BOAST 1.3.4 → 1.3.5

Sign up to get free protection for your applications and to get access to all the features.
Files changed (17) hide show
  1. checksums.yaml +4 -4
  2. data/BOAST.gemspec +1 -1
  3. data/lib/BOAST/Language/Algorithm.rb +3 -3
  4. data/lib/BOAST/Language/Architectures.rb +58 -0
  5. data/lib/BOAST/Language/CPUID_by_name.rb +1871 -1871
  6. data/lib/BOAST/Language/Config.rb +10 -0
  7. data/lib/BOAST/Language/Intrinsics.rb +87 -44
  8. data/lib/BOAST/Language/Operators.rb +92 -10
  9. data/lib/BOAST/Language/Variable.rb +11 -4
  10. data/lib/BOAST/Optimization/Optimization.rb +58 -7
  11. data/lib/BOAST/Runtime/CRuntime.rb +1 -1
  12. data/lib/BOAST/Runtime/CompiledRuntime.rb +1 -0
  13. data/lib/BOAST/Runtime/EnergyProbe.rb +18 -8
  14. data/lib/BOAST/Runtime/FFIRuntime.rb +12 -6
  15. data/lib/BOAST/Runtime/FORTRANRuntime.rb +7 -2
  16. data/lib/BOAST/Runtime/MPPARuntime.rb +1 -1
  17. metadata +2 -2
data/lib/BOAST/Language/CPUID_by_name.rb CHANGED
@@ -1,5 +1,5 @@
1
1
  module BOAST
2
- X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
2
+ X86CPUID_by_name = {"_m_from_int64"=>["MMX"],
3
3
  "_m_to_int64"=>["MMX"],
4
4
  "_m_empty"=>["MMX"],
5
5
  "_m_from_int"=>["MMX"],
@@ -134,7 +134,7 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
134
134
  "_MM_SET_ROUNDING_MODE"=>["SSE"],
135
135
  "_MM_GET_FLUSH_ZERO_MODE"=>["SSE"],
136
136
  "_MM_SET_FLUSH_ZERO_MODE"=>["SSE"],
137
- "_mm_prefetch"=>["PREFETCHWT1"],
137
+ "_mm_prefetch"=>["PREFETCHW"],
138
138
  "_mm_sfence"=>["SSE"],
139
139
  "_mm_max_pi16"=>["SSE"],
140
140
  "_m_pmaxsw"=>["SSE"],
@@ -1338,8 +1338,8 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
1338
1338
  "_xrstor64"=>["XSAVE"],
1339
1339
  "_xsave"=>["XSAVE"],
1340
1340
  "_xsave64"=>["XSAVE"],
1341
- "_xsaveopt"=>["XSAVE", "XSAVEOPT"],
1342
- "_xsaveopt64"=>["XSAVE", "XSAVEOPT"],
1341
+ "_xsaveopt"=>[["XSAVE", "XSAVEOPT"]],
1342
+ "_xsaveopt64"=>[["XSAVE", "XSAVEOPT"]],
1343
1343
  "_xsetbv"=>["XSAVE"],
1344
1344
  "_mm_loadu_si32"=>nil,
1345
1345
  "_mm_storeu_si32"=>nil,
@@ -1392,12 +1392,12 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
1392
1392
  "_storebe_i16"=>nil,
1393
1393
  "_storebe_i32"=>nil,
1394
1394
  "_storebe_i64"=>nil,
1395
- "_xsavec"=>["XSAVE", "XSAVEC"],
1396
- "_xsaves"=>["XSAVE", "XSS"],
1397
- "_xsavec64"=>["XSAVE", "XSAVEC"],
1398
- "_xsaves64"=>["XSAVE", "XSS"],
1399
- "_xrstors"=>["XSAVE", "XSS"],
1400
- "_xrstors64"=>["XSAVE", "XSS"],
1395
+ "_xsavec"=>[["XSAVE", "XSAVEC"]],
1396
+ "_xsaves"=>[["XSAVE", "XSS"]],
1397
+ "_xsavec64"=>[["XSAVE", "XSAVEC"]],
1398
+ "_xsaves64"=>[["XSAVE", "XSS"]],
1399
+ "_xrstors"=>[["XSAVE", "XSS"]],
1400
+ "_xrstors64"=>[["XSAVE", "XSS"]],
1401
1401
  "_mm512_kandn"=>["KNCNI"],
1402
1402
  "_mm512_kand"=>["KNCNI"],
1403
1403
  "_mm512_kmov"=>["KNCNI"],
@@ -3278,8 +3278,8 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
3278
3278
  "_mm512_kswapb"=>["KNCNI"],
3279
3279
  "_mm512_kortestz"=>["AVX512F"],
3280
3280
  "_mm512_kortestc"=>["AVX512F"],
3281
- "_mm512_mask2int"=>["KNCNI"],
3282
- "_mm512_int2mask"=>["KNCNI"],
3281
+ "_mm512_mask2int"=>["AVX512F"],
3282
+ "_mm512_int2mask"=>["AVX512F"],
3283
3283
  "_mm512_kconcathi_64"=>["KNCNI"],
3284
3284
  "_mm512_kconcatlo_64"=>["KNCNI"],
3285
3285
  "_mm512_kextract_64"=>["KNCNI"],
@@ -3323,903 +3323,903 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
3323
3323
  "_mm512_kmerge2l1l"=>["KNCNI"],
3324
3324
  "_mm512_kunpackd"=>["AVX512BW"],
3325
3325
  "_mm512_kunpackw"=>["AVX512BW"],
3326
- "_mm256_mask_add_pd"=>["AVX512F", "AVX512VL"],
3327
- "_mm256_maskz_add_pd"=>["AVX512F", "AVX512VL"],
3328
- "_mm_mask_add_pd"=>["AVX512F", "AVX512VL"],
3329
- "_mm_maskz_add_pd"=>["AVX512F", "AVX512VL"],
3330
- "_mm256_mask_add_ps"=>["AVX512F", "AVX512VL"],
3331
- "_mm256_maskz_add_ps"=>["AVX512F", "AVX512VL"],
3332
- "_mm_mask_add_ps"=>["AVX512F", "AVX512VL"],
3333
- "_mm_maskz_add_ps"=>["AVX512F", "AVX512VL"],
3334
- "_mm256_alignr_epi32"=>["AVX512F", "AVX512VL"],
3335
- "_mm256_mask_alignr_epi32"=>["AVX512F", "AVX512VL"],
3336
- "_mm256_maskz_alignr_epi32"=>["AVX512F", "AVX512VL"],
3337
- "_mm_alignr_epi32"=>["AVX512F", "AVX512VL"],
3338
- "_mm_mask_alignr_epi32"=>["AVX512F", "AVX512VL"],
3339
- "_mm_maskz_alignr_epi32"=>["AVX512F", "AVX512VL"],
3340
- "_mm256_alignr_epi64"=>["AVX512F", "AVX512VL"],
3341
- "_mm256_mask_alignr_epi64"=>["AVX512F", "AVX512VL"],
3342
- "_mm256_maskz_alignr_epi64"=>["AVX512F", "AVX512VL"],
3343
- "_mm_alignr_epi64"=>["AVX512F", "AVX512VL"],
3344
- "_mm_mask_alignr_epi64"=>["AVX512F", "AVX512VL"],
3345
- "_mm_maskz_alignr_epi64"=>["AVX512F", "AVX512VL"],
3346
- "_mm256_mask_andnot_pd"=>["AVX512VL", "AVX512DQ"],
3347
- "_mm256_maskz_andnot_pd"=>["AVX512VL", "AVX512DQ"],
3326
+ "_mm256_mask_add_pd"=>[["AVX512F", "AVX512VL"]],
3327
+ "_mm256_maskz_add_pd"=>[["AVX512F", "AVX512VL"]],
3328
+ "_mm_mask_add_pd"=>[["AVX512F", "AVX512VL"]],
3329
+ "_mm_maskz_add_pd"=>[["AVX512F", "AVX512VL"]],
3330
+ "_mm256_mask_add_ps"=>[["AVX512F", "AVX512VL"]],
3331
+ "_mm256_maskz_add_ps"=>[["AVX512F", "AVX512VL"]],
3332
+ "_mm_mask_add_ps"=>[["AVX512F", "AVX512VL"]],
3333
+ "_mm_maskz_add_ps"=>[["AVX512F", "AVX512VL"]],
3334
+ "_mm256_alignr_epi32"=>[["AVX512F", "AVX512VL"]],
3335
+ "_mm256_mask_alignr_epi32"=>[["AVX512F", "AVX512VL"]],
3336
+ "_mm256_maskz_alignr_epi32"=>[["AVX512F", "AVX512VL"]],
3337
+ "_mm_alignr_epi32"=>[["AVX512F", "AVX512VL"]],
3338
+ "_mm_mask_alignr_epi32"=>[["AVX512F", "AVX512VL"]],
3339
+ "_mm_maskz_alignr_epi32"=>[["AVX512F", "AVX512VL"]],
3340
+ "_mm256_alignr_epi64"=>[["AVX512F", "AVX512VL"]],
3341
+ "_mm256_mask_alignr_epi64"=>[["AVX512F", "AVX512VL"]],
3342
+ "_mm256_maskz_alignr_epi64"=>[["AVX512F", "AVX512VL"]],
3343
+ "_mm_alignr_epi64"=>[["AVX512F", "AVX512VL"]],
3344
+ "_mm_mask_alignr_epi64"=>[["AVX512F", "AVX512VL"]],
3345
+ "_mm_maskz_alignr_epi64"=>[["AVX512F", "AVX512VL"]],
3346
+ "_mm256_mask_andnot_pd"=>[["AVX512VL", "AVX512DQ"]],
3347
+ "_mm256_maskz_andnot_pd"=>[["AVX512VL", "AVX512DQ"]],
3348
3348
  "_mm512_andnot_pd"=>["AVX512DQ"],
3349
3349
  "_mm512_mask_andnot_pd"=>["AVX512DQ"],
3350
3350
  "_mm512_maskz_andnot_pd"=>["AVX512DQ"],
3351
- "_mm_mask_andnot_pd"=>["AVX512VL", "AVX512DQ"],
3352
- "_mm_maskz_andnot_pd"=>["AVX512VL", "AVX512DQ"],
3353
- "_mm256_mask_andnot_ps"=>["AVX512VL", "AVX512DQ"],
3354
- "_mm256_maskz_andnot_ps"=>["AVX512VL", "AVX512DQ"],
3351
+ "_mm_mask_andnot_pd"=>[["AVX512VL", "AVX512DQ"]],
3352
+ "_mm_maskz_andnot_pd"=>[["AVX512VL", "AVX512DQ"]],
3353
+ "_mm256_mask_andnot_ps"=>[["AVX512VL", "AVX512DQ"]],
3354
+ "_mm256_maskz_andnot_ps"=>[["AVX512VL", "AVX512DQ"]],
3355
3355
  "_mm512_andnot_ps"=>["AVX512DQ"],
3356
3356
  "_mm512_mask_andnot_ps"=>["AVX512DQ"],
3357
3357
  "_mm512_maskz_andnot_ps"=>["AVX512DQ"],
3358
- "_mm_mask_andnot_ps"=>["AVX512VL", "AVX512DQ"],
3359
- "_mm_maskz_andnot_ps"=>["AVX512VL", "AVX512DQ"],
3360
- "_mm256_mask_and_pd"=>["AVX512VL", "AVX512DQ"],
3361
- "_mm256_maskz_and_pd"=>["AVX512VL", "AVX512DQ"],
3358
+ "_mm_mask_andnot_ps"=>[["AVX512VL", "AVX512DQ"]],
3359
+ "_mm_maskz_andnot_ps"=>[["AVX512VL", "AVX512DQ"]],
3360
+ "_mm256_mask_and_pd"=>[["AVX512VL", "AVX512DQ"]],
3361
+ "_mm256_maskz_and_pd"=>[["AVX512VL", "AVX512DQ"]],
3362
3362
  "_mm512_and_pd"=>["AVX512DQ"],
3363
3363
  "_mm512_mask_and_pd"=>["AVX512DQ"],
3364
3364
  "_mm512_maskz_and_pd"=>["AVX512DQ"],
3365
- "_mm_mask_and_pd"=>["AVX512VL", "AVX512DQ"],
3366
- "_mm_maskz_and_pd"=>["AVX512VL", "AVX512DQ"],
3367
- "_mm256_mask_and_ps"=>["AVX512VL", "AVX512DQ"],
3368
- "_mm256_maskz_and_ps"=>["AVX512VL", "AVX512DQ"],
3365
+ "_mm_mask_and_pd"=>[["AVX512VL", "AVX512DQ"]],
3366
+ "_mm_maskz_and_pd"=>[["AVX512VL", "AVX512DQ"]],
3367
+ "_mm256_mask_and_ps"=>[["AVX512VL", "AVX512DQ"]],
3368
+ "_mm256_maskz_and_ps"=>[["AVX512VL", "AVX512DQ"]],
3369
3369
  "_mm512_and_ps"=>["AVX512DQ"],
3370
3370
  "_mm512_mask_and_ps"=>["AVX512DQ"],
3371
3371
  "_mm512_maskz_and_ps"=>["AVX512DQ"],
3372
- "_mm_mask_and_ps"=>["AVX512VL", "AVX512DQ"],
3373
- "_mm_maskz_and_ps"=>["AVX512VL", "AVX512DQ"],
3374
- "_mm256_mask_blend_pd"=>["AVX512VL", "AVX512F"],
3375
- "_mm_mask_blend_pd"=>["AVX512VL", "AVX512F"],
3376
- "_mm256_mask_blend_ps"=>["AVX512VL", "AVX512F"],
3377
- "_mm_mask_blend_ps"=>["AVX512VL", "AVX512F"],
3378
- "_mm256_broadcast_f32x2"=>["AVX512VL", "AVX512DQ"],
3379
- "_mm256_mask_broadcast_f32x2"=>["AVX512VL", "AVX512DQ"],
3380
- "_mm256_maskz_broadcast_f32x2"=>["AVX512VL", "AVX512DQ"],
3372
+ "_mm_mask_and_ps"=>[["AVX512VL", "AVX512DQ"]],
3373
+ "_mm_maskz_and_ps"=>[["AVX512VL", "AVX512DQ"]],
3374
+ "_mm256_mask_blend_pd"=>[["AVX512VL", "AVX512F"]],
3375
+ "_mm_mask_blend_pd"=>[["AVX512VL", "AVX512F"]],
3376
+ "_mm256_mask_blend_ps"=>[["AVX512VL", "AVX512F"]],
3377
+ "_mm_mask_blend_ps"=>[["AVX512VL", "AVX512F"]],
3378
+ "_mm256_broadcast_f32x2"=>[["AVX512VL", "AVX512DQ"]],
3379
+ "_mm256_mask_broadcast_f32x2"=>[["AVX512VL", "AVX512DQ"]],
3380
+ "_mm256_maskz_broadcast_f32x2"=>[["AVX512VL", "AVX512DQ"]],
3381
3381
  "_mm512_broadcast_f32x2"=>["AVX512DQ"],
3382
3382
  "_mm512_mask_broadcast_f32x2"=>["AVX512DQ"],
3383
3383
  "_mm512_maskz_broadcast_f32x2"=>["AVX512DQ"],
3384
- "_mm256_broadcast_f32x4"=>["AVX512VL", "AVX512F"],
3385
- "_mm256_mask_broadcast_f32x4"=>["AVX512VL", "AVX512F"],
3386
- "_mm256_maskz_broadcast_f32x4"=>["AVX512VL", "AVX512F"],
3384
+ "_mm256_broadcast_f32x4"=>[["AVX512VL", "AVX512F"]],
3385
+ "_mm256_mask_broadcast_f32x4"=>[["AVX512VL", "AVX512F"]],
3386
+ "_mm256_maskz_broadcast_f32x4"=>[["AVX512VL", "AVX512F"]],
3387
3387
  "_mm512_broadcast_f32x8"=>["AVX512DQ"],
3388
3388
  "_mm512_mask_broadcast_f32x8"=>["AVX512DQ"],
3389
3389
  "_mm512_maskz_broadcast_f32x8"=>["AVX512DQ"],
3390
- "_mm256_broadcast_f64x2"=>["AVX512VL", "AVX512DQ"],
3391
- "_mm256_mask_broadcast_f64x2"=>["AVX512VL", "AVX512DQ"],
3392
- "_mm256_maskz_broadcast_f64x2"=>["AVX512VL", "AVX512DQ"],
3390
+ "_mm256_broadcast_f64x2"=>[["AVX512VL", "AVX512DQ"]],
3391
+ "_mm256_mask_broadcast_f64x2"=>[["AVX512VL", "AVX512DQ"]],
3392
+ "_mm256_maskz_broadcast_f64x2"=>[["AVX512VL", "AVX512DQ"]],
3393
3393
  "_mm512_broadcast_f64x2"=>["AVX512DQ"],
3394
3394
  "_mm512_mask_broadcast_f64x2"=>["AVX512DQ"],
3395
3395
  "_mm512_maskz_broadcast_f64x2"=>["AVX512DQ"],
3396
- "_mm256_broadcast_i32x2"=>["AVX512VL", "AVX512DQ"],
3397
- "_mm256_mask_broadcast_i32x2"=>["AVX512VL", "AVX512DQ"],
3398
- "_mm256_maskz_broadcast_i32x2"=>["AVX512VL", "AVX512DQ"],
3396
+ "_mm256_broadcast_i32x2"=>[["AVX512VL", "AVX512DQ"]],
3397
+ "_mm256_mask_broadcast_i32x2"=>[["AVX512VL", "AVX512DQ"]],
3398
+ "_mm256_maskz_broadcast_i32x2"=>[["AVX512VL", "AVX512DQ"]],
3399
3399
  "_mm512_broadcast_i32x2"=>["AVX512DQ"],
3400
3400
  "_mm512_mask_broadcast_i32x2"=>["AVX512DQ"],
3401
3401
  "_mm512_maskz_broadcast_i32x2"=>["AVX512DQ"],
3402
- "_mm_broadcast_i32x2"=>["AVX512VL", "AVX512DQ"],
3403
- "_mm_mask_broadcast_i32x2"=>["AVX512VL", "AVX512DQ"],
3404
- "_mm_maskz_broadcast_i32x2"=>["AVX512VL", "AVX512DQ"],
3405
- "_mm256_broadcast_i32x4"=>["AVX512VL", "AVX512F"],
3406
- "_mm256_mask_broadcast_i32x4"=>["AVX512VL", "AVX512F"],
3407
- "_mm256_maskz_broadcast_i32x4"=>["AVX512VL", "AVX512F"],
3402
+ "_mm_broadcast_i32x2"=>[["AVX512VL", "AVX512DQ"]],
3403
+ "_mm_mask_broadcast_i32x2"=>[["AVX512VL", "AVX512DQ"]],
3404
+ "_mm_maskz_broadcast_i32x2"=>[["AVX512VL", "AVX512DQ"]],
3405
+ "_mm256_broadcast_i32x4"=>[["AVX512VL", "AVX512F"]],
3406
+ "_mm256_mask_broadcast_i32x4"=>[["AVX512VL", "AVX512F"]],
3407
+ "_mm256_maskz_broadcast_i32x4"=>[["AVX512VL", "AVX512F"]],
3408
3408
  "_mm512_broadcast_i32x8"=>["AVX512DQ"],
3409
3409
  "_mm512_mask_broadcast_i32x8"=>["AVX512DQ"],
3410
3410
  "_mm512_maskz_broadcast_i32x8"=>["AVX512DQ"],
