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
|
@@ -7,15 +7,15 @@ Used internally by packing routines for norm precomputation and by distance kern
|
|
|
7
7
|
|
|
8
8
|
Moments:
|
|
9
9
|
|
|
10
|
-
|
|
10
|
+
$$
|
|
11
11
|
\text{sum} = \sum a_i, \quad \text{sumsq} = \sum a_i^2
|
|
12
|
-
|
|
12
|
+
$$
|
|
13
13
|
|
|
14
14
|
Min-max:
|
|
15
15
|
|
|
16
|
-
|
|
16
|
+
$$
|
|
17
17
|
\text{min} = \min_i a_i, \quad \text{argmin} = \arg\min_i a_i
|
|
18
|
-
|
|
18
|
+
$$
|
|
19
19
|
|
|
20
20
|
Reformulating as Python pseudocode:
|
|
21
21
|
|
|
@@ -365,183 +365,188 @@ Measured with Wasmtime v42 (Cranelift backend).
|
|
|
365
365
|
| `nk_reduce_minmax_u64_serial` | 7.83 gb/s | 7.59 gb/s | 7.74 gb/s |
|
|
366
366
|
| `nk_reduce_minmax_u64_v128relaxed` | 2.54 gb/s | 2.59 gb/s | 2.64 gb/s |
|
|
367
367
|
|
|
368
|
-
### Apple
|
|
368
|
+
### Apple M5
|
|
369
369
|
|
|
370
370
|
#### Native
|
|
371
371
|
|
|
372
372
|
| Kernel | 256 | 1024 | 4096 |
|
|
373
373
|
| :--------------------------------- | -----------------------: | -----------------------: | -----------------------: |
|
|
374
374
|
| __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
375
|
-
| `nk_reduce_moments_f64_serial` |
|
|
376
|
-
| `nk_reduce_minmax_f64_serial` |
|
|
377
|
-
| `nk_reduce_moments_f64_neon` |
|
|
378
|
-
| `nk_reduce_minmax_f64_neon` |
|
|
375
|
+
| `nk_reduce_moments_f64_serial` | 6.41 gb/s, 0 ulp | 7.19 gb/s, 0 ulp | 6.71 gb/s, 0 ulp |
|
|
376
|
+
| `nk_reduce_minmax_f64_serial` | 16.4 gb/s, 0 ulp | 16.1 gb/s, 0 ulp | 16.1 gb/s, 0 ulp |
|
|
377
|
+
| `nk_reduce_moments_f64_neon` | 16.4 gb/s, 0 ulp | 16.3 gb/s, 0 ulp | 17.0 gb/s, 0 ulp |
|
|
378
|
+
| `nk_reduce_minmax_f64_neon` | 18.0 gb/s, 0 ulp | 17.0 gb/s, 0 ulp | 16.2 gb/s, 0 ulp |
|
|
379
379
|
| __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
380
|
-
| `nk_reduce_moments_f32_serial` |
|
|
381
|
-
| `nk_reduce_minmax_f32_serial` |
|
|
382
|
-
| `nk_reduce_moments_f32_neon` |
|
|
383
|
-
| `nk_reduce_minmax_f32_neon` |
|
|
380
|
+
| `nk_reduce_moments_f32_serial` | 3.06 gb/s, 0 ulp | 3.40 gb/s, 0 ulp | 3.15 gb/s, 0 ulp |
|
|
381
|
+
| `nk_reduce_minmax_f32_serial` | 8.29 gb/s, 0 ulp | 7.97 gb/s, 0 ulp | 8.00 gb/s, 0 ulp |
|
|
382
|
+
| `nk_reduce_moments_f32_neon` | 16.6 gb/s, 0.4 ulp | 10.8 gb/s, 1.8 ulp | 9.78 gb/s, 1.1 ulp |
|
|
383
|
+
| `nk_reduce_minmax_f32_neon` | 19.7 gb/s, 0 ulp | 17.3 gb/s, 0 ulp | 16.5 gb/s, 0 ulp |
|
|
384
384
|
| __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
385
|
-
| `nk_reduce_moments_bf16_serial` | 1.
|
|
386
|
-
| `nk_reduce_minmax_bf16_serial` |
|
|
387
|
-
| `nk_reduce_moments_bf16_neonbfdot` |
|
|
388
|
-
| `nk_reduce_minmax_bf16_neonbfdot` | 13.7 gb/s, 0 ulp | 14.5 gb/s, 0 ulp | 15.4 gb/s, 0 ulp |
|
|
385
|
+
| `nk_reduce_moments_bf16_serial` | 1.49 gb/s, 0 ulp | 1.62 gb/s, 0 ulp | 1.55 gb/s, 0 ulp |
|
|
386
|
+
| `nk_reduce_minmax_bf16_serial` | 2.19 gb/s, 0 ulp | 2.49 gb/s, 0 ulp | 2.69 gb/s, 0 ulp |
|
|
387
|
+
| `nk_reduce_moments_bf16_neonbfdot` | 26.1 gb/s, 0 ulp | 28.2 gb/s, 0.4 ulp | 29.9 gb/s, 0.3 ulp |
|
|
389
388
|
| __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
390
|
-
| `nk_reduce_moments_f16_serial` | 1.
