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
|
@@ -4,15 +4,15 @@ NumKong implements batched GEMM computing C = A × Bᵀ (packed) and C = A × A
|
|
|
4
4
|
|
|
5
5
|
Packed dot product computes the full cross-product matrix:
|
|
6
6
|
|
|
7
|
-
|
|
7
|
+
$$
|
|
8
8
|
C_{ij} = \sum_{k} A_{ik} \cdot B_{jk}^T
|
|
9
|
-
|
|
9
|
+
$$
|
|
10
10
|
|
|
11
11
|
Symmetric dot product uses the same matrix for both operands:
|
|
12
12
|
|
|
13
|
-
|
|
13
|
+
$$
|
|
14
14
|
C_{ij} = \sum_{k} A_{ik} \cdot A_{jk}
|
|
15
|
-
|
|
15
|
+
$$
|
|
16
16
|
|
|
17
17
|
Reformulating as Python pseudocode:
|
|
18
18
|
|
|
@@ -142,85 +142,83 @@ Workloads that significantly degrade CPU frequencies (Intel AMX, Apple SME) run
|
|
|
142
142
|
| `dots_packed_f64_with_blas` 🧩 | 58.7 gso/s, 16 ulp | 73.1 gso/s, 58 ulp | 73.8 gso/s, 56.2 ulp |
|
|
143
143
|
| `dots_packed_f64_with_mkl` 🧩 | 59.9 gso/s, 16 ulp | 73.7 gso/s, 58 ulp | 73.3 gso/s, 56.2 ulp |
|
|
144
144
|
| `dots_symmetric_f64_with_blas` 🧩 | 50.8 gso/s, 13 ulp | 70.4 gso/s, 30 ulp | 74 gso/s, 50.8 ulp |
|
|
145
|
-
| `nk_dots_packed_f64_serial` | 0.
|
|
146
|
-
| `nk_dots_symmetric_f64_serial` | 0.
|
|
147
|
-
| `nk_dots_packed_f64_haswell` | 5.
|
|
148
|
-
| `nk_dots_symmetric_f64_haswell` |
|
|
149
|
-
| `nk_dots_packed_f64_skylake` | 8.
|
|
150
|
-
| `nk_dots_symmetric_f64_skylake` | 7.
|
|
145
|
+
| `nk_dots_packed_f64_serial` | 0.850 gso/s, 2 ulp | 0.846 gso/s, 4.6 ulp | 0.862 gso/s, 5.9 ulp |
|
|
146
|
+
| `nk_dots_symmetric_f64_serial` | 0.484 gso/s, 2 ulp | 0.472 gso/s, 2.9 ulp | 0.471 gso/s, 3.9 ulp |
|
|
147
|
+
| `nk_dots_packed_f64_haswell` | 5.93 gso/s, 0 ulp | 6.11 gso/s, 0 ulp | 6.16 gso/s, 0 ulp |
|
|
148
|
+
| `nk_dots_symmetric_f64_haswell` | 5.68 gso/s, 0 ulp | 5.99 gso/s, 0 ulp | 5.86 gso/s, 0 ulp |
|
|
149
|
+
| `nk_dots_packed_f64_skylake` | 8.26 gso/s, 0 ulp | 9.27 gso/s, 0 ulp | 9.06 gso/s, 0 ulp |
|
|
150
|
+
| `nk_dots_symmetric_f64_skylake` | 7.53 gso/s, 0 ulp | 8.63 gso/s, 0 ulp | 8.58 gso/s, 0 ulp |
|
|
151
151
|
| __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
152
152
|
| `dots_packed_f32_with_blas` 🧩 | 113 gso/s, 18 ulp | 139 gso/s, 30 ulp | 147 gso/s, 267 ulp |
|
|
153
153
|
| `dots_symmetric_f32_with_blas` 🧩 | 94.5 gso/s, 23 ulp | 126 gso/s, 39 ulp | 146 gso/s, 260 ulp |
|
|
154
|
-
| `nk_dots_packed_f32_serial` | 9.
|
|
155
|
-
| `nk_dots_symmetric_f32_serial` |
|
|
156
|
-
| `nk_dots_packed_f32_haswell` | 30.
|
|
157
|
-
| `nk_dots_symmetric_f32_haswell` |
|
|
158
|
-
| `nk_dots_packed_f32_skylake` |
|
|
159
|
-
| `nk_dots_symmetric_f32_skylake` |
|
|
154
|
+
| `nk_dots_packed_f32_serial` | 9.98 gso/s, 5.3 ulp | 10.1 gso/s, 11.8 ulp | 10.1 gso/s, 14.5 ulp |
|
|
155
|
+
| `nk_dots_symmetric_f32_serial` | 4.96 gso/s, 11.1 ulp | 5.01 gso/s, 13.4 ulp | 5.01 gso/s, 14.1 ulp |
|
|
156
|
+
| `nk_dots_packed_f32_haswell` | 30.4 gso/s, 0 ulp | 32.5 gso/s, 0 ulp | 31.9 gso/s, 0 ulp |
|
|
157
|
+
| `nk_dots_symmetric_f32_haswell` | 15.5 gso/s, 0 ulp | 17.9 gso/s, 0 ulp | 18.4 gso/s, 0 ulp |
|
|
158
|
+
| `nk_dots_packed_f32_skylake` | 35.4 gso/s, 0 ulp | 41.4 gso/s, 0 ulp | 40.0 gso/s, 0 ulp |
|
|
159
|
+
| `nk_dots_symmetric_f32_skylake` | 22.4 gso/s, 0 ulp | 28.2 gso/s, 0 ulp | 28.1 gso/s, 0 ulp |
|
|
160
160
|
| __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
161
|
-
| `dots_packed_bf16_with_mkl` 🧩 |
|
|
162
|
-
| `nk_dots_packed_bf16_serial` |
|
|
163
|
-
| `nk_dots_symmetric_bf16_serial` |
|
|
164
|
-
| `nk_dots_packed_bf16_haswell` |
|
|
165
|
-
| `nk_dots_symmetric_bf16_haswell` |
|
|
166
|
-
| `nk_dots_packed_bf16_skylake` |
|
|
167
|
-
| `nk_dots_symmetric_bf16_skylake` |
|
|
168
|
-
| `nk_dots_packed_bf16_genoa` |
|
|
169
|
-
| `nk_dots_symmetric_bf16_genoa` |
|
|
170
|
-
| `nk_dots_packed_bf16_sapphireamx` |
|
|
171
|
-
| `nk_dots_symmetric_bf16_sapphireamx` |
|
|
161
|
+
| `dots_packed_bf16_with_mkl` 🧩 | 182 gso/s, 0 ulp | 523 gso/s, 0.7 ulp | 847 gso/s, 5.8 ulp |
|
|
162
|
+
| `nk_dots_packed_bf16_serial` | 1.20 gso/s, 0 ulp | 1.21 gso/s, 0.5 ulp | 1.22 gso/s, 5.4 ulp |
|
|
163
|
+
| `nk_dots_symmetric_bf16_serial` | 1.16 gso/s, 0 ulp | 1.19 gso/s, 0.9 ulp | 1.18 gso/s, 5.4 ulp |
|
|
164
|
+
| `nk_dots_packed_bf16_haswell` | 65.6 gso/s, 0 ulp | 73.3 gso/s, 0.3 ulp | 76.8 gso/s, 4.4 ulp |
|
|
165
|
+
| `nk_dots_symmetric_bf16_haswell` | 40.2 gso/s, 0 ulp | 55.6 gso/s, 0.5 ulp | 60.8 gso/s, 4.6 ulp |
|
|
166
|
+
| `nk_dots_packed_bf16_skylake` | 79.8 gso/s, 0 ulp | 92.1 gso/s, 0.3 ulp | 102 gso/s, 3.5 ulp |
|
|
167
|
+
| `nk_dots_symmetric_bf16_skylake` | 57.4 gso/s, 0 ulp | 78.9 gso/s, 0.5 ulp | 82.5 gso/s, 3.5 ulp |
|
|
168
|
+
| `nk_dots_packed_bf16_genoa` | 65.8 gso/s, 0 ulp | 83.2 gso/s, 0.3 ulp | 88.9 gso/s, 3.5 ulp |
|
|
169
|
+
| `nk_dots_symmetric_bf16_genoa` | 52.5 gso/s, 0 ulp | 70.5 gso/s, 0.5 ulp | 76.0 gso/s, 3.5 ulp |
|
|
170
|
+
| `nk_dots_packed_bf16_sapphireamx` | 348 gso/s, 0 ulp | 706 gso/s, 0.7 ulp | 667 gso/s, 5.8 ulp |
|
|
171
|
+
| `nk_dots_symmetric_bf16_sapphireamx` | 84.2 gso/s, 0 ulp | 120 gso/s, 0.5 ulp | 120 gso/s, 5.8 ulp |
|
|
172
172
|
| __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
173
173
|
| `dots_packed_f16_with_mkl` 🧩 | 123 gso/s, 17 ulp | 138 gso/s, 31 ulp | 138 gso/s, 39.5 ulp |
|
|
174
|
-
| `nk_dots_packed_f16_serial` |
|
|
175
|
-
| `nk_dots_symmetric_f16_serial` |
|
|
176
|
-
| `nk_dots_packed_f16_haswell` |
|
|
177
|
-
| `nk_dots_symmetric_f16_haswell` |
|
|
178
|
-
| `nk_dots_packed_f16_skylake` | 74.
