numkong 7.0.0 → 7.4.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +197 -124
- package/binding.gyp +34 -484
- 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
|
@@ -1,25 +1,25 @@
|
|
|
1
1
|
# Spatial Distances in NumKong
|
|
2
2
|
|
|
3
3
|
NumKong implements spatial distance functions for dense vectors: squared Euclidean distance, Euclidean distance, and angular (cosine) distance.
|
|
4
|
-
These are
|
|
4
|
+
These metrics are commonly used in nearest-neighbor search, clustering, and dimensionality reduction, and are implemented for every numeric type supported by the library.
|
|
5
5
|
|
|
6
6
|
Squared Euclidean distance measures the sum of squared element-wise differences:
|
|
7
7
|
|
|
8
|
-
|
|
8
|
+
$$
|
|
9
9
|
\text{sqeuclidean}(a, b) = \sum_{i=0}^{n-1} (a_i - b_i)^2
|
|
10
|
-
|
|
10
|
+
$$
|
|
11
11
|
|
|
12
12
|
Euclidean distance is the square root of the squared Euclidean distance:
|
|
13
13
|
|
|
14
|
-
|
|
14
|
+
$$
|
|
15
15
|
\text{euclidean}(a, b) = \sqrt{\sum_{i=0}^{n-1} (a_i - b_i)^2}
|
|
16
|
-
|
|
16
|
+
$$
|
|
17
17
|
|
|
18
18
|
Angular distance (cosine distance) measures the angle between two vectors:
|
|
19
19
|
|
|
20
|
-
|
|
20
|
+
$$
|
|
21
21
|
\text{angular}(a, b) = 1 - \frac{\sum_{i=0}^{n-1} a_i \cdot b_i}{\sqrt{\sum_{i=0}^{n-1} a_i^2} \cdot \sqrt{\sum_{i=0}^{n-1} b_i^2}}
|
|
22
|
-
|
|
22
|
+
$$
|
|
23
23
|
|
|
24
24
|
Reformulating as Python pseudocode:
|
|
25
25
|
|
|
@@ -111,127 +111,125 @@ Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run
|
|
|
111
111
|
|
|
112
112
|
#### Native
|
|
113
113
|
|
|
114
|
-
| Kernel
|
|
115
|
-
|
|
|
116
|
-
| __f64__
|
|
117
|
-
| `nk_sqeuclidean_f64_serial`
|
|
118
|
-
| `nk_euclidean_f64_serial`
|
|
119
|
-
| `nk_angular_f64_serial`
|
|
120
|
-
| `nk_sqeuclidean_f64_skylake`
|
|
121
|
-
| `nk_euclidean_f64_skylake`
|
|
122
|
-
| `nk_angular_f64_skylake`
|
|
123
|
-
| __f32__
|
|
124
|
-
| `nk_sqeuclidean_f32_serial`
|
|
125
|
-
| `nk_euclidean_f32_serial`
|
|
126
|
-
| `nk_angular_f32_serial`
|
|
127
|
-
| `nk_sqeuclidean_f32_skylake`
|
|
128
|
-
| `nk_euclidean_f32_skylake`
|
|
129
|
-
| `nk_angular_f32_skylake`
|
|
130
|
-
| __bf16__
|
|
131
|
-
| `nk_sqeuclidean_bf16_serial`
|
|
132
|
-
| `nk_euclidean_bf16_serial`
|
|
133
|
-
| `nk_angular_bf16_serial`
|
|
134
|
-
| `nk_sqeuclidean_bf16_haswell`
|
|
135
|
-
| `nk_euclidean_bf16_haswell`
|
|
136
|
-
| `nk_angular_bf16_haswell`
|
|
137
|
-
| `nk_sqeuclidean_bf16_genoa`
|
|
138
|
-
| `nk_euclidean_bf16_genoa`
|
|
139
|
-
| `nk_angular_bf16_genoa`
|
|
140
|
-
| __f16__
|
|
141
|
-
| `nk_sqeuclidean_f16_serial`
|
|
142
|
-
| `nk_euclidean_f16_serial`
|
|
143
|
-
| `nk_angular_f16_serial`
|
|
144
|
-
| `nk_sqeuclidean_f16_haswell`
|
|
145
|
-
| `nk_euclidean_f16_haswell`
|
|
146
|
-
| `nk_angular_f16_haswell`
|
|
147
|
-
| __e5m2__
|
|
148
|
-
| `nk_sqeuclidean_e5m2_serial`
|
|
149
|
-
| `nk_euclidean_e5m2_serial`
|
|
150
|
-
| `nk_angular_e5m2_serial`
|
|
151
|
-
| `nk_sqeuclidean_e5m2_skylake`
|
|
152
|
-
| `nk_euclidean_e5m2_skylake`
|
|
153
|
-
| `nk_angular_e5m2_skylake`
|
|
154
|
-
| `nk_sqeuclidean_e5m2_genoa`
|
|
155
|
-
| `nk_euclidean_e5m2_genoa`
|
|
156
|
-
| `nk_angular_e5m2_genoa`
|
|
157
|
-
| __e4m3__
|
|
158
|
-
| `nk_sqeuclidean_e4m3_serial`
|
|
159
|
-
| `nk_euclidean_e4m3_serial`
|
|
160
|
-
| `nk_angular_e4m3_serial`
|
|
161
|
-
| `nk_sqeuclidean_e4m3_skylake`
|
|
162
|
-
| `nk_euclidean_e4m3_skylake`
|
|
163
|
-
| `nk_angular_e4m3_skylake`
|
|
164
|
-
| `
|
|
165
|
-
| `
|
|
166
|
-
| `
|
|
167
|
-
|
|
|
168
|
-
| `
|
|
169
|
-
|
|
|
170
|
-
| `
|
|
171
|
-
| `
|
|
172
|
-
| `
|
|
173
|
-
| `
|
|
174
|
-
| `
|
|
175
|
-
| `
|
|
176
|
-
| `
|
|
177
|
-
| `
|
|
178
|
-
| `
|
|
179
|
-
| `
|
|
180
|
-
|
|
|
181
|
-
| `
|
|
182
|
-
|
|
|
183
|
-
| `
|
|
184
|
-
| `
|
|
185
|
-
| `
|
|
186
|
-
| `
|
|
187
|
-
| `
|
|
188
|
-
| `
|
|
189
|
-
| `
|
|
190
|
-
| `
|
|
191
|
-
| `
|
|
192
|
-
| `
|
|
193
|
-
|
|
|
194
|
-
| `
|
|
195
|
-
|
|
|
196
|
-
| `
|
|
197
|
-
| `
|
|
198
|
-
| `
|
|
199
|
-
| `
|
|
200
|
-
| `
|
|
201
|
-
| `
|
|
202
|