|
|
391
|
-
| `nk_reduce_minmax_f16_serial` | 1.
|
|
392
|
-
| `nk_reduce_moments_f16_neonhalf` |
|
|
389
|
+
| `nk_reduce_moments_f16_serial` | 1.46 gb/s, 0 ulp | 1.55 gb/s, 0 ulp | 1.51 gb/s, 0 ulp |
|
|
390
|
+
| `nk_reduce_minmax_f16_serial` | 1.62 gb/s, 0 ulp | 1.87 gb/s, 0 ulp | 2.02 gb/s, 0 ulp |
|
|
391
|
+
| `nk_reduce_moments_f16_neonhalf` | 21.2 gb/s, 0.1 ulp | 15.4 gb/s, 0.1 ulp | 10.6 gb/s, 0.8 ulp |
|
|
393
392
|
| __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
394
|
-
| `nk_reduce_moments_e5m2_serial` | 0.
|
|
395
|
-
| `nk_reduce_minmax_e5m2_serial` |
|
|
396
|
-
| `
|
|
397
|
-
| `
|
|
393
|
+
| `nk_reduce_moments_e5m2_serial` | 0.750 gb/s, 0 ulp | 0.830 gb/s, 0 ulp | 0.758 gb/s, 0 ulp |
|
|
394
|
+
| `nk_reduce_minmax_e5m2_serial` | 1.10 gb/s, 0 ulp | 1.34 gb/s, 0 ulp | 1.39 gb/s, 0 ulp |
|
|
395
|
+
| `nk_reduce_moments_e5m2_neon` | 10.4 gb/s, ? ulp | 7.47 gb/s, ? ulp | 5.22 gb/s, ? ulp |
|
|
396
|
+
| `nk_reduce_moments_e5m2_neonfhm` | 12.4 gb/s, 0 ulp | 7.31 gb/s, 0 ulp | 4.78 gb/s, 0 ulp |
|
|
397
|
+
| `nk_reduce_minmax_e5m2_neon` | 15.5 gb/s, 0 ulp | 17.4 gb/s, 0 ulp | 18.1 gb/s, 0 ulp |
|
|
398
398
|
| __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
399
|
-
| `nk_reduce_moments_e4m3_serial` | 0.
|
|
400
|
-
| `nk_reduce_minmax_e4m3_serial` |
|
|
401
|
-
| `
|
|
402
|
-
| `
|
|
399
|
+
| `nk_reduce_moments_e4m3_serial` | 0.569 gb/s, 0 ulp | 0.640 gb/s, 0 ulp | 0.599 gb/s, 0 ulp |
|
|
400
|
+
| `nk_reduce_minmax_e4m3_serial` | 1.14 gb/s, 0 ulp | 1.39 gb/s, 0 ulp | 1.42 gb/s, 0 ulp |
|
|
401
|
+
| `nk_reduce_moments_e4m3_neon` | 6.52 gb/s, ? ulp | 5.52 gb/s, ? ulp | 4.88 gb/s, ? ulp |
|
|
402
|
+
| `nk_reduce_moments_e4m3_neonfhm` | 4.21 gb/s, 0 ulp | 4.23 gb/s, 0 ulp | 4.11 gb/s, 0 ulp |
|
|
403
|
+
| `nk_reduce_minmax_e4m3_neon` | 15.5 gb/s, 0 ulp | 17.4 gb/s, 0 ulp | 17.5 gb/s, 0 ulp |
|
|
403
404
|
| __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
404
|
-
| `nk_reduce_moments_e3m2_serial` | 0.
|
|
405
|
-
| `nk_reduce_minmax_e3m2_serial` | 0.
|
|
405
|
+
| `nk_reduce_moments_e3m2_serial` | 0.750 gb/s, 0 ulp | 0.806 gb/s, 0 ulp | 0.773 gb/s, 0 ulp |
|
|
406
|
+
| `nk_reduce_minmax_e3m2_serial` | 0.640 gb/s, 0 ulp | 0.641 gb/s, 0 ulp | 0.634 gb/s, 0 ulp |
|
|
407
|
+
| `nk_reduce_moments_e3m2_neon` | 8.64 gb/s, ? ulp | 6.96 gb/s, ? ulp | 5.70 gb/s, ? ulp |
|
|
408
|
+
| `nk_reduce_minmax_e3m2_neon` | 16.6 gb/s, 0 ulp | 19.1 gb/s, 0 ulp | 18.9 gb/s, 0 ulp |
|
|
406
409
|
| __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
407
|
-
| `nk_reduce_moments_e2m3_serial` | 0.