|
|
179
|
-
| `nk_dots_symmetric_f16_skylake` |
|
|
174
|
+
| `nk_dots_packed_f16_serial` | 8.19 gso/s, 14 ulp | 8.21 gso/s, 40 ulp | 8.11 gso/s, 326 ulp |
|
|
175
|
+
| `nk_dots_symmetric_f16_serial` | 4.02 gso/s, 8.9 ulp | 4.04 gso/s, 25 ulp | 4.03 gso/s, 55.6 ulp |
|
|
176
|
+
| `nk_dots_packed_f16_haswell` | 65.1 gso/s, 12 ulp | 74.4 gso/s, 22 ulp | 71.5 gso/s, 374 ulp |
|
|
177
|
+
| `nk_dots_symmetric_f16_haswell` | 34.4 gso/s, 7.7 ulp | 44.0 gso/s, 32 ulp | 46.5 gso/s, 486 ulp |
|
|
178
|
+
| `nk_dots_packed_f16_skylake` | 74.7 gso/s, 7.3 ulp | 99.0 gso/s, 21 ulp | 94.0 gso/s, 138 ulp |
|
|
179
|
+
| `nk_dots_symmetric_f16_skylake` | 40.9 gso/s, 5.9 ulp | 56.8 gso/s, 25 ulp | 58.8 gso/s, 32 ulp |
|
|
180
180
|
| __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
181
|
-
| `nk_dots_packed_e5m2_serial` |
|
|
182
|
-
| `nk_dots_symmetric_e5m2_serial` |
|
|
183
|
-
| `nk_dots_packed_e5m2_haswell` |
|
|
184
|
-
| `nk_dots_symmetric_e5m2_haswell` | 15.
|
|
185
|
-
| `nk_dots_packed_e5m2_skylake` |
|
|
186
|
-
| `nk_dots_symmetric_e5m2_skylake` |
|
|
187
|
-
| `nk_dots_packed_e5m2_genoa` |
|
|
188
|
-
| `nk_dots_symmetric_e5m2_genoa` |
|
|
189
|
-
| `nk_dots_packed_e5m2_sapphireamx` |
|
|
181
|
+
| `nk_dots_packed_e5m2_serial` | 4.86 gso/s, 0 ulp | 4.75 gso/s, 0 ulp | 4.88 gso/s, 0 ulp |
|
|
182
|
+
| `nk_dots_symmetric_e5m2_serial` | 3.97 gso/s, 0 ulp | 4.28 gso/s, 0 ulp | 4.50 gso/s, 0 ulp |
|
|
183
|
+
| `nk_dots_packed_e5m2_haswell` | 29.1 gso/s, 0 ulp | 31.5 gso/s, 0 ulp | 30.6 gso/s, 0 ulp |
|
|
184
|
+
| `nk_dots_symmetric_e5m2_haswell` | 15.6 gso/s, 0 ulp | 16.4 gso/s, 0 ulp | 17.0 gso/s, 0 ulp |
|
|
185
|
+
| `nk_dots_packed_e5m2_skylake` | 34.6 gso/s, 0 ulp | 37.9 gso/s, 0 ulp | 38.9 gso/s, 0 ulp |
|
|
186
|
+
| `nk_dots_symmetric_e5m2_skylake` | 21.2 gso/s, 0 ulp | 22.7 gso/s, 0 ulp | 22.5 gso/s, 0 ulp |
|
|
187
|
+
| `nk_dots_packed_e5m2_genoa` | 41.7 gso/s, 0 ulp | 48.7 gso/s, 0 ulp | 49.1 gso/s, 0 ulp |
|
|
188
|
+
| `nk_dots_symmetric_e5m2_genoa` | 30.0 gso/s, 0 ulp | 33.3 gso/s, 0 ulp | 33.7 gso/s, 0 ulp |
|
|
189
|
+
| `nk_dots_packed_e5m2_sapphireamx` | 254 gso/s, 0 ulp | 407 gso/s, 0 ulp | 419 gso/s, 0 ulp |
|
|
190
|
+
| `nk_dots_symmetric_e5m2_sapphireamx` | 50.9 gso/s, 0 ulp | 69.9 gso/s, 0 ulp | 67.4 gso/s, 0 ulp |
|
|
190
191
|
| __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
191
|
-
| `nk_dots_packed_e4m3_serial` | 0.
|
|
192
|
-
| `nk_dots_symmetric_e4m3_serial` | 0.
