numkong 7.0.0 → 7.4.1
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.
- package/README.md +239 -122
- package/binding.gyp +25 -491
- package/c/dispatch_bf16.c +59 -1
- package/c/dispatch_e2m3.c +41 -8
- package/c/dispatch_e3m2.c +49 -8
- package/c/dispatch_e4m3.c +51 -9
- package/c/dispatch_e5m2.c +45 -1
- package/c/dispatch_f16.c +79 -26
- package/c/dispatch_f16c.c +5 -5
- package/c/dispatch_f32.c +56 -0
- package/c/dispatch_f64.c +52 -0
- package/c/dispatch_i4.c +3 -0
- package/c/dispatch_i8.c +62 -3
- package/c/dispatch_other.c +18 -0
- package/c/dispatch_u1.c +54 -9
- package/c/dispatch_u4.c +3 -0
- package/c/dispatch_u8.c +64 -3
- package/c/numkong.c +3 -0
- package/include/README.md +79 -9
- package/include/numkong/attention/sapphireamx.h +278 -276
- package/include/numkong/attention/sme.h +983 -977
- package/include/numkong/attention.h +1 -1
- package/include/numkong/capabilities.h +289 -94
- package/include/numkong/cast/README.md +40 -40
- package/include/numkong/cast/diamond.h +64 -0
- package/include/numkong/cast/haswell.h +42 -194
- package/include/numkong/cast/icelake.h +42 -37
- package/include/numkong/cast/loongsonasx.h +252 -0
- package/include/numkong/cast/neon.h +216 -249
- package/include/numkong/cast/powervsx.h +449 -0
- package/include/numkong/cast/rvv.h +223 -274
- package/include/numkong/cast/sapphire.h +18 -18
- package/include/numkong/cast/serial.h +1018 -944
- package/include/numkong/cast/skylake.h +82 -23
- package/include/numkong/cast/v128relaxed.h +462 -105
- package/include/numkong/cast.h +24 -0
- package/include/numkong/cast.hpp +44 -0
- package/include/numkong/curved/README.md +17 -17
- package/include/numkong/curved/neon.h +131 -7
- package/include/numkong/curved/neonbfdot.h +6 -7
- package/include/numkong/curved/rvv.h +26 -26
- package/include/numkong/curved/smef64.h +186 -182
- package/include/numkong/curved.h +14 -18
- package/include/numkong/dot/README.md +154 -137
- package/include/numkong/dot/alder.h +43 -43
- package/include/numkong/dot/diamond.h +158 -0
- package/include/numkong/dot/genoa.h +4 -30
- package/include/numkong/dot/haswell.h +215 -180
- package/include/numkong/dot/icelake.h +190 -76
- package/include/numkong/dot/loongsonasx.h +671 -0
- package/include/numkong/dot/neon.h +124 -73
- package/include/numkong/dot/neonbfdot.h +11 -12
- package/include/numkong/dot/neonfhm.h +44 -46
- package/include/numkong/dot/neonfp8.h +323 -0
- package/include/numkong/dot/neonsdot.h +190 -76
- package/include/numkong/dot/powervsx.h +752 -0
- package/include/numkong/dot/rvv.h +92 -84
- package/include/numkong/dot/rvvbf16.h +12 -12
- package/include/numkong/dot/rvvhalf.h +12 -12
- package/include/numkong/dot/sapphire.h +4 -4
- package/include/numkong/dot/serial.h +66 -30
- package/include/numkong/dot/sierra.h +31 -31
- package/include/numkong/dot/skylake.h +142 -110
- package/include/numkong/dot/sve.h +217 -177
- package/include/numkong/dot/svebfdot.h +10 -10
- package/include/numkong/dot/svehalf.h +85 -41
- package/include/numkong/dot/svesdot.h +89 -0
- package/include/numkong/dot/v128relaxed.h +124 -89
- package/include/numkong/dot.h +114 -48
- package/include/numkong/dots/README.md +203 -203
- package/include/numkong/dots/alder.h +12 -9
- package/include/numkong/dots/diamond.h +86 -0
- package/include/numkong/dots/genoa.h +10 -4
- package/include/numkong/dots/haswell.h +63 -48
- package/include/numkong/dots/icelake.h +27 -18
- package/include/numkong/dots/loongsonasx.h +176 -0
- package/include/numkong/dots/neon.h +14 -11
- package/include/numkong/dots/neonbfdot.h +4 -3
- package/include/numkong/dots/neonfhm.h +11 -9
- package/include/numkong/dots/neonfp8.h +99 -0
- package/include/numkong/dots/neonsdot.h +48 -12
- package/include/numkong/dots/powervsx.h +194 -0
- package/include/numkong/dots/rvv.h +451 -344
- package/include/numkong/dots/sapphireamx.h +1028 -984
- package/include/numkong/dots/serial.h +213 -197
- package/include/numkong/dots/sierra.h +10 -7
- package/include/numkong/dots/skylake.h +47 -36
- package/include/numkong/dots/sme.h +2001 -2364
- package/include/numkong/dots/smebi32.h +175 -162
- package/include/numkong/dots/smef64.h +328 -323
- package/include/numkong/dots/v128relaxed.h +64 -41
- package/include/numkong/dots.h +573 -293
- package/include/numkong/dots.hpp +45 -43
- package/include/numkong/each/README.md +133 -137
- package/include/numkong/each/haswell.h +6 -6
- package/include/numkong/each/icelake.h +7 -7
- package/include/numkong/each/neon.h +76 -42
- package/include/numkong/each/neonbfdot.h +11 -12
- package/include/numkong/each/neonhalf.h +24 -116
- package/include/numkong/each/rvv.h +28 -28
- package/include/numkong/each/sapphire.h +27 -161
- package/include/numkong/each/serial.h +6 -6
- package/include/numkong/each/skylake.h +7 -7
- package/include/numkong/each/v128relaxed.h +562 -0
- package/include/numkong/each.h +148 -62
- package/include/numkong/each.hpp +2 -2
- package/include/numkong/geospatial/README.md +18 -18
- package/include/numkong/geospatial/haswell.h +365 -325
- package/include/numkong/geospatial/neon.h +350 -306
- package/include/numkong/geospatial/rvv.h +4 -4
- package/include/numkong/geospatial/skylake.h +376 -340
- package/include/numkong/geospatial/v128relaxed.h +366 -327
- package/include/numkong/geospatial.h +17 -17
- package/include/numkong/matrix.hpp +4 -4
- package/include/numkong/maxsim/README.md +14 -14
- package/include/numkong/maxsim/alder.h +6 -6
- package/include/numkong/maxsim/genoa.h +4 -4
- package/include/numkong/maxsim/haswell.h +6 -6
- package/include/numkong/maxsim/icelake.h +18 -18
- package/include/numkong/maxsim/neonsdot.h +21 -21
- package/include/numkong/maxsim/sapphireamx.h +14 -14
- package/include/numkong/maxsim/serial.h +6 -6
- package/include/numkong/maxsim/sme.h +221 -196
- package/include/numkong/maxsim/v128relaxed.h +6 -6
- package/include/numkong/mesh/README.md +62 -56
- package/include/numkong/mesh/haswell.h +339 -464
- package/include/numkong/mesh/neon.h +1100 -519
- package/include/numkong/mesh/neonbfdot.h +36 -68
- package/include/numkong/mesh/rvv.h +530 -435
- package/include/numkong/mesh/serial.h +75 -91
- package/include/numkong/mesh/skylake.h +1627 -302
- package/include/numkong/mesh/v128relaxed.h +443 -330
- package/include/numkong/mesh.h +63 -49
- package/include/numkong/mesh.hpp +4 -4