3411
- "_mm256_broadcast_i64x2"=>["AVX512VL", "AVX512DQ"],
3412
- "_mm256_mask_broadcast_i64x2"=>["AVX512VL", "AVX512DQ"],
3413
- "_mm256_maskz_broadcast_i64x2"=>["AVX512VL", "AVX512DQ"],
3411
+ "_mm256_broadcast_i64x2"=>[["AVX512VL", "AVX512DQ"]],
3412
+ "_mm256_mask_broadcast_i64x2"=>[["AVX512VL", "AVX512DQ"]],
3413
+ "_mm256_maskz_broadcast_i64x2"=>[["AVX512VL", "AVX512DQ"]],
3414
3414
  "_mm512_broadcast_i64x2"=>["AVX512DQ"],
3415
3415
  "_mm512_mask_broadcast_i64x2"=>["AVX512DQ"],
3416
3416
  "_mm512_maskz_broadcast_i64x2"=>["AVX512DQ"],
3417
- "_mm256_mask_broadcastsd_pd"=>["AVX512VL", "AVX512F"],
3418
- "_mm256_maskz_broadcastsd_pd"=>["AVX512VL", "AVX512F"],
3419
- "_mm256_mask_broadcastss_ps"=>["AVX512VL", "AVX512F"],
3420
- "_mm256_maskz_broadcastss_ps"=>["AVX512VL", "AVX512F"],
3421
- "_mm_mask_broadcastss_ps"=>["AVX512VL", "AVX512F"],
3422
- "_mm_maskz_broadcastss_ps"=>["AVX512VL", "AVX512F"],
3423
- "_mm256_cmp_pd_mask"=>["AVX512VL", "AVX512F"],
3424
- "_mm256_mask_cmp_pd_mask"=>["AVX512VL", "AVX512F"],
3425
- "_mm_cmp_pd_mask"=>["AVX512VL", "AVX512F"],
3426
- "_mm_mask_cmp_pd_mask"=>["AVX512VL", "AVX512F"],
3427
- "_mm256_cmp_ps_mask"=>["AVX512VL", "AVX512F"],
3428
- "_mm256_mask_cmp_ps_mask"=>["AVX512VL", "AVX512F"],
3429
- "_mm_cmp_ps_mask"=>["AVX512VL", "AVX512F"],
3430
- "_mm_mask_cmp_ps_mask"=>["AVX512VL", "AVX512F"],
3431
- "_mm256_mask_compress_pd"=>["AVX512VL", "AVX512F"],
3432
- "_mm256_mask_compressstoreu_pd"=>["AVX512VL", "AVX512F"],
3433
- "_mm256_maskz_compress_pd"=>["AVX512VL", "AVX512F"],
3434
- "_mm_mask_compress_pd"=>["AVX512VL", "AVX512F"],
3435
- "_mm_mask_compressstoreu_pd"=>["AVX512VL", "AVX512F"],
3436
- "_mm_maskz_compress_pd"=>["AVX512VL", "AVX512F"],
3437
- "_mm256_mask_compress_ps"=>["AVX512VL", "AVX512F"],
3438
- "_mm256_mask_compressstoreu_ps"=>["AVX512VL", "AVX512F"],
3439
- "_mm256_maskz_compress_ps"=>["AVX512VL", "AVX512F"],
3440
- "_mm_mask_compress_ps"=>["AVX512VL", "AVX512F"],
3441
- "_mm_mask_compressstoreu_ps"=>["AVX512VL", "AVX512F"],
3442
- "_mm_maskz_compress_ps"=>["AVX512VL", "AVX512F"],
3443
- "_mm256_mask_cvtepi32_pd"=>["AVX512VL", "AVX512F"],
3444
- "_mm256_maskz_cvtepi32_pd"=>["AVX512VL", "AVX512F"],
3445
- "_mm_mask_cvtepi32_pd"=>["AVX512VL", "AVX512F"],
3446
- "_mm_maskz_cvtepi32_pd"=>["AVX512VL", "AVX512F"],
3447
- "_mm256_mask_cvtepi32_ps"=>["AVX512VL", "AVX512F"],
3448
- "_mm256_maskz_cvtepi32_ps"=>["AVX512VL", "AVX512F"],
3449
- "_mm_mask_cvtepi32_ps"=>["AVX512VL", "AVX512F"],
3450
- "_mm_maskz_cvtepi32_ps"=>["AVX512VL", "AVX512F"],
3451
- "_mm256_mask_cvtpd_epi32"=>["AVX512VL", "AVX512F"],
3452
- "_mm256_maskz_cvtpd_epi32"=>["AVX512VL", "AVX512F"],
3453
- "_mm_mask_cvtpd_epi32"=>["AVX512VL", "AVX512F"],
3454
- "_mm_maskz_cvtpd_epi32"=>["AVX512VL", "AVX512F"],
3455
- "_mm256_mask_cvtpd_ps"=>["AVX512VL", "AVX512F"],
3456
- "_mm256_maskz_cvtpd_ps"=>["AVX512VL", "AVX512F"],
3457
- "_mm_mask_cvtpd_ps"=>["AVX512VL", "AVX512F"],
3458
- "_mm_maskz_cvtpd_ps"=>["AVX512VL", "AVX512F"],
3459
- "_mm256_cvtpd_epi64"=>["AVX512VL", "AVX512DQ"],
3460
- "_mm256_mask_cvtpd_epi64"=>["AVX512VL", "AVX512DQ"],
3461
- "_mm256_maskz_cvtpd_epi64"=>["AVX512VL", "AVX512DQ"],
3417
+ "_mm256_mask_broadcastsd_pd"=>[["AVX512VL", "AVX512F"]],
3418
+ "_mm256_maskz_broadcastsd_pd"=>[["AVX512VL", "AVX512F"]],
3419
+ "_mm256_mask_broadcastss_ps"=>[["AVX512VL", "AVX512F"]],
3420
+ "_mm256_maskz_broadcastss_ps"=>[["AVX512VL", "AVX512F"]],
3421
+ "_mm_mask_broadcastss_ps"=>[["AVX512VL", "AVX512F"]],
3422
+ "_mm_maskz_broadcastss_ps"=>[["AVX512VL", "AVX512F"]],
3423
+ "_mm256_cmp_pd_mask"=>[["AVX512VL", "AVX512F"]],
3424
+ "_mm256_mask_cmp_pd_mask"=>[["AVX512VL", "AVX512F"]],
3425
+ "_mm_cmp_pd_mask"=>[["AVX512VL", "AVX512F"]],
3426
+ "_mm_mask_cmp_pd_mask"=>[["AVX512VL", "AVX512F"]],
3427
+ "_mm256_cmp_ps_mask"=>[["AVX512VL", "AVX512F"]],
3428
+ "_mm256_mask_cmp_ps_mask"=>[["AVX512VL", "AVX512F"]],
3429
+ "_mm_cmp_ps_mask"=>[["AVX512VL", "AVX512F"]],
3430
+ "_mm_mask_cmp_ps_mask"=>[["AVX512VL", "AVX512F"]],
3431
+ "_mm256_mask_compress_pd"=>[["AVX512VL", "AVX512F"]],
3432
+ "_mm256_mask_compressstoreu_pd"=>[["AVX512VL", "AVX512F"]],
3433
+ "_mm256_maskz_compress_pd"=>[["AVX512VL", "AVX512F"]],
3434
+ "_mm_mask_compress_pd"=>[["AVX512VL", "AVX512F"]],
3435
+ "_mm_mask_compressstoreu_pd"=>[["AVX512VL", "AVX512F"]],
3436
+ "_mm_maskz_compress_pd"=>[["AVX512VL", "AVX512F"]],
3437
+ "_mm256_mask_compress_ps"=>[["AVX512VL", "AVX512F"]],
3438
+ "_mm256_mask_compressstoreu_ps"=>[["AVX512VL", "AVX512F"]],
3439
+ "_mm256_maskz_compress_ps"=>[["AVX512VL", "AVX512F"]],
3440
+ "_mm_mask_compress_ps"=>[["AVX512VL", "AVX512F"]],
3441
+ "_mm_mask_compressstoreu_ps"=>[["AVX512VL", "AVX512F"]],
3442
+ "_mm_maskz_compress_ps"=>[["AVX512VL", "AVX512F"]],
3443
+ "_mm256_mask_cvtepi32_pd"=>[["AVX512VL", "AVX512F"]],
3444
+ "_mm256_maskz_cvtepi32_pd"=>[["AVX512VL", "AVX512F"]],
3445
+ "_mm_mask_cvtepi32_pd"=>[["AVX512VL", "AVX512F"]],
3446
+ "_mm_maskz_cvtepi32_pd"=>[["AVX512VL", "AVX512F"]],
3447
+ "_mm256_mask_cvtepi32_ps"=>[["AVX512VL", "AVX512F"]],
3448
+ "_mm256_maskz_cvtepi32_ps"=>[["AVX512VL", "AVX512F"]],
3449
+ "_mm_mask_cvtepi32_ps"=>[["AVX512VL", "AVX512F"]],
3450
+ "_mm_maskz_cvtepi32_ps"=>[["AVX512VL", "AVX512F"]],
3451
+ "_mm256_mask_cvtpd_epi32"=>[["AVX512VL", "AVX512F"]],
3452
+ "_mm256_maskz_cvtpd_epi32"=>[["AVX512VL", "AVX512F"]],
3453
+ "_mm_mask_cvtpd_epi32"=>[["AVX512VL", "AVX512F"]],
3454
+ "_mm_maskz_cvtpd_epi32"=>[["AVX512VL", "AVX512F"]],
3455
+ "_mm256_mask_cvtpd_ps"=>[["AVX512VL", "AVX512F"]],
3456
+ "_mm256_maskz_cvtpd_ps"=>[["AVX512VL", "AVX512F"]],
3457
+ "_mm_mask_cvtpd_ps"=>[["AVX512VL", "AVX512F"]],
3458
+ "_mm_maskz_cvtpd_ps"=>[["AVX512VL", "AVX512F"]],
3459
+ "_mm256_cvtpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3460
+ "_mm256_mask_cvtpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3461
+ "_mm256_maskz_cvtpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3462
3462
  "_mm512_cvt_roundpd_epi64"=>["AVX512DQ"],
3463
3463
  "_mm512_cvtpd_epi64"=>["AVX512DQ"],
3464
3464
  "_mm512_mask_cvt_roundpd_epi64"=>["AVX512DQ"],
3465
3465
  "_mm512_mask_cvtpd_epi64"=>["AVX512DQ"],
3466
3466
  "_mm512_maskz_cvt_roundpd_epi64"=>["AVX512DQ"],
3467
3467
  "_mm512_maskz_cvtpd_epi64"=>["AVX512DQ"],
3468
- "_mm_cvtpd_epi64"=>["AVX512VL", "AVX512DQ"],
3469
- "_mm_mask_cvtpd_epi64"=>["AVX512VL", "AVX512DQ"],
3470
- "_mm_maskz_cvtpd_epi64"=>["AVX512VL", "AVX512DQ"],
3471
- "_mm256_cvtpd_epu32"=>["AVX512VL", "AVX512F"],
3472
- "_mm256_mask_cvtpd_epu32"=>["AVX512VL", "AVX512F"],
3473
- "_mm256_maskz_cvtpd_epu32"=>["AVX512VL", "AVX512F"],
3474
- "_mm_cvtpd_epu32"=>["AVX512VL", "AVX512F"],
3475
- "_mm_mask_cvtpd_epu32"=>["AVX512VL", "AVX512F"],
3476
- "_mm_maskz_cvtpd_epu32"=>["AVX512VL", "AVX512F"],
3477
- "_mm256_cvtpd_epu64"=>["AVX512VL", "AVX512DQ"],
3478
- "_mm256_mask_cvtpd_epu64"=>["AVX512VL", "AVX512DQ"],
3479
- "_mm256_maskz_cvtpd_epu64"=>["AVX512VL", "AVX512DQ"],
3468
+ "_mm_cvtpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3469
+ "_mm_mask_cvtpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3470
+ "_mm_maskz_cvtpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3471
+ "_mm256_cvtpd_epu32"=>[["AVX512VL", "AVX512F"]],
3472
+ "_mm256_mask_cvtpd_epu32"=>[["AVX512VL", "AVX512F"]],
3473
+ "_mm256_maskz_cvtpd_epu32"=>[["AVX512VL", "AVX512F"]],
3474
+ "_mm_cvtpd_epu32"=>[["AVX512VL", "AVX512F"]],
3475
+ "_mm_mask_cvtpd_epu32"=>[["AVX512VL", "AVX512F"]],
3476
+ "_mm_maskz_cvtpd_epu32"=>[["AVX512VL", "AVX512F"]],
3477
+ "_mm256_cvtpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3478
+ "_mm256_mask_cvtpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3479
+ "_mm256_maskz_cvtpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3480
3480
  "_mm512_cvt_roundpd_epu64"=>["AVX512DQ"],
3481
3481
  "_mm512_cvtpd_epu64"=>["AVX512DQ"],
3482
3482
  "_mm512_mask_cvt_roundpd_epu64"=>["AVX512DQ"],
3483
3483
  "_mm512_mask_cvtpd_epu64"=>["AVX512DQ"],
3484
3484
  "_mm512_maskz_cvt_roundpd_epu64"=>["AVX512DQ"],
3485
3485
  "_mm512_maskz_cvtpd_epu64"=>["AVX512DQ"],
3486
- "_mm_cvtpd_epu64"=>["AVX512VL", "AVX512DQ"],
3487
- "_mm_mask_cvtpd_epu64"=>["AVX512VL", "AVX512DQ"],
3488
- "_mm_maskz_cvtpd_epu64"=>["AVX512VL", "AVX512DQ"],
3489
- "_mm256_mask_cvtph_ps"=>["AVX512VL", "AVX512F"],
3490
- "_mm256_maskz_cvtph_ps"=>["AVX512VL", "AVX512F"],
3491
- "_mm_mask_cvtph_ps"=>["AVX512VL", "AVX512F"],
3492
- "_mm_maskz_cvtph_ps"=>["AVX512VL", "AVX512F"],
3493
- "_mm256_mask_cvtps_epi32"=>["AVX512VL", "AVX512F"],
3494
- "_mm256_maskz_cvtps_epi32"=>["AVX512VL", "AVX512F"],
3495
- "_mm_mask_cvtps_epi32"=>["AVX512VL", "AVX512F"],
3496
- "_mm_maskz_cvtps_epi32"=>["AVX512VL", "AVX512F"],
3497
- "_mm256_mask_cvt_roundps_ph"=>["AVX512VL", "AVX512F"],
3498
- "_mm256_mask_cvtps_ph"=>["AVX512VL", "AVX512F"],
3499
- "_mm256_maskz_cvt_roundps_ph"=>["AVX512VL", "AVX512F"],
3500
- "_mm256_maskz_cvtps_ph"=>["AVX512VL", "AVX512F"],
3501
- "_mm_mask_cvt_roundps_ph"=>["AVX512VL", "AVX512F"],
3502
- "_mm_mask_cvtps_ph"=>["AVX512VL", "AVX512F"],
3503
- "_mm_maskz_cvt_roundps_ph"=>["AVX512VL", "AVX512F"],
3504
- "_mm_maskz_cvtps_ph"=>["AVX512VL", "AVX512F"],
3505
- "_mm256_cvtps_epi64"=>["AVX512VL", "AVX512DQ"],
3506
- "_mm256_mask_cvtps_epi64"=>["AVX512VL", "AVX512DQ"],
3507
- "_mm256_maskz_cvtps_epi64"=>["AVX512VL", "AVX512DQ"],
3486
+ "_mm_cvtpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3487
+ "_mm_mask_cvtpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3488
+ "_mm_maskz_cvtpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3489
+ "_mm256_mask_cvtph_ps"=>[["AVX512VL", "AVX512F"]],
3490
+ "_mm256_maskz_cvtph_ps"=>[["AVX512VL", "AVX512F"]],
3491
+ "_mm_mask_cvtph_ps"=>[["AVX512VL", "AVX512F"]],
3492
+ "_mm_maskz_cvtph_ps"=>[["AVX512VL", "AVX512F"]],
3493
+ "_mm256_mask_cvtps_epi32"=>[["AVX512VL", "AVX512F"]],
3494
+ "_mm256_maskz_cvtps_epi32"=>[["AVX512VL", "AVX512F"]],
3495
+ "_mm_mask_cvtps_epi32"=>[["AVX512VL", "AVX512F"]],
3496
+ "_mm_maskz_cvtps_epi32"=>[["AVX512VL", "AVX512F"]],
3497
+ "_mm256_mask_cvt_roundps_ph"=>[["AVX512VL", "AVX512F"]],
3498
+ "_mm256_mask_cvtps_ph"=>[["AVX512VL", "AVX512F"]],
3499
+ "_mm256_maskz_cvt_roundps_ph"=>[["AVX512VL", "AVX512F"]],
3500
+ "_mm256_maskz_cvtps_ph"=>[["AVX512VL", "AVX512F"]],
3501
+ "_mm_mask_cvt_roundps_ph"=>[["AVX512VL", "AVX512F"]],
3502
+ "_mm_mask_cvtps_ph"=>[["AVX512VL", "AVX512F"]],
3503
+ "_mm_maskz_cvt_roundps_ph"=>[["AVX512VL", "AVX512F"]],
3504
+ "_mm_maskz_cvtps_ph"=>[["AVX512VL", "AVX512F"]],
3505
+ "_mm256_cvtps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3506
+ "_mm256_mask_cvtps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3507
+ "_mm256_maskz_cvtps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3508
3508
  "_mm512_cvt_roundps_epi64"=>["AVX512DQ"],
3509
3509
  "_mm512_cvtps_epi64"=>["AVX512DQ"],
3510
3510
  "_mm512_mask_cvt_roundps_epi64"=>["AVX512DQ"],
3511
3511
  "_mm512_mask_cvtps_epi64"=>["AVX512DQ"],
3512
3512
  "_mm512_maskz_cvt_roundps_epi64"=>["AVX512DQ"],
3513
3513
  "_mm512_maskz_cvtps_epi64"=>["AVX512DQ"],
3514
- "_mm_cvtps_epi64"=>["AVX512VL", "AVX512DQ"],
3515
- "_mm_mask_cvtps_epi64"=>["AVX512VL", "AVX512DQ"],
3516
- "_mm_maskz_cvtps_epi64"=>["AVX512VL", "AVX512DQ"],
3517
- "_mm256_cvtps_epu32"=>["AVX512VL", "AVX512F"],
3518
- "_mm256_mask_cvtps_epu32"=>["AVX512VL", "AVX512F"],
3519
- "_mm256_maskz_cvtps_epu32"=>["AVX512VL", "AVX512F"],
3520
- "_mm_cvtps_epu32"=>["AVX512VL", "AVX512F"],
3521
- "_mm_mask_cvtps_epu32"=>["AVX512VL", "AVX512F"],
3522
- "_mm_maskz_cvtps_epu32"=>["AVX512VL", "AVX512F"],
3523
- "_mm256_cvtps_epu64"=>["AVX512VL", "AVX512DQ"],
3524
- "_mm256_mask_cvtps_epu64"=>["AVX512VL", "AVX512DQ"],
3525
- "_mm256_maskz_cvtps_epu64"=>["AVX512VL", "AVX512DQ"],