|
|
193
|
-
| `nk_dots_packed_e4m3_haswell` |
|
|
194
|
-
| `nk_dots_symmetric_e4m3_haswell` |
|
|
195
|
-
| `nk_dots_packed_e4m3_skylake` |
|
|
196
|
-
| `nk_dots_symmetric_e4m3_skylake` |
|
|
197
|
-
| `nk_dots_packed_e4m3_genoa` |
|
|
198
|
-
| `nk_dots_symmetric_e4m3_genoa` |
|
|
199
|
-
| `nk_dots_packed_e4m3_sapphireamx` |
|
|
192
|
+
| `nk_dots_packed_e4m3_serial` | 0.489 gso/s, 0 ulp | 0.499 gso/s, 0 ulp | 0.489 gso/s, 0 ulp |
|
|
193
|
+
| `nk_dots_symmetric_e4m3_serial` | 0.394 gso/s, 0 ulp | 0.390 gso/s, 0 ulp | 0.391 gso/s, 0 ulp |
|
|
194
|
+
| `nk_dots_packed_e4m3_haswell` | 24.3 gso/s, 0 ulp | 26.1 gso/s, 0 ulp | 25.2 gso/s, 0 ulp |
|
|
195
|
+
| `nk_dots_symmetric_e4m3_haswell` | 13.5 gso/s, 0 ulp | 14.0 gso/s, 0 ulp | 14.3 gso/s, 0 ulp |
|
|
196
|
+
| `nk_dots_packed_e4m3_skylake` | 31.6 gso/s, 0 ulp | 32.6 gso/s, 0 ulp | 34.0 gso/s, 0 ulp |
|
|
197
|
+
| `nk_dots_symmetric_e4m3_skylake` | 17.3 gso/s, 0 ulp | 18.2 gso/s, 0 ulp | 18.6 gso/s, 0 ulp |
|
|
198
|
+
| `nk_dots_packed_e4m3_genoa` | 38.6 gso/s, 0 ulp | 43.8 gso/s, 0 ulp | 43.7 gso/s, 0 ulp |
|
|
199
|
+
| `nk_dots_symmetric_e4m3_genoa` | 27.3 gso/s, 0 ulp | 29.4 gso/s, 0 ulp | 29.2 gso/s, 0 ulp |
|
|
200
|
+
| `nk_dots_packed_e4m3_sapphireamx` | 222 gso/s, 0 ulp | 333 gso/s, 0 ulp | 332 gso/s, 0 ulp |
|
|
201
|
+
| `nk_dots_symmetric_e4m3_sapphireamx` | 33.1 gso/s, 0 ulp | 36.3 gso/s, 0 ulp | 35.4 gso/s, 0 ulp |
|
|
200
202
|
| __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
201
|
-
| `nk_dots_packed_e3m2_serial` |
|
|
202
|
-
| `nk_dots_symmetric_e3m2_serial` |
|
|
203
|
-
| `nk_dots_packed_e3m2_haswell` |
|
|
204
|
-
| `nk_dots_symmetric_e3m2_haswell` |
|
|
205
|
-
| `nk_dots_packed_e3m2_skylake` |
|
|
206
|
-
| `nk_dots_symmetric_e3m2_skylake` |
|
|
207
|
-
| `
|
|
208
|
-
| `
|
|
209
|
-
| `nk_dots_packed_e3m2_sapphireamx` | 255 gso/s, 0 ulp | 448 gso/s, 0 ulp | 433 gso/s, 0 ulp |
|
|
210
|
-
| `nk_dots_symmetric_e3m2_sapphireamx` | 65.1 gso/s, 0 ulp | 101 gso/s, 0 ulp | 91 gso/s, 0 ulp |
|
|
203
|
+
| `nk_dots_packed_e3m2_serial` | 4.97 gso/s, 0 ulp | 4.90 gso/s, 0 ulp | 5.03 gso/s, 0 ulp |
|
|
204
|
+
| `nk_dots_symmetric_e3m2_serial` | 3.40 gso/s, 0 ulp | 3.81 gso/s, 0 ulp | 3.88 gso/s, 0 ulp |
|
|
205
|
+
| `nk_dots_packed_e3m2_haswell` | 31.0 gso/s, 0 ulp | 32.2 gso/s, 0 ulp | 33.9 gso/s, 0 ulp |
|
|
206
|
+
| `nk_dots_symmetric_e3m2_haswell` | 29.0 gso/s, 0 ulp | 31.7 gso/s, 0 ulp | 31.1 gso/s, 0 ulp |
|
|
207
|
+
| `nk_dots_packed_e3m2_skylake` | 39.3 gso/s, 0 ulp | 43.4 gso/s, 0 ulp | 44.1 gso/s, 0 ulp |
|
|
208
|
+
| `nk_dots_symmetric_e3m2_skylake` | 40.0 gso/s, 0 ulp | 46.6 gso/s, 0 ulp | 47.1 gso/s, 0 ulp |
|
|
209
|
+
| `nk_dots_packed_e3m2_sapphireamx` | 263 gso/s, 0 ulp | 471 gso/s, 0 ulp | 471 gso/s, 0 ulp |
|
|
210
|
+
| `nk_dots_symmetric_e3m2_sapphireamx` | 62.9 gso/s, 0 ulp | 101 gso/s, 0 ulp | 89.1 gso/s, 0 ulp |
|
|
211
211
|
| __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
212
|
-
| `nk_dots_packed_e2m3_serial` |
|
|
213
|
-
| `nk_dots_symmetric_e2m3_serial` |
|
|
214
|
-
| `nk_dots_packed_e2m3_haswell` |
|
|
215
|
-
| `nk_dots_symmetric_e2m3_haswell` | 50.
|
|
216
|
-
| `nk_dots_packed_e2m3_skylake` | 69.
|
|
217
|
-
| `nk_dots_symmetric_e2m3_skylake` |
|
|
218
|
-
| `
|
|
219
|
-
| `
|
|
220
|
-
| `
|
|
221
|
-
| `
|
|
222
|
-
| `nk_dots_packed_e2m3_alder` | 71.8 gso/s, 0 ulp | 80.8 gso/s, 0 ulp | 89.5 gso/s, 0 ulp |
|
|
223
|
-
| `nk_dots_symmetric_e2m3_alder` | 59.3 gso/s, 0 ulp | 73.4 gso/s, 0 ulp | 78.6 gso/s, 0 ulp |
|
|
212
|
+
| `nk_dots_packed_e2m3_serial` | 4.98 gso/s, 0 ulp | 4.95 gso/s, 0 ulp | 5.00 gso/s, 0 ulp |
|
|
213
|
+
| `nk_dots_symmetric_e2m3_serial` | 3.48 gso/s, 0 ulp | 3.83 gso/s, 0 ulp | 3.85 gso/s, 0 ulp |
|
|
214
|
+
| `nk_dots_packed_e2m3_haswell` | 58.6 gso/s, 0 ulp | 62.5 gso/s, 0 ulp | 65.3 gso/s, 0 ulp |
|
|
215
|
+
| `nk_dots_symmetric_e2m3_haswell` | 50.5 gso/s, 0 ulp | 61.2 gso/s, 0 ulp | 64.2 gso/s, 0 ulp |
|
|
216
|
+
| `nk_dots_packed_e2m3_skylake` | 69.8 gso/s, 0 ulp | 81.8 gso/s, 0 ulp | 88.4 gso/s, 0 ulp |
|
|
217
|
+
| `nk_dots_symmetric_e2m3_skylake` | 65.5 gso/s, 0 ulp | 83.4 gso/s, 0 ulp | 84.6 gso/s, 0 ulp |
|
|
218
|
+
| `nk_dots_packed_e2m3_sapphireamx` | 419 gso/s, 0 ulp | 1,195 gso/s, 0 ulp | 1,067 gso/s, 0 ulp |
|
|
219
|
+
| `nk_dots_symmetric_e2m3_sapphireamx` | 94.5 gso/s, 0 ulp | 213 gso/s, 0 ulp | 184 gso/s, 0 ulp |
|
|
220
|
+
| `nk_dots_packed_e2m3_alder` | 72.9 gso/s, 0 ulp | 78.6 gso/s, 0 ulp | 85.7 gso/s, 0 ulp |
|
|
221
|
+
| `nk_dots_symmetric_e2m3_alder` | 61.6 gso/s, 0 ulp | 75.2 gso/s, 0 ulp | 54.9 gso/s, 0 ulp |
|
|
224
222
|
| __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
225
223
|
| `dots_packed_i8u8_with_mkl` 🧩 | 250 gso/s | 627 gso/s | 1,670 gso/s |
|
|
226
224
|
| `nk_dots_packed_i8_serial` | 6.44 gso/s | 6.62 gso/s | 7.44 gso/s |
|
|
@@ -327,170 +325,172 @@ Measured with Wasmtime v42 (Cranelift backend).