- package/include/numkong/numkong.h +3 -3
- package/include/numkong/numkong.hpp +1 -0
- package/include/numkong/probability/README.md +23 -19
- package/include/numkong/probability/neon.h +82 -52
- package/include/numkong/probability/rvv.h +28 -23
- package/include/numkong/probability/serial.h +51 -39
- package/include/numkong/probability.h +20 -23
- package/include/numkong/random.h +1 -1
- package/include/numkong/reduce/README.md +143 -138
- package/include/numkong/reduce/alder.h +81 -77
- package/include/numkong/reduce/haswell.h +222 -220
- package/include/numkong/reduce/neon.h +629 -519
- package/include/numkong/reduce/neonbfdot.h +7 -218
- package/include/numkong/reduce/neonfhm.h +9 -381
- package/include/numkong/reduce/neonsdot.h +9 -9
- package/include/numkong/reduce/rvv.h +928 -802
- package/include/numkong/reduce/serial.h +23 -27
- package/include/numkong/reduce/sierra.h +20 -20
- package/include/numkong/reduce/skylake.h +326 -324
- package/include/numkong/reduce/v128relaxed.h +52 -52
- package/include/numkong/reduce.h +4 -23
- package/include/numkong/reduce.hpp +156 -11
- package/include/numkong/scalar/README.md +6 -6
- package/include/numkong/scalar/haswell.h +26 -17
- package/include/numkong/scalar/loongsonasx.h +74 -0
- package/include/numkong/scalar/neon.h +9 -9
- package/include/numkong/scalar/powervsx.h +96 -0
- package/include/numkong/scalar/rvv.h +2 -2
- package/include/numkong/scalar/sapphire.h +21 -10
- package/include/numkong/scalar/serial.h +21 -21
- package/include/numkong/scalar.h +13 -0
- package/include/numkong/set/README.md +28 -28
- package/include/numkong/set/haswell.h +12 -12
- package/include/numkong/set/icelake.h +14 -14
- package/include/numkong/set/loongsonasx.h +181 -0
- package/include/numkong/set/neon.h +17 -18
- package/include/numkong/set/powervsx.h +326 -0
- package/include/numkong/set/rvv.h +4 -4
- package/include/numkong/set/serial.h +6 -6
- package/include/numkong/set/sve.h +60 -59
- package/include/numkong/set/v128relaxed.h +6 -6
- package/include/numkong/set.h +21 -7
- package/include/numkong/sets/README.md +26 -26
- package/include/numkong/sets/loongsonasx.h +52 -0
- package/include/numkong/sets/powervsx.h +65 -0
- package/include/numkong/sets/smebi32.h +395 -364
- package/include/numkong/sets.h +83 -40
- package/include/numkong/sparse/README.md +4 -4
- package/include/numkong/sparse/icelake.h +101 -101
- package/include/numkong/sparse/serial.h +1 -1
- package/include/numkong/sparse/sve2.h +137 -141
- package/include/numkong/sparse/turin.h +12 -12
- package/include/numkong/sparse.h +10 -10
- package/include/numkong/spatial/README.md +230 -226
- package/include/numkong/spatial/alder.h +113 -116
- package/include/numkong/spatial/diamond.h +240 -0
- package/include/numkong/spatial/genoa.h +0 -68
- package/include/numkong/spatial/haswell.h +74 -55
- package/include/numkong/spatial/icelake.h +539 -58
- package/include/numkong/spatial/loongsonasx.h +483 -0
- package/include/numkong/spatial/neon.h +125 -52
- package/include/numkong/spatial/neonbfdot.h +8 -9
- package/include/numkong/spatial/neonfp8.h +258 -0
- package/include/numkong/spatial/neonsdot.h +180 -12
- package/include/numkong/spatial/powervsx.h +738 -0
- package/include/numkong/spatial/rvv.h +146 -139
- package/include/numkong/spatial/rvvbf16.h +17 -12
- package/include/numkong/spatial/rvvhalf.h +13 -10
- package/include/numkong/spatial/serial.h +13 -12
- package/include/numkong/spatial/sierra.h +232 -39
- package/include/numkong/spatial/skylake.h +73 -74
- package/include/numkong/spatial/sve.h +93 -72
- package/include/numkong/spatial/svebfdot.h +29 -29
- package/include/numkong/spatial/svehalf.h +52 -26
- package/include/numkong/spatial/svesdot.h +142 -0
- package/include/numkong/spatial/v128relaxed.h +293 -41
- package/include/numkong/spatial.h +338 -82
- package/include/numkong/spatials/README.md +194 -194
- package/include/numkong/spatials/diamond.h +82 -0
- package/include/numkong/spatials/haswell.h +2 -2
- package/include/numkong/spatials/loongsonasx.h +153 -0
- package/include/numkong/spatials/neonfp8.h +111 -0
- package/include/numkong/spatials/neonsdot.h +34 -0
- package/include/numkong/spatials/powervsx.h +153 -0
- package/include/numkong/spatials/rvv.h +259 -243
- package/include/numkong/spatials/sapphireamx.h +173 -173
- package/include/numkong/spatials/serial.h +2 -2
- package/include/numkong/spatials/skylake.h +2 -2
- package/include/numkong/spatials/sme.h +590 -605
- package/include/numkong/spatials/smef64.h +139 -130
- package/include/numkong/spatials/v128relaxed.h +2 -2
- package/include/numkong/spatials.h +820 -500
- package/include/numkong/spatials.hpp +49 -48
- package/include/numkong/tensor.hpp +406 -17
- package/include/numkong/trigonometry/README.md +19 -19
- package/include/numkong/trigonometry/haswell.h +402 -401
- package/include/numkong/trigonometry/neon.h +386 -387
- package/include/numkong/trigonometry/rvv.h +52 -51
- package/include/numkong/trigonometry/serial.h +13 -13
- package/include/numkong/trigonometry/skylake.h +373 -369
- package/include/numkong/trigonometry/v128relaxed.h +375 -374
- package/include/numkong/trigonometry.h +13 -13
- package/include/numkong/trigonometry.hpp +2 -2
- package/include/numkong/types.h +287 -49
- package/include/numkong/types.hpp +436 -12
- package/include/numkong/vector.hpp +82 -14
- package/javascript/dist/cjs/numkong-wasm.js +6 -12
- package/javascript/dist/cjs/numkong.d.ts +7 -1
- package/javascript/dist/cjs/numkong.js +37 -11
- package/javascript/dist/cjs/types.d.ts +9 -0
- package/javascript/dist/cjs/types.js +96 -0
- package/javascript/dist/esm/numkong-browser.d.ts +14 -0
- package/javascript/dist/esm/numkong-browser.js +23 -0
- package/javascript/dist/esm/numkong-wasm.js +6 -12
- package/javascript/dist/esm/numkong.d.ts +7 -1
- package/javascript/dist/esm/numkong.js +37 -11
- package/javascript/dist/esm/types.d.ts +9 -0
- package/javascript/dist/esm/types.js +96 -0
- package/javascript/node-gyp-build.d.ts +4 -1
- package/javascript/numkong-browser.ts +40 -0
- package/javascript/numkong-wasm.ts +7 -13
- package/javascript/numkong.c +5 -26
- package/javascript/numkong.ts +36 -11
- package/javascript/tsconfig-base.json +1 -0
- package/javascript/tsconfig-cjs.json +6 -1
- package/javascript/types.ts +110 -0
- package/numkong.gypi +101 -0