-
| `
|
|
203
|
-
| `
|
|
204
|
-
| `
|
|
205
|
-
| `
|
|
206
|
-
|
|
|
207
|
-
| `
|
|
208
|
-
|
|
|
209
|
-
| `
|
|
210
|
-
| `
|
|
211
|
-
| `
|
|
212
|
-
| `
|
|
213
|
-
| `
|
|
214
|
-
| `
|
|
215
|
-
| `
|
|
216
|
-
| `
|
|
217
|
-
| `
|
|
218
|
-
| `
|
|
219
|
-
|
|
|
220
|
-
| `
|
|
221
|
-
|
|
|
222
|
-
| `
|
|
223
|
-
| `
|
|
224
|
-
| `
|
|
225
|
-
| `
|
|
226
|
-
|
|
|
227
|
-
| `
|
|
228
|
-
|
|
|
229
|
-
| `
|
|
230
|
-
| `
|
|
231
|
-
| `
|
|
232
|
-
| `
|
|
233
|
-
| `nk_euclidean_u4_icelake` | 20.9 gb/s, 0 ulp | 18.8 gb/s, 0 ulp | 24.1 gb/s, 0 ulp |
|
|
234
|
-
| `nk_angular_u4_icelake` | 9.32 gb/s, 0.7 ulp | 27.4 gb/s, 0.6 ulp | 24.2 gb/s, 0.5 ulp |
|
|
114
|
+
| Kernel | 256 | 1024 | 4096 |
|
|
115
|
+
| :---------------------------- | -----------------------: | -----------------------: | -----------------------: |
|
|
116
|
+
| __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
117
|
+
| `nk_sqeuclidean_f64_serial` | 8.00 gb/s, 0.1 ulp | 8.32 gb/s, 0 ulp | 8.13 gb/s, 0 ulp |
|
|
118
|
+
| `nk_euclidean_f64_serial` | 7.81 gb/s, 0.6 ulp | 7.95 gb/s, 0.5 ulp | 8.34 gb/s, 0.5 ulp |
|
|
119
|
+
| `nk_angular_f64_serial` | 2.80 gb/s, 0 ulp | 3.03 gb/s, 0 ulp | 3.18 gb/s, 0 ulp |
|
|
120
|
+
| `nk_sqeuclidean_f64_skylake` | 32.4 gb/s, 0.4 ulp | 30.6 gb/s, 0.7 ulp | 22.2 gb/s, 1.3 ulp |
|
|
121
|
+
| `nk_euclidean_f64_skylake` | 31.7 gb/s, 0.3 ulp | 29.4 gb/s, 0.4 ulp | 22.9 gb/s, 0.7 ulp |
|
|
122
|
+
| `nk_angular_f64_skylake` | 26.5 gb/s, 0 ulp | 26.8 gb/s, 0 ulp | 17.8 gb/s, 0 ulp |
|
|
123
|
+
| __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
124
|
+
| `nk_sqeuclidean_f32_serial` | 4.01 gb/s, 0 ulp | 4.06 gb/s, 0 ulp | 4.19 gb/s, 0 ulp |
|
|
125
|
+
| `nk_euclidean_f32_serial` | 3.99 gb/s, 0.1 ulp | 4.07 gb/s, 0.1 ulp | 4.11 gb/s, 0.1 ulp |
|
|
126
|
+
| `nk_angular_f32_serial` | 1.29 gb/s, 0 ulp | 1.41 gb/s, 0 ulp | 1.53 gb/s, 0 ulp |
|
|
127
|
+
| `nk_sqeuclidean_f32_skylake` | 36.5 gb/s, 0 ulp | 27.0 gb/s, 0 ulp | 23.2 gb/s, 0 ulp |
|
|
128
|
+
| `nk_euclidean_f32_skylake` | 36.4 gb/s, 0.1 ulp | 28.1 gb/s, 0.1 ulp | 26.7 gb/s, 0.1 ulp |
|
|
129
|
+
| `nk_angular_f32_skylake` | 24.3 gb/s, 0 ulp | 23.2 gb/s, 0 ulp | 22.5 gb/s, 0 ulp |
|
|
130
|
+
| __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
131
|
+
| `nk_sqeuclidean_bf16_serial` | 0.582 gb/s, 0 ulp | 0.358 gb/s, 0 ulp | 0.390 gb/s, 0 ulp |
|
|
132
|
+
| `nk_euclidean_bf16_serial` | 0.569 gb/s, 0.5 ulp | 0.373 gb/s, 0.5 ulp | 0.372 gb/s, 0.4 ulp |
|
|
133
|
+
| `nk_angular_bf16_serial` | 0.455 gb/s, 0 ulp | 0.241 gb/s, 0 ulp | 0.259 gb/s, 0 ulp |
|
|
134
|
+
| `nk_sqeuclidean_bf16_haswell` | 27.7 gb/s, 0.5 ulp | 14.0 gb/s, 7.5 ulp | 11.8 gb/s, 27 ulp |
|
|
135
|
+
| `nk_euclidean_bf16_haswell` | 23.3 gb/s, 0.3 ulp | 13.4 gb/s, 4.1 ulp | 12.0 gb/s, 15 ulp |
|
|
136
|
+
| `nk_angular_bf16_haswell` | 20.1 gb/s, 0 ulp | 13.4 gb/s, 0 ulp | 10.6 gb/s, 0.2 ulp |
|
|
137
|
+
| `nk_sqeuclidean_bf16_genoa` | 50.1 gb/s, 0.3 ulp | 21.0 gb/s, 0.5 ulp | 20.5 gb/s, 10 ulp |
|
|
138
|
+
| `nk_euclidean_bf16_genoa` | 48.3 gb/s, 0.2 ulp | 23.1 gb/s, 0.3 ulp | 20.4 gb/s, 5.8 ulp |
|
|
139
|
+
| `nk_angular_bf16_genoa` | 36.4 gb/s, 0 ulp | 22.4 gb/s, 0 ulp | 21.0 gb/s, 0.1 ulp |
|
|
140
|
+
| __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
141
|
+
| `nk_sqeuclidean_f16_serial` | 0.950 gb/s, 0.1 ulp | 0.872 gb/s, 0.1 ulp | 0.864 gb/s, 0.1 ulp |
|
|
142
|
+
| `nk_euclidean_f16_serial` | 0.934 gb/s, 0.5 ulp | 0.913 gb/s, 0.5 ulp | 0.906 gb/s, 0.5 ulp |
|
|
143
|
+
| `nk_angular_f16_serial` | 0.881 gb/s, 0 ulp | 0.531 gb/s, 0 ulp | 0.543 gb/s, 0 ulp |
|
|
144
|
+
| `nk_sqeuclidean_f16_haswell` | 29.8 gb/s, 0.4 ulp | 14.8 gb/s, 1.4 ulp | 11.8 gb/s, 5.2 ulp |
|
|
145
|
+
| `nk_euclidean_f16_haswell` | 22.9 gb/s, 0.3 ulp | 12.9 gb/s, 0.8 ulp | 10.6 gb/s, 2.8 ulp |
|
|
146
|
+
| `nk_angular_f16_haswell` | 19.9 gb/s, 0.1 ulp | 17.5 gb/s, 0.1 ulp | 16.1 gb/s, 0.1 ulp |
|
|
147
|
+
| __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
148
|
+
| `nk_sqeuclidean_e5m2_serial` | 0.955 gb/s, 0 ulp | 1.01 gb/s, 0 ulp | 1.02 gb/s, 0 ulp |
