numkong 7.0.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/LICENSE +201 -0
- package/README.md +495 -0
- package/binding.gyp +540 -0
- package/c/dispatch.h +512 -0
- package/c/dispatch_bf16.c +389 -0
- package/c/dispatch_bf16c.c +52 -0
- package/c/dispatch_e2m3.c +263 -0
- package/c/dispatch_e3m2.c +243 -0
- package/c/dispatch_e4m3.c +276 -0
- package/c/dispatch_e5m2.c +272 -0
- package/c/dispatch_f16.c +376 -0
- package/c/dispatch_f16c.c +58 -0
- package/c/dispatch_f32.c +378 -0
- package/c/dispatch_f32c.c +99 -0
- package/c/dispatch_f64.c +296 -0
- package/c/dispatch_f64c.c +98 -0
- package/c/dispatch_i16.c +96 -0
- package/c/dispatch_i32.c +89 -0
- package/c/dispatch_i4.c +150 -0
- package/c/dispatch_i64.c +86 -0
- package/c/dispatch_i8.c +289 -0
- package/c/dispatch_other.c +330 -0
- package/c/dispatch_u1.c +148 -0
- package/c/dispatch_u16.c +124 -0
- package/c/dispatch_u32.c +118 -0
- package/c/dispatch_u4.c +150 -0
- package/c/dispatch_u64.c +102 -0
- package/c/dispatch_u8.c +303 -0
- package/c/numkong.c +950 -0
- package/include/README.md +573 -0
- package/include/module.modulemap +129 -0
- package/include/numkong/attention/sapphireamx.h +1361 -0
- package/include/numkong/attention/sme.h +2066 -0
- package/include/numkong/attention.h +49 -0
- package/include/numkong/capabilities.h +748 -0
- package/include/numkong/cast/README.md +262 -0
- package/include/numkong/cast/haswell.h +975 -0
- package/include/numkong/cast/icelake.h +470 -0
- package/include/numkong/cast/neon.h +1192 -0
- package/include/numkong/cast/rvv.h +1021 -0
- package/include/numkong/cast/sapphire.h +262 -0
- package/include/numkong/cast/serial.h +2262 -0
- package/include/numkong/cast/skylake.h +856 -0
- package/include/numkong/cast/v128relaxed.h +180 -0
- package/include/numkong/cast.h +230 -0
- package/include/numkong/curved/README.md +223 -0
- package/include/numkong/curved/genoa.h +182 -0
- package/include/numkong/curved/haswell.h +276 -0
- package/include/numkong/curved/neon.h +205 -0
- package/include/numkong/curved/neonbfdot.h +212 -0
- package/include/numkong/curved/neonhalf.h +212 -0
- package/include/numkong/curved/rvv.h +305 -0
- package/include/numkong/curved/serial.h +207 -0
- package/include/numkong/curved/skylake.h +457 -0
- package/include/numkong/curved/smef64.h +506 -0
- package/include/numkong/curved.h +517 -0
- package/include/numkong/curved.hpp +144 -0
- package/include/numkong/dot/README.md +425 -0
- package/include/numkong/dot/alder.h +563 -0
- package/include/numkong/dot/genoa.h +315 -0
- package/include/numkong/dot/haswell.h +1688 -0
- package/include/numkong/dot/icelake.h +883 -0
- package/include/numkong/dot/neon.h +818 -0
- package/include/numkong/dot/neonbfdot.h +244 -0
- package/include/numkong/dot/neonfhm.h +360 -0
- package/include/numkong/dot/neonhalf.h +198 -0
- package/include/numkong/dot/neonsdot.h +508 -0
- package/include/numkong/dot/rvv.h +714 -0
- package/include/numkong/dot/rvvbb.h +72 -0
- package/include/numkong/dot/rvvbf16.h +123 -0
- package/include/numkong/dot/rvvhalf.h +129 -0
- package/include/numkong/dot/sapphire.h +141 -0
- package/include/numkong/dot/serial.h +838 -0
- package/include/numkong/dot/sierra.h +405 -0
- package/include/numkong/dot/skylake.h +1084 -0
- package/include/numkong/dot/sve.h +379 -0
- package/include/numkong/dot/svebfdot.h +74 -0
- package/include/numkong/dot/svehalf.h +123 -0
- package/include/numkong/dot/v128relaxed.h +1258 -0
- package/include/numkong/dot.h +1070 -0
- package/include/numkong/dot.hpp +94 -0
- package/include/numkong/dots/README.md +496 -0
- package/include/numkong/dots/alder.h +114 -0
- package/include/numkong/dots/genoa.h +94 -0
- package/include/numkong/dots/haswell.h +295 -0
- package/include/numkong/dots/icelake.h +171 -0
- package/include/numkong/dots/neon.h +120 -0
- package/include/numkong/dots/neonbfdot.h +58 -0
- package/include/numkong/dots/neonfhm.h +94 -0
- package/include/numkong/dots/neonhalf.h +57 -0
- package/include/numkong/dots/neonsdot.h +108 -0
- package/include/numkong/dots/rvv.h +2486 -0
- package/include/numkong/dots/sapphireamx.h +3973 -0
- package/include/numkong/dots/serial.h +2844 -0
- package/include/numkong/dots/sierra.h +97 -0
- package/include/numkong/dots/skylake.h +196 -0
- package/include/numkong/dots/sme.h +5372 -0
- package/include/numkong/dots/smebi32.h +461 -0
- package/include/numkong/dots/smef64.h +1318 -0
- package/include/numkong/dots/smehalf.h +47 -0
- package/include/numkong/dots/v128relaxed.h +294 -0
- package/include/numkong/dots.h +2804 -0
- package/include/numkong/dots.hpp +639 -0
- package/include/numkong/each/README.md +469 -0
- package/include/numkong/each/haswell.h +1658 -0
- package/include/numkong/each/icelake.h +272 -0