- package/package.json +34 -13
- package/probes/arm_neon.c +8 -0
- package/probes/arm_neon_bfdot.c +9 -0
- package/probes/arm_neon_fhm.c +9 -0
- package/probes/arm_neon_half.c +8 -0
- package/probes/arm_neon_sdot.c +9 -0
- package/probes/arm_neonfp8.c +9 -0
- package/probes/arm_sme.c +16 -0
- package/probes/arm_sme2.c +16 -0
- package/probes/arm_sme2p1.c +16 -0
- package/probes/arm_sme_bf16.c +16 -0
- package/probes/arm_sme_bi32.c +16 -0
- package/probes/arm_sme_f64.c +16 -0
- package/probes/arm_sme_fa64.c +14 -0
- package/probes/arm_sme_half.c +16 -0
- package/probes/arm_sme_lut2.c +15 -0
- package/probes/arm_sve.c +18 -0
- package/probes/arm_sve2.c +20 -0
- package/probes/arm_sve2p1.c +18 -0
- package/probes/arm_sve_bfdot.c +20 -0
- package/probes/arm_sve_half.c +18 -0
- package/probes/arm_sve_sdot.c +21 -0
- package/probes/loongarch_lasx.c +12 -0
- package/probes/power_vsx.c +12 -0
- package/probes/probe.js +127 -0
- package/probes/riscv_rvv.c +14 -0
- package/probes/riscv_rvv_bb.c +15 -0
- package/probes/riscv_rvv_bf16.c +17 -0
- package/probes/riscv_rvv_half.c +14 -0
- package/probes/wasm_v128relaxed.c +11 -0
- package/probes/x86_alder.c +17 -0
- package/probes/x86_diamond.c +17 -0
- package/probes/x86_genoa.c +17 -0
- package/probes/x86_graniteamx.c +19 -0
- package/probes/x86_haswell.c +11 -0
- package/probes/x86_icelake.c +17 -0
- package/probes/x86_sapphire.c +16 -0
- package/probes/x86_sapphireamx.c +18 -0
- package/probes/x86_sierra.c +17 -0
- package/probes/x86_skylake.c +15 -0
- package/probes/x86_turin.c +17 -0
- package/wasm/numkong-emscripten.js +2 -0
- package/wasm/numkong.d.ts +14 -0
- package/wasm/numkong.js +1124 -0
- package/wasm/numkong.wasm +0 -0
- package/include/numkong/curved/neonhalf.h +0 -212
- package/include/numkong/dot/neonhalf.h +0 -198
- package/include/numkong/dots/neonhalf.h +0 -57
- package/include/numkong/mesh/neonhalf.h +0 -616
- package/include/numkong/reduce/neonhalf.h +0 -157
- package/include/numkong/spatial/neonhalf.h +0 -118
- package/include/numkong/spatial/sapphire.h +0 -343
- package/include/numkong/spatials/neonhalf.h +0 -58
- package/javascript/README.md +0 -246
package/include/numkong/curved.h
CHANGED
|
@@ -40,7 +40,7 @@
|
|
|
40
40
|
*
|
|
41
41
|
* @section references References
|
|
42
42
|
*
|
|
43
|
-
* - x86 intrinsics: https://www.intel.com/content/www/us/en/docs/intrinsics-guide/
|
|
43
|
+
* - x86 intrinsics: https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html
|
|
44
44
|
* - Arm intrinsics: https://developer.arm.com/architectures/instruction-sets/intrinsics/
|
|
45
45
|
* - Neumaier, A. (1974). "Rundungsfehleranalyse einiger Verfahren zur Summation endlicher Summen"
|
|
46
46
|
* - Ogita, T., Rump, S.M., Oishi, S. (2005). "Accurate Sum and Dot Product"
|
|
@@ -173,19 +173,16 @@ NK_PUBLIC void nk_bilinear_f32c_neon(nk_f32c_t const *a, nk_f32c_t const *b, nk_
|
|
|
173
173
|
/** @copydoc nk_mahalanobis_f32 */
|
|
174
174
|
NK_PUBLIC void nk_mahalanobis_f32_neon(nk_f32_t const *a, nk_f32_t const *b, nk_f32_t const *c, nk_size_t n,
|
|
175
175
|
nk_f64_t *result);
|
|
176
|
-
#endif // NK_TARGET_NEON
|
|
177
|
-
|
|
178
|
-
#if NK_TARGET_NEONHALF
|
|
179
176
|
/** @copydoc nk_bilinear_f16 */
|
|
180
|
-
NK_PUBLIC void
|
|
181
|
-
|
|
177
|
+
NK_PUBLIC void nk_bilinear_f16_neon(nk_f16_t const *a, nk_f16_t const *b, nk_f16_t const *c, nk_size_t n,
|
|
178
|
+
nk_f32_t *result);
|
|
182
179
|
/** @copydoc nk_bilinear_f16c */
|
|
183
|
-
NK_PUBLIC void
|
|
184
|
-
|
|
180
|
+
NK_PUBLIC void nk_bilinear_f16c_neon(nk_f16c_t const *a, nk_f16c_t const *b, nk_f16c_t const *c, nk_size_t n,
|
|
181
|
+
nk_f32c_t *results);
|
|
185
182
|
/** @copydoc nk_mahalanobis_f16 */
|
|
186
|
-
NK_PUBLIC void
|
|
187
|
-
|
|
188
|
-
#endif //
|
|
183
|
+
NK_PUBLIC void nk_mahalanobis_f16_neon(nk_f16_t const *a, nk_f16_t const *b, nk_f16_t const *c, nk_size_t n,
|
|
184
|
+
nk_f32_t *result);
|
|
185
|
+
#endif // NK_TARGET_NEON
|
|
189
186
|
|
|
190
187
|
#if NK_TARGET_NEONBFDOT
|
|
191
188
|
/** @copydoc nk_bilinear_bf16 */
|
|
@@ -337,7 +334,6 @@ NK_INTERNAL nk_dtype_t nk_mahalanobis_output_dtype(nk_dtype_t dtype) {
|
|
|
337
334
|
|
|
338
335
|
#include "numkong/curved/serial.h"
|
|
339
336
|
#include "numkong/curved/neon.h"
|
|
340
|
-
#include "numkong/curved/neonhalf.h"
|
|
341
337
|
#include "numkong/curved/neonbfdot.h"
|
|
342
338
|
#include "numkong/curved/smef64.h"
|
|
343
339
|
#include "numkong/curved/haswell.h"
|
|
@@ -382,8 +378,8 @@ NK_PUBLIC void nk_bilinear_f32(nk_f32_t const *a, nk_f32_t const *b, nk_f32_t co
|
|
|
382
378
|
NK_PUBLIC void nk_bilinear_f16(nk_f16_t const *a, nk_f16_t const *b, nk_f16_t const *c, nk_size_t n, nk_f32_t *result) {
|
|
383
379
|
#if NK_TARGET_HASWELL
|
|
384
380
|
nk_bilinear_f16_haswell(a, b, c, n, result);
|
|
385
|
-
#elif
|
|
386
|
-
|
|
381
|
+
#elif NK_TARGET_NEON
|
|
382
|
+
nk_bilinear_f16_neon(a, b, c, n, result);
|
|
387
383
|
#elif NK_TARGET_RVV
|
|
388
384
|
nk_bilinear_f16_rvv(a, b, c, n, result);
|
|
389
385
|
#else
|
|
@@ -432,8 +428,8 @@ NK_PUBLIC void nk_bilinear_f32c(nk_f32c_t const *a, nk_f32c_t const *b, nk_f32c_
|
|
|
432
428
|
|
|
433
429
|
NK_PUBLIC void nk_bilinear_f16c(nk_f16c_t const *a, nk_f16c_t const *b, nk_f16c_t const *c, nk_size_t n,
|
|
434
430
|
nk_f32c_t *results) {
|
|
435
|
-
#if
|
|
436
|
-
|
|
431
|
+
#if NK_TARGET_NEON
|
|
432
|
+
nk_bilinear_f16c_neon(a, b, c, n, results);
|
|
437
433
|
#else
|
|
438
434
|
nk_bilinear_f16c_serial(a, b, c, n, results);
|
|
439
435
|
#endif
|
|
@@ -484,8 +480,8 @@ NK_PUBLIC void nk_mahalanobis_f16(nk_f16_t const *a, nk_f16_t const *b, nk_f16_t
|
|
|
484
480
|
nk_f32_t *result) {
|
|
485
481
|
#if NK_TARGET_HASWELL
|
|
486
482
|
nk_mahalanobis_f16_haswell(a, b, c, n, result);
|
|
487
|
-
#elif
|
|
488
|
-
|
|
483
|
+
#elif NK_TARGET_NEON
|
|
484
|
+
nk_mahalanobis_f16_neon(a, b, c, n, result);
|
|
489
485
|
#elif NK_TARGET_RVV
|
|
490
486
|
nk_mahalanobis_f16_rvv(a, b, c, n, result);
|
|
491
487
|
#else
|
|
@@ -1,24 +1,24 @@
|
|
|
1
1
|
# Vector-Vector Dot Products in NumKong
|
|
2
2
|
|
|
3
|
-
NumKong implements dot products for every numeric type supported by the library, as
|
|
3
|
+
NumKong implements dot products for every numeric type supported by the library, as a core building block of higher-level functionality for vectors and higher rank tensors.