|
|
149
|
+
| `nk_euclidean_e5m2_serial` | 0.954 gb/s, 0.5 ulp | 0.985 gb/s, 0.5 ulp | 1.03 gb/s, 0.5 ulp |
|
|
150
|
+
| `nk_angular_e5m2_serial` | 0.336 gb/s, 0 ulp | 0.385 gb/s, 0 ulp | 0.407 gb/s, 0 ulp |
|
|
151
|
+
| `nk_sqeuclidean_e5m2_skylake` | 4.44 gb/s, 0 ulp | 4.65 gb/s, 0 ulp | 5.80 gb/s, 0 ulp |
|
|
152
|
+
| `nk_euclidean_e5m2_skylake` | 4.34 gb/s, 0 ulp | 4.65 gb/s, 0 ulp | 5.88 gb/s, 0 ulp |
|
|
153
|
+
| `nk_angular_e5m2_skylake` | 3.83 gb/s, 0 ulp | 4.39 gb/s, 0 ulp | 6.10 gb/s, 0 ulp |
|
|
154
|
+
| `nk_sqeuclidean_e5m2_genoa` | 7.12 gb/s, 0 ulp | 8.07 gb/s, 0 ulp | 8.05 gb/s, 0 ulp |
|
|
155
|
+
| `nk_euclidean_e5m2_genoa` | 7.01 gb/s, 0 ulp | 6.97 gb/s, 0 ulp | 8.16 gb/s, 0 ulp |
|
|
156
|
+
| `nk_angular_e5m2_genoa` | 6.33 gb/s, 0 ulp | 6.79 gb/s, 0 ulp | 7.99 gb/s, 0 ulp |
|
|
157
|
+
| __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
158
|
+
| `nk_sqeuclidean_e4m3_serial` | 0.569 gb/s, 0 ulp | 0.606 gb/s, 0 ulp | 0.609 gb/s, 0 ulp |
|
|
159
|
+
| `nk_euclidean_e4m3_serial` | 0.587 gb/s, 0.5 ulp | 0.602 gb/s, 0.5 ulp | 0.578 gb/s, 0.5 ulp |
|
|
160
|
+
| `nk_angular_e4m3_serial` | 0.326 gb/s, 0 ulp | 0.196 gb/s, 0 ulp | 0.366 gb/s, 0 ulp |
|
|
161
|
+
| `nk_sqeuclidean_e4m3_skylake` | 3.84 gb/s, 0 ulp | 3.62 gb/s, 0 ulp | 3.95 gb/s, 0.2 ulp |
|
|
162
|
+
| `nk_euclidean_e4m3_skylake` | 3.48 gb/s, 0 ulp | 3.69 gb/s, 0 ulp | 3.33 gb/s, 0.2 ulp |
|
|
163
|
+
| `nk_angular_e4m3_skylake` | 4.22 gb/s, 0 ulp | 3.38 gb/s, 0 ulp | 4.54 gb/s, 0 ulp |
|
|
164
|
+
| `nk_sqeuclidean_e4m3_icelake` | 10.2 gb/s, 0 ulp | 12.0 gb/s, 0 ulp | 12.0 gb/s, 0.2 ulp |
|
|
165
|
+
| `nk_euclidean_e4m3_icelake` | 10.3 gb/s, 0 ulp | 11.8 gb/s, 0 ulp | 11.9 gb/s, 0.2 ulp |
|
|
166
|
+
| `nk_angular_e4m3_icelake` | 8.78 gb/s, 0 ulp | 11.3 gb/s, 0 ulp | 11.9 gb/s, 0 ulp |
|
|
167
|
+
| __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
168
|
+
| `nk_sqeuclidean_e3m2_serial` | 1.01 gb/s, 0 ulp | 0.971 gb/s, 0 ulp | 1.03 gb/s, 0 ulp |
|
|
169
|
+
| `nk_euclidean_e3m2_serial` | 0.997 gb/s, 0.5 ulp | 0.990 gb/s, 0.5 ulp | 0.999 gb/s, 0.4 ulp |
|
|
170
|
+
| `nk_angular_e3m2_serial` | 0.332 gb/s, 0 ulp | 0.361 gb/s, 0 ulp | 0.437 gb/s, 0 ulp |
|
|
171
|
+
| `nk_sqeuclidean_e3m2_skylake` | 4.47 gb/s, 0 ulp | 5.46 gb/s, 0 ulp | 5.04 gb/s, 0 ulp |
|
|
172
|
+
| `nk_euclidean_e3m2_skylake` | 4.34 gb/s, 0 ulp | 6.20 gb/s, 0 ulp | 5.10 gb/s, 0 ulp |
|
|
173
|
+
| `nk_angular_e3m2_skylake` | 3.79 gb/s, 0 ulp | 4.41 gb/s, 0 ulp | 4.82 gb/s, 0 ulp |
|
|
174
|
+
| `nk_sqeuclidean_e3m2_genoa` | 8.79 gb/s, 0 ulp | 9.52 gb/s, 0 ulp | 10.6 gb/s, 0 ulp |
|
|
175
|
+
| `nk_euclidean_e3m2_genoa` | 8.68 gb/s, 0 ulp | 9.01 gb/s, 0 ulp | 12.8 gb/s, 0 ulp |
|
|
176
|
+
| `nk_angular_e3m2_genoa` | 6.89 gb/s, 0 ulp | 9.30 gb/s, 0 ulp | 10.3 gb/s, 0 ulp |
|
|
177
|
+
| `nk_sqeuclidean_e3m2_icelake` | 21.2 gb/s, 0 ulp | 22.1 gb/s, 0 ulp | 21.9 gb/s, 0 ulp |
|
|
178
|
+
| `nk_euclidean_e3m2_icelake` | 21.2 gb/s, 0 ulp | 22.9 gb/s, 0 ulp | 21.2 gb/s, 0 ulp |
|
|
179
|
+
| `nk_angular_e3m2_icelake` | 14.1 gb/s, 0 ulp | 18.0 gb/s, 0 ulp | 17.6 gb/s, 0 ulp |
|
|
180
|
+
| __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
181
|
+
| `nk_sqeuclidean_e2m3_serial` | 0.964 gb/s, 0 ulp | 0.981 gb/s, 0 ulp | 1.03 gb/s, 0 ulp |
|
|
182
|
+
| `nk_euclidean_e2m3_serial` | 0.979 gb/s, 0.5 ulp | 0.966 gb/s, 0.5 ulp | 1.02 gb/s, 0.5 ulp |
|
|
183
|
+
| `nk_angular_e2m3_serial` | 0.347 gb/s, 0 ulp | 0.389 gb/s, 0 ulp | 0.418 gb/s, 0 ulp |
|
|
184
|
+
| `nk_sqeuclidean_e2m3_skylake` | 4.58 gb/s, 0 ulp | 4.65 gb/s, 0 ulp | 5.08 gb/s, 0 ulp |
|
|
185
|
+
| `nk_euclidean_e2m3_skylake` | 4.48 gb/s, 0 ulp | 4.39 gb/s, 0 ulp | 4.96 gb/s, 0 ulp |
|
|
186
|
+
| `nk_angular_e2m3_skylake` | 3.94 gb/s, 0 ulp | 4.25 gb/s, 0 ulp | 4.90 gb/s, 0 ulp |
|
|
187
|
+
| `nk_sqeuclidean_e2m3_genoa` | 9.62 gb/s, 0 ulp | 10.9 gb/s, 0 ulp | 10.8 gb/s, 0 ulp |
|
|
188
|
+
| `nk_euclidean_e2m3_genoa` | 8.45 gb/s, 0 ulp | 9.80 gb/s, 0 ulp | 10.3 gb/s, 0 ulp |
|
|
189
|
+
| `nk_angular_e2m3_genoa` | 7.21 gb/s, 0 ulp | 10.1 gb/s, 0 ulp | 10.4 gb/s, 0 ulp |