3514
+ "_mm_cvtps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3515
+ "_mm_mask_cvtps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3516
+ "_mm_maskz_cvtps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3517
+ "_mm256_cvtps_epu32"=>[["AVX512VL", "AVX512F"]],
3518
+ "_mm256_mask_cvtps_epu32"=>[["AVX512VL", "AVX512F"]],
3519
+ "_mm256_maskz_cvtps_epu32"=>[["AVX512VL", "AVX512F"]],
3520
+ "_mm_cvtps_epu32"=>[["AVX512VL", "AVX512F"]],
3521
+ "_mm_mask_cvtps_epu32"=>[["AVX512VL", "AVX512F"]],
3522
+ "_mm_maskz_cvtps_epu32"=>[["AVX512VL", "AVX512F"]],
3523
+ "_mm256_cvtps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3524
+ "_mm256_mask_cvtps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3525
+ "_mm256_maskz_cvtps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3526
3526
  "_mm512_cvt_roundps_epu64"=>["AVX512DQ"],
3527
3527
  "_mm512_cvtps_epu64"=>["AVX512DQ"],
3528
3528
  "_mm512_mask_cvt_roundps_epu64"=>["AVX512DQ"],
3529
3529
  "_mm512_mask_cvtps_epu64"=>["AVX512DQ"],
3530
3530
  "_mm512_maskz_cvt_roundps_epu64"=>["AVX512DQ"],
3531
3531
  "_mm512_maskz_cvtps_epu64"=>["AVX512DQ"],
3532
- "_mm_cvtps_epu64"=>["AVX512VL", "AVX512DQ"],
3533
- "_mm_mask_cvtps_epu64"=>["AVX512VL", "AVX512DQ"],
3534
- "_mm_maskz_cvtps_epu64"=>["AVX512VL", "AVX512DQ"],
3535
- "_mm256_cvtepi64_pd"=>["AVX512VL", "AVX512DQ"],
3536
- "_mm256_mask_cvtepi64_pd"=>["AVX512VL", "AVX512DQ"],
3537
- "_mm256_maskz_cvtepi64_pd"=>["AVX512VL", "AVX512DQ"],
3532
+ "_mm_cvtps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3533
+ "_mm_mask_cvtps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3534
+ "_mm_maskz_cvtps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3535
+ "_mm256_cvtepi64_pd"=>[["AVX512VL", "AVX512DQ"]],
3536
+ "_mm256_mask_cvtepi64_pd"=>[["AVX512VL", "AVX512DQ"]],
3537
+ "_mm256_maskz_cvtepi64_pd"=>[["AVX512VL", "AVX512DQ"]],
3538
3538
  "_mm512_cvt_roundepi64_pd"=>["AVX512DQ"],
3539
3539
  "_mm512_cvtepi64_pd"=>["AVX512DQ"],
3540
3540
  "_mm512_mask_cvt_roundepi64_pd"=>["AVX512DQ"],
3541
3541
  "_mm512_mask_cvtepi64_pd"=>["AVX512DQ"],
3542
3542
  "_mm512_maskz_cvt_roundepi64_pd"=>["AVX512DQ"],
3543
3543
  "_mm512_maskz_cvtepi64_pd"=>["AVX512DQ"],
3544
- "_mm_cvtepi64_pd"=>["AVX512VL", "AVX512DQ"],
3545
- "_mm_mask_cvtepi64_pd"=>["AVX512VL", "AVX512DQ"],
3546
- "_mm_maskz_cvtepi64_pd"=>["AVX512VL", "AVX512DQ"],
3547
- "_mm256_cvtepi64_ps"=>["AVX512VL", "AVX512DQ"],
3548
- "_mm256_mask_cvtepi64_ps"=>["AVX512VL", "AVX512DQ"],
3549
- "_mm256_maskz_cvtepi64_ps"=>["AVX512VL", "AVX512DQ"],
3544
+ "_mm_cvtepi64_pd"=>[["AVX512VL", "AVX512DQ"]],
3545
+ "_mm_mask_cvtepi64_pd"=>[["AVX512VL", "AVX512DQ"]],
3546
+ "_mm_maskz_cvtepi64_pd"=>[["AVX512VL", "AVX512DQ"]],
3547
+ "_mm256_cvtepi64_ps"=>[["AVX512VL", "AVX512DQ"]],
3548
+ "_mm256_mask_cvtepi64_ps"=>[["AVX512VL", "AVX512DQ"]],
3549
+ "_mm256_maskz_cvtepi64_ps"=>[["AVX512VL", "AVX512DQ"]],
3550
3550
  "_mm512_cvt_roundepi64_ps"=>["AVX512DQ"],
3551
3551
  "_mm512_cvtepi64_ps"=>["AVX512DQ"],
3552
3552
  "_mm512_mask_cvt_roundepi64_ps"=>["AVX512DQ"],
3553
3553
  "_mm512_mask_cvtepi64_ps"=>["AVX512DQ"],
3554
3554
  "_mm512_maskz_cvt_roundepi64_ps"=>["AVX512DQ"],
3555
3555
  "_mm512_maskz_cvtepi64_ps"=>["AVX512DQ"],
3556
- "_mm_cvtepi64_ps"=>["AVX512VL", "AVX512DQ"],
3557
- "_mm_mask_cvtepi64_ps"=>["AVX512VL", "AVX512DQ"],
3558
- "_mm_maskz_cvtepi64_ps"=>["AVX512VL", "AVX512DQ"],
3559
- "_mm256_mask_cvttpd_epi32"=>["AVX512VL", "AVX512F"],
3560
- "_mm256_maskz_cvttpd_epi32"=>["AVX512VL", "AVX512F"],
3561
- "_mm_mask_cvttpd_epi32"=>["AVX512VL", "AVX512F"],
3562
- "_mm_maskz_cvttpd_epi32"=>["AVX512VL", "AVX512F"],
3563
- "_mm256_cvttpd_epi64"=>["AVX512VL", "AVX512DQ"],
3564
- "_mm256_mask_cvttpd_epi64"=>["AVX512VL", "AVX512DQ"],
3565
- "_mm256_maskz_cvttpd_epi64"=>["AVX512VL", "AVX512DQ"],
3556
+ "_mm_cvtepi64_ps"=>[["AVX512VL", "AVX512DQ"]],
3557
+ "_mm_mask_cvtepi64_ps"=>[["AVX512VL", "AVX512DQ"]],
3558
+ "_mm_maskz_cvtepi64_ps"=>[["AVX512VL", "AVX512DQ"]],
3559
+ "_mm256_mask_cvttpd_epi32"=>[["AVX512VL", "AVX512F"]],
3560
+ "_mm256_maskz_cvttpd_epi32"=>[["AVX512VL", "AVX512F"]],
3561
+ "_mm_mask_cvttpd_epi32"=>[["AVX512VL", "AVX512F"]],
3562
+ "_mm_maskz_cvttpd_epi32"=>[["AVX512VL", "AVX512F"]],
3563
+ "_mm256_cvttpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3564
+ "_mm256_mask_cvttpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3565
+ "_mm256_maskz_cvttpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3566
3566
  "_mm512_cvtt_roundpd_epi64"=>["AVX512DQ"],
3567
3567
  "_mm512_cvttpd_epi64"=>["AVX512DQ"],
3568
3568
  "_mm512_mask_cvtt_roundpd_epi64"=>["AVX512DQ"],
3569
3569
  "_mm512_mask_cvttpd_epi64"=>["AVX512DQ"],
3570
3570
  "_mm512_maskz_cvtt_roundpd_epi64"=>["AVX512DQ"],
3571
3571
  "_mm512_maskz_cvttpd_epi64"=>["AVX512DQ"],
3572
- "_mm_cvttpd_epi64"=>["AVX512VL", "AVX512DQ"],
3573
- "_mm_mask_cvttpd_epi64"=>["AVX512VL", "AVX512DQ"],
3574
- "_mm_maskz_cvttpd_epi64"=>["AVX512VL", "AVX512DQ"],
3575
- "_mm256_cvttpd_epu32"=>["AVX512VL", "AVX512F"],
3576
- "_mm256_mask_cvttpd_epu32"=>["AVX512VL", "AVX512F"],
3577
- "_mm256_maskz_cvttpd_epu32"=>["AVX512VL", "AVX512F"],
3578
- "_mm_cvttpd_epu32"=>["AVX512VL", "AVX512F"],
3579
- "_mm_mask_cvttpd_epu32"=>["AVX512VL", "AVX512F"],
3580
- "_mm_maskz_cvttpd_epu32"=>["AVX512VL", "AVX512F"],
3581
- "_mm256_cvttpd_epu64"=>["AVX512VL", "AVX512DQ"],
3582
- "_mm256_mask_cvttpd_epu64"=>["AVX512VL", "AVX512DQ"],
3583
- "_mm256_maskz_cvttpd_epu64"=>["AVX512VL", "AVX512DQ"],
3572
+ "_mm_cvttpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3573
+ "_mm_mask_cvttpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3574
+ "_mm_maskz_cvttpd_epi64"=>[["AVX512VL", "AVX512DQ"]],
3575
+ "_mm256_cvttpd_epu32"=>[["AVX512VL", "AVX512F"]],
3576
+ "_mm256_mask_cvttpd_epu32"=>[["AVX512VL", "AVX512F"]],
3577
+ "_mm256_maskz_cvttpd_epu32"=>[["AVX512VL", "AVX512F"]],
3578
+ "_mm_cvttpd_epu32"=>[["AVX512VL", "AVX512F"]],
3579
+ "_mm_mask_cvttpd_epu32"=>[["AVX512VL", "AVX512F"]],
3580
+ "_mm_maskz_cvttpd_epu32"=>[["AVX512VL", "AVX512F"]],
3581
+ "_mm256_cvttpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3582
+ "_mm256_mask_cvttpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3583
+ "_mm256_maskz_cvttpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3584
3584
  "_mm512_cvtt_roundpd_epu64"=>["AVX512DQ"],
3585
3585
  "_mm512_cvttpd_epu64"=>["AVX512DQ"],
3586
3586
  "_mm512_mask_cvtt_roundpd_epu64"=>["AVX512DQ"],
3587
3587
  "_mm512_mask_cvttpd_epu64"=>["AVX512DQ"],
3588
3588
  "_mm512_maskz_cvtt_roundpd_epu64"=>["AVX512DQ"],
3589
3589
  "_mm512_maskz_cvttpd_epu64"=>["AVX512DQ"],
3590
- "_mm_cvttpd_epu64"=>["AVX512VL", "AVX512DQ"],
3591
- "_mm_mask_cvttpd_epu64"=>["AVX512VL", "AVX512DQ"],
3592
- "_mm_maskz_cvttpd_epu64"=>["AVX512VL", "AVX512DQ"],
3593
- "_mm256_mask_cvttps_epi32"=>["AVX512VL", "AVX512F"],
3594
- "_mm256_maskz_cvttps_epi32"=>["AVX512VL", "AVX512F"],
3595
- "_mm_mask_cvttps_epi32"=>["AVX512VL", "AVX512F"],
3596
- "_mm_maskz_cvttps_epi32"=>["AVX512VL", "AVX512F"],
3597
- "_mm256_cvttps_epi64"=>["AVX512VL", "AVX512DQ"],
3598
- "_mm256_mask_cvttps_epi64"=>["AVX512VL", "AVX512DQ"],
3599
- "_mm256_maskz_cvttps_epi64"=>["AVX512VL", "AVX512DQ"],
3590
+ "_mm_cvttpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3591
+ "_mm_mask_cvttpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3592
+ "_mm_maskz_cvttpd_epu64"=>[["AVX512VL", "AVX512DQ"]],
3593
+ "_mm256_mask_cvttps_epi32"=>[["AVX512VL", "AVX512F"]],
3594
+ "_mm256_maskz_cvttps_epi32"=>[["AVX512VL", "AVX512F"]],
3595
+ "_mm_mask_cvttps_epi32"=>[["AVX512VL", "AVX512F"]],
3596
+ "_mm_maskz_cvttps_epi32"=>[["AVX512VL", "AVX512F"]],
3597
+ "_mm256_cvttps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3598
+ "_mm256_mask_cvttps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3599
+ "_mm256_maskz_cvttps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3600
3600
  "_mm512_cvtt_roundps_epi64"=>["AVX512DQ"],
3601
3601
  "_mm512_cvttps_epi64"=>["AVX512DQ"],
3602
3602
  "_mm512_mask_cvtt_roundps_epi64"=>["AVX512DQ"],
3603
3603
  "_mm512_mask_cvttps_epi64"=>["AVX512DQ"],
3604
3604
  "_mm512_maskz_cvtt_roundps_epi64"=>["AVX512DQ"],
3605
3605
  "_mm512_maskz_cvttps_epi64"=>["AVX512DQ"],
3606
- "_mm_cvttps_epi64"=>["AVX512VL", "AVX512DQ"],
3607
- "_mm_mask_cvttps_epi64"=>["AVX512VL", "AVX512DQ"],
3608
- "_mm_maskz_cvttps_epi64"=>["AVX512VL", "AVX512DQ"],
3609
- "_mm256_cvttps_epu32"=>["AVX512VL", "AVX512F"],
3610
- "_mm256_mask_cvttps_epu32"=>["AVX512VL", "AVX512F"],
3611
- "_mm256_maskz_cvttps_epu32"=>["AVX512VL", "AVX512F"],
3612
- "_mm_cvttps_epu32"=>["AVX512VL", "AVX512F"],
3613
- "_mm_mask_cvttps_epu32"=>["AVX512VL", "AVX512F"],
3614
- "_mm_maskz_cvttps_epu32"=>["AVX512VL", "AVX512F"],
3615
- "_mm256_cvttps_epu64"=>["AVX512VL", "AVX512DQ"],
3616
- "_mm256_mask_cvttps_epu64"=>["AVX512VL", "AVX512DQ"],
3617
- "_mm256_maskz_cvttps_epu64"=>["AVX512VL", "AVX512DQ"],
3606
+ "_mm_cvttps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3607
+ "_mm_mask_cvttps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3608
+ "_mm_maskz_cvttps_epi64"=>[["AVX512VL", "AVX512DQ"]],
3609
+ "_mm256_cvttps_epu32"=>[["AVX512VL", "AVX512F"]],
3610
+ "_mm256_mask_cvttps_epu32"=>[["AVX512VL", "AVX512F"]],
3611
+ "_mm256_maskz_cvttps_epu32"=>[["AVX512VL", "AVX512F"]],
3612
+ "_mm_cvttps_epu32"=>[["AVX512VL", "AVX512F"]],
3613
+ "_mm_mask_cvttps_epu32"=>[["AVX512VL", "AVX512F"]],
3614
+ "_mm_maskz_cvttps_epu32"=>[["AVX512VL", "AVX512F"]],
3615
+ "_mm256_cvttps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3616
+ "_mm256_mask_cvttps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3617
+ "_mm256_maskz_cvttps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3618
3618
  "_mm512_cvtt_roundps_epu64"=>["AVX512DQ"],
3619
3619
  "_mm512_cvttps_epu64"=>["AVX512DQ"],
3620
3620
  "_mm512_mask_cvtt_roundps_epu64"=>["AVX512DQ"],
3621
3621
  "_mm512_mask_cvttps_epu64"=>["AVX512DQ"],
3622
3622
  "_mm512_maskz_cvtt_roundps_epu64"=>["AVX512DQ"],
3623
3623
  "_mm512_maskz_cvttps_epu64"=>["AVX512DQ"],
3624
- "_mm_cvttps_epu64"=>["AVX512VL", "AVX512DQ"],
3625
- "_mm_mask_cvttps_epu64"=>["AVX512VL", "AVX512DQ"],
3626
- "_mm_maskz_cvttps_epu64"=>["AVX512VL", "AVX512DQ"],
3627
- "_mm256_cvtepu32_pd"=>["AVX512VL", "AVX512F"],
3628
- "_mm256_mask_cvtepu32_pd"=>["AVX512VL", "AVX512F"],
3629
- "_mm256_maskz_cvtepu32_pd"=>["AVX512VL", "AVX512F"],
3630
- "_mm_cvtepu32_pd"=>["AVX512VL", "AVX512F"],
3631
- "_mm_mask_cvtepu32_pd"=>["AVX512VL", "AVX512F"],
3632
- "_mm_maskz_cvtepu32_pd"=>["AVX512VL", "AVX512F"],
3633
- "_mm256_cvtepu64_pd"=>["AVX512VL", "AVX512DQ"],
3634
- "_mm256_mask_cvtepu64_pd"=>["AVX512VL", "AVX512DQ"],
3635
- "_mm256_maskz_cvtepu64_pd"=>["AVX512VL", "AVX512DQ"],
3624
+ "_mm_cvttps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3625
+ "_mm_mask_cvttps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3626
+ "_mm_maskz_cvttps_epu64"=>[["AVX512VL", "AVX512DQ"]],
3627
+ "_mm256_cvtepu32_pd"=>[["AVX512VL", "AVX512F"]],
3628
+ "_mm256_mask_cvtepu32_pd"=>[["AVX512VL", "AVX512F"]],
3629
+ "_mm256_maskz_cvtepu32_pd"=>[["AVX512VL", "AVX512F"]],
3630
+ "_mm_cvtepu32_pd"=>[["AVX512VL", "AVX512F"]],
3631
+ "_mm_mask_cvtepu32_pd"=>[["AVX512VL", "AVX512F"]],
3632
+ "_mm_maskz_cvtepu32_pd"=>[["AVX512VL", "AVX512F"]],
3633
+ "_mm256_cvtepu64_pd"=>[["AVX512VL", "AVX512DQ"]],
3634
+ "_mm256_mask_cvtepu64_pd"=>[["AVX512VL", "AVX512DQ"]],
3635
+ "_mm256_maskz_cvtepu64_pd"=>[["AVX512VL", "AVX512DQ"]],
3636
3636
  "_mm512_cvt_roundepu64_pd"=>["AVX512DQ"],
3637
3637
  "_mm512_cvtepu64_pd"=>["AVX512DQ"],
3638
3638
  "_mm512_mask_cvt_roundepu64_pd"=>["AVX512DQ"],
3639
3639
  "_mm512_mask_cvtepu64_pd"=>["AVX512DQ"],
3640
3640