|
|
4
4
|
|
|
5
5
|
Dot product for real numbers and integers is defined as:
|
|
6
6
|
|
|
7
|
-
|
|
7
|
+
$$
|
|
8
8
|
\text{dot}(a, b) = \sum_{i=0}^{n-1} a_i \cdot b_i
|
|
9
|
-
|
|
9
|
+
$$
|
|
10
10
|
|
|
11
11
|
For complex numbers, the dot product expands via the distributive property of complex multiplication:
|
|
12
12
|
|
|
13
|
-
|
|
13
|
+
$$
|
|
14
14
|
\text{dot}(a, b) = \sum_{i=0}^{n-1} (a_{i,re} \cdot b_{i,re} - a_{i,im} \cdot b_{i,im}) + j \sum_{i=0}^{n-1} (a_{i,re} \cdot b_{i,im} + a_{i,im} \cdot b_{i,re})
|
|
15
|
-
|
|
15
|
+
$$
|
|
16
16
|
|
|
17
17
|
The conjugate dot product negates the imaginary part of $b$:
|
|
18
18
|
|
|
19
|
-
|
|
19
|
+
$$
|
|
20
20
|
\text{vdot}(a, b) = \sum_{i=0}^{n-1} a_i \cdot \bar{b_i} = \sum_{i=0}^{n-1} (a_{i,re} \cdot b_{i,re} + a_{i,im} \cdot b_{i,im}) + j \sum_{i=0}^{n-1} (a_{i,im} \cdot b_{i,re} - a_{i,re} \cdot b_{i,im})
|
|
21
|
-
|
|
21
|
+
$$
|
|
22
22
|
|
|
23
23
|
Where $\bar{b_i}$ is the complex conjugate of $b_i$.
|
|
24
24
|
Reformulating as Python pseudocode for interleaved real/imaginary scalar arrays:
|
|
@@ -97,9 +97,19 @@ For Int8 × Int8, one operand is XORed with `0x80` to shift to unsigned, and the
|
|
|
97
97
|
For Int8 × Int8, the sign bit of $b$ is cleared to produce a 7-bit value, and a windowed correction $-128 \cdot \sum_{b_i < 0} a_i$ is accumulated in Int16 and flushed every 127 iterations to prevent overflow.
|
|
98
98
|
For UInt8 × UInt8, $b$ is XORed with `0x80` to shift into signed range, same as Ice Lake, with the correction $128 \cdot \sum a_i$ computed via pairwise widening adds.
|
|
99
99
|
|
|
100
|
+
### Octave Decomposition for E4M3 via VNNI
|
|
101
|
+
|
|
102
|
+
`nk_dot_e4m3_icelake` splits the 4-bit E4M3 exponent into 2 "octave" bits (top) and 2 "remainder" bits (bottom).
|
|
103
|
+
The bottom 5 bits (2 remainder + 3 mantissa) map via `VPERMB` to u8 integers in [0, 120] — identical structure to the E2M3 $\times 16$ LUT.
|
|
104
|
+
A subnormal fixup replaces LUT entries for magnitude < 8 with $2 \times \text{mantissa}$ via a second masked `VPERMB`, avoiding `VPADDB` on the VPDPBUSD execution ports.
|
|
105
|
+
Sign is computed via `VPTERNLOGD` with immediate 0x14, fusing `(a \oplus b) \wedge \lnot \text{0x7F}` in one instruction.
|
|
106
|
+
The 4 octave bins per operand produce $4 \times 4 = 16$ `VPDPBUSD` cross-products accumulated into 7 registers grouped by octave sum $k = o_a + o_b \in [0, 6]$.
|
|
107
|
+
Each accumulator is scaled by $2^{4k-20}$ — an exact power of two, introducing no rounding.
|
|
108
|
+
This processes 64 E4M3 bytes per iteration in u8, doubling the element density of the BF16 upcast path.
|
|
109
|
+
|
|
100
110
|
### Widening Fusion Through BFloat16 on x86
|
|
101
111
|
|
|
102
|
-
`
|
|
112
|
+
`nk_dot_e5m2_genoa` converts FP8 values to BF16, then accumulates via `VDPBF16PS`, reusing Genoa's BF16 dot-product instruction for FP8 types.
|
|
103
113
|
Each `VDPBF16PS` fuses two BF16 multiply-adds per 32-bit lane at 6-cycle throughput.
|
|
104
114
|
`nk_dot_bf16c_genoa` uses the same instruction for complex BF16, preparing operands with `VPSHUFB` for lane swapping and `VPXORD` with `0x80000000` for sign flips before feeding into `VDPBF16PS`.
|
|
105
115
|
|
|
@@ -170,66 +180,68 @@ Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run
|
|
|
170
180
|
| `nk_vdot_f16c_haswell` | 24.0 gb/s, 11.1 ulp | 20.0 gb/s, 17.4 ulp | 17.1 gb/s, 29.2 ulp |
|
|
171
181
|
| __f64__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
172
182
|
| `dot_f64_with_blas` 🧩 | 27.8 gb/s, 6.9 ulp | 30.1 gb/s, 9.3 ulp | 15.7 gb/s, 20 ulp |
|
|
173
|
-
| `nk_dot_f64_serial` |
|
|
174
|
-
| `nk_dot_f64_haswell` |
|
|
175
|
-
| `nk_dot_f64_skylake` |
|
|
183
|
+
| `nk_dot_f64_serial` | 4.28 gb/s, 2.2 ulp | 4.39 gb/s, 2.0 ulp | 4.42 gb/s, 3.3 ulp |
|
|
184
|
+
| `nk_dot_f64_haswell` | 24.2 gb/s, 0 ulp | 25.7 gb/s, 0 ulp | 18.3 gb/s, 0 ulp |
|
|
185
|
+
| `nk_dot_f64_skylake` | 29.0 gb/s, 0 ulp | 28.6 gb/s, 0 ulp | 24.9 gb/s, 0 ulp |
|
|
176
186
|
| __f32__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
177
187
|
| `dot_f32_with_blas` 🧩 | 47.8 gb/s, 14 ulp | 30.7 gb/s, 14 ulp | 29.7 gb/s, 15 ulp |
|
|
178
|
-
| `nk_dot_f32_serial` | 11.
|
|
179
|
-
| `nk_dot_f32_haswell` |
|
|
180
|
-
| `nk_dot_f32_skylake` |
|
|
188
|
+
| `nk_dot_f32_serial` | 11.0 gb/s, 0 ulp | 11.2 gb/s, 0 ulp | 11.5 gb/s, 0 ulp |
|
|
189
|
+
| `nk_dot_f32_haswell` | 30.5 gb/s, 0 ulp | 23.9 gb/s, 0 ulp | 24.4 gb/s, 0 ulp |
|
|
190
|
+
| `nk_dot_f32_skylake` | 44.2 gb/s, 0 ulp | 29.8 gb/s, 0 ulp | 30.0 gb/s, 0 ulp |
|
|
181
191
|
| __bf16__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
182
|
-
| `nk_dot_bf16_serial` | 0.
|
|
183
|
-
| `nk_dot_bf16_haswell` |
|
|
184
|
-
| `nk_dot_bf16_skylake` |
|
|
185
|
-
| `nk_dot_bf16_genoa` | 88.
|
|
192
|
+
| `nk_dot_bf16_serial` | 0.633 gb/s, 0 ulp | 0.630 gb/s, 0.5 ulp | 0.638 gb/s, 5.4 ulp |
|
|
193
|
+
| `nk_dot_bf16_haswell` | 39.3 gb/s, 0 ulp | 25.5 gb/s, 0.2 ulp | 20.2 gb/s, 25.3 ulp |
|
|
194
|
+
| `nk_dot_bf16_skylake` | 62.7 gb/s, 0 ulp | 30.2 gb/s, 0.2 ulp | 29.5 gb/s, 2.3 ulp |
|
|
195
|
+
| `nk_dot_bf16_genoa` | 88.8 gb/s, 0 ulp | 29.7 gb/s, 0.2 ulp | 31.2 gb/s, 2.2 ulp |
|
|
186
196
|
| __f16__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
187
|
-
| `nk_dot_f16_serial` | 1.
|
|
188
|
-
| `nk_dot_f16_haswell` |
|
|
189
|
-
| `nk_dot_f16_skylake` | 54.
|
|
197
|
+
| `nk_dot_f16_serial` | 1.31 gb/s, 11.5 ulp | 1.32 gb/s, 33.7 ulp | 1.30 gb/s, 59.7 ulp |
|
|
198
|
+
| `nk_dot_f16_haswell` | 31.3 gb/s, 7.0 ulp | 22.8 gb/s, 14.0 ulp | 19.8 gb/s, 29.8 ulp |
|
|
199
|
+
| `nk_dot_f16_skylake` | 54.9 gb/s, 6.2 ulp | 31.7 gb/s, 8.6 ulp | 30.9 gb/s, 22.8 ulp |
|
|
190
200
|
| __e5m2__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
191
|
-
| `nk_dot_e5m2_serial` | 1.