- package/include/numkong/each/neon.h +1104 -0
- package/include/numkong/each/neonbfdot.h +212 -0
- package/include/numkong/each/neonhalf.h +410 -0
- package/include/numkong/each/rvv.h +1121 -0
- package/include/numkong/each/sapphire.h +477 -0
- package/include/numkong/each/serial.h +260 -0
- package/include/numkong/each/skylake.h +1562 -0
- package/include/numkong/each.h +2146 -0
- package/include/numkong/each.hpp +434 -0
- package/include/numkong/geospatial/README.md +147 -0
- package/include/numkong/geospatial/haswell.h +593 -0
- package/include/numkong/geospatial/neon.h +571 -0
- package/include/numkong/geospatial/rvv.h +701 -0
- package/include/numkong/geospatial/serial.h +309 -0
- package/include/numkong/geospatial/skylake.h +577 -0
- package/include/numkong/geospatial/v128relaxed.h +613 -0
- package/include/numkong/geospatial.h +453 -0
- package/include/numkong/geospatial.hpp +235 -0
- package/include/numkong/matrix.hpp +336 -0
- package/include/numkong/maxsim/README.md +187 -0
- package/include/numkong/maxsim/alder.h +511 -0
- package/include/numkong/maxsim/genoa.h +115 -0
- package/include/numkong/maxsim/haswell.h +553 -0
- package/include/numkong/maxsim/icelake.h +480 -0
- package/include/numkong/maxsim/neonsdot.h +394 -0
- package/include/numkong/maxsim/sapphireamx.h +877 -0
- package/include/numkong/maxsim/serial.h +490 -0
- package/include/numkong/maxsim/sme.h +929 -0
- package/include/numkong/maxsim/v128relaxed.h +280 -0
- package/include/numkong/maxsim.h +571 -0
- package/include/numkong/maxsim.hpp +133 -0
- package/include/numkong/mesh/README.md +227 -0
- package/include/numkong/mesh/haswell.h +2235 -0
- package/include/numkong/mesh/neon.h +1329 -0
- package/include/numkong/mesh/neonbfdot.h +842 -0
- package/include/numkong/mesh/neonhalf.h +616 -0
- package/include/numkong/mesh/rvv.h +916 -0
- package/include/numkong/mesh/serial.h +742 -0
- package/include/numkong/mesh/skylake.h +1135 -0
- package/include/numkong/mesh/v128relaxed.h +1052 -0
- package/include/numkong/mesh.h +652 -0
- package/include/numkong/mesh.hpp +762 -0
- package/include/numkong/numkong.h +78 -0
- package/include/numkong/numkong.hpp +57 -0
- package/include/numkong/probability/README.md +173 -0
- package/include/numkong/probability/haswell.h +267 -0
- package/include/numkong/probability/neon.h +225 -0
- package/include/numkong/probability/rvv.h +409 -0
- package/include/numkong/probability/serial.h +169 -0
- package/include/numkong/probability/skylake.h +324 -0
- package/include/numkong/probability.h +383 -0
- package/include/numkong/probability.hpp +120 -0
- package/include/numkong/random.h +50 -0
- package/include/numkong/random.hpp +285 -0
- package/include/numkong/reduce/README.md +547 -0
- package/include/numkong/reduce/alder.h +632 -0
- package/include/numkong/reduce/genoa.h +201 -0
- package/include/numkong/reduce/haswell.h +3783 -0
- package/include/numkong/reduce/icelake.h +549 -0
- package/include/numkong/reduce/neon.h +3841 -0
- package/include/numkong/reduce/neonbfdot.h +353 -0
- package/include/numkong/reduce/neonfhm.h +665 -0
- package/include/numkong/reduce/neonhalf.h +157 -0
- package/include/numkong/reduce/neonsdot.h +357 -0
- package/include/numkong/reduce/rvv.h +3407 -0
- package/include/numkong/reduce/serial.h +757 -0
- package/include/numkong/reduce/sierra.h +338 -0
- package/include/numkong/reduce/skylake.h +3792 -0
- package/include/numkong/reduce/v128relaxed.h +2302 -0
- package/include/numkong/reduce.h +1597 -0
- package/include/numkong/reduce.hpp +633 -0
- package/include/numkong/scalar/README.md +89 -0
- package/include/numkong/scalar/haswell.h +113 -0
- package/include/numkong/scalar/neon.h +122 -0
- package/include/numkong/scalar/neonhalf.h +70 -0
- package/include/numkong/scalar/rvv.h +211 -0
- package/include/numkong/scalar/sapphire.h +63 -0
- package/include/numkong/scalar/serial.h +332 -0
- package/include/numkong/scalar/v128relaxed.h +56 -0
- package/include/numkong/scalar.h +683 -0
- package/include/numkong/set/README.md +179 -0
- package/include/numkong/set/haswell.h +334 -0
- package/include/numkong/set/icelake.h +485 -0
- package/include/numkong/set/neon.h +364 -0
- package/include/numkong/set/rvv.h +226 -0
- package/include/numkong/set/rvvbb.h +117 -0
- package/include/numkong/set/serial.h +174 -0
- package/include/numkong/set/sve.h +185 -0
- package/include/numkong/set/v128relaxed.h +240 -0
- package/include/numkong/set.h +457 -0
- package/include/numkong/set.hpp +114 -0
- package/include/numkong/sets/README.md +149 -0
- package/include/numkong/sets/haswell.h +63 -0
- package/include/numkong/sets/icelake.h +66 -0
- package/include/numkong/sets/neon.h +61 -0
- package/include/numkong/sets/serial.h +43 -0