|
|
|
327
325
|
| `nk_dots_symmetric_u1_serial` | 7.42 gso/s | 27.9 gso/s | 87.3 gso/s |
|
|
328
326
|
| `nk_dots_symmetric_u1_v128relaxed` | 7.35 gso/s | 27.5 gso/s | 81.9 gso/s |
|
|
329
327
|
|
|
330
|
-
### Apple
|
|
328
|
+
### Apple M5
|
|
331
329
|
|
|
332
330
|
#### Native
|
|
333
331
|
|
|
334
332
|
| Kernel | 256³ | 1024³ | 4096³ |
|
|
335
333
|
| :--------------------------------- | -----------------------: | -----------------------: | -----------------------: |
|
|
336
334
|
| __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
337
|
-
| `nk_dots_packed_f64_serial` |
|
|
338
|
-
| `nk_dots_symmetric_f64_serial` | 1.
|
|
339
|
-
| `nk_dots_packed_f64_neon` |
|
|
340
|
-
| `nk_dots_symmetric_f64_neon` |
|
|
341
|
-
| `nk_dots_packed_f64_smef64` |
|
|
342
|
-
| `nk_dots_symmetric_f64_smef64` | 5
|
|
335
|
+
| `nk_dots_packed_f64_serial` | 2.49 gso/s, 3 ulp | 2.36 gso/s, 5 ulp | 2.48 gso/s, 6 ulp |
|
|
336
|
+
| `nk_dots_symmetric_f64_serial` | 1.38 gso/s, 0 ulp | 1.36 gso/s, 0 ulp | 1.49 gso/s, 0 ulp |
|
|
337
|
+
| `nk_dots_packed_f64_neon` | 6.31 gso/s, 0 ulp | 6.00 gso/s, 0 ulp | 6.34 gso/s, 0 ulp |
|
|
338
|
+
| `nk_dots_symmetric_f64_neon` | 5.57 gso/s, 0 ulp | 5.41 gso/s, 0 ulp | 5.40 gso/s, 0 ulp |
|
|
339
|
+
| `nk_dots_packed_f64_smef64` | 45.9 gso/s, 1.5 ulp | 46.3 gso/s, 1.1 ulp | 46.2 gso/s, 0.9 ulp |
|
|
340
|
+
| `nk_dots_symmetric_f64_smef64` | 22.5 gso/s, 1.5 ulp | 24.3 gso/s, 1.2 ulp | 21.3 gso/s, 1.1 ulp |
|
|
343
341
|
| __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
344
|
-
| `nk_dots_packed_f32_serial` |
|
|
345
|
-
| `nk_dots_symmetric_f32_serial` | 8.
|
|
346
|
-
| `nk_dots_packed_f32_neon` |
|
|
347
|
-
| `nk_dots_symmetric_f32_neon` | 10.
|
|
348
|
-
| `nk_dots_packed_f32_smef64` |
|
|
349
|
-
| `nk_dots_symmetric_f32_smef64` |
|
|
342
|
+
| `nk_dots_packed_f32_serial` | 12.0 gso/s, 19 ulp | 11.4 gso/s, 30 ulp | 12.2 gso/s, 725 ulp |
|
|
343
|
+
| `nk_dots_symmetric_f32_serial` | 8.75 gso/s, 3.1 ulp | 9.15 gso/s, 12.8 ulp | 9.62 gso/s, 39.9 ulp |
|
|
344
|
+
| `nk_dots_packed_f32_neon` | 42.5 gso/s, 0 ulp | 40.6 gso/s, 0 ulp | 42.0 gso/s, 0 ulp |
|
|
345
|
+
| `nk_dots_symmetric_f32_neon` | 10.9 gso/s, 4.6 ulp | 10.5 gso/s, 17.7 ulp | 10.8 gso/s, 59 ulp |
|
|
346
|
+
| `nk_dots_packed_f32_smef64` | 236 gso/s, 0 ulp | 268 gso/s, 15 ulp | 221 gso/s, 0 ulp |
|
|
347
|
+
| `nk_dots_symmetric_f32_smef64` | 78.1 gso/s, 4.3 ulp | 94.1 gso/s, 19.0 ulp | 55.3 gso/s, 0 ulp |
|
|
350
348
|
| __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
351
|
-
| `nk_dots_packed_bf16_serial` |
|
|
352
|
-
| `nk_dots_symmetric_bf16_serial` |
|
|
353
|
-
| `
|
|
354
|
-
| `
|
|
355
|
-
| `
|
|
356
|
-
| `
|
|
349
|
+
| `nk_dots_packed_bf16_serial` | 20.4 gso/s, 0.1 ulp | 19.6 gso/s, 0.5 ulp | 20.3 gso/s, 5 ulp |
|
|
350
|
+
| `nk_dots_symmetric_bf16_serial` | 16.3 gso/s, 0.01 ulp | 16.9 gso/s, 0.7 ulp | 17.8 gso/s, 115 ulp |
|
|
351
|
+
| `nk_dots_packed_bf16_neon` | 83.0 gso/s, 0 ulp | 80.2 gso/s, 0 ulp | 84.0 gso/s, 0 ulp |
|
|
352
|
+
| `nk_dots_symmetric_bf16_neon` | 39.5 gso/s, 0 ulp | 41.2 gso/s, 0 ulp | 41.9 gso/s, 0 ulp |
|
|
353
|
+
| `nk_dots_packed_bf16_neonbfdot` | 57.9 gso/s, 0 ulp | 58.5 gso/s, 0.5 ulp | 63.4 gso/s, 7.2 ulp |
|
|
354
|
+
| `nk_dots_symmetric_bf16_neonbfdot` | 38.6 gso/s, 0 ulp | 41.1 gso/s, 0.5 ulp | 43.5 gso/s, 0 ulp |
|
|
355
|
+
| `nk_dots_packed_bf16_sme` | 1,106 gso/s, 0 ulp | 1,208 gso/s, 4.2 ulp | 1,190 gso/s, 3.8 ulp |
|
|
356
|
+
| `nk_dots_symmetric_bf16_sme` | 606 gso/s, 0.07 ulp | 650 gso/s, 1.2 ulp | 458 gso/s, 1.8 ulp |
|
|
357
357
|
| __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
358
|
-
| `nk_dots_packed_f16_serial` |
|
|
359
|
-
| `nk_dots_symmetric_f16_serial` |
|
|
360
|
-
| `nk_dots_packed_f16_neonhalf` |
|
|
361
|
-
| `nk_dots_symmetric_f16_neonhalf` | 20.