|
|
192
|
-
| `nk_dot_e5m2_haswell` | 4.
|
|
193
|
-
| `nk_dot_e5m2_skylake` | 6.
|
|
194
|
-
| `nk_dot_e5m2_genoa` | 12.
|
|
201
|
+
| `nk_dot_e5m2_serial` | 1.90 gb/s, 0 ulp | 1.07 gb/s, 0 ulp | 1.08 gb/s, 0 ulp |
|
|
202
|
+
| `nk_dot_e5m2_haswell` | 4.92 gb/s, 0 ulp | 4.95 gb/s, 0 ulp | 4.80 gb/s, 0 ulp |
|
|
203
|
+
| `nk_dot_e5m2_skylake` | 6.20 gb/s, 0 ulp | 6.36 gb/s, 0 ulp | 6.25 gb/s, 0 ulp |
|
|
204
|
+
| `nk_dot_e5m2_genoa` | 12.1 gb/s, 0 ulp | 12.6 gb/s, 0 ulp | 12.6 gb/s, 0 ulp |
|
|
195
205
|
| __e4m3__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
196
|
-
| `nk_dot_e4m3_serial` | 0.
|
|
197
|
-
| `nk_dot_e4m3_haswell` | 3.
|
|
198
|
-
| `nk_dot_e4m3_skylake` |
|
|
199
|
-
| `
|
|
206
|
+
| `nk_dot_e4m3_serial` | 0.762 gb/s, 0 ulp | 0.424 gb/s, 0 ulp | 0.420 gb/s, 0 ulp |
|
|
207
|
+
| `nk_dot_e4m3_haswell` | 3.78 gb/s, 0 ulp | 3.77 gb/s, 0 ulp | 3.75 gb/s, 0 ulp |
|
|
208
|
+
| `nk_dot_e4m3_skylake` | 5.10 gb/s, 0 ulp | 5.16 gb/s, 0 ulp | 5.21 gb/s, 0 ulp |
|
|
209
|
+
| `nk_dot_e4m3_icelake` | 13.2 gb/s, 0 ulp | 14.9 gb/s, 0 ulp | 14.7 gb/s, 0 ulp |
|
|
200
210
|
| __e3m2__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
201
|
-
| `nk_dot_e3m2_serial` |
|
|
202
|
-
| `nk_dot_e3m2_haswell` |
|
|
203
|
-
| `
|
|
211
|
+
| `nk_dot_e3m2_serial` | 1.47 gb/s, 0 ulp | 1.05 gb/s, 0 ulp | 1.04 gb/s, 0 ulp |
|
|
212
|
+
| `nk_dot_e3m2_haswell` | 12.0 gb/s, 0 ulp | 12.2 gb/s, 0 ulp | 12.2 gb/s, 0 ulp |
|
|
213
|
+
| `nk_dot_e3m2_skylake` | 21.6 gb/s, 0 ulp | 23.1 gb/s, 0 ulp | 23.2 gb/s, 0 ulp |
|
|
214
|
+
| `nk_dot_e3m2_icelake` | 23.1 gb/s, 0 ulp | 24.3 gb/s, 0 ulp | 23.9 gb/s, 0 ulp |
|
|
204
215
|
| __e2m3__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
205
|
-
| `nk_dot_e2m3_serial` |
|
|
206
|
-
| `nk_dot_e2m3_haswell` | 20.
|
|
207
|
-
| `
|
|
208
|
-
| `
|
|
216
|
+
| `nk_dot_e2m3_serial` | 1.87 gb/s, 0 ulp | 1.25 gb/s, 0 ulp | 1.96 gb/s, 0 ulp |
|
|
217
|
+
| `nk_dot_e2m3_haswell` | 20.5 gb/s, 0 ulp | 20.4 gb/s, 0 ulp | 19.3 gb/s, 0 ulp |
|
|
218
|
+
| `nk_dot_e2m3_skylake` | 35.7 gb/s, 0 ulp | 33.2 gb/s, 0 ulp | 30.7 gb/s, 0 ulp |
|
|
219
|
+
| `nk_dot_e2m3_icelake` | 58.0 gb/s, 0 ulp | 46.0 gb/s, 0 ulp | 31.5 gb/s, 0 ulp |
|
|
220
|
+
| `nk_dot_e2m3_alder` | 29.9 gb/s, 0 ulp | 30.8 gb/s, 0 ulp | 29.1 gb/s, 0 ulp |
|
|
209
221
|
| __i8__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
210
|
-
| `nk_dot_i8_serial` |
|
|
211
|
-
| `nk_dot_i8_haswell` |
|
|
212
|
-
| `nk_dot_i8_skylake` |
|
|
213
|
-
| `nk_dot_i8_icelake` |
|
|
214
|
-
| `nk_dot_i8_alder` |
|
|
222
|
+
| `nk_dot_i8_serial` | 16.9 gb/s | 16.8 gb/s | 15.6 gb/s |
|
|
223
|
+
| `nk_dot_i8_haswell` | 43.2 gb/s | 35.8 gb/s | 29.1 gb/s |
|
|
224
|
+
| `nk_dot_i8_skylake` | 52.9 gb/s | 36.5 gb/s | 28.5 gb/s |
|
|
225
|
+
| `nk_dot_i8_icelake` | 64.0 gb/s | 46.2 gb/s | 26.8 gb/s |
|
|
226
|
+
| `nk_dot_i8_alder` | 42.8 gb/s | 40.4 gb/s | 31.1 gb/s |
|
|
215
227
|
| __u8__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
216
|
-
| `nk_dot_u8_serial` |
|
|
217
|
-
| `nk_dot_u8_haswell` |
|
|
218
|
-
| `nk_dot_u8_skylake` |
|
|
219
|
-
| `nk_dot_u8_icelake` |
|
|
220
|
-
| `nk_dot_u8_alder` |
|
|
228
|
+
| `nk_dot_u8_serial` | 16.9 gb/s | 16.5 gb/s | 15.8 gb/s |
|
|
229
|
+
| `nk_dot_u8_haswell` | 47.7 gb/s | 37.7 gb/s | 29.1 gb/s |
|
|
230
|
+
| `nk_dot_u8_skylake` | 48.7 gb/s | 32.6 gb/s | 27.5 gb/s |
|
|
231
|
+
| `nk_dot_u8_icelake` | 68.4 gb/s | 46.9 gb/s | 30.2 gb/s |
|
|
232
|
+
| `nk_dot_u8_alder` | 42.1 gb/s | 41.8 gb/s | 31.6 gb/s |
|
|
221
233
|
| __i4__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
222
234
|
| `nk_dot_i4_serial` | 9.37 gb/s | 11.8 gb/s | 11.8 gb/s |
|
|
223
|
-
| `nk_dot_i4_haswell` | 8.
|
|
224
|
-
| `nk_dot_i4_icelake` | 24.
|
|
235
|
+
| `nk_dot_i4_haswell` | 8.22 gb/s | 8.53 gb/s | 8.23 gb/s |
|
|
236
|
+
| `nk_dot_i4_icelake` | 24.3 gb/s | 36.3 gb/s | 25.5 gb/s |
|
|
225
237
|
| __u4__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
226
238
|
| `nk_dot_u4_serial` | 10.6 gb/s | 12.0 gb/s | 11.9 gb/s |
|
|
227
|
-
| `nk_dot_u4_haswell` | 15.
|
|
228
|
-
| `nk_dot_u4_icelake` |
|
|
239
|
+
| `nk_dot_u4_haswell` | 15.0 gb/s | 16.4 gb/s | 14.3 gb/s |
|
|
240
|
+
| `nk_dot_u4_icelake` | 48.1 gb/s | 64.4 gb/s | 30.9 gb/s |
|
|
229
241
|
| __u1__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
230
242
|
| `nk_dot_u1_serial` | 3.92 gb/s | 5.04 gb/s | 4.97 gb/s |
|
|
231
|
-
| `nk_dot_u1_haswell` | 14.
|
|
232
|
-
| `nk_dot_u1_icelake` |
|
|
243
|
+
| `nk_dot_u1_haswell` | 14.7 gb/s | 43.2 gb/s | 69.4 gb/s |
|
|
244
|
+
| `nk_dot_u1_icelake` | 17.9 gb/s | 68.8 gb/s | 110 gb/s |
|
|
233
245
|
|
|
234
246
|
#### WASM
|
|
235
247
|
|
|
@@ -293,133 +305,138 @@ Measured with Wasmtime v42 (Cranelift backend).