- package/include/numkong/sets/smebi32.h +1099 -0
- package/include/numkong/sets/v128relaxed.h +58 -0
- package/include/numkong/sets.h +339 -0
- package/include/numkong/sparse/README.md +156 -0
- package/include/numkong/sparse/icelake.h +463 -0
- package/include/numkong/sparse/neon.h +288 -0
- package/include/numkong/sparse/serial.h +117 -0
- package/include/numkong/sparse/sve2.h +507 -0
- package/include/numkong/sparse/turin.h +322 -0
- package/include/numkong/sparse.h +363 -0
- package/include/numkong/sparse.hpp +113 -0
- package/include/numkong/spatial/README.md +435 -0
- package/include/numkong/spatial/alder.h +607 -0
- package/include/numkong/spatial/genoa.h +290 -0
- package/include/numkong/spatial/haswell.h +960 -0
- package/include/numkong/spatial/icelake.h +586 -0
- package/include/numkong/spatial/neon.h +773 -0
- package/include/numkong/spatial/neonbfdot.h +165 -0
- package/include/numkong/spatial/neonhalf.h +118 -0
- package/include/numkong/spatial/neonsdot.h +261 -0
- package/include/numkong/spatial/rvv.h +984 -0
- package/include/numkong/spatial/rvvbf16.h +123 -0
- package/include/numkong/spatial/rvvhalf.h +117 -0
- package/include/numkong/spatial/sapphire.h +343 -0
- package/include/numkong/spatial/serial.h +346 -0
- package/include/numkong/spatial/sierra.h +323 -0
- package/include/numkong/spatial/skylake.h +606 -0
- package/include/numkong/spatial/sve.h +224 -0
- package/include/numkong/spatial/svebfdot.h +122 -0
- package/include/numkong/spatial/svehalf.h +109 -0
- package/include/numkong/spatial/v128relaxed.h +717 -0
- package/include/numkong/spatial.h +1425 -0
- package/include/numkong/spatial.hpp +183 -0
- package/include/numkong/spatials/README.md +580 -0
- package/include/numkong/spatials/alder.h +94 -0
- package/include/numkong/spatials/genoa.h +94 -0
- package/include/numkong/spatials/haswell.h +219 -0
- package/include/numkong/spatials/icelake.h +113 -0
- package/include/numkong/spatials/neon.h +109 -0
- package/include/numkong/spatials/neonbfdot.h +60 -0
- package/include/numkong/spatials/neonfhm.h +92 -0
- package/include/numkong/spatials/neonhalf.h +58 -0
- package/include/numkong/spatials/neonsdot.h +109 -0
- package/include/numkong/spatials/rvv.h +1960 -0
- package/include/numkong/spatials/sapphireamx.h +1149 -0
- package/include/numkong/spatials/serial.h +226 -0
- package/include/numkong/spatials/sierra.h +96 -0
- package/include/numkong/spatials/skylake.h +184 -0
- package/include/numkong/spatials/sme.h +1901 -0
- package/include/numkong/spatials/smef64.h +465 -0
- package/include/numkong/spatials/v128relaxed.h +240 -0
- package/include/numkong/spatials.h +3021 -0
- package/include/numkong/spatials.hpp +508 -0
- package/include/numkong/tensor.hpp +1592 -0
- package/include/numkong/trigonometry/README.md +184 -0
- package/include/numkong/trigonometry/haswell.h +652 -0
- package/include/numkong/trigonometry/neon.h +639 -0
- package/include/numkong/trigonometry/rvv.h +699 -0
- package/include/numkong/trigonometry/serial.h +703 -0
- package/include/numkong/trigonometry/skylake.h +721 -0
- package/include/numkong/trigonometry/v128relaxed.h +666 -0
- package/include/numkong/trigonometry.h +467 -0
- package/include/numkong/trigonometry.hpp +166 -0
- package/include/numkong/types.h +1384 -0
- package/include/numkong/types.hpp +5603 -0
- package/include/numkong/vector.hpp +698 -0
- package/javascript/README.md +246 -0
- package/javascript/dist/cjs/numkong-wasm.d.ts +166 -0
- package/javascript/dist/cjs/numkong-wasm.js +617 -0
- package/javascript/dist/cjs/numkong.d.ts +343 -0
- package/javascript/dist/cjs/numkong.js +523 -0
- package/javascript/dist/cjs/package.json +3 -0
- package/javascript/dist/cjs/types.d.ts +284 -0
- package/javascript/dist/cjs/types.js +653 -0
- package/javascript/dist/esm/numkong-wasm.d.ts +166 -0
- package/javascript/dist/esm/numkong-wasm.js +595 -0
- package/javascript/dist/esm/numkong.d.ts +343 -0
- package/javascript/dist/esm/numkong.js +452 -0
- package/javascript/dist/esm/package.json +3 -0
- package/javascript/dist/esm/types.d.ts +284 -0
- package/javascript/dist/esm/types.js +630 -0
- package/javascript/dist-package-cjs.json +3 -0
- package/javascript/dist-package-esm.json +3 -0
- package/javascript/node-gyp-build.d.ts +1 -0
- package/javascript/numkong-wasm.ts +756 -0
- package/javascript/numkong.c +689 -0
- package/javascript/numkong.ts +575 -0
- package/javascript/tsconfig-base.json +39 -0
- package/javascript/tsconfig-cjs.json +8 -0
- package/javascript/tsconfig-esm.json +8 -0
- package/javascript/types.ts +674 -0
- package/package.json +87 -0
|
@@ -0,0 +1,263 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* @brief Dispatch Initialization for E2M3 Data Types.