|
|
190
|
+
| `nk_sqeuclidean_e2m3_icelake` | 50.7 gb/s, 0 ulp | 42.6 gb/s, 0 ulp | 31.0 gb/s, 0 ulp |
|
|
191
|
+
| `nk_euclidean_e2m3_icelake` | 50.2 gb/s, 0 ulp | 44.3 gb/s, 0 ulp | 31.0 gb/s, 0 ulp |
|
|
192
|
+
| `nk_angular_e2m3_icelake` | 27.2 gb/s, 0 ulp | 34.9 gb/s, 0 ulp | 30.5 gb/s, 0 ulp |
|
|
193
|
+
| __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
194
|
+
| `nk_sqeuclidean_i8_serial` | 34.0 gb/s | 18.4 gb/s | 16.5 gb/s |
|
|
195
|
+
| `nk_euclidean_i8_serial` | 29.0 gb/s, 0.4 ulp | 18.0 gb/s, 0.4 ulp | 15.6 gb/s, 0.4 ulp |
|
|
196
|
+
| `nk_angular_i8_serial` | 7.88 gb/s, 0 ulp | 6.31 gb/s, 0 ulp | 6.12 gb/s, 0 ulp |
|
|
197
|
+
| `nk_sqeuclidean_i8_haswell` | 38.4 gb/s | 17.9 gb/s | 18.4 gb/s |
|
|
198
|
+
| `nk_euclidean_i8_haswell` | 35.6 gb/s, 0 ulp | 17.0 gb/s, 0 ulp | 15.5 gb/s, 0 ulp |
|
|
199
|
+
| `nk_angular_i8_haswell` | 20.3 gb/s, 0.1 ulp | 12.9 gb/s, 0 ulp | 11.9 gb/s, 0 ulp |
|
|
200
|
+
| `nk_sqeuclidean_i8_icelake` | 60.2 gb/s | 24.5 gb/s | 23.5 gb/s |
|
|
201
|
+
| `nk_euclidean_i8_icelake` | 59.0 gb/s, 0 ulp | 23.0 gb/s, 0 ulp | 22.3 gb/s, 0 ulp |
|
|
202
|
+
| `nk_angular_i8_icelake` | 25.2 gb/s, 0.1 ulp | 18.4 gb/s, 0 ulp | 20.5 gb/s, 0 ulp |
|
|
203
|
+
| `nk_sqeuclidean_i8_alder` | 33.4 gb/s | 17.4 gb/s | 17.6 gb/s |
|
|
204
|
+
| `nk_euclidean_i8_alder` | 31.9 gb/s, 0 ulp | 19.1 gb/s, 0 ulp | 17.8 gb/s, 0 ulp |
|
|
205
|
+
| `nk_angular_i8_alder` | 26.2 gb/s, 0.1 ulp | 17.1 gb/s, 0 ulp | 17.8 gb/s, 0 ulp |
|
|
206
|
+
| __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
207
|
+
| `nk_sqeuclidean_u8_serial` | 11.7 gb/s | 8.77 gb/s | 7.07 gb/s |
|
|
208
|
+
| `nk_euclidean_u8_serial` | 11.6 gb/s, 0.5 ulp | 8.31 gb/s, 0.5 ulp | 8.36 gb/s, 0.6 ulp |
|
|
209
|
+
| `nk_angular_u8_serial` | 7.95 gb/s, 0.4 ulp | 6.68 gb/s, 0.4 ulp | 5.88 gb/s, 0.4 ulp |
|
|
210
|
+
| `nk_sqeuclidean_u8_haswell` | 45.4 gb/s | 17.7 gb/s | 18.5 gb/s |
|
|
211
|
+
| `nk_euclidean_u8_haswell` | 38.9 gb/s, 0 ulp | 18.8 gb/s, 0 ulp | 19.3 gb/s, 0 ulp |
|
|
212
|
+
| `nk_angular_u8_haswell` | 21.9 gb/s, 0.7 ulp | 11.7 gb/s, 0.6 ulp | 13.4 gb/s, 0.5 ulp |
|
|
213
|
+
| `nk_sqeuclidean_u8_icelake` | 70.1 gb/s | 28.8 gb/s | 21.0 gb/s |
|
|
214
|
+
| `nk_euclidean_u8_icelake` | 66.4 gb/s, 0 ulp | 27.6 gb/s, 0 ulp | 23.5 gb/s, 0 ulp |
|
|
215
|
+
| `nk_angular_u8_icelake` | 28.9 gb/s, 0.7 ulp | 21.2 gb/s, 0.6 ulp | 21.5 gb/s, 0.5 ulp |
|
|
216
|
+
| `nk_sqeuclidean_u8_alder` | 32.2 gb/s | 17.5 gb/s | 19.0 gb/s |
|
|
217
|
+
| `nk_euclidean_u8_alder` | 31.3 gb/s, 0 ulp | 17.0 gb/s, 0 ulp | 19.6 gb/s, 0 ulp |
|
|
218
|
+
| `nk_angular_u8_alder` | 26.5 gb/s, 0.7 ulp | 17.1 gb/s, 0.6 ulp | 17.5 gb/s, 0.5 ulp |
|
|
219
|
+
| __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
220
|
+
| `nk_sqeuclidean_i4_serial` | 15.4 gb/s | 16.5 gb/s | 15.6 gb/s |
|
|
221
|
+
| `nk_euclidean_i4_serial` | 12.2 gb/s, 0.5 ulp | 15.6 gb/s, 0.5 ulp | 15.2 gb/s, 0.6 ulp |
|
|
222
|
+
| `nk_angular_i4_serial` | 5.60 gb/s, 0.4 ulp | 6.42 gb/s, 0.4 ulp | 6.69 gb/s, 0.4 ulp |
|
|
223
|
+
| `nk_sqeuclidean_i4_icelake` | 23.6 gb/s | 51.5 gb/s | 29.3 gb/s |
|
|
224
|
+
| `nk_euclidean_i4_icelake` | 20.6 gb/s, 0 ulp | 45.2 gb/s, 0 ulp | 28.9 gb/s, 0 ulp |
|
|
225
|
+
| `nk_angular_i4_icelake` | 5.14 gb/s, 0.7 ulp | 18.0 gb/s, 0.6 ulp | 17.6 gb/s, 0.5 ulp |
|
|
226
|
+
| __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
227
|
+
| `nk_sqeuclidean_u4_serial` | 15.6 gb/s | 17.3 gb/s | 15.8 gb/s |
|
|
228
|
+
| `nk_euclidean_u4_serial` | 12.0 gb/s, 0.5 ulp | 15.9 gb/s, 0.5 ulp | 15.3 gb/s, 0.6 ulp |
|
|
229
|
+
| `nk_angular_u4_serial` | 5.20 gb/s, 0.4 ulp | 6.63 gb/s, 0.4 ulp | 7.01 gb/s, 0.4 ulp |
|
|
230
|
+
| `nk_sqeuclidean_u4_icelake` | 22.7 gb/s | 23.7 gb/s | 24.5 gb/s |
|
|
231
|
+
| `nk_euclidean_u4_icelake` | 20.9 gb/s, 0 ulp | 18.8 gb/s, 0 ulp | 24.1 gb/s, 0 ulp |
|
|
232
|
+
| `nk_angular_u4_icelake` | 9.32 gb/s, 0.7 ulp | 27.4 gb/s, 0.6 ulp | 24.2 gb/s, 0.5 ulp |
|
|
235
233
|
|
|
236
234
|
#### WASM
|
|
237
235
|
|
|
@@ -306,130 +304,136 @@ Measured with Wasmtime v42 (Cranelift backend).