|
|
362
|
-
| `nk_dots_packed_f16_neonfhm` |
|
|
363
|
-
| `nk_dots_symmetric_f16_neonfhm` |
|
|
364
|
-
| `nk_dots_packed_f16_sme` |
|
|
365
|
-
| `nk_dots_symmetric_f16_sme` |
|
|
358
|
+
| `nk_dots_packed_f16_serial` | 14.8 gso/s, 204 ulp | 14.2 gso/s, 36 ulp | 14.8 gso/s, 326 ulp |
|
|
359
|
+
| `nk_dots_symmetric_f16_serial` | 24.3 gso/s, 13 ulp | 24.9 gso/s, 24.6 ulp | 26.7 gso/s, 506 ulp |
|
|
360
|
+
| `nk_dots_packed_f16_neonhalf` | 77.0 gso/s, 16.8 ulp | 79.1 gso/s, 25.5 ulp | 84.2 gso/s, 618 ulp |
|
|
361
|
+
| `nk_dots_symmetric_f16_neonhalf` | 20.5 gso/s, 12.1 ulp | 20.4 gso/s, 25.0 ulp | 22.5 gso/s, 506 ulp |
|
|
362
|
+
| `nk_dots_packed_f16_neonfhm` | 104 gso/s, 16.7 ulp | 110 gso/s, 25.5 ulp | 118 gso/s, 618 ulp |
|
|
363
|
+
| `nk_dots_symmetric_f16_neonfhm` | 34.5 gso/s, 12.1 ulp | 40.4 gso/s, 25.0 ulp | 41.5 gso/s, 506 ulp |
|
|
364
|
+
| `nk_dots_packed_f16_sme` | 1,106 gso/s, 14.8 ulp | 1,213 gso/s, 28.2 ulp | 1,190 gso/s, 28.2 ulp |
|
|
365
|
+
| `nk_dots_symmetric_f16_sme` | 607 gso/s, 12.1 ulp | 636 gso/s, 23.8 ulp | 458 gso/s, 24.4 ulp |
|
|
366
366
|
| __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
367
|
-
| `nk_dots_packed_e5m2_serial` |
|
|
368
|
-
| `nk_dots_symmetric_e5m2_serial` | 7.
|
|
369
|
-
| `nk_dots_packed_e5m2_neonfhm` |
|
|
370
|
-
| `nk_dots_symmetric_e5m2_neonfhm` | 61.0 gso/s, 0 ulp |
|
|
371
|
-
| `nk_dots_packed_e5m2_sme` |
|
|
372
|
-
| `nk_dots_symmetric_e5m2_sme` |
|
|
367
|
+
| `nk_dots_packed_e5m2_serial` | 15.9 gso/s, 0 ulp | 16.7 gso/s, 0 ulp | 17.2 gso/s, 0 ulp |
|
|
368
|
+
| `nk_dots_symmetric_e5m2_serial` | 7.56 gso/s, 0 ulp | 8.37 gso/s, 0 ulp | 8.99 gso/s, 0 ulp |
|
|
369
|
+
| `nk_dots_packed_e5m2_neonfhm` | 88.1 gso/s, 0 ulp | 97.3 gso/s, 0 ulp | 103 gso/s, 0 ulp |
|
|
370
|
+
| `nk_dots_symmetric_e5m2_neonfhm` | 61.0 gso/s, 0 ulp | 73.2 gso/s, 0 ulp | 79.3 gso/s, 0 ulp |
|
|
371
|
+
| `nk_dots_packed_e5m2_sme` | 729 gso/s, 0 ulp | 800 gso/s, 0 ulp | 792 gso/s, 0 ulp |
|
|
372
|
+
| `nk_dots_symmetric_e5m2_sme` | 208 gso/s, 0 ulp | 227 gso/s, 0 ulp | 229 gso/s, 0 ulp |
|
|
373
373
|
| __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
374
|
-
| `nk_dots_packed_e4m3_serial` | 1.
|
|
375
|
-
| `nk_dots_symmetric_e4m3_serial` | 1.
|
|
376
|
-
| `nk_dots_packed_e4m3_neonfhm` |
|
|
377
|
-
| `nk_dots_symmetric_e4m3_neonfhm` |
|
|
378
|
-
| `nk_dots_packed_e4m3_sme` |
|
|
379
|
-
| `nk_dots_symmetric_e4m3_sme` |
|
|
374
|
+
| `nk_dots_packed_e4m3_serial` | 1.24 gso/s, 0 ulp | 1.20 gso/s, 0 ulp | 1.24 gso/s, 0 ulp |
|
|
375
|
+
| `nk_dots_symmetric_e4m3_serial` | 1.20 gso/s, 0 ulp | 1.24 gso/s, 0 ulp | 1.32 gso/s, 0 ulp |
|
|
376
|
+
| `nk_dots_packed_e4m3_neonfhm` | 29.6 gso/s, 0 ulp | 32.2 gso/s, 0 ulp | 34.1 gso/s, 0 ulp |
|
|
377
|
+
| `nk_dots_symmetric_e4m3_neonfhm` | 32.0 gso/s, 0 ulp | 36.6 gso/s, 0 ulp | 38.9 gso/s, 0 ulp |
|
|
378
|
+
| `nk_dots_packed_e4m3_sme` | 284 gso/s, 0 ulp | 314 gso/s, 0 ulp | 316 gso/s, 0 ulp |
|
|
379
|
+
| `nk_dots_symmetric_e4m3_sme` | 74.3 gso/s, 0 ulp | 80.9 gso/s, 0 ulp | 77.8 gso/s, 0 ulp |
|
|
380
380
|
| __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
381
|
-
| `nk_dots_packed_e3m2_serial` |
|
|
382
|
-
| `nk_dots_symmetric_e3m2_serial` | 7.
|
|
383
|
-
| `nk_dots_packed_e3m2_sme` |
|
|
384
|
-
| `nk_dots_symmetric_e3m2_sme` |
|
|
381
|
+
| `nk_dots_packed_e3m2_serial` | 14.0 gso/s, 0 ulp | 14.6 gso/s, 0 ulp | 15.5 gso/s, 0 ulp |
|
|
382
|
+
| `nk_dots_symmetric_e3m2_serial` | 7.51 gso/s, 0 ulp | 8.10 gso/s, 0 ulp | 9.05 gso/s, 0 ulp |
|
|
383
|
+
| `nk_dots_packed_e3m2_sme` | 671 gso/s, 0 ulp | 738 gso/s, 0 ulp | 730 gso/s, 0 ulp |
|
|
384
|
+
| `nk_dots_symmetric_e3m2_sme` | 191 gso/s, 0 ulp | 206 gso/s, 0 ulp | 207 gso/s, 0 ulp |
|
|
385
385
|
| __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
386
|
-
| `nk_dots_packed_e2m3_serial` |
|
|
387
|
-
| `nk_dots_symmetric_e2m3_serial` | 7.