|
|
|
293
305
|
| `nk_dot_u1_serial` | 1.95 gb/s | 1.53 gb/s | 0.09 gb/s |
|
|
294
306
|
| `nk_dot_u1_v128relaxed` | 0.548 gb/s | 1.88 gb/s | 0.13 gb/s |
|
|
295
307
|
|
|
296
|
-
### Apple
|
|
308
|
+
### Apple M5
|
|
297
309
|
|
|
298
310
|
#### Native
|
|
299
311
|
|
|
300
312
|
| Kernel | 256 | 1024 | 4096 |
|
|
301
313
|
| :------------------------ | -----------------------: | -----------------------: | -----------------------: |
|
|
302
314
|
| __f64c__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
303
|
-
| `nk_dot_f64c_serial` |
|
|
304
|
-
| `nk_vdot_f64c_serial` |
|
|
305
|
-
| `nk_dot_f64c_neon` |
|
|
306
|
-
| `nk_vdot_f64c_neon` |
|
|
315
|
+
| `nk_dot_f64c_serial` | 8.02 gb/s, 5 ulp | 7.30 gb/s, 3 ulp | 7.25 gb/s, 9.7 ulp |
|
|
316
|
+
| `nk_vdot_f64c_serial` | 8.29 gb/s, 4.2 ulp | 7.53 gb/s, 3.3 ulp | 7.38 gb/s, 3.3 ulp |
|
|
317
|
+
| `nk_dot_f64c_neon` | 23.7 gb/s, 0 ulp | 21.6 gb/s, 0 ulp | 21.3 gb/s, 0 ulp |
|
|
318
|
+
| `nk_vdot_f64c_neon` | 23.6 gb/s, 0 ulp | 21.8 gb/s, 0 ulp | 20.9 gb/s, 0 ulp |
|
|
307
319
|
| __f32c__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
308
|
-
| `nk_dot_f32c_serial` |
|
|
309
|
-
| `nk_vdot_f32c_serial` |
|
|
310
|
-
| `nk_dot_f32c_neon` |
|
|
311
|
-
| `nk_vdot_f32c_neon` |
|
|
320
|
+
| `nk_dot_f32c_serial` | 27.8 gb/s, 0 ulp | 24.6 gb/s, 0 ulp | 23.2 gb/s, 0 ulp |
|
|
321
|
+
| `nk_vdot_f32c_serial` | 27.2 gb/s, 0 ulp | 24.0 gb/s, 0 ulp | 22.6 gb/s, 0 ulp |
|
|
322
|
+
| `nk_dot_f32c_neon` | 22.8 gb/s, 0 ulp | 18.2 gb/s, 0 ulp | 16.9 gb/s, 0 ulp |
|
|
323
|
+
| `nk_vdot_f32c_neon` | 22.7 gb/s, 0 ulp | 17.5 gb/s, 0 ulp | 16.7 gb/s, 0 ulp |
|
|
312
324
|
| __bf16c__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
313
|
-
| `nk_dot_bf16c_serial` |
|
|
314
|
-
| `nk_vdot_bf16c_serial` |
|
|
315
|
-
| `nk_dot_bf16c_neonbfdot` |
|
|
316
|
-
| `nk_vdot_bf16c_neonbfdot` |
|
|
325
|
+
| `nk_dot_bf16c_serial` | 15.6 gb/s, 0.2 ulp | 12.5 gb/s, 2.8 ulp | 12.5 gb/s, 15.8 ulp |
|
|
326
|
+
| `nk_vdot_bf16c_serial` | 15.9 gb/s, 0.2 ulp | 12.9 gb/s, 2.6 ulp | 11.7 gb/s, 11.4 ulp |
|
|
327
|
+
| `nk_dot_bf16c_neonbfdot` | 26.3 gb/s, 0.1 ulp | 18.5 gb/s, 2 ulp | 17.6 gb/s, 8.8 ulp |
|
|
328
|
+
| `nk_vdot_bf16c_neonbfdot` | 26.5 gb/s, 0.1 ulp | 18.2 gb/s, 1.8 ulp | 17.3 gb/s, 8.8 ulp |
|
|
317
329
|
| __f16c__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
318
|
-
| `nk_dot_f16c_serial` |
|
|
319
|
-
| `nk_vdot_f16c_serial` |
|
|
320
|
-
| `nk_dot_f16c_neonhalf` |
|
|
321
|
-
| `nk_vdot_f16c_neonhalf` |
|
|
322
|
-
| `nk_dot_f16c_neonfhm` |
|
|
323
|
-
| `nk_vdot_f16c_neonfhm` |
|
|
330
|
+
| `nk_dot_f16c_serial` | 15.8 gb/s, 20.8 ulp | 13.0 gb/s, 64.1 ulp | 12.3 gb/s, 73.1 ulp |
|
|
331
|
+
| `nk_vdot_f16c_serial` | 15.8 gb/s, 24.8 ulp | 13.0 gb/s, 31.9 ulp | 12.3 gb/s, 137 ulp |
|
|
332
|
+
| `nk_dot_f16c_neonhalf` | 26.1 gb/s, 3.0 ulp | 18.4 gb/s, 6.5 ulp | 16.8 gb/s, 20.5 ulp |
|
|
333
|
+
| `nk_vdot_f16c_neonhalf` | 26.1 gb/s, 34.9 ulp | 18.5 gb/s, 40.7 ulp | 17.0 gb/s, 73.1 ulp |
|
|
334
|
+
| `nk_dot_f16c_neonfhm` | 25.3 gb/s, 3.0 ulp | 17.1 gb/s, 6.5 ulp | 15.9 gb/s, 20.5 ulp |
|
|
335
|
+
| `nk_vdot_f16c_neonfhm` | 25.0 gb/s, 31.4 ulp | 17.0 gb/s, 38.6 ulp | 15.8 gb/s, 67.6 ulp |
|
|
324
336
|
| __f64__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
325
|
-
| `nk_dot_f64_serial` |
|
|
326
|
-
| `nk_dot_f64_neon` |
|
|
337
|
+
| `nk_dot_f64_serial` | 8.11 gb/s, 2.4 ulp | 8.13 gb/s, 175 ulp | 8.09 gb/s, 2.7 ulp |
|
|
338
|
+
| `nk_dot_f64_neon` | 44.2 gb/s, 0 ulp | 42.3 gb/s, 0 ulp | 38.4 gb/s, 0 ulp |
|
|
327
339
|
| __f32__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
328
|
-
| `nk_dot_f32_serial` |
|
|
329
|
-
| `nk_dot_f32_neon` |
|
|
340
|
+
| `nk_dot_f32_serial` | 23.3 gb/s, 0 ulp | 15.8 gb/s, 0 ulp | 14.6 gb/s, 0 ulp |
|
|
341
|
+
| `nk_dot_f32_neon` | 46.4 gb/s, 0 ulp | 38.0 gb/s, 0 ulp | 34.8 gb/s, 0 ulp |
|
|
330
342
|
| __bf16__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
331
|
-
| `nk_dot_bf16_serial` |
|
|
332
|
-
| `
|
|
343
|
+
| `nk_dot_bf16_serial` | 12.4 gb/s, 0 ulp | 8.59 gb/s, 0.9 ulp | 7.36 gb/s, 6 ulp |
|
|
344
|
+
| `nk_dot_bf16_neon` | 39.0 gb/s, 3.7 ulp | 27.2 gb/s, 3.7 ulp | 19.9 gb/s, 3.7 ulp |
|
|
345
|
+
| `nk_dot_bf16_neonbfdot` | 70.8 gb/s, 0 ulp | 60.8 gb/s, 0.6 ulp | 47.8 gb/s, 4.5 ulp |
|
|
333
346
|
| __f16__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
334
|
-
| `nk_dot_f16_serial` |
|
|
335
|
-
| `
|
|
336
|
-
| `nk_dot_f16_neonfhm` |
|
|
347
|
+
| `nk_dot_f16_serial` | 12.0 gb/s, 19 ulp | 8.33 gb/s, 31.1 ulp | 7.11 gb/s, 57.8 ulp |
|
|
348
|
+
| `nk_dot_f16_neon` | 35.7 gb/s, 33.4 ulp | 25.8 gb/s, 37.4 ulp | 21.3 gb/s, 23.1 ulp |
|
|
349
|
+
| `nk_dot_f16_neonfhm` | 48.7 gb/s, 14.9 ulp | 27.5 gb/s, 26.7 ulp | 18.8 gb/s, 39.9 ulp |
|
|
337
350
|
| __e5m2__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
338
|
-
| `nk_dot_e5m2_serial` |
|
|
339
|
-
| `nk_dot_e5m2_neon` |
|
|
351
|
+
| `nk_dot_e5m2_serial` | 3.80 gb/s, 0 ulp | 3.41 gb/s, 0 ulp | 3.41 gb/s, 0 ulp |