|
|
|
306
304
|
| `nk_euclidean_u4_serial` | 2.78 gb/s, 0.5 ulp | 3.01 gb/s, 0.5 ulp | 0.10 gb/s, 0.0 ulp |
|
|
307
305
|
| `nk_angular_u4_serial` | 1.84 gb/s, 0.5 ulp | 2.03 gb/s, 0.5 ulp | 0.21 gb/s, 0.5 ulp |
|
|
308
306
|
|
|
309
|
-
### Apple
|
|
307
|
+
### Apple M5
|
|
310
308
|
|
|
311
309
|
#### Native
|
|
312
310
|
|
|
313
311
|
| Kernel | 256 | 1024 | 4096 |
|
|
314
312
|
| :------------------------------ | -----------------------: | -----------------------: | -----------------------: |
|
|
315
313
|
| __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
316
|
-
| `nk_sqeuclidean_f64_serial` | 12.
|
|
317
|
-
| `nk_euclidean_f64_serial` | 12.
|
|
318
|
-
| `nk_angular_f64_serial` | 8.
|
|
319
|
-
| `nk_sqeuclidean_f64_neon` |
|
|
320
|
-
| `nk_euclidean_f64_neon` |
|
|
321
|
-
| `nk_angular_f64_neon` |
|
|
314
|
+
| `nk_sqeuclidean_f64_serial` | 12.4 gb/s, 0.1 ulp | 12.8 gb/s, 0 ulp | 12.8 gb/s, 0 ulp |
|
|
315
|
+
| `nk_euclidean_f64_serial` | 12.7 gb/s, 0.6 ulp | 12.9 gb/s, 0.5 ulp | 12.6 gb/s, 0.5 ulp |
|
|
316
|
+
| `nk_angular_f64_serial` | 8.42 gb/s, 0 ulp | 8.57 gb/s, 0 ulp | 8.30 gb/s, 0 ulp |
|
|
317
|
+
| `nk_sqeuclidean_f64_neon` | 50.6 gb/s, 1.3 ulp | 40.0 gb/s, 2.6 ulp | 36.1 gb/s, 5.1 ulp |
|
|
318
|
+
| `nk_euclidean_f64_neon` | 48.4 gb/s, 0.7 ulp | 38.7 gb/s, 1.4 ulp | 35.1 gb/s, 2.8 ulp |
|
|
319
|
+
| `nk_angular_f64_neon` | 33.3 gb/s, 0.1 ulp | 33.4 gb/s, 0 ulp | 32.4 gb/s, 0 ulp |
|
|
322
320
|
| __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
323
|
-
| `nk_sqeuclidean_f32_serial` | 6.
|
|
324
|
-
| `nk_euclidean_f32_serial` | 6.
|
|
325
|
-
| `nk_angular_f32_serial` | 4.
|
|
326
|
-
| `nk_sqeuclidean_f32_neon` |
|
|
327
|
-
| `nk_euclidean_f32_neon` |
|
|
328
|
-
| `nk_angular_f32_neon` |
|
|
321
|
+
| `nk_sqeuclidean_f32_serial` | 6.32 gb/s, 0 ulp | 6.25 gb/s, 0 ulp | 6.30 gb/s, 0 ulp |
|
|
322
|
+
| `nk_euclidean_f32_serial` | 6.31 gb/s, 0.1 ulp | 6.37 gb/s, 0.1 ulp | 6.41 gb/s, 0.1 ulp |
|
|
323
|
+
| `nk_angular_f32_serial` | 4.03 gb/s, 0 ulp | 4.06 gb/s, 0 ulp | 4.07 gb/s, 0 ulp |
|
|
324
|
+
| `nk_sqeuclidean_f32_neon` | 25.3 gb/s, 0.1 ulp | 19.1 gb/s, 0 ulp | 17.5 gb/s, 0 ulp |
|
|
325
|
+
| `nk_euclidean_f32_neon` | 25.0 gb/s, 0.1 ulp | 20.8 gb/s, 0.1 ulp | 18.6 gb/s, 0.1 ulp |
|
|
326
|
+
| `nk_angular_f32_neon` | 22.2 gb/s, 0 ulp | 17.3 gb/s, 0 ulp | 16.6 gb/s, 0 ulp |
|
|
329
327
|
| __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
330
|
-
| `nk_sqeuclidean_bf16_serial` | 3.19 gb/s, 0 ulp |
|
|
331
|
-
| `nk_euclidean_bf16_serial` | 3.
|
|
332
|
-
| `nk_angular_bf16_serial` | 1.
|
|
333
|
-
| `nk_sqeuclidean_bf16_neonbfdot` |
|
|
334
|
-
| `nk_euclidean_bf16_neonbfdot` |
|
|
335
|
-
| `nk_angular_bf16_neonbfdot` |
|
|
328
|
+
| `nk_sqeuclidean_bf16_serial` | 3.19 gb/s, 0 ulp | 3.16 gb/s, 0 ulp | 3.14 gb/s, 0 ulp |
|
|
329
|
+
| `nk_euclidean_bf16_serial` | 3.16 gb/s, 0.5 ulp | 3.08 gb/s, 0.5 ulp | 3.11 gb/s, 0.5 ulp |
|
|
330
|
+
| `nk_angular_bf16_serial` | 1.88 gb/s, 0 ulp | 1.91 gb/s, 0 ulp | 1.93 gb/s, 0 ulp |
|
|
331
|
+
| `nk_sqeuclidean_bf16_neonbfdot` | 35.0 gb/s, 0.9 ulp | 22.7 gb/s, 13 ulp | 18.8 gb/s, 21 ulp |
|
|
332
|
+
| `nk_euclidean_bf16_neonbfdot` | 33.4 gb/s, 0.5 ulp | 23.0 gb/s, 7.0 ulp | 18.6 gb/s, 12 ulp |
|
|
333
|
+
| `nk_angular_bf16_neonbfdot` | 23.8 gb/s, 0 ulp | 32.7 gb/s, 0.1 ulp | 35.9 gb/s, 0 ulp |
|
|
336
334
|
| __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
337
|
-
| `nk_sqeuclidean_f16_serial` | 3.
|
|
338
|
-
| `nk_euclidean_f16_serial` | 3.
|
|
339
|
-
| `nk_angular_f16_serial` | 1.