|
|
408
|
-
| `nk_reduce_minmax_e2m3_serial` | 0.
|
|
409
|
-
| `
|
|
410
|
+
| `nk_reduce_moments_e2m3_serial` | 0.753 gb/s, 0 ulp | 0.821 gb/s, 0 ulp | 0.775 gb/s, 0 ulp |
|
|
411
|
+
| `nk_reduce_minmax_e2m3_serial` | 0.640 gb/s, 0 ulp | 0.643 gb/s, 0 ulp | 0.618 gb/s, 0 ulp |
|
|
412
|
+
| `nk_reduce_moments_e2m3_neon` | 17.1 gb/s, ? ulp | 15.2 gb/s, ? ulp | 11.9 gb/s, ? ulp |
|
|
413
|
+
| `nk_reduce_moments_e2m3_neonsdot` | 29.4 gb/s, 0 ulp | 29.3 gb/s, 0 ulp | 24.4 gb/s, 0 ulp |
|
|
414
|
+
| `nk_reduce_minmax_e2m3_neon` | 17.0 gb/s, 0 ulp | 19.3 gb/s, 0 ulp | 19.6 gb/s, 0 ulp |
|
|
410
415
|
| __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
411
|
-
| `nk_reduce_moments_i8_serial` |
|
|
412
|
-
| `nk_reduce_minmax_i8_serial` | 1.
|
|
413
|
-
| `nk_reduce_moments_i8_neon` |
|
|
414
|
-
| `nk_reduce_minmax_i8_neon` |
|
|
415
|
-
| `nk_reduce_moments_i8_neonsdot` |
|
|
416
|
+
| `nk_reduce_moments_i8_serial` | 3.00 gb/s | 3.50 gb/s | 3.21 gb/s |
|
|
417
|
+
| `nk_reduce_minmax_i8_serial` | 1.81 gb/s | 1.94 gb/s | 1.89 gb/s |
|
|
418
|
+
| `nk_reduce_moments_i8_neon` | 27.9 gb/s | 17.7 gb/s | 13.3 gb/s |
|
|
419
|
+
| `nk_reduce_minmax_i8_neon` | 25.7 gb/s | 30.3 gb/s | 27.8 gb/s |
|
|
420
|
+
| `nk_reduce_moments_i8_neonsdot` | 44.4 gb/s | 46.8 gb/s | 33.4 gb/s |
|
|
416
421
|
| __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
417
|
-
| `nk_reduce_moments_u8_serial` |
|
|
418
|
-
| `nk_reduce_minmax_u8_serial` | 1.
|
|
419
|
-
| `nk_reduce_moments_u8_neon` |
|
|
420
|
-
| `nk_reduce_minmax_u8_neon` |
|
|
421
|
-
| `nk_reduce_moments_u8_neonsdot` |
|
|
422
|
+
| `nk_reduce_moments_u8_serial` | 3.08 gb/s | 3.51 gb/s | 3.23 gb/s |
|
|
423
|
+
| `nk_reduce_minmax_u8_serial` | 1.84 gb/s | 1.94 gb/s | 1.90 gb/s |
|
|
424
|
+
| `nk_reduce_moments_u8_neon` | 29.2 gb/s | 18.0 gb/s | 13.6 gb/s |
|
|
425
|
+
| `nk_reduce_minmax_u8_neon` | 27.0 gb/s | 28.3 gb/s | 29.2 gb/s |
|
|
426
|
+
| `nk_reduce_moments_u8_neonsdot` | 45.5 gb/s | 43.5 gb/s | 33.3 gb/s |
|
|
422
427
|
| __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
423
|
-
| `nk_reduce_moments_i4_serial` |
|
|
424
|
-
| `nk_reduce_minmax_i4_serial` | 0.
|
|
428
|
+
| `nk_reduce_moments_i4_serial` | 2.06 gb/s | 2.57 gb/s | 2.39 gb/s |
|
|
429
|
+
| `nk_reduce_minmax_i4_serial` | 0.701 gb/s | 0.794 gb/s | 0.803 gb/s |
|
|
425
430
|
| __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
426
|
-
| `nk_reduce_moments_u4_serial` |
|
|
427
|
-
| `nk_reduce_minmax_u4_serial` | 0.
|
|
431
|
+
| `nk_reduce_moments_u4_serial` | 2.20 gb/s | 2.75 gb/s | 2.62 gb/s |
|
|
432
|
+
| `nk_reduce_minmax_u4_serial` | 0.741 gb/s | 0.845 gb/s | 0.856 gb/s |
|
|
428
433
|
| __u1__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
429
|
-
| `nk_reduce_moments_u1_serial` | 1.