  "_mm512_maskz_cvt_roundepu64_pd"=>["AVX512DQ"],
3641
3641
  "_mm512_maskz_cvtepu64_pd"=>["AVX512DQ"],
3642
- "_mm_cvtepu64_pd"=>["AVX512VL", "AVX512DQ"],
3643
- "_mm_mask_cvtepu64_pd"=>["AVX512VL", "AVX512DQ"],
3644
- "_mm_maskz_cvtepu64_pd"=>["AVX512VL", "AVX512DQ"],
3645
- "_mm256_cvtepu64_ps"=>["AVX512VL", "AVX512DQ"],
3646
- "_mm256_mask_cvtepu64_ps"=>["AVX512VL", "AVX512DQ"],
3647
- "_mm256_maskz_cvtepu64_ps"=>["AVX512VL", "AVX512DQ"],
3642
+ "_mm_cvtepu64_pd"=>[["AVX512VL", "AVX512DQ"]],
3643
+ "_mm_mask_cvtepu64_pd"=>[["AVX512VL", "AVX512DQ"]],
3644
+ "_mm_maskz_cvtepu64_pd"=>[["AVX512VL", "AVX512DQ"]],
3645
+ "_mm256_cvtepu64_ps"=>[["AVX512VL", "AVX512DQ"]],
3646
+ "_mm256_mask_cvtepu64_ps"=>[["AVX512VL", "AVX512DQ"]],
3647
+ "_mm256_maskz_cvtepu64_ps"=>[["AVX512VL", "AVX512DQ"]],
3648
3648
  "_mm512_cvt_roundepu64_ps"=>["AVX512DQ"],
3649
3649
  "_mm512_cvtepu64_ps"=>["AVX512DQ"],
3650
3650
  "_mm512_mask_cvt_roundepu64_ps"=>["AVX512DQ"],
3651
3651
  "_mm512_mask_cvtepu64_ps"=>["AVX512DQ"],
3652
3652
  "_mm512_maskz_cvt_roundepu64_ps"=>["AVX512DQ"],
3653
3653
  "_mm512_maskz_cvtepu64_ps"=>["AVX512DQ"],
3654
- "_mm_cvtepu64_ps"=>["AVX512VL", "AVX512DQ"],
3655
- "_mm_mask_cvtepu64_ps"=>["AVX512VL", "AVX512DQ"],
3656
- "_mm_maskz_cvtepu64_ps"=>["AVX512VL", "AVX512DQ"],
3657
- "_mm256_dbsad_epu8"=>["AVX512VL", "AVX512BW"],
3658
- "_mm256_mask_dbsad_epu8"=>["AVX512VL", "AVX512BW"],
3659
- "_mm256_maskz_dbsad_epu8"=>["AVX512VL", "AVX512BW"],
3654
+ "_mm_cvtepu64_ps"=>[["AVX512VL", "AVX512DQ"]],
3655
+ "_mm_mask_cvtepu64_ps"=>[["AVX512VL", "AVX512DQ"]],
3656
+ "_mm_maskz_cvtepu64_ps"=>[["AVX512VL", "AVX512DQ"]],
3657
+ "_mm256_dbsad_epu8"=>[["AVX512VL", "AVX512BW"]],
3658
+ "_mm256_mask_dbsad_epu8"=>[["AVX512VL", "AVX512BW"]],
3659
+ "_mm256_maskz_dbsad_epu8"=>[["AVX512VL", "AVX512BW"]],
3660
3660
  "_mm512_dbsad_epu8"=>["AVX512BW"],
3661
3661
  "_mm512_mask_dbsad_epu8"=>["AVX512BW"],
3662
3662
  "_mm512_maskz_dbsad_epu8"=>["AVX512BW"],
3663
- "_mm_dbsad_epu8"=>["AVX512VL", "AVX512BW"],
3664
- "_mm_mask_dbsad_epu8"=>["AVX512VL", "AVX512BW"],
3665
- "_mm_maskz_dbsad_epu8"=>["AVX512VL", "AVX512BW"],
3666
- "_mm256_mask_div_pd"=>["AVX512VL", "AVX512F"],
3667
- "_mm256_maskz_div_pd"=>["AVX512VL", "AVX512F"],
3668
- "_mm_mask_div_pd"=>["AVX512VL", "AVX512F"],
3669
- "_mm_maskz_div_pd"=>["AVX512VL", "AVX512F"],
3670
- "_mm256_mask_div_ps"=>["AVX512VL", "AVX512F"],
3671
- "_mm256_maskz_div_ps"=>["AVX512VL", "AVX512F"],
3672
- "_mm_mask_div_ps"=>["AVX512VL", "AVX512F"],
3673
- "_mm_maskz_div_ps"=>["AVX512VL", "AVX512F"],
3674
- "_mm256_mask_expand_pd"=>["AVX512VL", "AVX512F"],
3675
- "_mm256_mask_expandloadu_pd"=>["AVX512VL", "AVX512F"],
3676
- "_mm256_maskz_expand_pd"=>["AVX512VL", "AVX512F"],
3677
- "_mm256_maskz_expandloadu_pd"=>["AVX512VL", "AVX512F"],
3678
- "_mm_mask_expand_pd"=>["AVX512VL", "AVX512F"],
3679
- "_mm_mask_expandloadu_pd"=>["AVX512VL", "AVX512F"],
3680
- "_mm_maskz_expand_pd"=>["AVX512VL", "AVX512F"],
3681
- "_mm_maskz_expandloadu_pd"=>["AVX512VL", "AVX512F"],
3682
- "_mm256_mask_expand_ps"=>["AVX512VL", "AVX512F"],
3683
- "_mm256_mask_expandloadu_ps"=>["AVX512VL", "AVX512F"],
3684
- "_mm256_maskz_expand_ps"=>["AVX512VL", "AVX512F"],
3685
- "_mm256_maskz_expandloadu_ps"=>["AVX512VL", "AVX512F"],
3686
- "_mm_mask_expand_ps"=>["AVX512VL", "AVX512F"],
3687
- "_mm_mask_expandloadu_ps"=>["AVX512VL", "AVX512F"],
3688
- "_mm_maskz_expand_ps"=>["AVX512VL", "AVX512F"],
3689
- "_mm_maskz_expandloadu_ps"=>["AVX512VL", "AVX512F"],
3690
- "_mm256_extractf32x4_ps"=>["AVX512VL", "AVX512F"],
3691
- "_mm256_mask_extractf32x4_ps"=>["AVX512VL", "AVX512F"],
3692
- "_mm256_maskz_extractf32x4_ps"=>["AVX512VL", "AVX512F"],
3663
+ "_mm_dbsad_epu8"=>[["AVX512VL", "AVX512BW"]],
3664
+ "_mm_mask_dbsad_epu8"=>[["AVX512VL", "AVX512BW"]],
3665
+ "_mm_maskz_dbsad_epu8"=>[["AVX512VL", "AVX512BW"]],
3666
+ "_mm256_mask_div_pd"=>[["AVX512VL", "AVX512F"]],
3667
+ "_mm256_maskz_div_pd"=>[["AVX512VL", "AVX512F"]],
3668
+ "_mm_mask_div_pd"=>[["AVX512VL", "AVX512F"]],
3669
+ "_mm_maskz_div_pd"=>[["AVX512VL", "AVX512F"]],
3670
+ "_mm256_mask_div_ps"=>[["AVX512VL", "AVX512F"]],
3671
+ "_mm256_maskz_div_ps"=>[["AVX512VL", "AVX512F"]],
3672
+ "_mm_mask_div_ps"=>[["AVX512VL", "AVX512F"]],
3673
+ "_mm_maskz_div_ps"=>[["AVX512VL", "AVX512F"]],
3674
+ "_mm256_mask_expand_pd"=>[["AVX512VL", "AVX512F"]],
3675
+ "_mm256_mask_expandloadu_pd"=>[["AVX512VL", "AVX512F"]],
3676
+ "_mm256_maskz_expand_pd"=>[["AVX512VL", "AVX512F"]],
3677
+ "_mm256_maskz_expandloadu_pd"=>[["AVX512VL", "AVX512F"]],
3678
+ "_mm_mask_expand_pd"=>[["AVX512VL", "AVX512F"]],
3679
+ "_mm_mask_expandloadu_pd"=>[["AVX512VL", "AVX512F"]],
3680
+ "_mm_maskz_expand_pd"=>[["AVX512VL", "AVX512F"]],
3681
+ "_mm_maskz_expandloadu_pd"=>[["AVX512VL", "AVX512F"]],
3682
+ "_mm256_mask_expand_ps"=>[["AVX512VL", "AVX512F"]],
3683
+ "_mm256_mask_expandloadu_ps"=>[["AVX512VL", "AVX512F"]],
3684
+ "_mm256_maskz_expand_ps"=>[["AVX512VL", "AVX512F"]],
3685
+ "_mm256_maskz_expandloadu_ps"=>[["AVX512VL", "AVX512F"]],
3686
+ "_mm_mask_expand_ps"=>[["AVX512VL", "AVX512F"]],
3687
+ "_mm_mask_expandloadu_ps"=>[["AVX512VL", "AVX512F"]],
3688
+ "_mm_maskz_expand_ps"=>[["AVX512VL", "AVX512F"]],
3689
+ "_mm_maskz_expandloadu_ps"=>[["AVX512VL", "AVX512F"]],
3690
+ "_mm256_extractf32x4_ps"=>[["AVX512VL", "AVX512F"]],
3691
+ "_mm256_mask_extractf32x4_ps"=>[["AVX512VL", "AVX512F"]],
3692
+ "_mm256_maskz_extractf32x4_ps"=>[["AVX512VL", "AVX512F"]],
3693
3693
  "_mm512_extractf32x8_ps"=>["AVX512DQ"],
3694
3694
  "_mm512_mask_extractf32x8_ps"=>["AVX512DQ"],
3695
3695
  "_mm512_maskz_extractf32x8_ps"=>["AVX512DQ"],
3696
- "_mm256_extractf64x2_pd"=>["AVX512VL", "AVX512DQ"],
3697
- "_mm256_mask_extractf64x2_pd"=>["AVX512VL", "AVX512DQ"],
3698
- "_mm256_maskz_extractf64x2_pd"=>["AVX512VL", "AVX512DQ"],
3696
+ "_mm256_extractf64x2_pd"=>[["AVX512VL", "AVX512DQ"]],
3697
+ "_mm256_mask_extractf64x2_pd"=>[["AVX512VL", "AVX512DQ"]],
3698
+ "_mm256_maskz_extractf64x2_pd"=>[["AVX512VL", "AVX512DQ"]],
3699
3699
  "_mm512_extractf64x2_pd"=>["AVX512DQ"],
3700
3700
  "_mm512_mask_extractf64x2_pd"=>["AVX512DQ"],
3701
3701
  "_mm512_maskz_extractf64x2_pd"=>["AVX512DQ"],
3702
- "_mm256_extracti32x4_epi32"=>["AVX512VL", "AVX512F"],
3703
- "_mm256_mask_extracti32x4_epi32"=>["AVX512VL", "AVX512F"],
3704
- "_mm256_maskz_extracti32x4_epi32"=>["AVX512VL", "AVX512F"],
3702
+ "_mm256_extracti32x4_epi32"=>[["AVX512VL", "AVX512F"]],
3703
+ "_mm256_mask_extracti32x4_epi32"=>[["AVX512VL", "AVX512F"]],
3704
+ "_mm256_maskz_extracti32x4_epi32"=>[["AVX512VL", "AVX512F"]],
3705
3705
  "_mm512_extracti32x8_epi32"=>["AVX512DQ"],
3706
3706
  "_mm512_mask_extracti32x8_epi32"=>["AVX512DQ"],
3707
3707
  "_mm512_maskz_extracti32x8_epi32"=>["AVX512DQ"],
3708
- "_mm256_extracti64x2_epi64"=>["AVX512VL", "AVX512DQ"],
3709
- "_mm256_mask_extracti64x2_epi64"=>["AVX512VL", "AVX512DQ"],
3710
- "_mm256_maskz_extracti64x2_epi64"=>["AVX512VL", "AVX512DQ"],
3708
+ "_mm256_extracti64x2_epi64"=>[["AVX512VL", "AVX512DQ"]],
3709
+ "_mm256_mask_extracti64x2_epi64"=>[["AVX512VL", "AVX512DQ"]],
3710
+ "_mm256_maskz_extracti64x2_epi64"=>[["AVX512VL", "AVX512DQ"]],
3711
3711
  "_mm512_extracti64x2_epi64"=>["AVX512DQ"],
3712
3712
  "_mm512_mask_extracti64x2_epi64"=>["AVX512DQ"],
3713
3713
  "_mm512_maskz_extracti64x2_epi64"=>["AVX512DQ"],
3714
- "_mm256_fixupimm_pd"=>["AVX512VL", "AVX512F"],
3715
- "_mm256_mask_fixupimm_pd"=>["AVX512VL", "AVX512F"],
3716
- "_mm256_maskz_fixupimm_pd"=>["AVX512VL", "AVX512F"],
3717
- "_mm_fixupimm_pd"=>["AVX512VL", "AVX512F"],
3718
- "_mm_mask_fixupimm_pd"=>["AVX512VL", "AVX512F"],
3719
- "_mm_maskz_fixupimm_pd"=>["AVX512VL", "AVX512F"],
3720
- "_mm256_fixupimm_ps"=>["AVX512VL", "AVX512F"],
3721
- "_mm256_mask_fixupimm_ps"=>["AVX512VL", "AVX512F"],
3722
- "_mm256_maskz_fixupimm_ps"=>["AVX512VL", "AVX512F"],
3723
- "_mm_fixupimm_ps"=>["AVX512VL", "AVX512F"],
3724
- "_mm_mask_fixupimm_ps"=>["AVX512VL", "AVX512F"],
3725
- "_mm_maskz_fixupimm_ps"=>["AVX512VL", "AVX512F"],
3726
- "_mm256_mask3_fmadd_pd"=>["AVX512VL", "AVX512F"],
3727
- "_mm256_mask_fmadd_pd"=>["AVX512VL", "AVX512F"],
3728
- "_mm256_maskz_fmadd_pd"=>["AVX512VL", "AVX512F"],
3729
- "_mm_mask3_fmadd_pd"=>["AVX512VL", "AVX512F"],
3730
- "_mm_mask_fmadd_pd"=>["AVX512VL", "AVX512F"],
3731
- "_mm_maskz_fmadd_pd"=>["AVX512VL", "AVX512F"],
3732
- "_mm256_mask3_fmadd_ps"=>["AVX512VL", "AVX512F"],
3733
- "_mm256_mask_fmadd_ps"=>["AVX512VL", "AVX512F"],
3734
- "_mm256_maskz_fmadd_ps"=>["AVX512VL", "AVX512F"],
3735
- "_mm_mask3_fmadd_ps"=>["AVX512VL", "AVX512F"],
3736
- "_mm_mask_fmadd_ps"=>["AVX512VL", "AVX512F"],
3737
- "_mm_maskz_fmadd_ps"=>["AVX512VL", "AVX512F"],
3738
- "_mm256_mask3_fmaddsub_pd"=>["AVX512VL", "AVX512F"],
3739
- "_mm256_mask_fmaddsub_pd"=>["AVX512VL", "AVX512F"],
3740
- "_mm256_maskz_fmaddsub_pd"=>["AVX512VL", "AVX512F"],
3741
- "_mm_mask3_fmaddsub_pd"=>["AVX512VL", "AVX512F"],
3742
- "_mm_mask_fmaddsub_pd"=>["AVX512VL", "AVX512F"],
3743
- "_mm_maskz_fmaddsub_pd"=>["AVX512VL", "AVX512F"],
3744
- "_mm256_mask3_fmaddsub_ps"=>["AVX512VL", "AVX512F"],
3745
- "_mm256_mask_fmaddsub_ps"=>["AVX512VL", "AVX512F"],
3746
- "_mm256_maskz_fmaddsub_ps"=>["AVX512VL", "AVX512F"],
3747
- "_mm_mask3_fmaddsub_ps"=>["AVX512VL", "AVX512F"],
3748
- "_mm_mask_fmaddsub_ps"=>["AVX512VL", "AVX512F"],
3749
- "_mm_maskz_fmaddsub_ps"=>["AVX512VL", "AVX512F"],
3750
- "_mm256_mask3_fmsub_pd"=>["AVX512VL", "AVX512F"],
3751
- "_mm256_mask_fmsub_pd"=>["AVX512VL", "AVX512F"],
3752
- "_mm256_maskz_fmsub_pd"=>["AVX512VL", "AVX512F"],
3753
- "_mm_mask3_fmsub_pd"=>["AVX512VL", "AVX512F"],
3754
- "_mm_mask_fmsub_pd"=>["AVX512VL", "AVX512F"],
3755
- "_mm_maskz_fmsub_pd"=>["AVX512VL", "AVX512F"],
3756
- "_mm256_mask3_fmsub_ps"=>["AVX512VL", "AVX512F"],
3757
- "_mm256_mask_fmsub_ps"=>["AVX512VL", "AVX512F"],
3758
- "_mm256_maskz_fmsub_ps"=>["AVX512VL", "AVX512F"],
3759
- "_mm_mask3_fmsub_ps"=>["AVX512VL", "AVX512F"],
3760
- "_mm_mask_fmsub_ps"=>["AVX512VL", "AVX512F"],
3761
- "_mm_maskz_fmsub_ps"=>["AVX512VL", "AVX512F"],
3762
- "_mm256_mask3_fmsubadd_pd"=>["AVX512VL", "AVX512F"],
3763
- "_mm256_mask_fmsubadd_pd"=>["AVX512VL", "AVX512F"],
3764
- "_mm256_maskz_fmsubadd_pd"=>["AVX512VL", "AVX512F"],
3765
- "_mm_mask3_fmsubadd_pd"=>["AVX512VL", "AVX512F"],
3766
- "_mm_mask_fmsubadd_pd"=>["AVX512VL", "AVX512F"],
3767
- "_mm_maskz_fmsubadd_pd"=>["AVX512VL", "AVX512F"],
3768
- "_mm256_mask3_fmsubadd_ps"=>["AVX512VL", "AVX512F"],
3769
- "_mm256_mask_fmsubadd_ps"=>["AVX512VL", "AVX512F"],
3770
- "_mm256_maskz_fmsubadd_ps"=>["AVX512VL", "AVX512F"],
3771
- "_mm_mask3_fmsubadd_ps"=>["AVX512VL", "AVX512F"],
3772
- "_mm_mask_fmsubadd_ps"=>["AVX512VL", "AVX512F"],
3773
- "_mm_maskz_fmsubadd_ps"=>["AVX512VL", "AVX512F"],
3774
- "_mm256_mask3_fnmadd_pd"=>["AVX512VL", "AVX512F"],
3775
- "_mm256_mask_fnmadd_pd"=>["AVX512VL", "AVX512F"],
3776
- "_mm256_maskz_fnmadd_pd"=>["AVX512VL", "AVX512F"],
3777
- "_mm_mask3_fnmadd_pd"=>["AVX512VL", "AVX512F"],
3778
- "_mm_mask_fnmadd_pd"=>["AVX512VL", "AVX512F"],
3779
- "_mm_maskz_fnmadd_pd"=>["AVX512VL", "AVX512F"],
3780
- "_mm256_mask3_fnmadd_ps"=>["AVX512VL", "AVX512F"],
3781
- "_mm256_mask_fnmadd_ps"=>["AVX512VL", "AVX512F"],
3782
- "_mm256_maskz_fnmadd_ps"=>["AVX512VL", "AVX512F"],
3783
- "_mm_mask3_fnmadd_ps"=>["AVX512VL", "AVX512F"],
3784
- "_mm_mask_fnmadd_ps"=>["AVX512VL", "AVX512F"],
3785
- "_mm_maskz_fnmadd_ps"=>["AVX512VL", "AVX512F"],
3786
- "_mm256_mask3_fnmsub_pd"=>["AVX512VL", "AVX512F"],
3787
- "_mm256_mask_fnmsub_pd"=>["AVX512VL", "AVX512F"],
3788
- "_mm256_maskz_fnmsub_pd"=>["AVX512VL", "AVX512F"],
3789
- "_mm_mask3_fnmsub_pd"=>["AVX512VL", "AVX512F"],
3790
- "_mm_mask_fnmsub_pd"=>["AVX512VL", "AVX512F"],
3791
- "_mm_maskz_fnmsub_pd"=>["AVX512VL", "AVX512F"],
3792