|
|
3
|
+
* @file c/dispatch_e2m3.c
|
|
4
|
+
* @author Ash Vardanian
|
|
5
|
+
* @date February 3, 2026
|
|
6
|
+
*/
|
|
7
|
+
#include "dispatch.h"
|
|
8
|
+
|
|
9
|
+
#ifdef __cplusplus
|
|
10
|
+
extern "C" {
|
|
11
|
+
#endif
|
|
12
|
+
|
|
13
|
+
void nk_dispatch_e2m3_find_(nk_capability_t v, nk_kernel_kind_t k, nk_kernel_punned_t *m, nk_capability_t *c) {
|
|
14
|
+
typedef nk_kernel_punned_t m_t;
|
|
15
|
+
#if NK_TARGET_V128RELAXED
|
|
16
|
+
if (v & nk_cap_v128relaxed_k) switch (k) {
|
|
17
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k; return;
|
|
18
|
+
case nk_kernel_reduce_moments_k:
|
|
19
|
+
*m = (m_t)&nk_reduce_moments_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
20
|
+
return;
|
|
21
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k; return;
|
|
22
|
+
case nk_kernel_dots_packed_size_k:
|
|
23
|
+
*m = (m_t)&nk_dots_packed_size_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
24
|
+
return;
|
|
25
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k; return;
|
|
26
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k; return;
|
|
27
|
+
case nk_kernel_dots_symmetric_k:
|
|
28
|
+
*m = (m_t)&nk_dots_symmetric_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
29
|
+
return;
|
|
30
|
+
case nk_kernel_angulars_packed_k:
|
|
31
|
+
*m = (m_t)&nk_angulars_packed_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
32
|
+
return;
|
|
33
|
+
case nk_kernel_angulars_symmetric_k:
|
|
34
|
+
*m = (m_t)&nk_angulars_symmetric_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
35
|
+
return;
|
|
36
|
+
case nk_kernel_euclideans_packed_k:
|
|
37
|
+
*m = (m_t)&nk_euclideans_packed_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
38
|
+
return;
|
|
39
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
40
|
+
*m = (m_t)&nk_euclideans_symmetric_e2m3_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
41
|
+
return;
|
|
42
|
+
default: break;
|
|
43
|
+
}
|
|
44
|
+
#endif
|
|
45
|
+
#if NK_TARGET_SME
|
|
46
|
+
if (v & nk_cap_sme_k) switch (k) {
|
|
47
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e2m3_sme, *c = nk_cap_sme_k; return;
|
|
48
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_sme, *c = nk_cap_sme_k; return;
|
|
49
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_sme, *c = nk_cap_sme_k; return;
|
|
50
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e2m3_sme, *c = nk_cap_sme_k; return;
|
|
51
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e2m3_sme, *c = nk_cap_sme_k; return;
|
|
52
|
+
case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_e2m3_sme, *c = nk_cap_sme_k; return;
|
|
53
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e2m3_sme, *c = nk_cap_sme_k; return;
|
|
54
|
+
case nk_kernel_euclideans_symmetric_k: *m = (m_t)&nk_euclideans_symmetric_e2m3_sme, *c = nk_cap_sme_k; return;
|
|
55
|
+
default: break;
|
|
56
|
+
}
|
|
57
|
+
#endif
|
|
58
|
+
#if NK_TARGET_NEONSDOT
|
|
59
|
+
if (v & nk_cap_neonsdot_k) switch (k) {
|
|
60
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_neonsdot, *c = nk_cap_neonsdot_k; return;
|
|
61
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_neonsdot, *c = nk_cap_neonsdot_k; return;
|
|
62
|
+
default: break;
|
|
63
|
+
}
|
|
64
|
+
#endif
|
|
65
|
+
#if NK_TARGET_NEON
|
|
66
|
+
if (v & nk_cap_neon_k) switch (k) {
|
|
67
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_neon, *c = nk_cap_neon_k; return;
|
|
68
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e2m3_neon, *c = nk_cap_neon_k; return;
|
|
69
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e2m3_neon, *c = nk_cap_neon_k; return;
|
|
70
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e2m3_neon, *c = nk_cap_neon_k; return;
|
|
71
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_neon, *c = nk_cap_neon_k; return;
|
|
72
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e2m3_neon, *c = nk_cap_neon_k; return;
|
|
73
|
+
default: break;
|
|
74
|
+
}
|
|
75
|
+
#endif
|
|
76
|
+
#if NK_TARGET_SAPPHIREAMX
|
|
77
|
+
if (v & nk_cap_sapphireamx_k) switch (k) {
|
|
78
|
+
case nk_kernel_dots_packed_size_k:
|
|
79
|
+
*m = (m_t)&nk_dots_packed_size_e2m3_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
80
|
+
return;
|
|
81
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_sapphireamx, *c = nk_cap_sapphireamx_k; return;
|
|
82
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_sapphireamx, *c = nk_cap_sapphireamx_k; return;
|
|
83
|
+
case nk_kernel_dots_symmetric_k:
|
|
84
|
+
*m = (m_t)&nk_dots_symmetric_e2m3_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
85
|
+
return;
|
|
86
|
+
case nk_kernel_angulars_packed_k:
|
|
87
|
+
*m = (m_t)&nk_angulars_packed_e2m3_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
88
|
+
return;
|
|
89
|
+
case nk_kernel_angulars_symmetric_k:
|
|
90
|
+
*m = (m_t)&nk_angulars_symmetric_e2m3_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
91
|
+
return;
|
|
92
|
+
case nk_kernel_euclideans_packed_k:
|
|
93
|
+
*m = (m_t)&nk_euclideans_packed_e2m3_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
94
|
+
return;
|
|
95
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
96
|
+
*m = (m_t)&nk_euclideans_symmetric_e2m3_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
97
|
+
return;
|
|
98
|
+
default: break;
|
|
99
|
+
}
|
|
100
|
+
#endif
|
|
101
|
+
#if NK_TARGET_SAPPHIRE
|
|
102
|
+
if (v & nk_cap_sapphire_k) switch (k) {
|
|
103
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e2m3_sapphire, *c = nk_cap_sapphire_k; return;
|
|
104
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e2m3_sapphire, *c = nk_cap_sapphire_k; return;
|
|
105
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e2m3_sapphire, *c = nk_cap_sapphire_k; return;
|
|
106
|
+
default: break;
|
|
107
|
+
}
|
|
108
|
+
#endif
|
|
109
|
+
#if NK_TARGET_ICELAKE
|
|
110
|
+
if (v & nk_cap_icelake_k) switch (k) {
|
|
111
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_icelake, *c = nk_cap_icelake_k; return;
|
|
112