|
|
388
|
-
| `nk_dots_packed_e2m3_sme` |
|
|
389
|
-
| `nk_dots_symmetric_e2m3_sme` |
|
|
386
|
+
| `nk_dots_packed_e2m3_serial` | 14.4 gso/s, 0 ulp | 14.8 gso/s, 0 ulp | 15.5 gso/s, 0 ulp |
|
|
387
|
+
| `nk_dots_symmetric_e2m3_serial` | 7.58 gso/s, 0 ulp | 8.21 gso/s, 0 ulp | 9.09 gso/s, 0 ulp |
|
|
388
|
+
| `nk_dots_packed_e2m3_sme` | 1,211 gso/s, 0 ulp | 1,404 gso/s, 0 ulp | 1,313 gso/s, 0 ulp |
|
|
389
|
+
| `nk_dots_symmetric_e2m3_sme` | 372 gso/s, 0 ulp | 410 gso/s, 0 ulp | 416 gso/s, 0 ulp |
|
|
390
390
|
| __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
391
|
-
| `nk_dots_packed_i8_serial` |
|
|
392
|
-
| `nk_dots_symmetric_i8_serial` |
|
|
393
|
-
| `nk_dots_packed_i8_neonsdot` |
|
|
394
|
-
| `nk_dots_symmetric_i8_neonsdot` |
|
|
395
|
-
| `nk_dots_packed_i8_sme` |
|
|
396
|
-
| `nk_dots_symmetric_i8_sme` |
|
|
391
|
+
| `nk_dots_packed_i8_serial` | 18.9 gso/s | 20.0 gso/s | 20.2 gso/s |
|
|
392
|
+
| `nk_dots_symmetric_i8_serial` | 12.6 gso/s | 13.9 gso/s | 14.8 gso/s |
|
|
393
|
+
| `nk_dots_packed_i8_neonsdot` | 345 gso/s | 419 gso/s | 477 gso/s |
|
|
394
|
+
| `nk_dots_symmetric_i8_neonsdot` | 76.6 gso/s | 86.9 gso/s | 87.2 gso/s |
|
|
395
|
+
| `nk_dots_packed_i8_sme` | 2,348 gso/s | 2,687 gso/s | 2,570 gso/s |
|
|
396
|
+
| `nk_dots_symmetric_i8_sme` | 1,390 gso/s | 1,531 gso/s | 1,369 gso/s |
|
|
397
397
|
| __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
398
|
-
| `nk_dots_packed_u8_serial` |
|
|
399
|
-
| `nk_dots_symmetric_u8_serial` |
|
|
400
|
-
| `nk_dots_packed_u8_neonsdot` |
|
|
401
|
-
| `nk_dots_symmetric_u8_neonsdot` |
|
|
402
|
-
| `nk_dots_packed_u8_sme` |
|
|
403
|
-
| `nk_dots_symmetric_u8_sme` |
|
|
398
|
+
| `nk_dots_packed_u8_serial` | 16.3 gso/s | 16.3 gso/s | 17.4 gso/s |
|
|
399
|
+
| `nk_dots_symmetric_u8_serial` | 14.8 gso/s | 16.2 gso/s | 17.5 gso/s |
|
|
400
|
+
| `nk_dots_packed_u8_neonsdot` | 343 gso/s | 413 gso/s | 470 gso/s |
|
|
401
|
+
| `nk_dots_symmetric_u8_neonsdot` | 76.1 gso/s | 87.4 gso/s | 87.7 gso/s |
|
|
402
|
+
| `nk_dots_packed_u8_sme` | 2,351 gso/s | 2,684 gso/s | 2,570 gso/s |
|
|
403
|
+
| `nk_dots_symmetric_u8_sme` | 1,390 gso/s | 1,543 gso/s | 1,371 gso/s |
|
|
404
404
|
| __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
405
|
-
| `nk_dots_packed_i4_serial` |
|
|
406
|
-
| `nk_dots_symmetric_i4_serial` |
|
|
407
|
-
| `nk_dots_packed_i4_neonsdot` |
|
|
408
|
-
| `nk_dots_symmetric_i4_neonsdot` |
|
|
409
|
-
| `nk_dots_packed_i4_sme` |
|
|
410
|
-
| `nk_dots_symmetric_i4_sme` |
|
|
405
|
+
| `nk_dots_packed_i4_serial` | 18.3 gso/s | 18.2 gso/s | 19.6 gso/s |
|
|
406
|
+
| `nk_dots_symmetric_i4_serial` | 13.7 gso/s | 14.9 gso/s | 15.6 gso/s |
|
|
407
|
+
| `nk_dots_packed_i4_neonsdot` | 259 gso/s | 284 gso/s | 291 gso/s |
|
|
408
|
+
| `nk_dots_symmetric_i4_neonsdot` | 129 gso/s | 162 gso/s | 171 gso/s |
|
|
409
|
+
| `nk_dots_packed_i4_sme` | 2,269 gso/s | 2,455 gso/s | 2,396 gso/s |
|
|
410
|
+
| `nk_dots_symmetric_i4_sme` | 1,585 gso/s | 1,692 gso/s | 1,737 gso/s |
|
|
411
411
|
| __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
412
|
-
| `nk_dots_packed_u4_serial` | 19.
|
|
413
|
-
| `nk_dots_symmetric_u4_serial` |
|
|
414
|
-
| `nk_dots_packed_u4_neonsdot` |
|
|
415
|
-
| `nk_dots_symmetric_u4_neonsdot` |
|
|
416
|
-
| `nk_dots_packed_u4_sme` |
|
|
417
|
-
| `nk_dots_symmetric_u4_sme` |
|
|
412
|
+
| `nk_dots_packed_u4_serial` | 19.4 gso/s | 19.4 gso/s | 20.6 gso/s |
|
|
413
|
+
| `nk_dots_symmetric_u4_serial` | 14.9 gso/s | 16.4 gso/s | 17.4 gso/s |
|
|
414
|
+
| `nk_dots_packed_u4_neonsdot` | 300 gso/s | 319 gso/s | 340 gso/s |
|
|
415
|
+
| `nk_dots_symmetric_u4_neonsdot` | 128 gso/s | 166 gso/s | 173 gso/s |
|
|
416
|
+
| `nk_dots_packed_u4_sme` | 2,342 gso/s | 2,503 gso/s | 2,471 gso/s |
|
|
417
|
+
| `nk_dots_symmetric_u4_sme` | 1,695 gso/s | 1,925 gso/s | 2,055 gso/s |
|
|
418
418
|
| __u1__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
419
|
-
| `nk_dots_packed_u1_serial` |
|
|
420
|
-
| `nk_dots_symmetric_u1_serial` |
|
|
421
|
-
| `nk_dots_packed_u1_neon` |
|
|
422
|
-
| `nk_dots_symmetric_u1_neon` |
|
|
423
|
-
| `nk_dots_packed_u1_smebi32` |
|
|
424
|
-
| `nk_dots_symmetric_u1_smebi32` |
|
|
419
|
+
| `nk_dots_packed_u1_serial` | 405 gso/s | 467 gso/s | 534 gso/s |
|
|
420
|
+
| `nk_dots_symmetric_u1_serial` | 254 gso/s | 430 gso/s | 519 gso/s |
|
|
421
|
+
| `nk_dots_packed_u1_neon` | 849 gso/s | 932 gso/s | 1,014 gso/s |
|
|
422
|
+
| `nk_dots_symmetric_u1_neon` | 318 gso/s | 580 gso/s | 664 gso/s |
|
|
423
|
+
| `nk_dots_packed_u1_smebi32` | 1,903 gso/s | 12,029 gso/s | 26,354 gso/s |
|
|
424
|
+
| `nk_dots_symmetric_u1_smebi32` | 176 gso/s | 768 gso/s | 2,153 gso/s |
|
|
425
425
|
|
|
426
426
|
#### WASM
|
|
427
427
|
|
|
428
|
-
Measured with Wasmtime
|
|
428
|
+
Measured with Wasmtime v43 (Cranelift backend).