- "_mm256_mask3_fnmsub_ps"=>["AVX512VL", "AVX512F"],
3793
- "_mm256_mask_fnmsub_ps"=>["AVX512VL", "AVX512F"],
3794
- "_mm256_maskz_fnmsub_ps"=>["AVX512VL", "AVX512F"],
3795
- "_mm_mask3_fnmsub_ps"=>["AVX512VL", "AVX512F"],
3796
- "_mm_mask_fnmsub_ps"=>["AVX512VL", "AVX512F"],
3797
- "_mm_maskz_fnmsub_ps"=>["AVX512VL", "AVX512F"],
3798
- "_mm256_fpclass_pd_mask"=>["AVX512VL", "AVX512DQ"],
3799
- "_mm256_mask_fpclass_pd_mask"=>["AVX512VL", "AVX512DQ"],
3714
+ "_mm256_fixupimm_pd"=>[["AVX512VL", "AVX512F"]],
3715
+ "_mm256_mask_fixupimm_pd"=>[["AVX512VL", "AVX512F"]],
3716
+ "_mm256_maskz_fixupimm_pd"=>[["AVX512VL", "AVX512F"]],
3717
+ "_mm_fixupimm_pd"=>[["AVX512VL", "AVX512F"]],
3718
+ "_mm_mask_fixupimm_pd"=>[["AVX512VL", "AVX512F"]],
3719
+ "_mm_maskz_fixupimm_pd"=>[["AVX512VL", "AVX512F"]],
3720
+ "_mm256_fixupimm_ps"=>[["AVX512VL", "AVX512F"]],
3721
+ "_mm256_mask_fixupimm_ps"=>[["AVX512VL", "AVX512F"]],
3722
+ "_mm256_maskz_fixupimm_ps"=>[["AVX512VL", "AVX512F"]],
3723
+ "_mm_fixupimm_ps"=>[["AVX512VL", "AVX512F"]],
3724
+ "_mm_mask_fixupimm_ps"=>[["AVX512VL", "AVX512F"]],
3725
+ "_mm_maskz_fixupimm_ps"=>[["AVX512VL", "AVX512F"]],
3726
+ "_mm256_mask3_fmadd_pd"=>[["AVX512VL", "AVX512F"]],
3727
+ "_mm256_mask_fmadd_pd"=>[["AVX512VL", "AVX512F"]],
3728
+ "_mm256_maskz_fmadd_pd"=>[["AVX512VL", "AVX512F"]],
3729
+ "_mm_mask3_fmadd_pd"=>[["AVX512VL", "AVX512F"]],
3730
+ "_mm_mask_fmadd_pd"=>[["AVX512VL", "AVX512F"]],
3731
+ "_mm_maskz_fmadd_pd"=>[["AVX512VL", "AVX512F"]],
3732
+ "_mm256_mask3_fmadd_ps"=>[["AVX512VL", "AVX512F"]],
3733
+ "_mm256_mask_fmadd_ps"=>[["AVX512VL", "AVX512F"]],
3734
+ "_mm256_maskz_fmadd_ps"=>[["AVX512VL", "AVX512F"]],
3735
+ "_mm_mask3_fmadd_ps"=>[["AVX512VL", "AVX512F"]],
3736
+ "_mm_mask_fmadd_ps"=>[["AVX512VL", "AVX512F"]],
3737
+ "_mm_maskz_fmadd_ps"=>[["AVX512VL", "AVX512F"]],
3738
+ "_mm256_mask3_fmaddsub_pd"=>[["AVX512VL", "AVX512F"]],
3739
+ "_mm256_mask_fmaddsub_pd"=>[["AVX512VL", "AVX512F"]],
3740
+ "_mm256_maskz_fmaddsub_pd"=>[["AVX512VL", "AVX512F"]],
3741
+ "_mm_mask3_fmaddsub_pd"=>[["AVX512VL", "AVX512F"]],
3742
+ "_mm_mask_fmaddsub_pd"=>[["AVX512VL", "AVX512F"]],
3743
+ "_mm_maskz_fmaddsub_pd"=>[["AVX512VL", "AVX512F"]],
3744
+ "_mm256_mask3_fmaddsub_ps"=>[["AVX512VL", "AVX512F"]],
3745
+ "_mm256_mask_fmaddsub_ps"=>[["AVX512VL", "AVX512F"]],
3746
+ "_mm256_maskz_fmaddsub_ps"=>[["AVX512VL", "AVX512F"]],
3747
+ "_mm_mask3_fmaddsub_ps"=>[["AVX512VL", "AVX512F"]],
3748
+ "_mm_mask_fmaddsub_ps"=>[["AVX512VL", "AVX512F"]],
3749
+ "_mm_maskz_fmaddsub_ps"=>[["AVX512VL", "AVX512F"]],
3750
+ "_mm256_mask3_fmsub_pd"=>[["AVX512VL", "AVX512F"]],
3751
+ "_mm256_mask_fmsub_pd"=>[["AVX512VL", "AVX512F"]],
3752
+ "_mm256_maskz_fmsub_pd"=>[["AVX512VL", "AVX512F"]],
3753
+ "_mm_mask3_fmsub_pd"=>[["AVX512VL", "AVX512F"]],
3754
+ "_mm_mask_fmsub_pd"=>[["AVX512VL", "AVX512F"]],
3755
+ "_mm_maskz_fmsub_pd"=>[["AVX512VL", "AVX512F"]],
3756
+ "_mm256_mask3_fmsub_ps"=>[["AVX512VL", "AVX512F"]],
3757
+ "_mm256_mask_fmsub_ps"=>[["AVX512VL", "AVX512F"]],
3758
+ "_mm256_maskz_fmsub_ps"=>[["AVX512VL", "AVX512F"]],
3759
+ "_mm_mask3_fmsub_ps"=>[["AVX512VL", "AVX512F"]],
3760
+ "_mm_mask_fmsub_ps"=>[["AVX512VL", "AVX512F"]],
3761
+ "_mm_maskz_fmsub_ps"=>[["AVX512VL", "AVX512F"]],
3762
+ "_mm256_mask3_fmsubadd_pd"=>[["AVX512VL", "AVX512F"]],
3763
+ "_mm256_mask_fmsubadd_pd"=>[["AVX512VL", "AVX512F"]],
3764
+ "_mm256_maskz_fmsubadd_pd"=>[["AVX512VL", "AVX512F"]],
3765
+ "_mm_mask3_fmsubadd_pd"=>[["AVX512VL", "AVX512F"]],
3766
+ "_mm_mask_fmsubadd_pd"=>[["AVX512VL", "AVX512F"]],
3767
+ "_mm_maskz_fmsubadd_pd"=>[["AVX512VL", "AVX512F"]],
3768
+ "_mm256_mask3_fmsubadd_ps"=>[["AVX512VL", "AVX512F"]],
3769
+ "_mm256_mask_fmsubadd_ps"=>[["AVX512VL", "AVX512F"]],
3770
+ "_mm256_maskz_fmsubadd_ps"=>[["AVX512VL", "AVX512F"]],
3771
+ "_mm_mask3_fmsubadd_ps"=>[["AVX512VL", "AVX512F"]],
3772
+ "_mm_mask_fmsubadd_ps"=>[["AVX512VL", "AVX512F"]],
3773
+ "_mm_maskz_fmsubadd_ps"=>[["AVX512VL", "AVX512F"]],
3774
+ "_mm256_mask3_fnmadd_pd"=>[["AVX512VL", "AVX512F"]],
3775
+ "_mm256_mask_fnmadd_pd"=>[["AVX512VL", "AVX512F"]],
3776
+ "_mm256_maskz_fnmadd_pd"=>[["AVX512VL", "AVX512F"]],
3777
+ "_mm_mask3_fnmadd_pd"=>[["AVX512VL", "AVX512F"]],
3778
+ "_mm_mask_fnmadd_pd"=>[["AVX512VL", "AVX512F"]],
3779
+ "_mm_maskz_fnmadd_pd"=>[["AVX512VL", "AVX512F"]],
3780
+ "_mm256_mask3_fnmadd_ps"=>[["AVX512VL", "AVX512F"]],
3781
+ "_mm256_mask_fnmadd_ps"=>[["AVX512VL", "AVX512F"]],
3782
+ "_mm256_maskz_fnmadd_ps"=>[["AVX512VL", "AVX512F"]],
3783
+ "_mm_mask3_fnmadd_ps"=>[["AVX512VL", "AVX512F"]],
3784
+ "_mm_mask_fnmadd_ps"=>[["AVX512VL", "AVX512F"]],
3785
+ "_mm_maskz_fnmadd_ps"=>[["AVX512VL", "AVX512F"]],
3786
+ "_mm256_mask3_fnmsub_pd"=>[["AVX512VL", "AVX512F"]],
3787
+ "_mm256_mask_fnmsub_pd"=>[["AVX512VL", "AVX512F"]],
3788
+ "_mm256_maskz_fnmsub_pd"=>[["AVX512VL", "AVX512F"]],
3789
+ "_mm_mask3_fnmsub_pd"=>[["AVX512VL", "AVX512F"]],
3790
+ "_mm_mask_fnmsub_pd"=>[["AVX512VL", "AVX512F"]],
3791
+ "_mm_maskz_fnmsub_pd"=>[["AVX512VL", "AVX512F"]],
3792
+ "_mm256_mask3_fnmsub_ps"=>[["AVX512VL", "AVX512F"]],
3793
+ "_mm256_mask_fnmsub_ps"=>[["AVX512VL", "AVX512F"]],
3794
+ "_mm256_maskz_fnmsub_ps"=>[["AVX512VL", "AVX512F"]],
3795
+ "_mm_mask3_fnmsub_ps"=>[["AVX512VL", "AVX512F"]],
3796
+ "_mm_mask_fnmsub_ps"=>[["AVX512VL", "AVX512F"]],
3797
+ "_mm_maskz_fnmsub_ps"=>[["AVX512VL", "AVX512F"]],
3798
+ "_mm256_fpclass_pd_mask"=>[["AVX512VL", "AVX512DQ"]],
3799
+ "_mm256_mask_fpclass_pd_mask"=>[["AVX512VL", "AVX512DQ"]],
3800
3800
  "_mm512_fpclass_pd_mask"=>["AVX512DQ"],
3801
3801
  "_mm512_mask_fpclass_pd_mask"=>["AVX512DQ"],
3802
- "_mm_fpclass_pd_mask"=>["AVX512VL", "AVX512DQ"],
3803
- "_mm_mask_fpclass_pd_mask"=>["AVX512VL", "AVX512DQ"],
3804
- "_mm256_fpclass_ps_mask"=>["AVX512VL", "AVX512DQ"],
3805
- "_mm256_mask_fpclass_ps_mask"=>["AVX512VL", "AVX512DQ"],
3802
+ "_mm_fpclass_pd_mask"=>[["AVX512VL", "AVX512DQ"]],
3803
+ "_mm_mask_fpclass_pd_mask"=>[["AVX512VL", "AVX512DQ"]],
3804
+ "_mm256_fpclass_ps_mask"=>[["AVX512VL", "AVX512DQ"]],
3805
+ "_mm256_mask_fpclass_ps_mask"=>[["AVX512VL", "AVX512DQ"]],
3806
3806
  "_mm512_fpclass_ps_mask"=>["AVX512DQ"],
3807
3807
  "_mm512_mask_fpclass_ps_mask"=>["AVX512DQ"],
3808
- "_mm_fpclass_ps_mask"=>["AVX512VL", "AVX512DQ"],
3809
- "_mm_mask_fpclass_ps_mask"=>["AVX512VL", "AVX512DQ"],
3808
+ "_mm_fpclass_ps_mask"=>[["AVX512VL", "AVX512DQ"]],
3809
+ "_mm_mask_fpclass_ps_mask"=>[["AVX512VL", "AVX512DQ"]],
3810
3810
  "_mm_fpclass_sd_mask"=>["AVX512DQ"],
3811
3811
  "_mm_mask_fpclass_sd_mask"=>["AVX512DQ"],
3812
3812
  "_mm_fpclass_ss_mask"=>["AVX512DQ"],
3813
3813
  "_mm_mask_fpclass_ss_mask"=>["AVX512DQ"],
3814
- "_mm256_mmask_i32gather_pd"=>["AVX512VL", "AVX512F"],
3815
- "_mm_mmask_i32gather_pd"=>["AVX512VL", "AVX512F"],
3816
- "_mm256_mmask_i32gather_ps"=>["AVX512VL", "AVX512F"],
3817
- "_mm_mmask_i32gather_ps"=>["AVX512VL", "AVX512F"],
3818
- "_mm256_mmask_i64gather_pd"=>["AVX512VL", "AVX512F"],
3819
- "_mm_mmask_i64gather_pd"=>["AVX512VL", "AVX512F"],
3820
- "_mm256_mmask_i64gather_ps"=>["AVX512VL", "AVX512F"],
3821
- "_mm_mmask_i64gather_ps"=>["AVX512VL", "AVX512F"],
3822
- "_mm256_getexp_pd"=>["AVX512VL", "AVX512F"],
3823
- "_mm256_mask_getexp_pd"=>["AVX512VL", "AVX512F"],
3824
- "_mm256_maskz_getexp_pd"=>["AVX512VL", "AVX512F"],
3825
- "_mm_getexp_pd"=>["AVX512VL", "AVX512F"],
3826
- "_mm_mask_getexp_pd"=>["AVX512VL", "AVX512F"],
3827
- "_mm_maskz_getexp_pd"=>["AVX512VL", "AVX512F"],
3828
- "_mm256_getexp_ps"=>["AVX512VL", "AVX512F"],
3829
- "_mm256_mask_getexp_ps"=>["AVX512VL", "AVX512F"],
3830
- "_mm256_maskz_getexp_ps"=>["AVX512VL", "AVX512F"],
3831
- "_mm_getexp_ps"=>["AVX512VL", "AVX512F"],
3832
- "_mm_mask_getexp_ps"=>["AVX512VL", "AVX512F"],
3833
- "_mm_maskz_getexp_ps"=>["AVX512VL", "AVX512F"],
3834
- "_mm256_getmant_pd"=>["AVX512VL", "AVX512F"],
3835
- "_mm256_mask_getmant_pd"=>["AVX512VL", "AVX512F"],
3836
- "_mm256_maskz_getmant_pd"=>["AVX512VL", "AVX512F"],
3837
- "_mm_getmant_pd"=>["AVX512VL", "AVX512F"],
3838
- "_mm_mask_getmant_pd"=>["AVX512VL", "AVX512F"],
3839
- "_mm_maskz_getmant_pd"=>["AVX512VL", "AVX512F"],
3840
- "_mm256_getmant_ps"=>["AVX512VL", "AVX512F"],
3841
- "_mm256_mask_getmant_ps"=>["AVX512VL", "AVX512F"],
3842
- "_mm256_maskz_getmant_ps"=>["AVX512VL", "AVX512F"],
3843
- "_mm_getmant_ps"=>["AVX512VL", "AVX512F"],
3844
- "_mm_mask_getmant_ps"=>["AVX512VL", "AVX512F"],
3845
- "_mm_maskz_getmant_ps"=>["AVX512VL", "AVX512F"],
3846
- "_mm256_insertf32x4"=>["AVX512VL", "AVX512F"],
3847
- "_mm256_mask_insertf32x4"=>["AVX512VL", "AVX512F"],
3848
- "_mm256_maskz_insertf32x4"=>["AVX512VL", "AVX512F"],
3814
+ "_mm256_mmask_i32gather_pd"=>[["AVX512VL", "AVX512F"]],
3815
+ "_mm_mmask_i32gather_pd"=>[["AVX512VL", "AVX512F"]],
3816
+ "_mm256_mmask_i32gather_ps"=>[["AVX512VL", "AVX512F"]],
3817
+ "_mm_mmask_i32gather_ps"=>[["AVX512VL", "AVX512F"]],
3818
+ "_mm256_mmask_i64gather_pd"=>[["AVX512VL", "AVX512F"]],
3819
+ "_mm_mmask_i64gather_pd"=>[["AVX512VL", "AVX512F"]],
3820
+ "_mm256_mmask_i64gather_ps"=>[["AVX512VL", "AVX512F"]],
3821
+ "_mm_mmask_i64gather_ps"=>[["AVX512VL", "AVX512F"]],
3822
+ "_mm256_getexp_pd"=>[["AVX512VL", "AVX512F"]],
3823
+ "_mm256_mask_getexp_pd"=>[["AVX512VL", "AVX512F"]],
3824
+ "_mm256_maskz_getexp_pd"=>[["AVX512VL", "AVX512F"]],
3825
+ "_mm_getexp_pd"=>[["AVX512VL", "AVX512F"]],
3826
+ "_mm_mask_getexp_pd"=>[["AVX512VL", "AVX512F"]],
3827
+ "_mm_maskz_getexp_pd"=>[["AVX512VL", "AVX512F"]],
3828
+ "_mm256_getexp_ps"=>[["AVX512VL", "AVX512F"]],
3829
+ "_mm256_mask_getexp_ps"=>[["AVX512VL", "AVX512F"]],
3830
+ "_mm256_maskz_getexp_ps"=>[["AVX512VL", "AVX512F"]],
3831
+ "_mm_getexp_ps"=>[["AVX512VL", "AVX512F"]],
3832
+ "_mm_mask_getexp_ps"=>[["AVX512VL", "AVX512F"]],
3833
+ "_mm_maskz_getexp_ps"=>[["AVX512VL", "AVX512F"]],
3834
+ "_mm256_getmant_pd"=>[["AVX512VL", "AVX512F"]],
3835
+ "_mm256_mask_getmant_pd"=>[["AVX512VL", "AVX512F"]],
3836
+ "_mm256_maskz_getmant_pd"=>[["AVX512VL", "AVX512F"]],
3837
+ "_mm_getmant_pd"=>[["AVX512VL", "AVX512F"]],
3838
+ "_mm_mask_getmant_pd"=>[["AVX512VL", "AVX512F"]],
3839
+ "_mm_maskz_getmant_pd"=>[["AVX512VL", "AVX512F"]],
3840
+ "_mm256_getmant_ps"=>[["AVX512VL", "AVX512F"]],
3841
+ "_mm256_mask_getmant_ps"=>[["AVX512VL", "AVX512F"]],
3842
+ "_mm256_maskz_getmant_ps"=>[["AVX512VL", "AVX512F"]],
3843
+ "_mm_getmant_ps"=>[["AVX512VL", "AVX512F"]],
3844
+ "_mm_mask_getmant_ps"=>[["AVX512VL", "AVX512F"]],
3845
+ "_mm_maskz_getmant_ps"=>[["AVX512VL", "AVX512F"]],
3846
+ "_mm256_insertf32x4"=>[["AVX512VL", "AVX512F"]],
3847
+ "_mm256_mask_insertf32x4"=>[["AVX512VL", "AVX512F"]],
3848
+ "_mm256_maskz_insertf32x4"=>[["AVX512VL", "AVX512F"]],
3849
3849
  "_mm512_insertf32x8"=>["AVX512DQ"],
3850
3850
  "_mm512_mask_insertf32x8"=>["AVX512DQ"],
3851
3851
  "_mm512_maskz_insertf32x8"=>["AVX512DQ"],
3852
- "_mm256_insertf64x2"=>["AVX512VL", "AVX512DQ"],
3853
- "_mm256_mask_insertf64x2"=>["AVX512VL", "AVX512DQ"],
3854
- "_mm256_maskz_insertf64x2"=>["AVX512VL", "AVX512DQ"],
3852
+ "_mm256_insertf64x2"=>[["AVX512VL", "AVX512DQ"]],
3853
+ "_mm256_mask_insertf64x2"=>[["AVX512VL", "AVX512DQ"]],
3854
+ "_mm256_maskz_insertf64x2"=>[["AVX512VL", "AVX512DQ"]],
3855
3855
  "_mm512_insertf64x2"=>["AVX512DQ"],
3856
3856
  "_mm512_mask_insertf64x2"=>["AVX512DQ"],
3857
3857
  "_mm512_maskz_insertf64x2"=>["AVX512DQ"],
3858
- "_mm256_inserti32x4"=>["AVX512VL", "AVX512F"],
3859
- "_mm256_mask_inserti32x4"=>["AVX512VL", "AVX512F"],
3860
- "_mm256_maskz_inserti32x4"=>["AVX512VL", "AVX512F"],
3858
+ "_mm256_inserti32x4"=>[["AVX512VL", "AVX512F"]],
3859
+ "_mm256_mask_inserti32x4"=>[["AVX512VL", "AVX512F"]],
3860
+ "_mm256_maskz_inserti32x4"=>[["AVX512VL", "AVX512F"]],
3861
3861
  "_mm512_inserti32x8"=>["AVX512DQ"],
3862
3862