|
|
352
|
+
| `nk_dot_e5m2_neon` | 19.0 gb/s, 0 ulp | 13.2 gb/s, 0 ulp | 10.5 gb/s, 0 ulp |
|
|
353
|
+
| `nk_dot_e5m2_neonfhm` | 25.6 gb/s, 0 ulp | 15.3 gb/s, 0 ulp | 9.55 gb/s, 0 ulp |
|
|
354
|
+
| `nk_dot_e5m2_neonbfdot` | 3.65 gb/s, 0 ulp | 3.82 gb/s, 0 ulp | 3.68 gb/s, 0 ulp |
|
|
340
355
|
| __e4m3__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
341
|
-
| `nk_dot_e4m3_serial` |
|
|
342
|
-
| `nk_dot_e4m3_neon` |
|
|
356
|
+
| `nk_dot_e4m3_serial` | 1.74 gb/s, 0 ulp | 1.72 gb/s, 0 ulp | 1.71 gb/s, 0 ulp |
|
|
357
|
+
| `nk_dot_e4m3_neon` | 4.44 gb/s, 0 ulp | 4.51 gb/s, 0 ulp | 4.57 gb/s, 0 ulp |
|
|
358
|
+
| `nk_dot_e4m3_neonfhm` | 10.1 gb/s, 0 ulp | 8.51 gb/s, 0 ulp | 7.96 gb/s, 0 ulp |
|
|
359
|
+
| `nk_dot_e4m3_neonbfdot` | 3.59 gb/s, 0 ulp | 3.68 gb/s, 0 ulp | 3.64 gb/s, 0 ulp |
|
|
343
360
|
| __e3m2__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
344
|
-
| `nk_dot_e3m2_serial` |
|
|
345
|
-
| `nk_dot_e3m2_neonsdot` |
|
|
361
|
+
| `nk_dot_e3m2_serial` | 2.51 gb/s, 0 ulp | 2.33 gb/s, 0 ulp | 2.24 gb/s, 0 ulp |
|
|
362
|
+
| `nk_dot_e3m2_neonsdot` | 20.5 gb/s, 0 ulp | 20.7 gb/s, 0 ulp | 20.1 gb/s, 0 ulp |
|
|
346
363
|
| __e2m3__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
347
|
-
| `nk_dot_e2m3_serial` |
|
|
348
|
-
| `nk_dot_e2m3_neonsdot` |
|
|
364
|
+
| `nk_dot_e2m3_serial` | 2.54 gb/s, 0 ulp | 2.27 gb/s, 0 ulp | 2.29 gb/s, 0 ulp |
|
|
365
|
+
| `nk_dot_e2m3_neonsdot` | 47.3 gb/s, 0 ulp | 47.5 gb/s, 0 ulp | 43.4 gb/s, 0 ulp |
|
|
349
366
|
| __i8__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
350
|
-
| `nk_dot_i8_serial` |
|
|
351
|
-
| `nk_dot_i8_neonsdot` |
|
|
367
|
+
| `nk_dot_i8_serial` | 115 gb/s | 102 gb/s | 92.3 gb/s |
|
|
368
|
+
| `nk_dot_i8_neonsdot` | 92.8 gb/s | 87.4 gb/s | 59.9 gb/s |
|
|
352
369
|
| __u8__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
353
|
-
| `nk_dot_u8_serial` |
|
|
354
|
-
| `nk_dot_u8_neonsdot` |
|
|
370
|
+
| `nk_dot_u8_serial` | 110 gb/s | 99.2 gb/s | 94.9 gb/s |
|
|
371
|
+
| `nk_dot_u8_neonsdot` | 92.5 gb/s | 86.6 gb/s | 59.5 gb/s |
|
|
355
372
|
| __i4__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
356
|
-
| `nk_dot_i4_serial` |
|
|
357
|
-
| `nk_dot_i4_neonsdot` |
|
|
373
|
+
| `nk_dot_i4_serial` | 23.0 gb/s | 24.4 gb/s | 24.2 gb/s |
|
|
374
|
+
| `nk_dot_i4_neonsdot` | 58.2 gb/s | 44.7 gb/s | 30.4 gb/s |
|
|
358
375
|
| __u4__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
359
|
-
| `nk_dot_u4_serial` |
|
|
360
|
-
| `nk_dot_u4_neonsdot` |
|
|
376
|
+
| `nk_dot_u4_serial` | 25.3 gb/s | 27.2 gb/s | 26.9 gb/s |
|
|
377
|
+
| `nk_dot_u4_neonsdot` | 67.3 gb/s | 47.4 gb/s | 29.4 gb/s |
|
|
361
378
|
| __u1__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
362
|
-
| `nk_dot_u1_serial` |
|
|
363
|
-
| `nk_dot_u1_neon` |
|
|
379
|
+
| `nk_dot_u1_serial` | 7.01 gb/s | 7.63 gb/s | 7.19 gb/s |
|
|
380
|
+
| `nk_dot_u1_neon` | 33.3 gb/s | 64.6 gb/s | 88.0 gb/s |
|
|
364
381
|
|
|
365
382
|
#### WASM
|
|
366
383
|
|
|
367
|
-
Measured with Wasmtime
|
|
384
|
+
Measured with Wasmtime v43 (Cranelift backend).
|
|
368
385
|
|
|
369
386
|
| Kernel | 256 | 1024 | 4096 |
|
|
370
387
|
| :------------------------- | -----------------------: | -----------------------: | -----------------------: |
|
|
371
388
|
| __f64c__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
372
|
-
| `nk_dot_f64c_serial` |
|
|
373
|
-
| `nk_vdot_f64c_serial` |
|
|
374
|
-
| `nk_dot_f64c_v128relaxed` |
|
|
375
|
-
| `nk_vdot_f64c_v128relaxed` |
|
|
389
|
+
| `nk_dot_f64c_serial` | 4.17 gb/s, 3.8 ulp | 6.03 gb/s, 3.9 ulp | 6.37 gb/s, 3.2 ulp |
|
|
390
|
+
| `nk_vdot_f64c_serial` | 6.00 gb/s, 3.8 ulp | 6.55 gb/s, 3.4 ulp | 6.83 gb/s, 15.1 ulp |
|
|
391
|
+
| `nk_dot_f64c_v128relaxed` | 46.7 gb/s, 26 ulp | 38.2 gb/s, 42 ulp | 40.5 gb/s, 88 ulp |
|
|
392
|
+
| `nk_vdot_f64c_v128relaxed` | 46.1 gb/s, 22.8 ulp | 39.8 gb/s, 37.3 ulp | 39.9 gb/s, 43.6 ulp |
|
|
376
393
|
| __f32c__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
377
|
-
| `nk_dot_f32c_serial` |
|
|
378
|
-
| `nk_vdot_f32c_serial` |
|
|
379
|
-
| `nk_dot_f32c_v128relaxed` |
|
|
380
|
-
| `nk_vdot_f32c_v128relaxed` |
|
|
394
|
+
| `nk_dot_f32c_serial` | 20.9 gb/s, 0 ulp | 21.3 gb/s, 0 ulp | 22.5 gb/s, 0 ulp |
|
|
395
|
+
| `nk_vdot_f32c_serial` | 19.9 gb/s, 0 ulp | 21.5 gb/s, 0 ulp | 22.4 gb/s, 0 ulp |
|
|
396
|
+
| `nk_dot_f32c_v128relaxed` | 22.4 gb/s, 0 ulp | 20.3 gb/s, 0 ulp | 19.8 gb/s, 0 ulp |
|
|
397
|
+
| `nk_vdot_f32c_v128relaxed` | 21.9 gb/s, 0 ulp | 20.3 gb/s, 0 ulp | 19.9 gb/s, 0 ulp |
|
|
381
398
|
| __bf16c__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
382
|
-
| `nk_dot_bf16c_serial` |
|
|
383
|
-
| `nk_vdot_bf16c_serial` |
|
|
399
|
+
| `nk_dot_bf16c_serial` | 10.3 gb/s, 0.1 ulp | 11.6 gb/s, 2.5 ulp | 11.6 gb/s, 10 ulp |
|
|
400
|
+
| `nk_vdot_bf16c_serial` | 10.7 gb/s, 0.2 ulp | 11.7 gb/s, 2.1 ulp | 11.6 gb/s, 11.4 ulp |
|
|
384
401
|
| __f16c__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
385
|
-
| `nk_dot_f16c_serial` | 3.