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_icelake, *c = nk_cap_icelake_k; return;
|
|
113
|
+
default: break;
|
|
114
|
+
}
|
|
115
|
+
#endif
|
|
116
|
+
#if NK_TARGET_SKYLAKE
|
|
117
|
+
if (v & nk_cap_skylake_k) switch (k) {
|
|
118
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
119
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
120
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
121
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
122
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
123
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
124
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
125
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
126
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
127
|
+
case nk_kernel_angulars_symmetric_k:
|
|
128
|
+
*m = (m_t)&nk_angulars_symmetric_e2m3_skylake, *c = nk_cap_skylake_k;
|
|
129
|
+
return;
|
|
130
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
131
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
132
|
+
*m = (m_t)&nk_euclideans_symmetric_e2m3_skylake, *c = nk_cap_skylake_k;
|
|
133
|
+
return;
|
|
134
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
135
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e2m3_skylake, *c = nk_cap_skylake_k; return;
|
|
136
|
+
default: break;
|
|
137
|
+
}
|
|
138
|
+
#endif
|
|
139
|
+
#if NK_TARGET_SIERRA
|
|
140
|
+
if (v & nk_cap_sierra_k) switch (k) {
|
|
141
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
142
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
143
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
144
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
145
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
146
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
147
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
148
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
149
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_sierra, *c = nk_cap_sierra_k; return;
|
|
150
|
+
default: break;
|
|
151
|
+
}
|
|
152
|
+
#endif
|
|
153
|
+
#if NK_TARGET_ALDER
|
|
154
|
+
if (v & nk_cap_alder_k) switch (k) {
|
|
155
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
156
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
157
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
158
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
159
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
160
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
161
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
162
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
163
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e2m3_alder, *c = nk_cap_alder_k; return;
|
|
164
|
+
default: break;
|
|
165
|
+
}
|
|
166
|
+
#endif
|
|
167
|
+
#if NK_TARGET_HASWELL
|
|
168
|
+
if (v & nk_cap_haswell_k) switch (k) {
|
|
169
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
170
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
171
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
172
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
173
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
174
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
175
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
176
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
177
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
178
|
+
case nk_kernel_angulars_symmetric_k:
|
|
179
|
+
*m = (m_t)&nk_angulars_symmetric_e2m3_haswell, *c = nk_cap_haswell_k;
|
|
180
|
+
return;
|
|
181
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
182
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
183
|
+
*m = (m_t)&nk_euclideans_symmetric_e2m3_haswell, *c = nk_cap_haswell_k;
|
|
184
|
+
return;
|
|
185
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
186
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e2m3_haswell, *c = nk_cap_haswell_k; return;
|
|
187
|
+
default: break;
|
|
188
|
+
}
|
|
189
|
+
#endif
|
|
190
|
+
#if NK_TARGET_RVV
|
|
191
|
+
if (v & nk_cap_rvv_k) switch (k) {
|
|
192
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
193
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
194
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
195
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
196
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
197
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
198
|
+
case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
199
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
200
|
+
case nk_kernel_euclideans_symmetric_k: *m = (m_t)&nk_euclideans_symmetric_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
201
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
202
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e2m3_rvv, *c = nk_cap_rvv_k; return;
|
|
203
|
+
default: break;
|
|
204
|
+
}
|
|
205
|
+
#endif
|
|
206
|
+
if (v & nk_cap_serial_k) switch (k) {
|
|
207
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
208
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
209
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
210
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
211
|
+
case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
212
|
+
case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
213
|
+
case nk_kernel_each_blend_k: *m = (m_t)&nk_each_blend_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
214
|
+
case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
215
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
216
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
217
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
218
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
219
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
220
|
+
case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
221
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
222
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
223
|
+
*m = (m_t)&nk_euclideans_symmetric_e2m3_serial, *c = nk_cap_serial_k;
|
|
224
|
+
return;
|
|
225