|
|
429
429
|
|
|
430
430
|
| Kernel | 256³ | 1024³ | 4096³ |
|
|
431
431
|
| :----------------------------------- | -----------------------: | -----------------------: | -----------------------: |
|
|
432
432
|
| __f64__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
433
|
-
| `nk_dots_packed_f64_serial` | 2.
|
|
434
|
-
| `nk_dots_symmetric_f64_serial` |
|
|
435
|
-
| `nk_dots_packed_f64_v128relaxed` |
|
|
436
|
-
| `nk_dots_symmetric_f64_v128relaxed` |
|
|
433
|
+
| `nk_dots_packed_f64_serial` | 2.15 gso/s, 3 ulp | 2.07 gso/s, 5 ulp | 2.23 gso/s, 2.2 ulp |
|
|
434
|
+
| `nk_dots_symmetric_f64_serial` | 2.35 gso/s, 4 ulp | 2.24 gso/s, 3 ulp | 2.46 gso/s, 2.4 ulp |
|
|
435
|
+
| `nk_dots_packed_f64_v128relaxed` | 5.59 gso/s, 32.4 ulp | 6.10 gso/s, 32.4 ulp | 6.24 gso/s, 32.4 ulp |
|
|
436
|
+
| `nk_dots_symmetric_f64_v128relaxed` | 5.26 gso/s, 37.6 ulp | 5.89 gso/s, 37.6 ulp | 6.04 gso/s, 37.6 ulp |
|
|
437
437
|
| __f32__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
438
|
-
| `nk_dots_packed_f32_serial` |
|
|
439
|
-
| `nk_dots_symmetric_f32_serial` |
|
|
440
|
-
| `nk_dots_packed_f32_v128relaxed` |
|
|
441
|
-
| `nk_dots_symmetric_f32_v128relaxed` |
|
|
438
|
+
| `nk_dots_packed_f32_serial` | 8.95 gso/s, 19 ulp | 8.71 gso/s, 30 ulp | 9.17 gso/s, 41.7 ulp |
|
|
439
|
+
| `nk_dots_symmetric_f32_serial` | 10.9 gso/s, 20 ulp | 10.5 gso/s, 29 ulp | 11.6 gso/s, 58.8 ulp |
|
|
440
|
+
| `nk_dots_packed_f32_v128relaxed` | 27.4 gso/s, 44.1 ulp | 31.6 gso/s, 44.1 ulp | 32.7 gso/s, 44.1 ulp |
|
|
441
|
+
| `nk_dots_symmetric_f32_v128relaxed` | 10.0 gso/s, 48.2 ulp | 10.9 gso/s, 48.2 ulp | 11.2 gso/s, 48.2 ulp |
|
|
442
442
|
| __bf16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
443
|
-
| `nk_dots_packed_bf16_serial` |
|
|
444
|
-
| `nk_dots_symmetric_bf16_serial` |
|
|
445
|
-
| `nk_dots_packed_bf16_v128relaxed` |
|
|
446
|
-
| `nk_dots_symmetric_bf16_v128relaxed` |
|
|
443
|
+
| `nk_dots_packed_bf16_serial` | 23.1 gso/s, 0.1 ulp | 21.6 gso/s, 0.5 ulp | 24.3 gso/s, 1.3 ulp |
|
|
444
|
+
| `nk_dots_symmetric_bf16_serial` | 24.3 gso/s, 0 ulp | 24.9 gso/s, 0.6 ulp | 28.0 gso/s, 1.1 ulp |
|
|
445
|
+
| `nk_dots_packed_bf16_v128relaxed` | 70.4 gso/s, 1.4 ulp | 86.2 gso/s, 1.4 ulp | 90.3 gso/s, 1.4 ulp |
|
|
446
|
+
| `nk_dots_symmetric_bf16_v128relaxed` | 37.2 gso/s, 1.3 ulp | 45.5 gso/s, 1.3 ulp | 47.7 gso/s, 1.3 ulp |
|
|
447
447
|
| __f16__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
448
|
-
| `nk_dots_packed_f16_serial` |
|
|
449
|
-
| `nk_dots_symmetric_f16_serial` |
|
|
450
|
-
| `nk_dots_packed_f16_v128relaxed` |
|
|
451
|
-
| `nk_dots_symmetric_f16_v128relaxed` |
|
|
448
|
+
| `nk_dots_packed_f16_serial` | 12.2 gso/s, 204 ulp | 11.6 gso/s, 36 ulp | 12.4 gso/s, 25.9 ulp |
|
|
449
|
+
| `nk_dots_symmetric_f16_serial` | 1.65 gso/s, 13 ulp | 1.54 gso/s, 29 ulp | 1.70 gso/s, 27.9 ulp |
|
|
450
|
+
| `nk_dots_packed_f16_v128relaxed` | 35.4 gso/s, ? ulp | 40.7 gso/s, ? ulp | 39.3 gso/s, ? ulp |
|
|
451
|
+
| `nk_dots_symmetric_f16_v128relaxed` | 14.7 gso/s, ? ulp | 17.1 gso/s, ? ulp | 17.3 gso/s, ? ulp |
|
|
452
452
|
| __e5m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
453
|
-
| `nk_dots_packed_e5m2_serial` |
|
|
454
|
-
| `nk_dots_symmetric_e5m2_serial` |
|
|
455
|
-
| `nk_dots_packed_e5m2_v128relaxed` |
|
|
456
|
-
| `nk_dots_symmetric_e5m2_v128relaxed` |
|
|
453
|
+
| `nk_dots_packed_e5m2_serial` | 5.95 gso/s, 0 ulp | 5.59 gso/s, 0 ulp | 6.31 gso/s, 0 ulp |
|
|
454
|
+
| `nk_dots_symmetric_e5m2_serial` | 8.98 gso/s, 0 ulp | 9.09 gso/s, 0 ulp | 10.2 gso/s, 0 ulp |
|
|
455
|
+
| `nk_dots_packed_e5m2_v128relaxed` | 23.0 gso/s, 0 ulp | 25.5 gso/s, 0 ulp | 25.9 gso/s, 0 ulp |
|
|
456
|
+
| `nk_dots_symmetric_e5m2_v128relaxed` | 12.3 gso/s, 0 ulp | 13.8 gso/s, 0 ulp | 14.2 gso/s, 0 ulp |
|
|
457
457
|
| __e4m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
458
|
-
| `nk_dots_packed_e4m3_serial` | 0.
|
|
459
|
-
| `nk_dots_symmetric_e4m3_serial` | 0.