|
|
340
|
-
| `nk_sqeuclidean_f16_neonhalf` |
|
|
341
|
-
| `nk_euclidean_f16_neonhalf` |
|
|
342
|
-
| `nk_angular_f16_neonhalf` |
|
|
335
|
+
| `nk_sqeuclidean_f16_serial` | 3.09 gb/s, 0.1 ulp | 3.16 gb/s, 0.1 ulp | 3.10 gb/s, 0.1 ulp |
|
|
336
|
+
| `nk_euclidean_f16_serial` | 3.13 gb/s, 0.6 ulp | 3.14 gb/s, 0.5 ulp | 3.11 gb/s, 0.5 ulp |
|
|
337
|
+
| `nk_angular_f16_serial` | 1.84 gb/s, 0 ulp | 1.92 gb/s, 0 ulp | 1.88 gb/s, 0 ulp |
|
|
338
|
+
| `nk_sqeuclidean_f16_neonhalf` | 34.7 gb/s, 0.9 ulp | 21.5 gb/s, 3.6 ulp | 18.3 gb/s, 9.7 ulp |
|
|
339
|
+
| `nk_euclidean_f16_neonhalf` | 32.7 gb/s, 0.5 ulp | 21.7 gb/s, 2.0 ulp | 18.4 gb/s, 5.3 ulp |
|
|
340
|
+
| `nk_angular_f16_neonhalf` | 25.2 gb/s, 0.1 ulp | 19.6 gb/s, 0.1 ulp | 17.3 gb/s, 0.1 ulp |
|
|
343
341
|
| __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
344
|
-
| `nk_sqeuclidean_e5m2_serial` | 2.
|
|
345
|
-
| `nk_euclidean_e5m2_serial` | 2.
|
|
346
|
-
| `nk_angular_e5m2_serial` | 0.
|
|
342
|
+
| `nk_sqeuclidean_e5m2_serial` | 2.09 gb/s, 0 ulp | 2.08 gb/s, 0 ulp | 2.10 gb/s, 0 ulp |
|
|
343
|
+
| `nk_euclidean_e5m2_serial` | 2.06 gb/s, 0.5 ulp | 2.10 gb/s, 0.5 ulp | 2.05 gb/s, 0.5 ulp |
|
|
344
|
+
| `nk_angular_e5m2_serial` | 0.921 gb/s, 0 ulp | 0.956 gb/s, 0 ulp | 0.938 gb/s, 0 ulp |
|
|
345
|
+
| `nk_sqeuclidean_e5m2_neon` | 18.2 gb/s, 0 ulp | 12.8 gb/s, 0 ulp | 9.84 gb/s, 0 ulp |
|
|
346
|
+
| `nk_euclidean_e5m2_neon` | 18.0 gb/s, 0.5 ulp | 11.8 gb/s, 0.5 ulp | 9.33 gb/s, 0.5 ulp |
|
|
347
|
+
| `nk_angular_e5m2_neon` | 13.7 gb/s, 0 ulp | 10.9 gb/s, 0 ulp | 9.83 gb/s, 0 ulp |
|
|
347
348
|
| __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
348
|
-
| `nk_sqeuclidean_e4m3_serial` | 1.
|
|
349
|
-
| `nk_euclidean_e4m3_serial` | 1.
|
|
350
|
-
| `nk_angular_e4m3_serial` | 0.
|
|
349
|
+
| `nk_sqeuclidean_e4m3_serial` | 1.07 gb/s, 0 ulp | 1.12 gb/s, 0 ulp | 1.11 gb/s, 0 ulp |
|
|
350
|
+
| `nk_euclidean_e4m3_serial` | 1.01 gb/s, 0.5 ulp | 1.12 gb/s, 0.5 ulp | 1.09 gb/s, 0.5 ulp |
|
|
351
|
+
| `nk_angular_e4m3_serial` | 0.711 gb/s, 0 ulp | 0.732 gb/s, 0 ulp | 0.729 gb/s, 0 ulp |
|
|
352
|
+
| `nk_sqeuclidean_e4m3_neon` | 4.29 gb/s, 0.2 ulp | 4.36 gb/s, 0.2 ulp | 4.33 gb/s, 0.2 ulp |
|
|
353
|
+
| `nk_euclidean_e4m3_neon` | 4.20 gb/s, 0.5 ulp | 4.11 gb/s, 0.5 ulp | 4.17 gb/s, 0.5 ulp |
|
|
354
|
+
| `nk_angular_e4m3_neon` | 4.13 gb/s, 0 ulp | 4.21 gb/s, 0 ulp | 4.16 gb/s, 0 ulp |
|
|
351
355
|
| __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
352
|
-
| `nk_sqeuclidean_e3m2_serial` |
|
|
353
|
-
| `nk_euclidean_e3m2_serial` |
|
|
354
|
-
| `nk_angular_e3m2_serial` | 0.
|
|
355
|
-
| `nk_sqeuclidean_e3m2_neon` |
|
|
356
|
-
| `nk_euclidean_e3m2_neon` |
|
|
357
|
-
| `nk_angular_e3m2_neon` |
|
|
356
|
+
| `nk_sqeuclidean_e3m2_serial` | 1.95 gb/s, 0 ulp | 2.15 gb/s, 0 ulp | 2.08 gb/s, 0 ulp |
|
|
357
|
+
| `nk_euclidean_e3m2_serial` | 1.97 gb/s, 0.5 ulp | 2.18 gb/s, 0.5 ulp | 2.09 gb/s, 0.5 ulp |
|
|
358
|
+
| `nk_angular_e3m2_serial` | 0.900 gb/s, 0 ulp | 0.985 gb/s, 0 ulp | 0.943 gb/s, 0 ulp |
|
|
359
|
+
| `nk_sqeuclidean_e3m2_neon` | 4.73 gb/s, 0 ulp | 5.19 gb/s, 0 ulp | 5.03 gb/s, 0 ulp |
|
|
360
|
+
| `nk_euclidean_e3m2_neon` | 4.78 gb/s, 0 ulp | 5.23 gb/s, 0 ulp | 5.05 gb/s, 0 ulp |
|
|
361
|
+
| `nk_angular_e3m2_neon` | 4.24 gb/s, 0 ulp | 4.85 gb/s, 0 ulp | 4.73 gb/s, 0 ulp |
|
|
358
362
|
| __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
359
|
-
| `nk_sqeuclidean_e2m3_serial` |
|
|
360
|
-
| `nk_euclidean_e2m3_serial` |
|
|
361
|
-
| `nk_angular_e2m3_serial` | 0.