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
226
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e2m3_serial, *c = nk_cap_serial_k; return;
|
|
227
|
+
default: break;
|
|
228
|
+
}
|
|
229
|
+
|
|
230
|
+
// Error fallback - zero capability signals lookup failure
|
|
231
|
+
*m = (m_t)nk_error_dense_, *c = 0;
|
|
232
|
+
}
|
|
233
|
+
|
|
234
|
+
void nk_dispatch_e2m3_init_(nk_capability_t caps) {
|
|
235
|
+
nk_implementations_t *t = &nk_dispatch_table;
|
|
236
|
+
nk_capability_t used;
|
|
237
|
+
|
|
238
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_dot_k, (nk_kernel_punned_t *)&t->dot_e2m3, &used);
|
|
239
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_angular_k, (nk_kernel_punned_t *)&t->angular_e2m3, &used);
|
|
240
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_euclidean_k, (nk_kernel_punned_t *)&t->euclidean_e2m3, &used);
|
|
241
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_sqeuclidean_k, (nk_kernel_punned_t *)&t->sqeuclidean_e2m3, &used);
|
|
242
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_each_scale_k, (nk_kernel_punned_t *)&t->each_scale_e2m3, &used);
|
|
243
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_each_sum_k, (nk_kernel_punned_t *)&t->each_sum_e2m3, &used);
|
|
244
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_each_blend_k, (nk_kernel_punned_t *)&t->each_blend_e2m3, &used);
|
|
245
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_each_fma_k, (nk_kernel_punned_t *)&t->each_fma_e2m3, &used);
|
|
246
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_reduce_moments_k, (nk_kernel_punned_t *)&t->reduce_moments_e2m3, &used);
|
|
247
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_reduce_minmax_k, (nk_kernel_punned_t *)&t->reduce_minmax_e2m3, &used);
|
|
248
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_dots_packed_size_k, (nk_kernel_punned_t *)&t->dots_packed_size_e2m3, &used);
|
|
249
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_dots_pack_k, (nk_kernel_punned_t *)&t->dots_pack_e2m3, &used);
|
|
250
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_dots_packed_k, (nk_kernel_punned_t *)&t->dots_packed_e2m3, &used);
|
|
251
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_dots_symmetric_k, (nk_kernel_punned_t *)&t->dots_symmetric_e2m3, &used);
|
|
252
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_angulars_packed_k, (nk_kernel_punned_t *)&t->angulars_packed_e2m3, &used);
|
|
253
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_angulars_symmetric_k, (nk_kernel_punned_t *)&t->angulars_symmetric_e2m3,
|
|
254
|
+
&used);
|
|
255
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_euclideans_packed_k, (nk_kernel_punned_t *)&t->euclideans_packed_e2m3,
|
|
256
|
+
&used);
|
|
257
|
+
nk_dispatch_e2m3_find_(caps, nk_kernel_euclideans_symmetric_k, (nk_kernel_punned_t *)&t->euclideans_symmetric_e2m3,
|
|
258
|
+
&used);
|
|
259
|
+
}
|
|
260
|
+
|
|
261
|
+
#ifdef __cplusplus
|
|
262
|
+
}
|
|
263
|
+
#endif
|
|
@@ -0,0 +1,243 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* @brief Dispatch Initialization for E3M2 Data Types.
|
|
3
|
+
* @file c/dispatch_e3m2.c
|
|
4
|
+
* @author Ash Vardanian
|
|
5
|
+
* @date February 3, 2026
|
|
6
|
+
*/
|
|
7
|
+
#include "dispatch.h"
|
|
8
|
+
|
|
9
|
+
#ifdef __cplusplus
|
|
10
|
+
extern "C" {
|
|
11
|
+
#endif
|
|
12
|
+
|
|
13
|
+
void nk_dispatch_e3m2_find_(nk_capability_t v, nk_kernel_kind_t k, nk_kernel_punned_t *m, nk_capability_t *c) {
|
|
14
|
+
typedef nk_kernel_punned_t m_t;
|
|
15
|
+
#if NK_TARGET_V128RELAXED
|
|
16
|
+
if (v & nk_cap_v128relaxed_k) switch (k) {
|
|
17
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k; return;
|
|
18
|
+
case nk_kernel_reduce_moments_k:
|
|
19
|
+
*m = (m_t)&nk_reduce_moments_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
20
|
+
return;
|
|
21
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k; return;
|
|
22
|
+
case nk_kernel_dots_packed_size_k:
|
|
23
|
+
*m = (m_t)&nk_dots_packed_size_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
24
|
+
return;
|
|
25
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k; return;
|
|
26
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k; return;
|
|
27
|
+
case nk_kernel_dots_symmetric_k:
|
|
28
|
+
*m = (m_t)&nk_dots_symmetric_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
29
|
+
return;
|
|
30
|
+
case nk_kernel_angulars_packed_k:
|
|
31
|
+
*m = (m_t)&nk_angulars_packed_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
32
|
+
return;
|
|
33
|
+
case nk_kernel_angulars_symmetric_k:
|
|
34
|
+
*m = (m_t)&nk_angulars_symmetric_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
35
|
+
return;
|
|
36
|
+
case nk_kernel_euclideans_packed_k:
|
|
37
|
+
*m = (m_t)&nk_euclideans_packed_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
38
|
+
return;
|
|
39
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
40
|
+
*m = (m_t)&nk_euclideans_symmetric_e3m2_v128relaxed, *c = nk_cap_v128relaxed_k;
|
|
41
|
+
return;
|
|
42
|
+
default: break;
|
|
43
|
+
}
|
|
44
|
+
#endif
|
|
45
|
+
#if NK_TARGET_SME
|
|
46
|
+
if (v & nk_cap_sme_k) switch (k) {
|
|
47
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e3m2_sme, *c = nk_cap_sme_k; return;
|
|
48
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e3m2_sme, *c = nk_cap_sme_k; return;
|
|
49
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e3m2_sme, *c = nk_cap_sme_k; return;
|
|
50
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e3m2_sme, *c = nk_cap_sme_k; return;
|
|
51
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e3m2_sme, *c = nk_cap_sme_k; return;
|
|
52
|
+
case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_e3m2_sme, *c = nk_cap_sme_k; return;
|
|
53
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e3m2_sme, *c = nk_cap_sme_k; return;
|
|
54
|
+
case nk_kernel_euclideans_symmetric_k: *m = (m_t)&nk_euclideans_symmetric_e3m2_sme, *c = nk_cap_sme_k; return;
|
|
55
|
+
default: break;
|
|
56
|
+
}
|
|
57
|
+
#endif
|
|
58
|
+
#if NK_TARGET_NEONSDOT