|
|
460
|
-
| `nk_dots_packed_e4m3_v128relaxed` |
|
|
461
|
-
| `nk_dots_symmetric_e4m3_v128relaxed` |
|
|
458
|
+
| `nk_dots_packed_e4m3_serial` | 0.884 gso/s, 0 ulp | 0.840 gso/s, 0 ulp | 0.911 gso/s, 0 ulp |
|
|
459
|
+
| `nk_dots_symmetric_e4m3_serial` | 0.868 gso/s, 0 ulp | 0.826 gso/s, 0 ulp | 0.915 gso/s, 0 ulp |
|
|
460
|
+
| `nk_dots_packed_e4m3_v128relaxed` | 19.2 gso/s, 0 ulp | 20.8 gso/s, 0 ulp | 22.5 gso/s, 0 ulp |
|
|
461
|
+
| `nk_dots_symmetric_e4m3_v128relaxed` | 10.7 gso/s, 0 ulp | 11.7 gso/s, 0 ulp | 12.1 gso/s, 0 ulp |
|
|
462
462
|
| __e3m2__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
463
|
-
| `nk_dots_packed_e3m2_serial` |
|
|
464
|
-
| `nk_dots_symmetric_e3m2_serial` |
|
|
465
|
-
| `nk_dots_packed_e3m2_v128relaxed` |
|
|
466
|
-
| `nk_dots_symmetric_e3m2_v128relaxed` |
|
|
463
|
+
| `nk_dots_packed_e3m2_serial` | 5.89 gso/s, 0 ulp | 5.73 gso/s, 0 ulp | 6.25 gso/s, 0 ulp |
|
|
464
|
+
| `nk_dots_symmetric_e3m2_serial` | 7.69 gso/s, 0 ulp | 7.45 gso/s, 0 ulp | 8.68 gso/s, 0 ulp |
|
|
465
|
+
| `nk_dots_packed_e3m2_v128relaxed` | 35.2 gso/s, 0 ulp | 38.9 gso/s, 0 ulp | 40.1 gso/s, 0 ulp |
|
|
466
|
+
| `nk_dots_symmetric_e3m2_v128relaxed` | 32.0 gso/s, 0 ulp | 38.1 gso/s, 0 ulp | 39.7 gso/s, 0 ulp |
|
|
467
467
|
| __e2m3__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
468
|
-
| `nk_dots_packed_e2m3_serial` |
|
|
469
|
-
| `nk_dots_symmetric_e2m3_serial` |
|
|
470
|
-
| `nk_dots_packed_e2m3_v128relaxed` |
|
|
471
|
-
| `nk_dots_symmetric_e2m3_v128relaxed` |
|
|
468
|
+
| `nk_dots_packed_e2m3_serial` | 5.97 gso/s, 0 ulp | 5.69 gso/s, 0 ulp | 6.32 gso/s, 0 ulp |
|
|
469
|
+
| `nk_dots_symmetric_e2m3_serial` | 7.65 gso/s, 0 ulp | 7.71 gso/s, 0 ulp | 8.66 gso/s, 0 ulp |
|
|
470
|
+
| `nk_dots_packed_e2m3_v128relaxed` | 35.4 gso/s, 0 ulp | 39.0 gso/s, 0 ulp | 40.1 gso/s, 0 ulp |
|
|
471
|
+
| `nk_dots_symmetric_e2m3_v128relaxed` | 31.6 gso/s, 0 ulp | 37.6 gso/s, 0 ulp | 39.7 gso/s, 0 ulp |
|
|
472
472
|
| __i8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
473
|
-
| `nk_dots_packed_i8_serial` |
|
|
474
|
-
| `nk_dots_symmetric_i8_serial` |
|
|
475
|
-
| `nk_dots_packed_i8_v128relaxed` |
|
|
476
|
-
| `nk_dots_symmetric_i8_v128relaxed` |
|
|
473
|
+
| `nk_dots_packed_i8_serial` | 16.5 gso/s | 16.0 gso/s | 16.7 gso/s |
|
|
474
|
+
| `nk_dots_symmetric_i8_serial` | 12.5 gso/s | 11.8 gso/s | 13.6 gso/s |
|
|
475
|
+
| `nk_dots_packed_i8_v128relaxed` | 44.0 gso/s | 50.0 gso/s | 52.1 gso/s |
|
|
476
|
+
| `nk_dots_symmetric_i8_v128relaxed` | 37.7 gso/s | 45.5 gso/s | 50.6 gso/s |
|
|
477
477
|
| __u8__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
478
|
-
| `nk_dots_packed_u8_serial` |
|
|
479
|
-
| `nk_dots_symmetric_u8_serial` |
|
|
480
|
-
| `nk_dots_packed_u8_v128relaxed` |
|
|
481
|
-
| `nk_dots_symmetric_u8_v128relaxed` |
|
|
478
|
+
| `nk_dots_packed_u8_serial` | 17.2 gso/s | 16.7 gso/s | 17.7 gso/s |
|
|
479
|
+
| `nk_dots_symmetric_u8_serial` | 13.0 gso/s | 12.1 gso/s | 14.1 gso/s |
|
|
480
|
+
| `nk_dots_packed_u8_v128relaxed` | 43.3 gso/s | 47.7 gso/s | 50.8 gso/s |
|
|
481
|
+
| `nk_dots_symmetric_u8_v128relaxed` | 34.6 gso/s | 42.2 gso/s | 48.6 gso/s |
|
|
482
482
|
| __i4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
483
|
-
| `nk_dots_packed_i4_serial` |
|
|
484
|
-
| `nk_dots_symmetric_i4_serial` |
|
|
485
|
-
| `nk_dots_packed_i4_v128relaxed` |
|
|
486
|
-
| `nk_dots_symmetric_i4_v128relaxed` | 54.
|
|
483
|
+
| `nk_dots_packed_i4_serial` | 15.0 gso/s | 14.3 gso/s | 15.9 gso/s |
|
|
484
|
+
| `nk_dots_symmetric_i4_serial` | 12.8 gso/s | 12.6 gso/s | 14.0 gso/s |
|
|
485
|
+
| `nk_dots_packed_i4_v128relaxed` | 29.3 gso/s | 26.7 gso/s | 25.8 gso/s |
|
|
486
|
+
| `nk_dots_symmetric_i4_v128relaxed` | 54.0 gso/s | 70.9 gso/s | 80.8 gso/s |
|
|
487
487
|
| __u4__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
488
|
-
| `nk_dots_packed_u4_serial` |
|
|
489
|
-
| `nk_dots_symmetric_u4_serial` |
|
|
490
|
-
| `nk_dots_packed_u4_v128relaxed` |
|
|
491
|
-
| `nk_dots_symmetric_u4_v128relaxed` |
|
|
488
|
+
| `nk_dots_packed_u4_serial` | 14.6 gso/s | 14.1 gso/s | 15.4 gso/s |
|
|
489
|
+
| `nk_dots_symmetric_u4_serial` | 11.9 gso/s | 11.8 gso/s | 13.0 gso/s |
|
|
490
|
+
| `nk_dots_packed_u4_v128relaxed` | 84.9 gso/s | 92.5 gso/s | 96.2 gso/s |
|
|
491
|
+
| `nk_dots_symmetric_u4_v128relaxed` | 67.4 gso/s | 87.7 gso/s | 93.7 gso/s |
|
|
492
492
|
| __u1__ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ | ░░░░░░░░░░░░░░░░░░░░░░░░ |
|
|
493
|
-
| `nk_dots_packed_u1_serial` |
|
|
494
|
-
| `nk_dots_symmetric_u1_serial` |
|
|
495
|
-
| `nk_dots_packed_u1_v128relaxed` |
|
|
496
|
-
| `nk_dots_symmetric_u1_v128relaxed` |
|
|
493
|
+
| `nk_dots_packed_u1_serial` | 236 gso/s | 265 gso/s | 311 gso/s |
|
|
494
|
+
| `nk_dots_symmetric_u1_serial` | 173 gso/s | 321 gso/s | 443 gso/s |
|
|
495
|
+
| `nk_dots_packed_u1_v128relaxed` | 598 gso/s | 804 gso/s | 871 gso/s |
|
|
496
|
+
| `nk_dots_symmetric_u1_v128relaxed` | 183 gso/s | 390 gso/s | 543 gso/s |
|