|
|
430
|
-
| `nk_reduce_minmax_u1_serial` |
|
|
434
|
+
| `nk_reduce_moments_u1_serial` | 1.79 gb/s | 1.99 gb/s | 2.03 gb/s |
|
|
435
|
+
| `nk_reduce_minmax_u1_serial` | 9.74 gb/s | 38.5 gb/s | 121 gb/s |
|
|
431
436
|
| __i16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
432
|
-
| `nk_reduce_moments_i16_serial` |
|
|
433
|
-
| `nk_reduce_minmax_i16_serial` |
|
|
434
|
-
| `nk_reduce_moments_i16_neon` |
|
|
435
|
-
| `nk_reduce_minmax_i16_neon` |
|
|
437
|
+
| `nk_reduce_moments_i16_serial` | 6.02 gb/s | 7.03 gb/s | 6.70 gb/s |
|
|
438
|
+
| `nk_reduce_minmax_i16_serial` | 3.66 gb/s | 3.88 gb/s | 3.82 gb/s |
|
|
439
|
+
| `nk_reduce_moments_i16_neon` | 23.0 gb/s | 16.2 gb/s | 11.9 gb/s |
|
|
440
|
+
| `nk_reduce_minmax_i16_neon` | 26.8 gb/s | 28.5 gb/s | 25.9 gb/s |
|
|
436
441
|
| __u16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
437
|
-
| `nk_reduce_moments_u16_serial` |
|
|
438
|
-
| `nk_reduce_minmax_u16_serial` | 2.
|
|
439
|
-
| `nk_reduce_moments_u16_neon` |
|
|
440
|
-
| `nk_reduce_minmax_u16_neon` |
|
|
442
|
+
| `nk_reduce_moments_u16_serial` | 6.13 gb/s | 6.79 gb/s | 6.45 gb/s |
|
|
443
|
+
| `nk_reduce_minmax_u16_serial` | 2.72 gb/s | 2.68 gb/s | 2.67 gb/s |
|
|
444
|
+
| `nk_reduce_moments_u16_neon` | 23.1 gb/s | 16.3 gb/s | 11.6 gb/s |
|
|
445
|
+
| `nk_reduce_minmax_u16_neon` | 26.6 gb/s | 28.5 gb/s | 25.8 gb/s |
|
|
441
446
|
| __i32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
442
|
-
| `nk_reduce_moments_i32_serial` |
|
|
443
|
-
| `nk_reduce_minmax_i32_serial` |
|
|
444
|
-
| `nk_reduce_moments_i32_neon` |
|
|
445
|
-
| `nk_reduce_minmax_i32_neon` |
|
|
447
|
+
| `nk_reduce_moments_i32_serial` | 7.70 gb/s | 8.09 gb/s | 7.52 gb/s |
|
|
448
|
+
| `nk_reduce_minmax_i32_serial` | 7.35 gb/s | 7.49 gb/s | 7.65 gb/s |
|
|
449
|
+
| `nk_reduce_moments_i32_neon` | 8.88 gb/s | 6.74 gb/s | 6.60 gb/s |
|
|
450
|
+
| `nk_reduce_minmax_i32_neon` | 27.3 gb/s | 27.9 gb/s | 25.8 gb/s |
|
|
446
451
|
| __u32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
447
|
-
| `nk_reduce_moments_u32_serial` |
|
|
448
|
-
| `nk_reduce_minmax_u32_serial` |
|
|
449
|
-
| `nk_reduce_moments_u32_neon` |
|
|
450
|
-
| `nk_reduce_minmax_u32_neon` |
|
|
452
|
+
| `nk_reduce_moments_u32_serial` | 7.51 gb/s | 8.10 gb/s | 7.73 gb/s |
|
|
453
|
+
| `nk_reduce_minmax_u32_serial` | 7.28 gb/s | 7.69 gb/s | 7.68 gb/s |
|
|
454
|
+
| `nk_reduce_moments_u32_neon` | 19.0 gb/s | 12.8 gb/s | 11.5 gb/s |
|
|
455
|
+
| `nk_reduce_minmax_u32_neon` | 27.6 gb/s | 28.3 gb/s | 26.4 gb/s |
|
|
451
456
|
| __i64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
452
|
-
| `nk_reduce_moments_i64_serial` |
|
|
453
|
-
| `nk_reduce_minmax_i64_serial` |
|
|
454
|
-
| `nk_reduce_moments_i64_neon` |
|
|
455
|
-
| `nk_reduce_minmax_i64_neon` |
|
|
457
|
+
| `nk_reduce_moments_i64_serial` | 10.4 gb/s | 12.9 gb/s | 11.9 gb/s |
|
|
458
|
+
| `nk_reduce_minmax_i64_serial` | 13.7 gb/s | 14.1 gb/s | 14.0 gb/s |
|
|
459
|
+
| `nk_reduce_moments_i64_neon` | 15.6 gb/s | 13.1 gb/s | 12.9 gb/s |
|
|
460
|
+
| `nk_reduce_minmax_i64_neon` | 18.0 gb/s | 16.6 gb/s | 15.6 gb/s |
|
|
456
461
|
| __u64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
457
|
-
| `nk_reduce_moments_u64_serial` |
|
|
458
|
-
| `nk_reduce_minmax_u64_serial` |
|
|
459
|
-
| `nk_reduce_moments_u64_neon` |
|
|
460
|
-
| `nk_reduce_minmax_u64_neon` |
|
|
462
|
+
| `nk_reduce_moments_u64_serial` | 10.4 gb/s | 11.9 gb/s | 11.2 gb/s |
|
|
463
|
+
| `nk_reduce_minmax_u64_serial` | 13.4 gb/s | 14.1 gb/s | 13.8 gb/s |
|
|
464
|
+
| `nk_reduce_moments_u64_neon` | 29.1 gb/s | 22.5 gb/s | 22.2 gb/s |
|
|
465
|
+
| `nk_reduce_minmax_u64_neon` | 18.3 gb/s | 16.7 gb/s | 16.1 gb/s |
|
|
461
466
|
|
|
462
467
|
#### WASM
|
|
463
468
|
|
|
464
|
-
Measured with Wasmtime
|
|
469
|
+
Measured with Wasmtime v43 (Cranelift backend).