|
|
386
|
-
| `nk_vdot_f16c_serial` |
|
|
402
|
+
| `nk_dot_f16c_serial` | 3.76 gb/s, 13 ulp | 3.83 gb/s, 20 ulp | 3.86 gb/s, 90 ulp |
|
|
403
|
+
| `nk_vdot_f16c_serial` | 3.81 gb/s, 13.9 ulp | 3.89 gb/s, 35.5 ulp | 3.85 gb/s, 42.4 ulp |
|
|
387
404
|
| __f64__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
388
|
-
| `nk_dot_f64_serial` |
|
|
389
|
-
| `nk_dot_f64_v128relaxed` |
|
|
405
|
+
| `nk_dot_f64_serial` | 7.26 gb/s, 2.4 ulp | 7.45 gb/s, 2.6 ulp | 7.96 gb/s, 2.2 ulp |
|
|
406
|
+
| `nk_dot_f64_v128relaxed` | 38.7 gb/s, 2.6 ulp | 42.0 gb/s, 3.2 ulp | 43.9 gb/s, 2.6 ulp |
|
|
390
407
|
| __f32__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
391
|
-
| `nk_dot_f32_serial` |
|
|
392
|
-
| `nk_dot_f32_v128relaxed` |
|
|
408
|
+
| `nk_dot_f32_serial` | 19.0 gb/s, 16 ulp | 14.6 gb/s, 69 ulp | 14.0 gb/s, 104 ulp |
|
|
409
|
+
| `nk_dot_f32_v128relaxed` | 20.4 gb/s, 0 ulp | 18.9 gb/s, 0 ulp | 18.7 gb/s, 0 ulp |
|
|
393
410
|
| __bf16__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
394
|
-
| `nk_dot_bf16_serial` |
|
|
395
|
-
| `nk_dot_bf16_v128relaxed` | 9
|
|
411
|
+
| `nk_dot_bf16_serial` | 9.53 gb/s, 0 ulp | 7.42 gb/s, 0.6 ulp | 7.20 gb/s, 5.9 ulp |
|
|
412
|
+
| `nk_dot_bf16_v128relaxed` | 41.9 gb/s, 0 ulp | 28.3 gb/s, 0.4 ulp | 21.5 gb/s, 3.7 ulp |
|
|
396
413
|
| __f16__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
397
|
-
| `nk_dot_f16_serial` | 3.
|
|
398
|
-
| `nk_dot_f16_v128relaxed` | 4
|
|
414
|
+
| `nk_dot_f16_serial` | 3.31 gb/s, 16 ulp | 3.63 gb/s, 26 ulp | 3.66 gb/s, 53 ulp |
|
|
415
|
+
| `nk_dot_f16_v128relaxed` | 11.4 gb/s, 9.0 ulp | 11.2 gb/s, 23 ulp | 12.0 gb/s, 39 ulp |
|
|
399
416
|
| __e5m2__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
400
|
-
| `nk_dot_e5m2_serial` |
|
|
401
|
-
| `nk_dot_e5m2_v128relaxed` | 3.
|
|
417
|
+
| `nk_dot_e5m2_serial` | 3.02 gb/s, 0 ulp | 2.95 gb/s, 0 ulp | 3.16 gb/s, 0 ulp |
|
|
418
|
+
| `nk_dot_e5m2_v128relaxed` | 3.47 gb/s, 0 ulp | 3.45 gb/s, 0 ulp | 3.48 gb/s, 0 ulp |
|
|
402
419
|
| __e4m3__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
403
|
-
| `nk_dot_e4m3_serial` | 0.
|
|
404
|
-
| `nk_dot_e4m3_v128relaxed` | 2.
|
|
420
|
+
| `nk_dot_e4m3_serial` | 0.978 gb/s, 0 ulp | 0.893 gb/s, 0 ulp | 0.936 gb/s, 0 ulp |
|
|
421
|
+
| `nk_dot_e4m3_v128relaxed` | 2.78 gb/s, 0 ulp | 2.75 gb/s, 0 ulp | 2.78 gb/s, 0 ulp |
|
|
405
422
|
| __e3m2__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
406
|
-
| `nk_dot_e3m2_serial` |
|
|
407
|
-
| `nk_dot_e3m2_v128relaxed` |
|
|
423
|
+
| `nk_dot_e3m2_serial` | 3.13 gb/s, 0 ulp | 2.95 gb/s, 0 ulp | 3.16 gb/s, 0 ulp |
|
|
424
|
+
| `nk_dot_e3m2_v128relaxed` | 12.1 gb/s, 0 ulp | 11.9 gb/s, 0 ulp | 12.6 gb/s, 0 ulp |
|
|
408
425
|
| __e2m3__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
409
|
-
| `nk_dot_e2m3_serial` | 2.
|
|
410
|
-
| `nk_dot_e2m3_v128relaxed` | 20.
|
|
426
|
+
| `nk_dot_e2m3_serial` | 2.99 gb/s, 0 ulp | 3.00 gb/s, 0 ulp | 3.17 gb/s, 0 ulp |
|
|
427
|
+
| `nk_dot_e2m3_v128relaxed` | 20.4 gb/s, 0 ulp | 20.6 gb/s, 0 ulp | 21.7 gb/s, 0 ulp |
|
|
411
428
|
| __i8__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
412
|
-
| `nk_dot_i8_serial` |
|
|
413
|
-
| `nk_dot_i8_v128relaxed` | 42.0 gb/s |
|
|
429
|
+
| `nk_dot_i8_serial` | 22.2 gb/s | 19.6 gb/s | 17.8 gb/s |
|
|
430
|
+
| `nk_dot_i8_v128relaxed` | 42.0 gb/s | 49.0 gb/s | 49.7 gb/s |
|
|
414
431
|
| __u8__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
415
|
-
| `nk_dot_u8_serial` |
|
|
416
|
-
| `nk_dot_u8_v128relaxed` |
|
|
432
|
+
| `nk_dot_u8_serial` | 23.0 gb/s | 19.9 gb/s | 17.8 gb/s |
|
|
433
|
+
| `nk_dot_u8_v128relaxed` | 29.3 gb/s | 33.0 gb/s | 35.1 gb/s |
|
|
417
434
|
| __i4__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
418
|
-
| `nk_dot_i4_serial` | 0.
|
|
419
|
-
| `nk_dot_i4_v128relaxed` |
|
|
435
|
+
| `nk_dot_i4_serial` | 0.990 gb/s | 0.923 gb/s | 0.985 gb/s |
|
|
436
|
+
| `nk_dot_i4_v128relaxed` | 15.2 gb/s | 17.9 gb/s | 19.2 gb/s |
|
|
420
437
|
| __u4__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
421
|
-
| `nk_dot_u4_serial` | 0.
|
|
422
|
-
| `nk_dot_u4_v128relaxed` | 30.
|
|
438
|
+
| `nk_dot_u4_serial` | 0.992 gb/s | 0.933 gb/s | 0.988 gb/s |
|
|
439
|
+
| `nk_dot_u4_v128relaxed` | 30.2 gb/s | 32.1 gb/s | 33.8 gb/s |
|
|
423
440
|
| __u1__ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ | â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘â–‘ |
|
|
424
|
-
| `nk_dot_u1_serial` |
|
|
425
|
-
| `nk_dot_u1_v128relaxed` | 21.
|
|
441
|
+
| `nk_dot_u1_serial` | 5.26 gb/s | 5.80 gb/s | 6.48 gb/s |
|
|
442
|
+
| `nk_dot_u1_v128relaxed` | 21.2 gb/s | 47.4 gb/s | 67.3 gb/s |
|