|
|
362
|
-
| `nk_sqeuclidean_e2m3_neon` |
|
|
363
|
-
| `nk_euclidean_e2m3_neon` |
|
|
364
|
-
| `nk_angular_e2m3_neon` |
|
|
363
|
+
| `nk_sqeuclidean_e2m3_serial` | 1.98 gb/s, 0 ulp | 2.20 gb/s, 0 ulp | 2.11 gb/s, 0 ulp |
|
|
364
|
+
| `nk_euclidean_e2m3_serial` | 1.91 gb/s, 0.5 ulp | 2.16 gb/s, 0.5 ulp | 2.09 gb/s, 0.4 ulp |
|
|
365
|
+
| `nk_angular_e2m3_serial` | 0.885 gb/s, 0 ulp | 0.985 gb/s, 0 ulp | 0.953 gb/s, 0 ulp |
|
|
366
|
+
| `nk_sqeuclidean_e2m3_neon` | 4.67 gb/s, 0 ulp | 5.06 gb/s, 0 ulp | 5.07 gb/s, 0 ulp |
|
|
367
|
+
| `nk_euclidean_e2m3_neon` | 4.84 gb/s, 0 ulp | 5.17 gb/s, 0 ulp | 4.98 gb/s, 0 ulp |
|
|
368
|
+
| `nk_angular_e2m3_neon` | 4.45 gb/s, 0 ulp | 4.88 gb/s, 0 ulp | 4.73 gb/s, 0 ulp |
|
|
365
369
|
| __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
366
|
-
| `nk_sqeuclidean_i8_serial` |
|
|
367
|
-
| `nk_euclidean_i8_serial` |
|
|
368
|
-
| `nk_angular_i8_serial` |
|
|
369
|
-
| `nk_sqeuclidean_i8_neonsdot` |
|
|
370
|
-
| `nk_euclidean_i8_neonsdot` |
|
|
371
|
-
| `nk_angular_i8_neonsdot` |
|
|
370
|
+
| `nk_sqeuclidean_i8_serial` | 61.6 gb/s | 75.2 gb/s | 64.9 gb/s |
|
|
371
|
+
| `nk_euclidean_i8_serial` | 43.5 gb/s | 54.0 gb/s | 64.3 gb/s |
|
|
372
|
+
| `nk_angular_i8_serial` | 55.8 gb/s | 63.7 gb/s | 49.6 gb/s |
|
|
373
|
+
| `nk_sqeuclidean_i8_neonsdot` | 89.1 gb/s | 85.9 gb/s | 58.8 gb/s |
|
|
374
|
+
| `nk_euclidean_i8_neonsdot` | 86.9 gb/s | 78.9 gb/s | 57.4 gb/s |
|
|
375
|
+
| `nk_angular_i8_neonsdot` | 66.5 gb/s | 68.9 gb/s | 50.9 gb/s |
|
|
372
376
|
| __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
373
|
-
| `nk_sqeuclidean_u8_serial` |
|
|
374
|
-
| `nk_euclidean_u8_serial` |
|
|
375
|
-
| `nk_angular_u8_serial` |
|
|
376
|
-
| `nk_sqeuclidean_u8_neonsdot` |
|
|
377
|
-
| `nk_euclidean_u8_neonsdot` |
|
|
378
|
-
| `nk_angular_u8_neonsdot` |
|
|
377
|
+
| `nk_sqeuclidean_u8_serial` | 62.9 gb/s | 77.2 gb/s | 66.5 gb/s |
|
|
378
|
+
| `nk_euclidean_u8_serial` | 45.7 gb/s | 52.3 gb/s | 61.4 gb/s |
|
|
379
|
+
| `nk_angular_u8_serial` | 17.8 gb/s | 18.5 gb/s | 16.0 gb/s |
|
|
380
|
+
| `nk_sqeuclidean_u8_neonsdot` | 91.7 gb/s | 83.1 gb/s | 56.6 gb/s |
|
|
381
|
+
| `nk_euclidean_u8_neonsdot` | 87.9 gb/s | 79.3 gb/s | 56.5 gb/s |
|
|
382
|
+
| `nk_angular_u8_neonsdot` | 68.0 gb/s | 64.8 gb/s | 49.5 gb/s |
|
|
379
383
|
| __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
380
|
-
| `nk_sqeuclidean_i4_serial` |
|
|
381
|
-
| `nk_euclidean_i4_serial` | 20.
|
|
382
|
-
| `nk_angular_i4_serial` | 9.
|
|
384
|
+
| `nk_sqeuclidean_i4_serial` | 22.9 gb/s | 25.2 gb/s | 25.1 gb/s |
|
|
385
|
+
| `nk_euclidean_i4_serial` | 20.1 gb/s | 23.6 gb/s | 24.2 gb/s |
|
|
386
|
+
| `nk_angular_i4_serial` | 9.11 gb/s | 10.4 gb/s | 10.4 gb/s |
|
|
383
387
|
| __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
384
|
-
| `nk_sqeuclidean_u4_serial` |
|
|
385
|
-
| `nk_euclidean_u4_serial` |
|
|
386
|
-
| `nk_angular_u4_serial` | 9.
|
|
388
|
+
| `nk_sqeuclidean_u4_serial` | 26.7 gb/s | 26.8 gb/s | 22.4 gb/s |
|
|
389
|
+
| `nk_euclidean_u4_serial` | 20.9 gb/s | 22.5 gb/s | 21.2 gb/s |
|
|
390
|
+
| `nk_angular_u4_serial` | 9.00 gb/s | 9.57 gb/s | 9.62 gb/s |
|
|
387
391
|
|
|
388
392
|
#### WASM
|
|
389
393
|
|
|
390
|
-
Measured with Wasmtime
|
|
394
|
+
Measured with Wasmtime v43 (Cranelift backend).
|
|
391
395
|
|
|
392
396
|
| Kernel | 256 | 1024 | 4096 |
|
|
393
397
|
| :-------------------------------- | -----------------------: | -----------------------: | -----------------------: |
|
|
394
398
|
| __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
395
|
-
| `nk_sqeuclidean_f64_serial` |
|
|
396
|
-
| `nk_euclidean_f64_serial` |
|
|
397
|
-
| `nk_angular_f64_serial` |
|
|
398
|
-
| `nk_sqeuclidean_f64_v128relaxed` |
|
|
399
|
-
| `nk_euclidean_f64_v128relaxed` |
|
|
400
|
-
| `nk_angular_f64_v128relaxed` |
|
|
399
|
+
| `nk_sqeuclidean_f64_serial` | 20.3 gb/s, 0.1 ulp | 19.3 gb/s, 0 ulp | 20.3 gb/s, 0 ulp |
|
|
400
|
+
| `nk_euclidean_f64_serial` | 20.0 gb/s, 0.6 ulp | 19.4 gb/s, 0.6 ulp | 20.3 gb/s, 0.5 ulp |
|
|
401
|
+
| `nk_angular_f64_serial` | 9.28 gb/s, 0 ulp | 8.83 gb/s, 0 ulp | 9.29 gb/s, 0 ulp |
|
|
402
|
+
| `nk_sqeuclidean_f64_v128relaxed` | 48.2 gb/s, 1.3 ulp | 35.7 gb/s, 2.6 ulp | 37.0 gb/s, 5.0 ulp |
|
|
403
|
+
| `nk_euclidean_f64_v128relaxed` | 50.1 gb/s, 0.7 ulp | 36.5 gb/s, 1.4 ulp | 36.7 gb/s, 2.8 ulp |
|
|
404
|
+
| `nk_angular_f64_v128relaxed` | 31.5 gb/s, 0.1 ulp | 22.3 gb/s, 0.1 ulp | 22.4 gb/s, 0.1 ulp |
|
|
401
405
|
| __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
402
|
-
| `nk_sqeuclidean_f32_serial` |
|
|
403
|
-
| `nk_euclidean_f32_serial` |
|
|
404
|
-
| `nk_angular_f32_serial` |
|
|
405
|
-
| `nk_sqeuclidean_f32_v128relaxed` |
|
|
406
|
-
| `nk_euclidean_f32_v128relaxed` |
|
|
407
|
-
| `nk_angular_f32_v128relaxed` |
|
|
406
|
+
| `nk_sqeuclidean_f32_serial` | 8.90 gb/s, 0 ulp | 8.54 gb/s, 0 ulp | 8.86 gb/s, 0 ulp |
|
|
407
|
+
| `nk_euclidean_f32_serial` | 8.86 gb/s, 0.1 ulp | 8.58 gb/s, 0.1 ulp | 8.86 gb/s, 0.1 ulp |
|
|
408
|
+
| `nk_angular_f32_serial` | 4.33 gb/s, 0 ulp | 4.17 gb/s, 0 ulp | 4.30 gb/s, 0 ulp |
|
|
409
|
+
| `nk_sqeuclidean_f32_v128relaxed` | 20.4 gb/s, 0.7 ulp | 17.9 gb/s, 1.3 ulp | 18.3 gb/s, 2.6 ulp |
|
|
410
|
+
| `nk_euclidean_f32_v128relaxed` | 20.3 gb/s, 0.4 ulp | 18.0 gb/s, 0.7 ulp | 18.3 gb/s, 1.4 ulp |
|
|
411
|
+
| `nk_angular_f32_v128relaxed` | 19.7 gb/s, 0 ulp | 17.8 gb/s, 0 ulp | 18.5 gb/s, 0 ulp |
|
|
408
412
|
| __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
409
|
-
| `nk_sqeuclidean_bf16_serial` | 5.