|
|
59
|
+
if (v & nk_cap_neonsdot_k) switch (k) {
|
|
60
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e3m2_neonsdot, *c = nk_cap_neonsdot_k; return;
|
|
61
|
+
default: break;
|
|
62
|
+
}
|
|
63
|
+
#endif
|
|
64
|
+
#if NK_TARGET_NEON
|
|
65
|
+
if (v & nk_cap_neon_k) switch (k) {
|
|
66
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e3m2_neon, *c = nk_cap_neon_k; return;
|
|
67
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e3m2_neon, *c = nk_cap_neon_k; return;
|
|
68
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e3m2_neon, *c = nk_cap_neon_k; return;
|
|
69
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e3m2_neon, *c = nk_cap_neon_k; return;
|
|
70
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e3m2_neon, *c = nk_cap_neon_k; return;
|
|
71
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e3m2_neon, *c = nk_cap_neon_k; return;
|
|
72
|
+
default: break;
|
|
73
|
+
}
|
|
74
|
+
#endif
|
|
75
|
+
#if NK_TARGET_SAPPHIREAMX
|
|
76
|
+
if (v & nk_cap_sapphireamx_k) switch (k) {
|
|
77
|
+
case nk_kernel_dots_packed_size_k:
|
|
78
|
+
*m = (m_t)&nk_dots_packed_size_e3m2_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
79
|
+
return;
|
|
80
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e3m2_sapphireamx, *c = nk_cap_sapphireamx_k; return;
|
|
81
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e3m2_sapphireamx, *c = nk_cap_sapphireamx_k; return;
|
|
82
|
+
case nk_kernel_dots_symmetric_k:
|
|
83
|
+
*m = (m_t)&nk_dots_symmetric_e3m2_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
84
|
+
return;
|
|
85
|
+
case nk_kernel_angulars_packed_k:
|
|
86
|
+
*m = (m_t)&nk_angulars_packed_e3m2_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
87
|
+
return;
|
|
88
|
+
case nk_kernel_angulars_symmetric_k:
|
|
89
|
+
*m = (m_t)&nk_angulars_symmetric_e3m2_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
90
|
+
return;
|
|
91
|
+
case nk_kernel_euclideans_packed_k:
|
|
92
|
+
*m = (m_t)&nk_euclideans_packed_e3m2_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
93
|
+
return;
|
|
94
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
95
|
+
*m = (m_t)&nk_euclideans_symmetric_e3m2_sapphireamx, *c = nk_cap_sapphireamx_k;
|
|
96
|
+
return;
|
|
97
|
+
default: break;
|
|
98
|
+
}
|
|
99
|
+
#endif
|
|
100
|
+
#if NK_TARGET_SAPPHIRE
|
|
101
|
+
if (v & nk_cap_sapphire_k) switch (k) {
|
|
102
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e3m2_sapphire, *c = nk_cap_sapphire_k; return;
|
|
103
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e3m2_sapphire, *c = nk_cap_sapphire_k; return;
|
|
104
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e3m2_sapphire, *c = nk_cap_sapphire_k; return;
|
|
105
|
+
default: break;
|
|
106
|
+
}
|
|
107
|
+
#endif
|
|
108
|
+
#if NK_TARGET_ICELAKE
|
|
109
|
+
if (v & nk_cap_icelake_k) switch (k) {
|
|
110
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e3m2_icelake, *c = nk_cap_icelake_k; return;
|
|
111
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e3m2_icelake, *c = nk_cap_icelake_k; return;
|
|
112
|
+
default: break;
|
|
113
|
+
}
|
|
114
|
+
#endif
|
|
115
|
+
#if NK_TARGET_SKYLAKE
|
|
116
|
+
if (v & nk_cap_skylake_k) switch (k) {
|
|
117
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
118
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
119
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
120
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
121
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
122
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
123
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
124
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
125
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
126
|
+
case nk_kernel_angulars_symmetric_k:
|
|
127
|
+
*m = (m_t)&nk_angulars_symmetric_e3m2_skylake, *c = nk_cap_skylake_k;
|
|
128
|
+
return;
|
|
129
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
130
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
131
|
+
*m = (m_t)&nk_euclideans_symmetric_e3m2_skylake, *c = nk_cap_skylake_k;
|
|
132
|
+
return;
|
|
133
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
134
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e3m2_skylake, *c = nk_cap_skylake_k; return;
|
|
135
|
+
default: break;
|
|
136
|
+
}
|
|
137
|
+
#endif
|
|
138
|
+
#if NK_TARGET_ALDER
|
|
139
|
+
if (v & nk_cap_alder_k) switch (k) {
|
|
140
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e3m2_alder, *c = nk_cap_alder_k; return;
|
|
141
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e3m2_alder, *c = nk_cap_alder_k; return;
|
|
142
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e3m2_alder, *c = nk_cap_alder_k; return;
|
|
143
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e3m2_alder, *c = nk_cap_alder_k; return;
|
|
144
|
+
default: break;
|
|
145
|
+
}
|
|
146
|
+
#endif
|
|
147
|
+
#if NK_TARGET_HASWELL
|
|
148
|
+
if (v & nk_cap_haswell_k) switch (k) {
|
|
149
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
150
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
151
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
152
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
153
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
154
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
155
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
156
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
157
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
158
|
+
case nk_kernel_angulars_symmetric_k:
|
|
159
|
+
*m = (m_t)&nk_angulars_symmetric_e3m2_haswell, *c = nk_cap_haswell_k;
|
|
160
|
+
return;
|
|
161
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
162
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
163
|
+
*m = (m_t)&nk_euclideans_symmetric_e3m2_haswell, *c = nk_cap_haswell_k;
|
|
164
|
+
return;
|
|
165
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
166
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e3m2_haswell, *c = nk_cap_haswell_k; return;