  "_mm512_mask_inserti32x8"=>["AVX512DQ"],
3863
3863
  "_mm512_maskz_inserti32x8"=>["AVX512DQ"],
3864
- "_mm256_inserti64x2"=>["AVX512VL", "AVX512DQ"],
3865
- "_mm256_mask_inserti64x2"=>["AVX512VL", "AVX512DQ"],
3866
- "_mm256_maskz_inserti64x2"=>["AVX512VL", "AVX512DQ"],
3864
+ "_mm256_inserti64x2"=>[["AVX512VL", "AVX512DQ"]],
3865
+ "_mm256_mask_inserti64x2"=>[["AVX512VL", "AVX512DQ"]],
3866
+ "_mm256_maskz_inserti64x2"=>[["AVX512VL", "AVX512DQ"]],
3867
3867
  "_mm512_inserti64x2"=>["AVX512DQ"],
3868
3868
  "_mm512_mask_inserti64x2"=>["AVX512DQ"],
3869
3869
  "_mm512_maskz_inserti64x2"=>["AVX512DQ"],
3870
- "_mm256_mask_max_pd"=>["AVX512VL", "AVX512F"],
3871
- "_mm256_maskz_max_pd"=>["AVX512VL", "AVX512F"],
3872
- "_mm_mask_max_pd"=>["AVX512VL", "AVX512F"],
3873
- "_mm_maskz_max_pd"=>["AVX512VL", "AVX512F"],
3874
- "_mm256_mask_max_ps"=>["AVX512VL", "AVX512F"],
3875
- "_mm256_maskz_max_ps"=>["AVX512VL", "AVX512F"],
3876
- "_mm_mask_max_ps"=>["AVX512VL", "AVX512F"],
3877
- "_mm_maskz_max_ps"=>["AVX512VL", "AVX512F"],
3878
- "_mm256_mask_min_pd"=>["AVX512VL", "AVX512F"],
3879
- "_mm256_maskz_min_pd"=>["AVX512VL", "AVX512F"],
3880
- "_mm_mask_min_pd"=>["AVX512VL", "AVX512F"],
3881
- "_mm_maskz_min_pd"=>["AVX512VL", "AVX512F"],
3882
- "_mm256_mask_min_ps"=>["AVX512VL", "AVX512F"],
3883
- "_mm256_maskz_min_ps"=>["AVX512VL", "AVX512F"],
3884
- "_mm_mask_min_ps"=>["AVX512VL", "AVX512F"],
3885
- "_mm_maskz_min_ps"=>["AVX512VL", "AVX512F"],
3886
- "_mm256_mask_load_pd"=>["AVX512VL", "AVX512F"],
3887
- "_mm256_mask_mov_pd"=>["AVX512VL", "AVX512F"],
3888
- "_mm256_mask_store_pd"=>["AVX512VL", "AVX512F"],
3889
- "_mm256_maskz_load_pd"=>["AVX512VL", "AVX512F"],
3890
- "_mm256_maskz_mov_pd"=>["AVX512VL", "AVX512F"],
3891
- "_mm_mask_load_pd"=>["AVX512VL", "AVX512F"],
3892
- "_mm_mask_mov_pd"=>["AVX512VL", "AVX512F"],
3893
- "_mm_mask_store_pd"=>["AVX512VL", "AVX512F"],
3894
- "_mm_maskz_load_pd"=>["AVX512VL", "AVX512F"],
3895
- "_mm_maskz_mov_pd"=>["AVX512VL", "AVX512F"],
3896
- "_mm256_mask_load_ps"=>["AVX512VL", "AVX512F"],
3897
- "_mm256_mask_mov_ps"=>["AVX512VL", "AVX512F"],
3898
- "_mm256_mask_store_ps"=>["AVX512VL", "AVX512F"],
3899
- "_mm256_maskz_load_ps"=>["AVX512VL", "AVX512F"],
3900
- "_mm256_maskz_mov_ps"=>["AVX512VL", "AVX512F"],
3901
- "_mm_mask_load_ps"=>["AVX512VL", "AVX512F"],
3902
- "_mm_mask_mov_ps"=>["AVX512VL", "AVX512F"],
3903
- "_mm_mask_store_ps"=>["AVX512VL", "AVX512F"],
3904
- "_mm_maskz_load_ps"=>["AVX512VL", "AVX512F"],
3905
- "_mm_maskz_mov_ps"=>["AVX512VL", "AVX512F"],
3906
- "_mm256_mask_movedup_pd"=>["AVX512VL", "AVX512F"],
3907
- "_mm256_maskz_movedup_pd"=>["AVX512VL", "AVX512F"],
3908
- "_mm_mask_movedup_pd"=>["AVX512VL", "AVX512F"],
3909
- "_mm_maskz_movedup_pd"=>["AVX512VL", "AVX512F"],
3910
- "_mm256_mask_load_epi32"=>["AVX512VL", "AVX512F"],
3911
- "_mm256_mask_mov_epi32"=>["AVX512VL", "AVX512F"],
3912
- "_mm256_mask_store_epi32"=>["AVX512VL", "AVX512F"],
3913
- "_mm256_maskz_load_epi32"=>["AVX512VL", "AVX512F"],
3914
- "_mm256_maskz_mov_epi32"=>["AVX512VL", "AVX512F"],
3915
- "_mm_mask_load_epi32"=>["AVX512VL", "AVX512F"],
3916
- "_mm_mask_mov_epi32"=>["AVX512VL", "AVX512F"],
3917
- "_mm_mask_store_epi32"=>["AVX512VL", "AVX512F"],
3918
- "_mm_maskz_load_epi32"=>["AVX512VL", "AVX512F"],
3919
- "_mm_maskz_mov_epi32"=>["AVX512VL", "AVX512F"],
3920
- "_mm256_mask_load_epi64"=>["AVX512VL", "AVX512F"],
3921
- "_mm256_mask_mov_epi64"=>["AVX512VL", "AVX512F"],
3922
- "_mm256_mask_store_epi64"=>["AVX512VL", "AVX512F"],
3923
- "_mm256_maskz_load_epi64"=>["AVX512VL", "AVX512F"],
3924
- "_mm256_maskz_mov_epi64"=>["AVX512VL", "AVX512F"],
3925
- "_mm_mask_load_epi64"=>["AVX512VL", "AVX512F"],
3926
- "_mm_mask_mov_epi64"=>["AVX512VL", "AVX512F"],
3927
- "_mm_mask_store_epi64"=>["AVX512VL", "AVX512F"],
3928
- "_mm_maskz_load_epi64"=>["AVX512VL", "AVX512F"],
3929
- "_mm_maskz_mov_epi64"=>["AVX512VL", "AVX512F"],
3930
- "_mm256_mask_loadu_epi16"=>["AVX512VL", "AVX512BW"],
3931
- "_mm256_mask_mov_epi16"=>["AVX512VL", "AVX512BW"],
3932
- "_mm256_mask_storeu_epi16"=>["AVX512VL", "AVX512BW"],
3933
- "_mm256_maskz_loadu_epi16"=>["AVX512VL", "AVX512BW"],
3934
- "_mm256_maskz_mov_epi16"=>["AVX512VL", "AVX512BW"],
3870
+ "_mm256_mask_max_pd"=>[["AVX512VL", "AVX512F"]],
3871
+ "_mm256_maskz_max_pd"=>[["AVX512VL", "AVX512F"]],
3872
+ "_mm_mask_max_pd"=>[["AVX512VL", "AVX512F"]],
3873
+ "_mm_maskz_max_pd"=>[["AVX512VL", "AVX512F"]],
3874
+ "_mm256_mask_max_ps"=>[["AVX512VL", "AVX512F"]],
3875
+ "_mm256_maskz_max_ps"=>[["AVX512VL", "AVX512F"]],
3876
+ "_mm_mask_max_ps"=>[["AVX512VL", "AVX512F"]],
3877
+ "_mm_maskz_max_ps"=>[["AVX512VL", "AVX512F"]],
3878
+ "_mm256_mask_min_pd"=>[["AVX512VL", "AVX512F"]],
3879
+ "_mm256_maskz_min_pd"=>[["AVX512VL", "AVX512F"]],
3880
+ "_mm_mask_min_pd"=>[["AVX512VL", "AVX512F"]],
3881
+ "_mm_maskz_min_pd"=>[["AVX512VL", "AVX512F"]],
3882
+ "_mm256_mask_min_ps"=>[["AVX512VL", "AVX512F"]],
3883
+ "_mm256_maskz_min_ps"=>[["AVX512VL", "AVX512F"]],
3884
+ "_mm_mask_min_ps"=>[["AVX512VL", "AVX512F"]],
3885
+ "_mm_maskz_min_ps"=>[["AVX512VL", "AVX512F"]],
3886
+ "_mm256_mask_load_pd"=>[["AVX512VL", "AVX512F"]],
3887
+ "_mm256_mask_mov_pd"=>[["AVX512VL", "AVX512F"]],
3888
+ "_mm256_mask_store_pd"=>[["AVX512VL", "AVX512F"]],
3889
+ "_mm256_maskz_load_pd"=>[["AVX512VL", "AVX512F"]],
3890
+ "_mm256_maskz_mov_pd"=>[["AVX512VL", "AVX512F"]],
3891
+ "_mm_mask_load_pd"=>[["AVX512VL", "AVX512F"]],
3892
+ "_mm_mask_mov_pd"=>[["AVX512VL", "AVX512F"]],
3893
+ "_mm_mask_store_pd"=>[["AVX512VL", "AVX512F"]],
3894
+ "_mm_maskz_load_pd"=>[["AVX512VL", "AVX512F"]],
3895
+ "_mm_maskz_mov_pd"=>[["AVX512VL", "AVX512F"]],
3896
+ "_mm256_mask_load_ps"=>[["AVX512VL", "AVX512F"]],
3897
+ "_mm256_mask_mov_ps"=>[["AVX512VL", "AVX512F"]],
3898
+ "_mm256_mask_store_ps"=>[["AVX512VL", "AVX512F"]],
3899
+ "_mm256_maskz_load_ps"=>[["AVX512VL", "AVX512F"]],
3900
+ "_mm256_maskz_mov_ps"=>[["AVX512VL", "AVX512F"]],
3901
+ "_mm_mask_load_ps"=>[["AVX512VL", "AVX512F"]],
3902
+ "_mm_mask_mov_ps"=>[["AVX512VL", "AVX512F"]],
3903
+ "_mm_mask_store_ps"=>[["AVX512VL", "AVX512F"]],
3904
+ "_mm_maskz_load_ps"=>[["AVX512VL", "AVX512F"]],
3905
+ "_mm_maskz_mov_ps"=>[["AVX512VL", "AVX512F"]],
3906
+ "_mm256_mask_movedup_pd"=>[["AVX512VL", "AVX512F"]],
3907
+ "_mm256_maskz_movedup_pd"=>[["AVX512VL", "AVX512F"]],
3908
+ "_mm_mask_movedup_pd"=>[["AVX512VL", "AVX512F"]],
3909
+ "_mm_maskz_movedup_pd"=>[["AVX512VL", "AVX512F"]],
3910
+ "_mm256_mask_load_epi32"=>[["AVX512VL", "AVX512F"]],
3911
+ "_mm256_mask_mov_epi32"=>[["AVX512VL", "AVX512F"]],
3912
+ "_mm256_mask_store_epi32"=>[["AVX512VL", "AVX512F"]],
3913
+ "_mm256_maskz_load_epi32"=>[["AVX512VL", "AVX512F"]],
3914
+ "_mm256_maskz_mov_epi32"=>[["AVX512VL", "AVX512F"]],
3915
+ "_mm_mask_load_epi32"=>[["AVX512VL", "AVX512F"]],
3916
+ "_mm_mask_mov_epi32"=>[["AVX512VL", "AVX512F"]],
3917
+ "_mm_mask_store_epi32"=>[["AVX512VL", "AVX512F"]],
3918
+ "_mm_maskz_load_epi32"=>[["AVX512VL", "AVX512F"]],
3919
+ "_mm_maskz_mov_epi32"=>[["AVX512VL", "AVX512F"]],
3920
+ "_mm256_mask_load_epi64"=>[["AVX512VL", "AVX512F"]],
3921
+ "_mm256_mask_mov_epi64"=>[["AVX512VL", "AVX512F"]],
3922
+ "_mm256_mask_store_epi64"=>[["AVX512VL", "AVX512F"]],
3923
+ "_mm256_maskz_load_epi64"=>[["AVX512VL", "AVX512F"]],
3924
+ "_mm256_maskz_mov_epi64"=>[["AVX512VL", "AVX512F"]],
3925
+ "_mm_mask_load_epi64"=>[["AVX512VL", "AVX512F"]],
3926
+ "_mm_mask_mov_epi64"=>[["AVX512VL", "AVX512F"]],
3927
+ "_mm_mask_store_epi64"=>[["AVX512VL", "AVX512F"]],
3928
+ "_mm_maskz_load_epi64"=>[["AVX512VL", "AVX512F"]],
3929
+ "_mm_maskz_mov_epi64"=>[["AVX512VL", "AVX512F"]],
3930
+ "_mm256_mask_loadu_epi16"=>[["AVX512VL", "AVX512BW"]],
3931
+ "_mm256_mask_mov_epi16"=>[["AVX512VL", "AVX512BW"]],
3932
+ "_mm256_mask_storeu_epi16"=>[["AVX512VL", "AVX512BW"]],
3933
+ "_mm256_maskz_loadu_epi16"=>[["AVX512VL", "AVX512BW"]],
3934
+ "_mm256_maskz_mov_epi16"=>[["AVX512VL", "AVX512BW"]],
3935
3935
  "_mm512_mask_loadu_epi16"=>["AVX512BW"],
3936
3936
  "_mm512_mask_mov_epi16"=>["AVX512BW"],
3937
3937
  "_mm512_mask_storeu_epi16"=>["AVX512BW"],
3938
3938
  "_mm512_maskz_loadu_epi16"=>["AVX512BW"],
3939
3939
  "_mm512_maskz_mov_epi16"=>["AVX512BW"],
3940
- "_mm_mask_loadu_epi16"=>["AVX512VL", "AVX512BW"],
3941
- "_mm_mask_mov_epi16"=>["AVX512VL", "AVX512BW"],
3942
- "_mm_mask_storeu_epi16"=>["AVX512VL", "AVX512BW"],
3943
- "_mm_maskz_loadu_epi16"=>["AVX512VL", "AVX512BW"],
3944
- "_mm_maskz_mov_epi16"=>["AVX512VL", "AVX512BW"],
3945
- "_mm256_mask_loadu_epi32"=>["AVX512VL", "AVX512F"],
3946
- "_mm256_mask_storeu_epi32"=>["AVX512VL", "AVX512F"],
3947
- "_mm256_maskz_loadu_epi32"=>["AVX512VL", "AVX512F"],
3948
- "_mm_mask_loadu_epi32"=>["AVX512VL", "AVX512F"],
3949
- "_mm_mask_storeu_epi32"=>["AVX512VL", "AVX512F"],
3950
- "_mm_maskz_loadu_epi32"=>["AVX512VL", "AVX512F"],
3951
- "_mm256_mask_loadu_epi64"=>["AVX512VL", "AVX512F"],
3952
- "_mm256_mask_storeu_epi64"=>["AVX512VL", "AVX512F"],
3953
- "_mm256_maskz_loadu_epi64"=>["AVX512VL", "AVX512F"],
3954
- "_mm_mask_loadu_epi64"=>["AVX512VL", "AVX512F"],
3955
- "_mm_mask_storeu_epi64"=>["AVX512VL", "AVX512F"],
3956
- "_mm_maskz_loadu_epi64"=>["AVX512VL", "AVX512F"],
3957
- "_mm256_mask_loadu_epi8"=>["AVX512VL", "AVX512BW"],
3958
- "_mm256_mask_mov_epi8"=>["AVX512VL", "AVX512BW"],
3959
- "_mm256_mask_storeu_epi8"=>["AVX512VL", "AVX512BW"],
3960
- "_mm256_maskz_loadu_epi8"=>["AVX512VL", "AVX512BW"],
3961
- "_mm256_maskz_mov_epi8"=>["AVX512VL", "AVX512BW"],
3940
+ "_mm_mask_loadu_epi16"=>[["AVX512VL", "AVX512BW"]],
3941
+ "_mm_mask_mov_epi16"=>[["AVX512VL", "AVX512BW"]],
3942
+ "_mm_mask_storeu_epi16"=>[["AVX512VL", "AVX512BW"]],
3943
+ "_mm_maskz_loadu_epi16"=>[["AVX512VL", "AVX512BW"]],
3944
+ "_mm_maskz_mov_epi16"=>[["AVX512VL", "AVX512BW"]],
3945
+ "_mm256_mask_loadu_epi32"=>[["AVX512VL", "AVX512F"]],
3946
+ "_mm256_mask_storeu_epi32"=>[["AVX512VL", "AVX512F"]],
3947
+ "_mm256_maskz_loadu_epi32"=>[["AVX512VL", "AVX512F"]],
3948
+ "_mm_mask_loadu_epi32"=>[["AVX512VL", "AVX512F"]],
3949
+ "_mm_mask_storeu_epi32"=>[["AVX512VL", "AVX512F"]],
3950
+ "_mm_maskz_loadu_epi32"=>[["AVX512VL", "AVX512F"]],
3951
+ "_mm256_mask_loadu_epi64"=>[["AVX512VL", "AVX512F"]],
3952
+ "_mm256_mask_storeu_epi64"=>[["AVX512VL", "AVX512F"]],
3953
+ "_mm256_maskz_loadu_epi64"=>[["AVX512VL", "AVX512F"]],
3954
+ "_mm_mask_loadu_epi64"=>[["AVX512VL", "AVX512F"]],
3955
+ "_mm_mask_storeu_epi64"=>[["AVX512VL", "AVX512F"]],
3956
+ "_mm_maskz_loadu_epi64"=>[["AVX512VL", "AVX512F"]],
3957
+ "_mm256_mask_loadu_epi8"=>[["AVX512VL", "AVX512BW"]],
3958
+ "_mm256_mask_mov_epi8"=>[["AVX512VL", "AVX512BW"]],
3959
+ "_mm256_mask_storeu_epi8"=>[["AVX512VL", "AVX512BW"]],
3960
+ "_mm256_maskz_loadu_epi8"=>[["AVX512VL", "AVX512BW"]],
3961
+ "_mm256_maskz_mov_epi8"=>[["AVX512VL", "AVX512BW"]],
3962
3962
  "_mm512_mask_loadu_epi8"=>["AVX512BW"],
3963
3963
  "_mm512_mask_mov_epi8"=>["AVX512BW"],
3964
3964
  "_mm512_mask_storeu_epi8"=>["AVX512BW"],
3965
3965
  "_mm512_maskz_loadu_epi8"=>["AVX512BW"],
3966
3966
  "_mm512_maskz_mov_epi8"=>["AVX512BW"],
3967
- "_mm_mask_loadu_epi8"=>["AVX512VL", "AVX512BW"],
3968
- "_mm_mask_mov_epi8"=>["AVX512VL", "AVX512BW"],
3969
- "_mm_mask_storeu_epi8"=>["AVX512VL", "AVX512BW"],
3970
- "_mm_maskz_loadu_epi8"=>["AVX512VL", "AVX512BW"],
3971
- "_mm_maskz_mov_epi8"=>["AVX512VL", "AVX512BW"],
3972
- "_mm256_mask_movehdup_ps"=>["AVX512VL", "AVX512F"],
3973
- "_mm256_maskz_movehdup_ps"=>["AVX512VL", "AVX512F"],
3974
- "_mm_mask_movehdup_ps"=>["AVX512VL", "AVX512F"],
3975
- "_mm_maskz_movehdup_ps"=>["AVX512VL", "AVX512F"],
3976
- "_mm256_mask_moveldup_ps"=>["AVX512VL", "AVX512F"],
3977
- "_mm256_maskz_moveldup_ps"=>["AVX512VL", "AVX512F"],
3978
- "_mm_mask_moveldup_ps"=>["AVX512VL", "AVX512F"],
3979