|
|
465
470
|
|
|
466
471
|
| Kernel | 256 | 1024 | 4096 |
|
|
467
472
|
| :----------------------------------- | -----------------------: | -----------------------: | -----------------------: |
|
|
468
473
|
| __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
469
|
-
| `nk_reduce_moments_f64_serial` |
|
|
470
|
-
| `nk_reduce_moments_f64_v128relaxed` |
|
|
471
|
-
| `nk_reduce_minmax_f64_serial` |
|
|
472
|
-
| `nk_reduce_minmax_f64_v128relaxed` |
|
|
474
|
+
| `nk_reduce_moments_f64_serial` | 7.05 gb/s, 0 ulp | 6.89 gb/s, 0 ulp | 7.04 gb/s, 0 ulp |
|
|
475
|
+
| `nk_reduce_moments_f64_v128relaxed` | 17.9 gb/s, 0 ulp | 17.9 gb/s, 0 ulp | 17.9 gb/s, 0 ulp |
|
|
476
|
+
| `nk_reduce_minmax_f64_serial` | 11.8 gb/s, 0 ulp | 11.5 gb/s, 0 ulp | 11.4 gb/s, 0 ulp |
|
|
477
|
+
| `nk_reduce_minmax_f64_v128relaxed` | 16.4 gb/s, 0 ulp | 16.8 gb/s, 0 ulp | 16.9 gb/s, 0 ulp |
|
|
473
478
|
| __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
474
|
-
| `nk_reduce_moments_f32_serial` |
|
|
475
|
-
| `nk_reduce_moments_f32_v128relaxed` |
|
|
476
|
-
| `nk_reduce_minmax_f32_serial` |
|
|
477
|
-
| `nk_reduce_minmax_f32_v128relaxed` |
|
|
479
|
+
| `nk_reduce_moments_f32_serial` | 3.34 gb/s, 0 ulp | 3.32 gb/s, 0 ulp | 3.34 gb/s, 0 ulp |
|
|
480
|
+
| `nk_reduce_moments_f32_v128relaxed` | 16.7 gb/s, 0.1 ulp | 11.9 gb/s, 0.5 ulp | 10.3 gb/s, 0 ulp |
|
|
481
|
+
| `nk_reduce_minmax_f32_serial` | 4.40 gb/s, 0 ulp | 4.31 gb/s, 0 ulp | 4.28 gb/s, 0 ulp |
|
|
482
|
+
| `nk_reduce_minmax_f32_v128relaxed` | 14.5 gb/s, 0 ulp | 15.8 gb/s, 0 ulp | 16.7 gb/s, 0 ulp |
|
|
478
483
|
| __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
479
|
-
| `nk_reduce_moments_bf16_serial` |
|
|
480
|
-
| `nk_reduce_moments_bf16_v128relaxed` |
|
|
481
|
-
| `nk_reduce_minmax_bf16_serial` |
|
|
482
|
-
| `nk_reduce_minmax_bf16_v128relaxed` |
|
|
484
|
+
| `nk_reduce_moments_bf16_serial` | 1.76 gb/s, 0 ulp | 1.72 gb/s, 0 ulp | 1.72 gb/s, 0 ulp |
|
|
485
|
+
| `nk_reduce_moments_bf16_v128relaxed` | 20.8 gb/s, 0 ulp | 14.8 gb/s, 0.2 ulp | 11.1 gb/s, 1.6 ulp |
|
|
486
|
+
| `nk_reduce_minmax_bf16_serial` | 1.80 gb/s, 0 ulp | 2.01 gb/s, 0 ulp | 2.10 gb/s, 0 ulp |
|
|
487
|
+
| `nk_reduce_minmax_bf16_v128relaxed` | 9.22 gb/s, 0 ulp | 9.53 gb/s, 0 ulp | 10.5 gb/s, 0 ulp |
|
|
483
488
|
| __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
484
|
-
| `nk_reduce_moments_f16_serial` |
|
|
485
|
-
| `nk_reduce_moments_f16_v128relaxed` |
|
|
486
|
-
| `nk_reduce_minmax_f16_serial` |
|
|
487
|
-
| `nk_reduce_minmax_f16_v128relaxed` |
|
|
489
|
+
| `nk_reduce_moments_f16_serial` | 1.77 gb/s, 0 ulp | 1.73 gb/s, 0 ulp | 1.73 gb/s, 0 ulp |
|
|
490
|