|
|
167
|
+
default: break;
|
|
168
|
+
}
|
|
169
|
+
#endif
|
|
170
|
+
#if NK_TARGET_RVV
|
|
171
|
+
if (v & nk_cap_rvv_k) switch (k) {
|
|
172
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
173
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
174
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
175
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
176
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
177
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
178
|
+
case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
179
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
180
|
+
case nk_kernel_euclideans_symmetric_k: *m = (m_t)&nk_euclideans_symmetric_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
181
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
182
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e3m2_rvv, *c = nk_cap_rvv_k; return;
|
|
183
|
+
default: break;
|
|
184
|
+
}
|
|
185
|
+
#endif
|
|
186
|
+
if (v & nk_cap_serial_k) switch (k) {
|
|
187
|
+
case nk_kernel_dot_k: *m = (m_t)&nk_dot_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
188
|
+
case nk_kernel_euclidean_k: *m = (m_t)&nk_euclidean_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
189
|
+
case nk_kernel_sqeuclidean_k: *m = (m_t)&nk_sqeuclidean_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
190
|
+
case nk_kernel_angular_k: *m = (m_t)&nk_angular_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
191
|
+
case nk_kernel_each_sum_k: *m = (m_t)&nk_each_sum_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
192
|
+
case nk_kernel_each_scale_k: *m = (m_t)&nk_each_scale_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
193
|
+
case nk_kernel_each_blend_k: *m = (m_t)&nk_each_blend_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
194
|
+
case nk_kernel_each_fma_k: *m = (m_t)&nk_each_fma_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
195
|
+
case nk_kernel_dots_packed_size_k: *m = (m_t)&nk_dots_packed_size_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
196
|
+
case nk_kernel_dots_pack_k: *m = (m_t)&nk_dots_pack_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
197
|
+
case nk_kernel_dots_packed_k: *m = (m_t)&nk_dots_packed_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
198
|
+
case nk_kernel_dots_symmetric_k: *m = (m_t)&nk_dots_symmetric_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
199
|
+
case nk_kernel_angulars_packed_k: *m = (m_t)&nk_angulars_packed_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
200
|
+
case nk_kernel_angulars_symmetric_k: *m = (m_t)&nk_angulars_symmetric_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
201
|
+
case nk_kernel_euclideans_packed_k: *m = (m_t)&nk_euclideans_packed_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
202
|
+
case nk_kernel_euclideans_symmetric_k:
|
|
203
|
+
*m = (m_t)&nk_euclideans_symmetric_e3m2_serial, *c = nk_cap_serial_k;
|
|
204
|
+
return;
|
|
205
|
+
case nk_kernel_reduce_moments_k: *m = (m_t)&nk_reduce_moments_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
206
|
+
case nk_kernel_reduce_minmax_k: *m = (m_t)&nk_reduce_minmax_e3m2_serial, *c = nk_cap_serial_k; return;
|
|
207
|
+
default: break;
|
|
208
|
+
}
|
|
209
|
+
|
|
210
|
+
// Error fallback - zero capability signals lookup failure
|
|
211
|
+
*m = (m_t)nk_error_dense_, *c = 0;
|
|
212
|
+
}
|
|
213
|
+
|
|
214
|
+
void nk_dispatch_e3m2_init_(nk_capability_t caps) {
|
|
215
|
+
nk_implementations_t *t = &nk_dispatch_table;
|
|
216
|
+
nk_capability_t used;
|
|
217
|
+
|
|
218
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_dot_k, (nk_kernel_punned_t *)&t->dot_e3m2, &used);
|
|
219
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_angular_k, (nk_kernel_punned_t *)&t->angular_e3m2, &used);
|
|
220
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_euclidean_k, (nk_kernel_punned_t *)&t->euclidean_e3m2, &used);
|
|
221
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_sqeuclidean_k, (nk_kernel_punned_t *)&t->sqeuclidean_e3m2, &used);
|
|
222
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_each_scale_k, (nk_kernel_punned_t *)&t->each_scale_e3m2, &used);
|
|
223
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_each_sum_k, (nk_kernel_punned_t *)&t->each_sum_e3m2, &used);
|
|
224
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_each_blend_k, (nk_kernel_punned_t *)&t->each_blend_e3m2, &used);
|
|
225
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_each_fma_k, (nk_kernel_punned_t *)&t->each_fma_e3m2, &used);
|
|
226
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_reduce_moments_k, (nk_kernel_punned_t *)&t->reduce_moments_e3m2, &used);
|
|
227
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_reduce_minmax_k, (nk_kernel_punned_t *)&t->reduce_minmax_e3m2, &used);
|
|
228
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_dots_packed_size_k, (nk_kernel_punned_t *)&t->dots_packed_size_e3m2, &used);
|
|
229
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_dots_pack_k, (nk_kernel_punned_t *)&t->dots_pack_e3m2, &used);
|
|
230
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_dots_packed_k, (nk_kernel_punned_t *)&t->dots_packed_e3m2, &used);
|
|
231
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_dots_symmetric_k, (nk_kernel_punned_t *)&t->dots_symmetric_e3m2, &used);
|
|
232
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_angulars_packed_k, (nk_kernel_punned_t *)&t->angulars_packed_e3m2, &used);
|
|
233
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_angulars_symmetric_k, (nk_kernel_punned_t *)&t->angulars_symmetric_e3m2,
|
|
234
|
+
&used);
|
|
235
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_euclideans_packed_k, (nk_kernel_punned_t *)&t->euclideans_packed_e3m2,
|
|
236
|
+
&used);
|
|
237
|
+
nk_dispatch_e3m2_find_(caps, nk_kernel_euclideans_symmetric_k, (nk_kernel_punned_t *)&t->euclideans_symmetric_e3m2,
|
|
238
|
+
&used);
|
|
239
|
+
}
|
|
240
|
+
|
|
241
|
+
#ifdef __cplusplus
|
|
242
|
+
}
|
|
243
|
+
#endif
|