- "_mm_maskz_moveldup_ps"=>["AVX512VL", "AVX512F"],
3980
- "_mm256_mask_loadu_pd"=>["AVX512VL", "AVX512F"],
3981
- "_mm256_mask_storeu_pd"=>["AVX512VL", "AVX512F"],
3982
- "_mm256_maskz_loadu_pd"=>["AVX512VL", "AVX512F"],
3983
- "_mm_mask_loadu_pd"=>["AVX512VL", "AVX512F"],
3984
- "_mm_mask_storeu_pd"=>["AVX512VL", "AVX512F"],
3985
- "_mm_maskz_loadu_pd"=>["AVX512VL", "AVX512F"],
3986
- "_mm256_mask_loadu_ps"=>["AVX512VL", "AVX512F"],
3987
- "_mm256_mask_storeu_ps"=>["AVX512VL", "AVX512F"],
3988
- "_mm256_maskz_loadu_ps"=>["AVX512VL", "AVX512F"],
3989
- "_mm_mask_loadu_ps"=>["AVX512VL", "AVX512F"],
3990
- "_mm_mask_storeu_ps"=>["AVX512VL", "AVX512F"],
3991
- "_mm_maskz_loadu_ps"=>["AVX512VL", "AVX512F"],
3992
- "_mm256_mask_mul_pd"=>["AVX512VL", "AVX512F"],
3993
- "_mm256_maskz_mul_pd"=>["AVX512VL", "AVX512F"],
3994
- "_mm_mask_mul_pd"=>["AVX512VL", "AVX512F"],
3995
- "_mm_maskz_mul_pd"=>["AVX512VL", "AVX512F"],
3996
- "_mm256_mask_mul_ps"=>["AVX512VL", "AVX512F"],
3997
- "_mm256_maskz_mul_ps"=>["AVX512VL", "AVX512F"],
3998
- "_mm_mask_mul_ps"=>["AVX512VL", "AVX512F"],
3999
- "_mm_maskz_mul_ps"=>["AVX512VL", "AVX512F"],
4000
- "_mm256_mask_or_pd"=>["AVX512VL", "AVX512DQ"],
4001
- "_mm256_maskz_or_pd"=>["AVX512VL", "AVX512DQ"],
3967
+ "_mm_mask_loadu_epi8"=>[["AVX512VL", "AVX512BW"]],
3968
+ "_mm_mask_mov_epi8"=>[["AVX512VL", "AVX512BW"]],
3969
+ "_mm_mask_storeu_epi8"=>[["AVX512VL", "AVX512BW"]],
3970
+ "_mm_maskz_loadu_epi8"=>[["AVX512VL", "AVX512BW"]],
3971
+ "_mm_maskz_mov_epi8"=>[["AVX512VL", "AVX512BW"]],
3972
+ "_mm256_mask_movehdup_ps"=>[["AVX512VL", "AVX512F"]],
3973
+ "_mm256_maskz_movehdup_ps"=>[["AVX512VL", "AVX512F"]],
3974
+ "_mm_mask_movehdup_ps"=>[["AVX512VL", "AVX512F"]],
3975
+ "_mm_maskz_movehdup_ps"=>[["AVX512VL", "AVX512F"]],
3976
+ "_mm256_mask_moveldup_ps"=>[["AVX512VL", "AVX512F"]],
3977
+ "_mm256_maskz_moveldup_ps"=>[["AVX512VL", "AVX512F"]],
3978
+ "_mm_mask_moveldup_ps"=>[["AVX512VL", "AVX512F"]],
3979
+ "_mm_maskz_moveldup_ps"=>[["AVX512VL", "AVX512F"]],
3980
+ "_mm256_mask_loadu_pd"=>[["AVX512VL", "AVX512F"]],
3981
+ "_mm256_mask_storeu_pd"=>[["AVX512VL", "AVX512F"]],
3982
+ "_mm256_maskz_loadu_pd"=>[["AVX512VL", "AVX512F"]],
3983
+ "_mm_mask_loadu_pd"=>[["AVX512VL", "AVX512F"]],
3984
+ "_mm_mask_storeu_pd"=>[["AVX512VL", "AVX512F"]],
3985
+ "_mm_maskz_loadu_pd"=>[["AVX512VL", "AVX512F"]],
3986
+ "_mm256_mask_loadu_ps"=>[["AVX512VL", "AVX512F"]],
3987
+ "_mm256_mask_storeu_ps"=>[["AVX512VL", "AVX512F"]],
3988
+ "_mm256_maskz_loadu_ps"=>[["AVX512VL", "AVX512F"]],
3989
+ "_mm_mask_loadu_ps"=>[["AVX512VL", "AVX512F"]],
3990
+ "_mm_mask_storeu_ps"=>[["AVX512VL", "AVX512F"]],
3991
+ "_mm_maskz_loadu_ps"=>[["AVX512VL", "AVX512F"]],
3992
+ "_mm256_mask_mul_pd"=>[["AVX512VL", "AVX512F"]],
3993
+ "_mm256_maskz_mul_pd"=>[["AVX512VL", "AVX512F"]],
3994
+ "_mm_mask_mul_pd"=>[["AVX512VL", "AVX512F"]],
3995
+ "_mm_maskz_mul_pd"=>[["AVX512VL", "AVX512F"]],
3996
+ "_mm256_mask_mul_ps"=>[["AVX512VL", "AVX512F"]],
3997
+ "_mm256_maskz_mul_ps"=>[["AVX512VL", "AVX512F"]],
3998
+ "_mm_mask_mul_ps"=>[["AVX512VL", "AVX512F"]],
3999
+ "_mm_maskz_mul_ps"=>[["AVX512VL", "AVX512F"]],
4000
+ "_mm256_mask_or_pd"=>[["AVX512VL", "AVX512DQ"]],
4001
+ "_mm256_maskz_or_pd"=>[["AVX512VL", "AVX512DQ"]],
4002
4002
  "_mm512_mask_or_pd"=>["AVX512DQ"],
4003
4003
  "_mm512_maskz_or_pd"=>["AVX512DQ"],
4004
4004
  "_mm512_or_pd"=>["AVX512DQ"],
4005
- "_mm_mask_or_pd"=>["AVX512VL", "AVX512DQ"],
4006
- "_mm_maskz_or_pd"=>["AVX512VL", "AVX512DQ"],
4007
- "_mm256_mask_or_ps"=>["AVX512VL", "AVX512DQ"],
4008
- "_mm256_maskz_or_ps"=>["AVX512VL", "AVX512DQ"],
4005
+ "_mm_mask_or_pd"=>[["AVX512VL", "AVX512DQ"]],
4006
+ "_mm_maskz_or_pd"=>[["AVX512VL", "AVX512DQ"]],
4007
+ "_mm256_mask_or_ps"=>[["AVX512VL", "AVX512DQ"]],
4008
+ "_mm256_maskz_or_ps"=>[["AVX512VL", "AVX512DQ"]],
4009
4009
  "_mm512_mask_or_ps"=>["AVX512DQ"],
4010
4010
  "_mm512_maskz_or_ps"=>["AVX512DQ"],
4011
4011
  "_mm512_or_ps"=>["AVX512DQ"],
4012
- "_mm_mask_or_ps"=>["AVX512VL", "AVX512DQ"],
4013
- "_mm_maskz_or_ps"=>["AVX512VL", "AVX512DQ"],
4014
- "_mm256_mask_abs_epi8"=>["AVX512VL", "AVX512BW"],
4015
- "_mm256_maskz_abs_epi8"=>["AVX512VL", "AVX512BW"],
4012
+ "_mm_mask_or_ps"=>[["AVX512VL", "AVX512DQ"]],
4013
+ "_mm_maskz_or_ps"=>[["AVX512VL", "AVX512DQ"]],
4014
+ "_mm256_mask_abs_epi8"=>[["AVX512VL", "AVX512BW"]],
4015
+ "_mm256_maskz_abs_epi8"=>[["AVX512VL", "AVX512BW"]],
4016
4016
  "_mm512_abs_epi8"=>["AVX512BW"],
4017
4017
  "_mm512_mask_abs_epi8"=>["AVX512BW"],
4018
4018
  "_mm512_maskz_abs_epi8"=>["AVX512BW"],
4019
- "_mm_mask_abs_epi8"=>["AVX512VL", "AVX512BW"],
4020
- "_mm_maskz_abs_epi8"=>["AVX512VL", "AVX512BW"],
4021
- "_mm256_mask_abs_epi32"=>["AVX512VL", "AVX512F"],
4022
- "_mm256_maskz_abs_epi32"=>["AVX512VL", "AVX512F"],
4023
- "_mm_mask_abs_epi32"=>["AVX512VL", "AVX512F"],
4024
- "_mm_maskz_abs_epi32"=>["AVX512VL", "AVX512F"],
4025
- "_mm256_abs_epi64"=>["AVX512VL", "AVX512F"],
4026
- "_mm256_mask_abs_epi64"=>["AVX512VL", "AVX512F"],
4027
- "_mm256_maskz_abs_epi64"=>["AVX512VL", "AVX512F"],
4028
- "_mm_abs_epi64"=>["AVX512VL", "AVX512F"],
4029
- "_mm_mask_abs_epi64"=>["AVX512VL", "AVX512F"],
4030
- "_mm_maskz_abs_epi64"=>["AVX512VL", "AVX512F"],
4031
- "_mm256_mask_abs_epi16"=>["AVX512VL", "AVX512BW"],
4032
- "_mm256_maskz_abs_epi16"=>["AVX512VL", "AVX512BW"],
4019
+ "_mm_mask_abs_epi8"=>[["AVX512VL", "AVX512BW"]],
4020
+ "_mm_maskz_abs_epi8"=>[["AVX512VL", "AVX512BW"]],
4021
+ "_mm256_mask_abs_epi32"=>[["AVX512VL", "AVX512F"]],
4022
+ "_mm256_maskz_abs_epi32"=>[["AVX512VL", "AVX512F"]],
4023
+ "_mm_mask_abs_epi32"=>[["AVX512VL", "AVX512F"]],
4024
+ "_mm_maskz_abs_epi32"=>[["AVX512VL", "AVX512F"]],
4025
+ "_mm256_abs_epi64"=>[["AVX512VL", "AVX512F"]],
4026
+ "_mm256_mask_abs_epi64"=>[["AVX512VL", "AVX512F"]],
4027
+ "_mm256_maskz_abs_epi64"=>[["AVX512VL", "AVX512F"]],
4028
+ "_mm_abs_epi64"=>[["AVX512VL", "AVX512F"]],
4029
+ "_mm_mask_abs_epi64"=>[["AVX512VL", "AVX512F"]],
4030
+ "_mm_maskz_abs_epi64"=>[["AVX512VL", "AVX512F"]],
4031
+ "_mm256_mask_abs_epi16"=>[["AVX512VL", "AVX512BW"]],
4032
+ "_mm256_maskz_abs_epi16"=>[["AVX512VL", "AVX512BW"]],
4033
4033
  "_mm512_abs_epi16"=>["AVX512BW"],
4034
4034
  "_mm512_mask_abs_epi16"=>["AVX512BW"],
4035
4035
  "_mm512_maskz_abs_epi16"=>["AVX512BW"],
4036
- "_mm_mask_abs_epi16"=>["AVX512VL", "AVX512BW"],
4037
- "_mm_maskz_abs_epi16"=>["AVX512VL", "AVX512BW"],
4038
- "_mm256_mask_packs_epi32"=>["AVX512VL", "AVX512BW"],
4039
- "_mm256_maskz_packs_epi32"=>["AVX512VL", "AVX512BW"],
4036
+ "_mm_mask_abs_epi16"=>[["AVX512VL", "AVX512BW"]],
4037
+ "_mm_maskz_abs_epi16"=>[["AVX512VL", "AVX512BW"]],
4038
+ "_mm256_mask_packs_epi32"=>[["AVX512VL", "AVX512BW"]],
4039
+ "_mm256_maskz_packs_epi32"=>[["AVX512VL", "AVX512BW"]],
4040
4040
  "_mm512_mask_packs_epi32"=>["AVX512BW"],
4041
4041
  "_mm512_maskz_packs_epi32"=>["AVX512BW"],
4042
4042
  "_mm512_packs_epi32"=>["AVX512BW"],
4043
- "_mm_mask_packs_epi32"=>["AVX512VL", "AVX512BW"],
4044
- "_mm_maskz_packs_epi32"=>["AVX512VL", "AVX512BW"],
4045
- "_mm256_mask_packs_epi16"=>["AVX512VL", "AVX512BW"],
4046
- "_mm256_maskz_packs_epi16"=>["AVX512VL", "AVX512BW"],
4043
+ "_mm_mask_packs_epi32"=>[["AVX512VL", "AVX512BW"]],
4044
+ "_mm_maskz_packs_epi32"=>[["AVX512VL", "AVX512BW"]],
4045
+ "_mm256_mask_packs_epi16"=>[["AVX512VL", "AVX512BW"]],
4046
+ "_mm256_maskz_packs_epi16"=>[["AVX512VL", "AVX512BW"]],
4047
4047
  "_mm512_mask_packs_epi16"=>["AVX512BW"],
4048
4048
  "_mm512_maskz_packs_epi16"=>["AVX512BW"],
4049
4049
  "_mm512_packs_epi16"=>["AVX512BW"],
4050
- "_mm_mask_packs_epi16"=>["AVX512VL", "AVX512BW"],
4051
- "_mm_maskz_packs_epi16"=>["AVX512VL", "AVX512BW"],
4052
- "_mm256_mask_packus_epi32"=>["AVX512VL", "AVX512BW"],
4053
- "_mm256_maskz_packus_epi32"=>["AVX512VL", "AVX512BW"],
4050
+ "_mm_mask_packs_epi16"=>[["AVX512VL", "AVX512BW"]],
4051
+ "_mm_maskz_packs_epi16"=>[["AVX512VL", "AVX512BW"]],
4052
+ "_mm256_mask_packus_epi32"=>[["AVX512VL", "AVX512BW"]],
4053
+ "_mm256_maskz_packus_epi32"=>[["AVX512VL", "AVX512BW"]],
4054
4054
  "_mm512_mask_packus_epi32"=>["AVX512BW"],
4055
4055
  "_mm512_maskz_packus_epi32"=>["AVX512BW"],
4056
4056
  "_mm512_packus_epi32"=>["AVX512BW"],
4057
- "_mm_mask_packus_epi32"=>["AVX512VL", "AVX512BW"],
4058
- "_mm_maskz_packus_epi32"=>["AVX512VL", "AVX512BW"],
4059
- "_mm256_mask_packus_epi16"=>["AVX512VL", "AVX512BW"],
4060
- "_mm256_maskz_packus_epi16"=>["AVX512VL", "AVX512BW"],
4057
+ "_mm_mask_packus_epi32"=>[["AVX512VL", "AVX512BW"]],
4058
+ "_mm_maskz_packus_epi32"=>[["AVX512VL", "AVX512BW"]],
4059
+ "_mm256_mask_packus_epi16"=>[["AVX512VL", "AVX512BW"]],
4060
+ "_mm256_maskz_packus_epi16"=>[["AVX512VL", "AVX512BW"]],
4061
4061
  "_mm512_mask_packus_epi16"=>["AVX512BW"],
4062
4062
  "_mm512_maskz_packus_epi16"=>["AVX512BW"],
4063
4063
  "_mm512_packus_epi16"=>["AVX512BW"],
4064
- "_mm_mask_packus_epi16"=>["AVX512VL", "AVX512BW"],
4065
- "_mm_maskz_packus_epi16"=>["AVX512VL", "AVX512BW"],
4066
- "_mm256_mask_add_epi8"=>["AVX512VL", "AVX512BW"],
4067
- "_mm256_maskz_add_epi8"=>["AVX512VL", "AVX512BW"],
4064
+ "_mm_mask_packus_epi16"=>[["AVX512VL", "AVX512BW"]],
4065
+ "_mm_maskz_packus_epi16"=>[["AVX512VL", "AVX512BW"]],
4066
+ "_mm256_mask_add_epi8"=>[["AVX512VL", "AVX512BW"]],
4067
+ "_mm256_maskz_add_epi8"=>[["AVX512VL", "AVX512BW"]],
4068
4068
  "_mm512_add_epi8"=>["AVX512BW"],
4069
4069
  "_mm512_mask_add_epi8"=>["AVX512BW"],
4070
4070
  "_mm512_maskz_add_epi8"=>["AVX512BW"],
4071
- "_mm_mask_add_epi8"=>["AVX512VL", "AVX512BW"],
4072
- "_mm_maskz_add_epi8"=>["AVX512VL", "AVX512BW"],
4073
- "_mm256_mask_add_epi32"=>["AVX512VL", "AVX512F"],
4074
- "_mm256_maskz_add_epi32"=>["AVX512VL", "AVX512F"],
4075
- "_mm_mask_add_epi32"=>["AVX512VL", "AVX512F"],
4076
- "_mm_maskz_add_epi32"=>["AVX512VL", "AVX512F"],
4077
- "_mm256_mask_add_epi64"=>["AVX512VL", "AVX512F"],
4078
- "_mm256_maskz_add_epi64"=>["AVX512VL", "AVX512F"],
4079
- "_mm_mask_add_epi64"=>["AVX512VL", "AVX512F"],
4080
- "_mm_maskz_add_epi64"=>["AVX512VL", "AVX512F"],
4081
- "_mm256_mask_adds_epi8"=>["AVX512VL", "AVX512BW"],
4082
- "_mm256_maskz_adds_epi8"=>["AVX512VL", "AVX512BW"],
4071
+ "_mm_mask_add_epi8"=>[["AVX512VL", "AVX512BW"]],
4072
+ "_mm_maskz_add_epi8"=>[["AVX512VL", "AVX512BW"]],
4073
+ "_mm256_mask_add_epi32"=>[["AVX512VL", "AVX512F"]],
4074
+ "_mm256_maskz_add_epi32"=>[["AVX512VL", "AVX512F"]],
4075
+ "_mm_mask_add_epi32"=>[["AVX512VL", "AVX512F"]],
4076
+ "_mm_maskz_add_epi32"=>[["AVX512VL", "AVX512F"]],
4077
+ "_mm256_mask_add_epi64"=>[["AVX512VL", "AVX512F"]],
4078
+ "_mm256_maskz_add_epi64"=>[["AVX512VL", "AVX512F"]],
4079
+ "_mm_mask_add_epi64"=>[["AVX512VL", "AVX512F"]],
4080
+ "_mm_maskz_add_epi64"=>[["AVX512VL", "AVX512F"]],
4081
+ "_mm256_mask_adds_epi8"=>[["AVX512VL", "AVX512BW"]],
4082
+ "_mm256_maskz_adds_epi8"=>[["AVX512VL", "AVX512BW"]],
4083
4083
  "_mm512_adds_epi8"=>["AVX512BW"],
4084
4084
  "_mm512_mask_adds_epi8"=>["AVX512BW"],
4085
4085
  "_mm512_maskz_adds_epi8"=>["AVX512BW"],
4086
- "_mm_mask_adds_epi8"=>["AVX512VL", "AVX512BW"],
4087
- "_mm_maskz_adds_epi8"=>["AVX512VL", "AVX512BW"],
4088
- "_mm256_mask_adds_epi16"=>["AVX512VL", "AVX512BW"],
4089
- "_mm256_maskz_adds_epi16"=>["AVX512VL", "AVX512BW"],
4086
+ "_mm_mask_adds_epi8"=>[["AVX512VL", "AVX512BW"]],
4087
+ "_mm_maskz_adds_epi8"=>[["AVX512VL", "AVX512BW"]],
4088
+ "_mm256_mask_adds_epi16"=>[["AVX512VL", "AVX512BW"]],
4089
+ "_mm256_maskz_adds_epi16"=>[["AVX512VL", "AVX512BW"]],
4090
4090
  "_mm512_adds_epi16"=>["AVX512BW"],
4091
4091
  "_mm512_mask_adds_epi16"=>["AVX512BW"],
4092
4092
  "_mm512_maskz_adds_epi16"=>["AVX512BW"],
4093
- "_mm_mask_adds_epi16"=>["AVX512VL", "AVX512BW"],
4094
- "_mm_maskz_adds_epi16"=>["AVX512VL", "AVX512BW"],
4095