|
|
410
|
-
| `nk_euclidean_bf16_serial` | 5.
|
|
411
|
-
| `nk_angular_bf16_serial` | 2.
|
|
412
|
-
| `nk_sqeuclidean_bf16_v128relaxed` |
|
|
413
|
-
| `nk_euclidean_bf16_v128relaxed` |
|
|
414
|
-
| `nk_angular_bf16_v128relaxed` |
|
|
413
|
+
| `nk_sqeuclidean_bf16_serial` | 5.16 gb/s, 0 ulp | 4.88 gb/s, 0 ulp | 5.10 gb/s, 0 ulp |
|
|
414
|
+
| `nk_euclidean_bf16_serial` | 5.12 gb/s, 0.6 ulp | 4.89 gb/s, 0.5 ulp | 5.10 gb/s, 0.5 ulp |
|
|
415
|
+
| `nk_angular_bf16_serial` | 2.24 gb/s, 0 ulp | 2.15 gb/s, 0 ulp | 2.24 gb/s, 0 ulp |
|
|
416
|
+
| `nk_sqeuclidean_bf16_v128relaxed` | 39.9 gb/s, 0.9 ulp | 27.0 gb/s, 13 ulp | 20.3 gb/s, 21 ulp |
|
|
417
|
+
| `nk_euclidean_bf16_v128relaxed` | 38.6 gb/s, 0.5 ulp | 27.1 gb/s, 7.0 ulp | 21.2 gb/s, 12 ulp |
|
|
418
|
+
| `nk_angular_bf16_v128relaxed` | 27.9 gb/s, 0 ulp | 22.6 gb/s, 0.2 ulp | 20.5 gb/s, 0.6 ulp |
|
|
415
419
|
| __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
416
|
-
| `nk_sqeuclidean_f16_serial` | 3.
|
|
417
|
-
| `nk_euclidean_f16_serial` |
|
|
418
|
-
| `nk_angular_f16_serial` | 2.
|
|
419
|
-
| `nk_sqeuclidean_f16_v128relaxed` |
|
|
420
|
-
| `nk_euclidean_f16_v128relaxed` |
|
|
421
|
-
| `nk_angular_f16_v128relaxed` |
|
|
420
|
+
| `nk_sqeuclidean_f16_serial` | 3.22 gb/s, 0.1 ulp | 3.06 gb/s, 0.1 ulp | 3.09 gb/s, 0.1 ulp |
|
|
421
|
+
| `nk_euclidean_f16_serial` | 3.19 gb/s, 0.6 ulp | 2.92 gb/s, 0.5 ulp | 3.26 gb/s, 0.5 ulp |
|
|
422
|
+
| `nk_angular_f16_serial` | 2.33 gb/s, 0 ulp | 2.21 gb/s, 0 ulp | 2.32 gb/s, 0 ulp |
|
|
423
|
+
| `nk_sqeuclidean_f16_v128relaxed` | 11.2 gb/s, 0.9 ulp | 11.0 gb/s, 3.6 ulp | 11.8 gb/s, 9.6 ulp |
|
|
424
|
+
| `nk_euclidean_f16_v128relaxed` | 11.3 gb/s, 0.5 ulp | 11.0 gb/s, 2.0 ulp | 11.8 gb/s, 5.3 ulp |
|
|
425
|
+
| `nk_angular_f16_v128relaxed` | 9.41 gb/s, 0.1 ulp | 9.56 gb/s, 0.1 ulp | 10.4 gb/s, 0.1 ulp |
|
|
422
426
|
| __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
423
|
-
| `nk_sqeuclidean_i8_serial` |
|
|
424
|
-
| `nk_euclidean_i8_serial` | 14.7 gb/s, 0.5 ulp |
|
|
425
|
-
| `nk_angular_i8_serial` | 8.
|
|
426
|
-
| `nk_sqeuclidean_i8_v128relaxed` |
|
|
427
|
-
| `nk_euclidean_i8_v128relaxed` |
|
|
428
|
-
| `nk_angular_i8_v128relaxed` |
|
|
427
|
+
| `nk_sqeuclidean_i8_serial` | 14.9 gb/s | 14.7 gb/s | 16.5 gb/s |
|
|
428
|
+
| `nk_euclidean_i8_serial` | 14.7 gb/s, 0.5 ulp | 14.8 gb/s, 0.4 ulp | 16.3 gb/s, 0.4 ulp |
|
|
429
|
+
| `nk_angular_i8_serial` | 8.06 gb/s, 0 ulp | 8.42 gb/s, 0 ulp | 10.7 gb/s, 0 ulp |
|
|
430
|
+
| `nk_sqeuclidean_i8_v128relaxed` | 30.7 gb/s | 22.9 gb/s | 18.0 gb/s |
|
|
431
|
+
| `nk_euclidean_i8_v128relaxed` | 27.4 gb/s, 0 ulp | 22.6 gb/s, 0 ulp | 17.9 gb/s, 0 ulp |
|
|
432
|
+
| `nk_angular_i8_v128relaxed` | 17.2 gb/s, 0 ulp | 18.1 gb/s, 0 ulp | 19.7 gb/s, 0 ulp |
|
|
429
433
|
| __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
430
|
-
| `nk_sqeuclidean_u8_serial` | 14.
|
|
431
|
-
| `nk_euclidean_u8_serial` | 14.
|
|
432
|
-
| `nk_angular_u8_serial` |
|
|
433
|
-
| `nk_sqeuclidean_u8_v128relaxed` |
|
|
434
|
-
| `nk_euclidean_u8_v128relaxed` |
|
|
435
|
-
| `nk_angular_u8_v128relaxed` |
|
|
434
|
+
| `nk_sqeuclidean_u8_serial` | 14.8 gb/s | 14.5 gb/s | 16.3 gb/s |
|
|
435
|
+
| `nk_euclidean_u8_serial` | 14.5 gb/s, 0.5 ulp | 14.5 gb/s, 0.5 ulp | 16.0 gb/s, 0.6 ulp |
|
|
436
|
+
| `nk_angular_u8_serial` | 7.86 gb/s, 0.5 ulp | 8.25 gb/s, 0.5 ulp | 10.7 gb/s, 0.4 ulp |
|
|
437
|
+
| `nk_sqeuclidean_u8_v128relaxed` | 33.2 gb/s | 24.6 gb/s | 18.3 gb/s |
|
|
438
|
+
| `nk_euclidean_u8_v128relaxed` | 28.6 gb/s, 0 ulp | 23.7 gb/s, 0 ulp | 18.2 gb/s, 0 ulp |
|
|
439
|
+
| `nk_angular_u8_v128relaxed` | 14.1 gb/s, 0 ulp | 15.0 gb/s, 0 ulp | 16.1 gb/s, 0 ulp |
|