+
| `nk_reduce_moments_f16_v128relaxed` | 11.8 gb/s, 0 ulp | 10.8 gb/s, 0 ulp | 9.84 gb/s, 0.3 ulp |
|
|
491
|
+
| `nk_reduce_minmax_f16_serial` | 1.79 gb/s, 0 ulp | 1.99 gb/s, 0 ulp | 2.09 gb/s, 0 ulp |
|
|
492
|
+
| `nk_reduce_minmax_f16_v128relaxed` | 6.04 gb/s, 0 ulp | 7.62 gb/s, 0 ulp | 8.11 gb/s, 0 ulp |
|
|
488
493
|
| __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
489
|
-
| `nk_reduce_moments_e5m2_serial` |
|
|
490
|
-
| `nk_reduce_moments_e5m2_v128relaxed` |
|
|
491
|
-
| `nk_reduce_minmax_e5m2_serial` |
|
|
492
|
-
| `nk_reduce_minmax_e5m2_v128relaxed` |
|
|
494
|
+
| `nk_reduce_moments_e5m2_serial` | 0.886 gb/s, 0 ulp | 0.865 gb/s, 0 ulp | 0.870 gb/s, 0 ulp |
|
|
495
|
+
| `nk_reduce_moments_e5m2_v128relaxed` | 3.53 gb/s, 0 ulp | 3.55 gb/s, 0 ulp | 3.56 gb/s, 0 ulp |
|
|
496
|
+
| `nk_reduce_minmax_e5m2_serial` | 1.12 gb/s, 0 ulp | 1.18 gb/s, 0 ulp | 1.22 gb/s, 0 ulp |
|
|
497
|
+
| `nk_reduce_minmax_e5m2_v128relaxed` | 5.50 gb/s, 0 ulp | 10.8 gb/s, 0 ulp | 14.6 gb/s, 0 ulp |
|
|
493
498
|
| __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
494
|
-
| `nk_reduce_moments_e4m3_serial` | 0.
|
|
495
|
-
| `nk_reduce_moments_e4m3_v128relaxed` |
|
|
496
|
-
| `nk_reduce_minmax_e4m3_serial` |
|
|
497
|
-
| `nk_reduce_minmax_e4m3_v128relaxed` |
|
|
499
|
+
| `nk_reduce_moments_e4m3_serial` | 0.617 gb/s, 0 ulp | 0.606 gb/s, 0 ulp | 0.610 gb/s, 0 ulp |
|
|
500
|
+
| `nk_reduce_moments_e4m3_v128relaxed` | 3.01 gb/s, 0 ulp | 3.04 gb/s, 0 ulp | 3.05 gb/s, 0 ulp |
|
|
501
|
+
| `nk_reduce_minmax_e4m3_serial` | 1.09 gb/s, 0 ulp | 1.21 gb/s, 0 ulp | 1.27 gb/s, 0 ulp |
|
|
502
|
+
| `nk_reduce_minmax_e4m3_v128relaxed` | 6.46 gb/s, 0 ulp | 10.5 gb/s, 0 ulp | 15.0 gb/s, 0 ulp |
|
|
498
503
|
| __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
499
|
-
| `nk_reduce_moments_e3m2_serial` |
|
|
500
|
-
| `nk_reduce_moments_e3m2_v128relaxed` |
|
|
501
|
-
| `nk_reduce_minmax_e3m2_serial` |
|
|
502
|
-
| `nk_reduce_minmax_e3m2_v128relaxed` |
|
|
504
|
+
| `nk_reduce_moments_e3m2_serial` | 0.881 gb/s, 0 ulp | 0.873 gb/s, 0 ulp | 0.868 gb/s, 0 ulp |
|
|
505
|
+
| `nk_reduce_moments_e3m2_v128relaxed` | 5.41 gb/s, 0 ulp | 4.67 gb/s, 0 ulp | 4.34 gb/s, 0 ulp |
|
|
506
|
+
| `nk_reduce_minmax_e3m2_serial` | 0.671 gb/s, 0 ulp | 0.667 gb/s, 0 ulp | 0.666 gb/s, 0 ulp |
|
|
507
|
+
| `nk_reduce_minmax_e3m2_v128relaxed` | 7.95 gb/s, 0 ulp | 13.0 gb/s, 0 ulp | 16.3 gb/s, 0 ulp |
|
|
503
508
|
| __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
504
|
-
| `nk_reduce_moments_e2m3_serial` |
|
|
505
|
-
| `nk_reduce_moments_e2m3_v128relaxed` |
|
|
506
|
-
| `nk_reduce_minmax_e2m3_serial` |
|
|
507
|
-
