1.50 gb/s | 1.43 gb/s | 1.41 gb/s |
435
+ | `nk_each_fma_u32_neon` | 28.2 gb/s | 28.9 gb/s | 27.0 gb/s |
436
+ | __i64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
437
+ | `nk_each_sum_i64_serial` | 55.2 gb/s | 41.2 gb/s | 53.9 gb/s |
438
+ | `nk_each_sum_i64_neon` | 52.9 gb/s | 41.7 gb/s | 51.4 gb/s |
439
+ | `nk_each_scale_i64_serial` | 7.90 gb/s | 7.64 gb/s | 8.13 gb/s |
440
+ | `nk_each_scale_i64_neon` | 39.4 gb/s | 22.8 gb/s | 34.7 gb/s |
441
+ | `nk_each_blend_i64_serial` | 13.8 gb/s | 13.8 gb/s | 14.1 gb/s |
442
+ | `nk_each_fma_i64_serial` | 13.3 gb/s | 9.94 gb/s | 12.8 gb/s |
443
+ | `nk_each_fma_i64_neon` | 57.7 gb/s | 34.2 gb/s | 49.6 gb/s |
444
+ | __u64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
445
+ | `nk_each_sum_u64_serial` | 52.6 gb/s | 31.5 gb/s | 50.9 gb/s |
446
+ | `nk_each_sum_u64_neon` | 54.3 gb/s | 41.3 gb/s | 48.8 gb/s |
447
+ | `nk_each_scale_u64_serial` | 9.29 gb/s | 9.48 gb/s | 9.11 gb/s |
448
+ | `nk_each_scale_u64_neon` | 38.4 gb/s | 35.1 gb/s | 35.1 gb/s |
449
+ | `nk_each_blend_u64_serial` | 14.3 gb/s | 16.1 gb/s | 14.1 gb/s |
450
+ | `nk_each_fma_u64_serial` | 14.2 gb/s | 13.9 gb/s | 15.2 gb/s |
451
+ | `nk_each_fma_u64_neon` | 52.4 gb/s | 52.0 gb/s | 53.2 gb/s |
452
+ | __f64c__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
453
+ | `nk_each_scale_f64c_serial` | 28.1 gb/s, 2.2 ulp | 39.2 gb/s, 1.9 ulp | 37.8 gb/s, 2.2 ulp |
454
+ | `nk_each_scale_f64c_neon` | 25.6 gb/s, 1.5 ulp | 51.0 gb/s, 1.5 ulp | 42.1 gb/s, 1.3 ulp |
455
+ | `nk_each_blend_f64c_serial` | 44.9 gb/s, 4.2 ulp | 35.5 gb/s, 3.0 ulp | 48.8 gb/s, 2.6 ulp |
456
+ | `nk_each_blend_f64c_neon` | 43.7 gb/s, 3.2 ulp | 46.3 gb/s, 3.0 ulp | 48.3 gb/s, 2.2 ulp |
457
+ | `nk_each_fma_f64c_serial` | 58.2 gb/s, 3.4 ulp | 43.9 gb/s, 5.2 ulp | 52.2 gb/s, 2.5 ulp |
458
+ | `nk_each_fma_f64c_neon` | 56.1 gb/s, 3.2 ulp | 52.2 gb/s, 3.0 ulp | 53.2 gb/s, 2.4 ulp |
459
+ | __f32c__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
460
+ | `nk_each_scale_f32c_serial` | 44.8 gb/s, 17 ulp | 37.7 gb/s, 2.4 ulp | 44.3 gb/s, 17 ulp |
461
+ | `nk_each_scale_f32c_neon` | 50.8 gb/s, 1.8 ulp | 52.7 gb/s, 1.6 ulp | 45.7 gb/s, 1.6 ulp |
462
+ | `nk_each_blend_f32c_serial` | 50.3 gb/s, 2.4 ulp | 49.0 gb/s, 2.6 ulp | 53.5 gb/s, 28 ulp |
463
+ | `nk_each_blend_f32c_neon` | 49.0 gb/s, 2.2 ulp | 58.3 gb/s, 2.2 ulp | 51.0 gb/s, 3.2 ulp |
464
+ | `nk_each_fma_f32c_serial` | 56.4 gb/s, 3.1 ulp | 49.8 gb/s, 78 ulp | 51.5 gb/s, 3.5 ulp |
465
+ | `nk_each_fma_f32c_neon` | 62.7 gb/s, 2.9 ulp | 62.4 gb/s, 4.1 ulp | 58.0 gb/s, 81 ulp |