| `nk_reduce_minmax_e2m3_v128relaxed` |
|
|
509
|
+
| `nk_reduce_moments_e2m3_serial` | 0.884 gb/s, 0 ulp | 0.877 gb/s, 0 ulp | 0.857 gb/s, 0 ulp |
|
|
510
|
+
| `nk_reduce_moments_e2m3_v128relaxed` | 9.54 gb/s, 0 ulp | 9.49 gb/s, 0 ulp | 8.72 gb/s, 0 ulp |
|
|
511
|
+
| `nk_reduce_minmax_e2m3_serial` | 0.670 gb/s, 0 ulp | 0.667 gb/s, 0 ulp | 0.666 gb/s, 0 ulp |
|
|
512
|
+
| `nk_reduce_minmax_e2m3_v128relaxed` | 9.35 gb/s, 0 ulp | 12.9 gb/s, 0 ulp | 16.4 gb/s, 0 ulp |
|
|
508
513
|
| __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
509
|
-
| `nk_reduce_moments_i8_serial` |
|
|
510
|
-
| `nk_reduce_moments_i8_v128relaxed` | 6
|
|
511
|
-
| `nk_reduce_minmax_i8_serial` |
|
|
512
|
-
| `nk_reduce_minmax_i8_v128relaxed` |
|
|
514
|
+
| `nk_reduce_moments_i8_serial` | 3.78 gb/s | 3.68 gb/s | 3.62 gb/s |
|
|
515
|
+
| `nk_reduce_moments_i8_v128relaxed` | 21.6 gb/s | 22.3 gb/s | 18.7 gb/s |
|
|
516
|
+
| `nk_reduce_minmax_i8_serial` | 1.47 gb/s | 1.49 gb/s | 1.49 gb/s |
|
|
517
|
+
| `nk_reduce_minmax_i8_v128relaxed` | 13.0 gb/s | 19.1 gb/s | 23.1 gb/s |
|
|
513
518
|
| __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
514
|
-
| `nk_reduce_moments_u8_serial` |
|
|
515
|
-
| `nk_reduce_moments_u8_v128relaxed` | 6
|
|
516
|
-
| `nk_reduce_minmax_u8_serial` |
|
|
517
|
-
| `nk_reduce_minmax_u8_v128relaxed` |
|
|
519
|
+
| `nk_reduce_moments_u8_serial` | 3.71 gb/s | 3.68 gb/s | 3.62 gb/s |
|
|
520
|
+
| `nk_reduce_moments_u8_v128relaxed` | 21.6 gb/s | 22.3 gb/s | 18.8 gb/s |
|
|
521
|
+
| `nk_reduce_minmax_u8_serial` | 1.51 gb/s | 1.55 gb/s | 1.55 gb/s |
|
|
522
|
+
| `nk_reduce_minmax_u8_v128relaxed` | 12.6 gb/s | 16.6 gb/s | 22.5 gb/s |
|
|
518
523
|
| __i16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
519
|
-
| `nk_reduce_moments_i16_serial` |
|
|
520
|
-
| `nk_reduce_moments_i16_v128relaxed` |
|
|
521
|
-
| `nk_reduce_minmax_i16_serial` |
|
|
522
|
-
| `nk_reduce_minmax_i16_v128relaxed` |
|
|
524
|
+
| `nk_reduce_moments_i16_serial` | 7.52 gb/s | 7.35 gb/s | 7.33 gb/s |
|
|
525
|
+
| `nk_reduce_moments_i16_v128relaxed` | 13.3 gb/s | 9.97 gb/s | 8.83 gb/s |
|
|
526
|
+
| `nk_reduce_minmax_i16_serial` | 2.96 gb/s | 2.99 gb/s | 2.99 gb/s |
|
|
527
|
+
| `nk_reduce_minmax_i16_v128relaxed` | 18.8 gb/s | 14.5 gb/s | 16.1 gb/s |
|
|
523
528
|
| __u16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
524
|
-
| `nk_reduce_moments_u16_serial` |
|
|
525
|
-
| `nk_reduce_moments_u16_v128relaxed` |
|
|
526
|
-
| `nk_reduce_minmax_u16_serial` |
|
|
527
|
-
| `nk_reduce_minmax_u16_v128relaxed` |
|
|
529
|
+
| `nk_reduce_moments_u16_serial` | 7.51 gb/s | 7.28 gb/s | 7.23 gb/s |
|
|
530
|
+
| `nk_reduce_moments_u16_v128relaxed` | 13.2 gb/s | 9.98 gb/s | 8.84 gb/s |
|
|
531
|
+
| `nk_reduce_minmax_u16_serial` | 3.04 gb/s | 3.11 gb/s | 3.11 gb/s |
|
|
532
|
+
| `nk_reduce_minmax_u16_v128relaxed` | 16.5 gb/s | 14.5 gb/s | 16.2 gb/s |
|
|
528
533
|
| __i32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
529
|
-
| `nk_reduce_moments_i32_serial` |
|
|
530
|
-
| `nk_reduce_moments_i32_v128relaxed` |
|
|
531
|
-
| `nk_reduce_minmax_i32_serial` |
|
|
532
|
-
| `nk_reduce_minmax_i32_v128relaxed` |
|
|
534
|
+
| `nk_reduce_moments_i32_serial` | 5.73 gb/s | 5.66 gb/s | 5.58 gb/s |
|
|
535
|
+
| `nk_reduce_moments_i32_v128relaxed` | 6.40 gb/s | 6.35 gb/s | 6.39 gb/s |
|
|
536
|
+
| `nk_reduce_minmax_i32_serial` | 6.85 gb/s | 6.88 gb/s | 6.90 gb/s |
|
|
537
|
+
| `nk_reduce_minmax_i32_v128relaxed` | 14.0 gb/s | 15.9 gb/s | 16.7 gb/s |
|
|
533
538
|
| __u32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
534
|
-
| `nk_reduce_moments_u32_serial` |
|
|
535
|
-
| `nk_reduce_moments_u32_v128relaxed` | 1.
|
|
536
|
-
| `nk_reduce_minmax_u32_serial` |
|
|
537
|
-
| `nk_reduce_minmax_u32_v128relaxed` |
|
|
539
|
+
| `nk_reduce_moments_u32_serial` | 6.01 gb/s | 5.70 gb/s | 5.54 gb/s |
|
|
540
|
+
| `nk_reduce_moments_u32_v128relaxed` | 1.65 gb/s | 1.58 gb/s | 1.56 gb/s |
|
|
541
|
+
| `nk_reduce_minmax_u32_serial` | 6.80 gb/s | 6.87 gb/s | 6.90 gb/s |
|
|
542
|
+
| `nk_reduce_minmax_u32_v128relaxed` | 14.4 gb/s | 15.9 gb/s | 16.7 gb/s |
|
|
538
543
|
| __i64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
539
|
-
| `nk_reduce_moments_i64_serial` |
|
|
540
|
-
| `nk_reduce_moments_i64_v128relaxed` |
|
|
541
|
-
| `nk_reduce_minmax_i64_serial` |
|
|
542
|
-
| `nk_reduce_minmax_i64_v128relaxed` |
|
|
544
|
+
| `nk_reduce_moments_i64_serial` | 9.52 gb/s | 9.73 gb/s | 9.62 gb/s |
|
|
545
|
+
| `nk_reduce_moments_i64_v128relaxed` | 7.42 gb/s | 7.40 gb/s | 7.50 gb/s |
|
|
546
|
+
| `nk_reduce_minmax_i64_serial` | 13.6 gb/s | 13.7 gb/s | 13.7 gb/s |
|
|
547
|
+
| `nk_reduce_minmax_i64_v128relaxed` | 16.4 gb/s | 16.9 gb/s | 17.0 gb/s |
|
|
543
548
|
| __u64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
544
|
-
| `nk_reduce_moments_u64_serial` |
|
|
545
|
-
| `nk_reduce_moments_u64_v128relaxed` |
|
|
546
|
-
| `nk_reduce_minmax_u64_serial` |
|
|
547
|
-
| `nk_reduce_minmax_u64_v128relaxed` |
|
|
549
|
+
| `nk_reduce_moments_u64_serial` | 10.2 gb/s | 10.3 gb/s | 10.5 gb/s |
|
|
550
|
+
| `nk_reduce_moments_u64_v128relaxed` | 3.04 gb/s | 2.98 gb/s | 2.94 gb/s |
|
|
551
|
+
| `nk_reduce_minmax_u64_serial` | 13.6 gb/s | 13.7 gb/s | 13.7 gb/s |
|
|
552
|
+
| `nk_reduce_minmax_u64_v128relaxed` | 3.31 gb/s | 3.33 gb/s | 3.35 gb/s |
|