- "_mm256_mask_adds_epu8"=>["AVX512VL", "AVX512BW"],
4096
- "_mm256_maskz_adds_epu8"=>["AVX512VL", "AVX512BW"],
4093
+ "_mm_mask_adds_epi16"=>[["AVX512VL", "AVX512BW"]],
4094
+ "_mm_maskz_adds_epi16"=>[["AVX512VL", "AVX512BW"]],
4095
+ "_mm256_mask_adds_epu8"=>[["AVX512VL", "AVX512BW"]],
4096
+ "_mm256_maskz_adds_epu8"=>[["AVX512VL", "AVX512BW"]],
4097
4097
  "_mm512_adds_epu8"=>["AVX512BW"],
4098
4098
  "_mm512_mask_adds_epu8"=>["AVX512BW"],
4099
4099
  "_mm512_maskz_adds_epu8"=>["AVX512BW"],
4100
- "_mm_mask_adds_epu8"=>["AVX512VL", "AVX512BW"],
4101
- "_mm_maskz_adds_epu8"=>["AVX512VL", "AVX512BW"],
4102
- "_mm256_mask_adds_epu16"=>["AVX512VL", "AVX512BW"],
4103
- "_mm256_maskz_adds_epu16"=>["AVX512VL", "AVX512BW"],
4100
+ "_mm_mask_adds_epu8"=>[["AVX512VL", "AVX512BW"]],
4101
+ "_mm_maskz_adds_epu8"=>[["AVX512VL", "AVX512BW"]],
4102
+ "_mm256_mask_adds_epu16"=>[["AVX512VL", "AVX512BW"]],
4103
+ "_mm256_maskz_adds_epu16"=>[["AVX512VL", "AVX512BW"]],
4104
4104
  "_mm512_adds_epu16"=>["AVX512BW"],
4105
4105
  "_mm512_mask_adds_epu16"=>["AVX512BW"],
4106
4106
  "_mm512_maskz_adds_epu16"=>["AVX512BW"],
4107
- "_mm_mask_adds_epu16"=>["AVX512VL", "AVX512BW"],
4108
- "_mm_maskz_adds_epu16"=>["AVX512VL", "AVX512BW"],
4109
- "_mm256_mask_add_epi16"=>["AVX512VL", "AVX512BW"],
4110
- "_mm256_maskz_add_epi16"=>["AVX512VL", "AVX512BW"],
4107
+ "_mm_mask_adds_epu16"=>[["AVX512VL", "AVX512BW"]],
4108
+ "_mm_maskz_adds_epu16"=>[["AVX512VL", "AVX512BW"]],
4109
+ "_mm256_mask_add_epi16"=>[["AVX512VL", "AVX512BW"]],
4110
+ "_mm256_maskz_add_epi16"=>[["AVX512VL", "AVX512BW"]],
4111
4111
  "_mm512_add_epi16"=>["AVX512BW"],
4112
4112
  "_mm512_mask_add_epi16"=>["AVX512BW"],
4113
4113
  "_mm512_maskz_add_epi16"=>["AVX512BW"],
4114
- "_mm_mask_add_epi16"=>["AVX512VL", "AVX512BW"],
4115
- "_mm_maskz_add_epi16"=>["AVX512VL", "AVX512BW"],
4116
- "_mm256_mask_alignr_epi8"=>["AVX512VL", "AVX512BW"],
4117
- "_mm256_maskz_alignr_epi8"=>["AVX512VL", "AVX512BW"],
4114
+ "_mm_mask_add_epi16"=>[["AVX512VL", "AVX512BW"]],
4115
+ "_mm_maskz_add_epi16"=>[["AVX512VL", "AVX512BW"]],
4116
+ "_mm256_mask_alignr_epi8"=>[["AVX512VL", "AVX512BW"]],
4117
+ "_mm256_maskz_alignr_epi8"=>[["AVX512VL", "AVX512BW"]],
4118
4118
  "_mm512_alignr_epi8"=>["AVX512BW"],
4119
4119
  "_mm512_mask_alignr_epi8"=>["AVX512BW"],
4120
4120
  "_mm512_maskz_alignr_epi8"=>["AVX512BW"],
4121
- "_mm_mask_alignr_epi8"=>["AVX512VL", "AVX512BW"],
4122
- "_mm_maskz_alignr_epi8"=>["AVX512VL", "AVX512BW"],
4123
- "_mm256_mask_and_epi32"=>["AVX512VL", "AVX512F"],
4124
- "_mm256_maskz_and_epi32"=>["AVX512VL", "AVX512F"],
4125
- "_mm_mask_and_epi32"=>["AVX512VL", "AVX512F"],
4126
- "_mm_maskz_and_epi32"=>["AVX512VL", "AVX512F"],
4127
- "_mm256_mask_andnot_epi32"=>["AVX512VL", "AVX512F"],
4128
- "_mm256_maskz_andnot_epi32"=>["AVX512VL", "AVX512F"],
4129
- "_mm_mask_andnot_epi32"=>["AVX512VL", "AVX512F"],
4130
- "_mm_maskz_andnot_epi32"=>["AVX512VL", "AVX512F"],
4131
- "_mm256_mask_andnot_epi64"=>["AVX512VL", "AVX512F"],
4132
- "_mm256_maskz_andnot_epi64"=>["AVX512VL", "AVX512F"],
4133
- "_mm_mask_andnot_epi64"=>["AVX512VL", "AVX512F"],
4134
- "_mm_maskz_andnot_epi64"=>["AVX512VL", "AVX512F"],
4135
- "_mm256_mask_and_epi64"=>["AVX512VL", "AVX512F"],
4136
- "_mm256_maskz_and_epi64"=>["AVX512VL", "AVX512F"],
4137
- "_mm_mask_and_epi64"=>["AVX512VL", "AVX512F"],
4138
- "_mm_maskz_and_epi64"=>["AVX512VL", "AVX512F"],
4139
- "_mm256_mask_avg_epu8"=>["AVX512VL", "AVX512BW"],
4140
- "_mm256_maskz_avg_epu8"=>["AVX512VL", "AVX512BW"],
4121
+ "_mm_mask_alignr_epi8"=>[["AVX512VL", "AVX512BW"]],
4122
+ "_mm_maskz_alignr_epi8"=>[["AVX512VL", "AVX512BW"]],
4123
+ "_mm256_mask_and_epi32"=>[["AVX512VL", "AVX512F"]],
4124
+ "_mm256_maskz_and_epi32"=>[["AVX512VL", "AVX512F"]],
4125
+ "_mm_mask_and_epi32"=>[["AVX512VL", "AVX512F"]],
4126
+ "_mm_maskz_and_epi32"=>[["AVX512VL", "AVX512F"]],
4127
+ "_mm256_mask_andnot_epi32"=>[["AVX512VL", "AVX512F"]],
4128
+ "_mm256_maskz_andnot_epi32"=>[["AVX512VL", "AVX512F"]],
4129
+ "_mm_mask_andnot_epi32"=>[["AVX512VL", "AVX512F"]],
4130
+ "_mm_maskz_andnot_epi32"=>[["AVX512VL", "AVX512F"]],
4131
+ "_mm256_mask_andnot_epi64"=>[["AVX512VL", "AVX512F"]],
4132
+ "_mm256_maskz_andnot_epi64"=>[["AVX512VL", "AVX512F"]],
4133
+ "_mm_mask_andnot_epi64"=>[["AVX512VL", "AVX512F"]],
4134
+ "_mm_maskz_andnot_epi64"=>[["AVX512VL", "AVX512F"]],
4135
+ "_mm256_mask_and_epi64"=>[["AVX512VL", "AVX512F"]],
4136
+ "_mm256_maskz_and_epi64"=>[["AVX512VL", "AVX512F"]],
4137
+ "_mm_mask_and_epi64"=>[["AVX512VL", "AVX512F"]],
4138
+ "_mm_maskz_and_epi64"=>[["AVX512VL", "AVX512F"]],
4139
+ "_mm256_mask_avg_epu8"=>[["AVX512VL", "AVX512BW"]],
4140
+ "_mm256_maskz_avg_epu8"=>[["AVX512VL", "AVX512BW"]],
4141
4141
  "_mm512_avg_epu8"=>["AVX512BW"],
4142
4142
  "_mm512_mask_avg_epu8"=>["AVX512BW"],
4143
4143
  "_mm512_maskz_avg_epu8"=>["AVX512BW"],
4144
- "_mm_mask_avg_epu8"=>["AVX512VL", "AVX512BW"],
4145
- "_mm_maskz_avg_epu8"=>["AVX512VL", "AVX512BW"],
4146
- "_mm256_mask_avg_epu16"=>["AVX512VL", "AVX512BW"],
4147
- "_mm256_maskz_avg_epu16"=>["AVX512VL", "AVX512BW"],
4144
+ "_mm_mask_avg_epu8"=>[["AVX512VL", "AVX512BW"]],
4145
+ "_mm_maskz_avg_epu8"=>[["AVX512VL", "AVX512BW"]],
4146
+ "_mm256_mask_avg_epu16"=>[["AVX512VL", "AVX512BW"]],
4147
+ "_mm256_maskz_avg_epu16"=>[["AVX512VL", "AVX512BW"]],
4148
4148
  "_mm512_avg_epu16"=>["AVX512BW"],
4149
4149
  "_mm512_mask_avg_epu16"=>["AVX512BW"],
4150
4150
  "_mm512_maskz_avg_epu16"=>["AVX512BW"],
4151
- "_mm_mask_avg_epu16"=>["AVX512VL", "AVX512BW"],
4152
- "_mm_maskz_avg_epu16"=>["AVX512VL", "AVX512BW"],
4153
- "_mm256_mask_blend_epi8"=>["AVX512VL", "AVX512BW"],
4151
+ "_mm_mask_avg_epu16"=>[["AVX512VL", "AVX512BW"]],
4152
+ "_mm_maskz_avg_epu16"=>[["AVX512VL", "AVX512BW"]],
4153
+ "_mm256_mask_blend_epi8"=>[["AVX512VL", "AVX512BW"]],
4154
4154
  "_mm512_mask_blend_epi8"=>["AVX512BW"],
4155
- "_mm_mask_blend_epi8"=>["AVX512VL", "AVX512BW"],
4156
- "_mm256_mask_blend_epi32"=>["AVX512VL", "AVX512F"],
4157
- "_mm_mask_blend_epi32"=>["AVX512VL", "AVX512F"],
4158
- "_mm256_mask_blend_epi64"=>["AVX512VL", "AVX512F"],
4159
- "_mm_mask_blend_epi64"=>["AVX512VL", "AVX512F"],
4160
- "_mm256_mask_blend_epi16"=>["AVX512VL", "AVX512BW"],
4155
+ "_mm_mask_blend_epi8"=>[["AVX512VL", "AVX512BW"]],
4156
+ "_mm256_mask_blend_epi32"=>[["AVX512VL", "AVX512F"]],
4157
+ "_mm_mask_blend_epi32"=>[["AVX512VL", "AVX512F"]],
4158
+ "_mm256_mask_blend_epi64"=>[["AVX512VL", "AVX512F"]],
4159
+ "_mm_mask_blend_epi64"=>[["AVX512VL", "AVX512F"]],
4160
+ "_mm256_mask_blend_epi16"=>[["AVX512VL", "AVX512BW"]],
4161
4161
  "_mm512_mask_blend_epi16"=>["AVX512BW"],
4162
- "_mm_mask_blend_epi16"=>["AVX512VL", "AVX512BW"],
4163
- "_mm256_mask_broadcastb_epi8"=>["AVX512VL", "AVX512BW"],
4164
- "_mm256_mask_set1_epi8"=>["AVX512VL", "AVX512BW"],
4165
- "_mm256_maskz_broadcastb_epi8"=>["AVX512VL", "AVX512BW"],
4166
- "_mm256_maskz_set1_epi8"=>["AVX512VL", "AVX512BW"],
4162
+ "_mm_mask_blend_epi16"=>[["AVX512VL", "AVX512BW"]],
4163
+ "_mm256_mask_broadcastb_epi8"=>[["AVX512VL", "AVX512BW"]],
4164
+ "_mm256_mask_set1_epi8"=>[["AVX512VL", "AVX512BW"]],
4165
+ "_mm256_maskz_broadcastb_epi8"=>[["AVX512VL", "AVX512BW"]],
4166
+ "_mm256_maskz_set1_epi8"=>[["AVX512VL", "AVX512BW"]],
4167
4167
  "_mm512_broadcastb_epi8"=>["AVX512BW"],
4168
4168
  "_mm512_mask_broadcastb_epi8"=>["AVX512BW"],
4169
4169
  "_mm512_mask_set1_epi8"=>["AVX512BW"],
4170
4170
  "_mm512_maskz_broadcastb_epi8"=>["AVX512BW"],
4171
4171
  "_mm512_maskz_set1_epi8"=>["AVX512BW"],
4172
- "_mm_mask_broadcastb_epi8"=>["AVX512VL", "AVX512BW"],
4173
- "_mm_mask_set1_epi8"=>["AVX512VL", "AVX512BW"],
4174
- "_mm_maskz_broadcastb_epi8"=>["AVX512VL", "AVX512BW"],
4175
- "_mm_maskz_set1_epi8"=>["AVX512VL", "AVX512BW"],
4176
- "_mm256_mask_broadcastd_epi32"=>["AVX512VL", "AVX512F"],
4177
- "_mm256_mask_set1_epi32"=>["AVX512VL", "AVX512F"],
4178
- "_mm256_maskz_broadcastd_epi32"=>["AVX512VL", "AVX512F"],
4179
- "_mm256_maskz_set1_epi32"=>["AVX512VL", "AVX512F"],
4180
- "_mm_mask_broadcastd_epi32"=>["AVX512VL", "AVX512F"],
4181
- "_mm_mask_set1_epi32"=>["AVX512VL", "AVX512F"],
4182
- "_mm_maskz_broadcastd_epi32"=>["AVX512VL", "AVX512F"],
4183
- "_mm_maskz_set1_epi32"=>["AVX512VL", "AVX512F"],
4184
- "_mm256_broadcastmb_epi64"=>["AVX512VL", "AVX512CD"],
4185
- "_mm_broadcastmb_epi64"=>["AVX512VL", "AVX512CD"],
4186
- "_mm256_broadcastmw_epi32"=>["AVX512VL", "AVX512CD"],
4187
- "_mm_broadcastmw_epi32"=>["AVX512VL", "AVX512CD"],
4188
- "_mm256_mask_broadcastq_epi64"=>["AVX512VL", "AVX512F"],
4189
- "_mm256_mask_set1_epi64"=>["AVX512VL", "AVX512F"],
4190
- "_mm256_maskz_broadcastq_epi64"=>["AVX512VL", "AVX512F"],
4191
- "_mm256_maskz_set1_epi64"=>["AVX512VL", "AVX512F"],
4192
- "_mm_mask_broadcastq_epi64"=>["AVX512VL", "AVX512F"],
4193
- "_mm_mask_set1_epi64"=>["AVX512VL", "AVX512F"],
4194
- "_mm_maskz_broadcastq_epi64"=>["AVX512VL", "AVX512F"],
4195
- "_mm_maskz_set1_epi64"=>["AVX512VL", "AVX512F"],
4196
- "_mm256_mask_broadcastw_epi16"=>["AVX512VL", "AVX512BW"],
4197
- "_mm256_mask_set1_epi16"=>["AVX512VL", "AVX512BW"],
4198
- "_mm256_maskz_broadcastw_epi16"=>["AVX512VL", "AVX512BW"],
4199
- "_mm256_maskz_set1_epi16"=>["AVX512VL", "AVX512BW"],
4172
+ "_mm_mask_broadcastb_epi8"=>[["AVX512VL", "AVX512BW"]],
4173
+ "_mm_mask_set1_epi8"=>[["AVX512VL", "AVX512BW"]],
4174
+ "_mm_maskz_broadcastb_epi8"=>[["AVX512VL", "AVX512BW"]],
4175
+ "_mm_maskz_set1_epi8"=>[["AVX512VL", "AVX512BW"]],
4176
+ "_mm256_mask_broadcastd_epi32"=>[["AVX512VL", "AVX512F"]],
4177
+ "_mm256_mask_set1_epi32"=>[["AVX512VL", "AVX512F"]],
4178
+ "_mm256_maskz_broadcastd_epi32"=>[["AVX512VL", "AVX512F"]],
4179
+ "_mm256_maskz_set1_epi32"=>[["AVX512VL", "AVX512F"]],
4180
+ "_mm_mask_broadcastd_epi32"=>[["AVX512VL", "AVX512F"]],
4181
+ "_mm_mask_set1_epi32"=>[["AVX512VL", "AVX512F"]],
4182
+ "_mm_maskz_broadcastd_epi32"=>[["AVX512VL", "AVX512F"]],
4183
+ "_mm_maskz_set1_epi32"=>[["AVX512VL", "AVX512F"]],
4184
+ "_mm256_broadcastmb_epi64"=>[["AVX512VL", "AVX512CD"]],
4185
+ "_mm_broadcastmb_epi64"=>[["AVX512VL", "AVX512CD"]],
4186
+ "_mm256_broadcastmw_epi32"=>[["AVX512VL", "AVX512CD"]],
4187
+ "_mm_broadcastmw_epi32"=>[["AVX512VL", "AVX512CD"]],
4188
+ "_mm256_mask_broadcastq_epi64"=>[["AVX512VL", "AVX512F"]],
4189
+ "_mm256_mask_set1_epi64"=>[["AVX512VL", "AVX512F"]],
4190
+ "_mm256_maskz_broadcastq_epi64"=>[["AVX512VL", "AVX512F"]],
4191
+ "_mm256_maskz_set1_epi64"=>[["AVX512VL", "AVX512F"]],
4192
+ "_mm_mask_broadcastq_epi64"=>[["AVX512VL", "AVX512F"]],
4193
+ "_mm_mask_set1_epi64"=>[["AVX512VL", "AVX512F"]],
4194
+ "_mm_maskz_broadcastq_epi64"=>[["AVX512VL", "AVX512F"]],
4195
+ "_mm_maskz_set1_epi64"=>[["AVX512VL", "AVX512F"]],
4196
+ "_mm256_mask_broadcastw_epi16"=>[["AVX512VL", "AVX512BW"]],
4197
+ "_mm256_mask_set1_epi16"=>[["AVX512VL", "AVX512BW"]],
4198
+ "_mm256_maskz_broadcastw_epi16"=>[["AVX512VL", "AVX512BW"]],
4199
+ "_mm256_maskz_set1_epi16"=>[["AVX512VL", "AVX512BW"]],
4200
4200
  "_mm512_broadcastw_epi16"=>["AVX512BW"],
4201
4201
  "_mm512_mask_broadcastw_epi16"=>["AVX512BW"],
4202
4202
  "_mm512_mask_set1_epi16"=>["AVX512BW"],
4203
4203
  "_mm512_maskz_broadcastw_epi16"=>["AVX512BW"],
4204
4204
  "_mm512_maskz_set1_epi16"=>["AVX512BW"],
4205
- "_mm_mask_broadcastw_epi16"=>["AVX512VL", "AVX512BW"],
4206
- "_mm_mask_set1_epi16"=>["AVX512VL", "AVX512BW"],
4207
- "_mm_maskz_broadcastw_epi16"=>["AVX512VL", "AVX512BW"],
4208
- "_mm_maskz_set1_epi16"=>["AVX512VL", "AVX512BW"],
4209
- "_mm256_cmp_epi8_mask"=>["AVX512VL", "AVX512BW"],
4210
- "_mm256_cmpeq_epi8_mask"=>["AVX512VL", "AVX512BW"],
4211
- "_mm256_cmpge_epi8_mask"=>["AVX512VL", "AVX512BW"],
4212
- "_mm256_cmpgt_epi8_mask"=>["AVX512VL", "AVX512BW"],
4213
- "_mm256_cmple_epi8_mask"=>["AVX512VL", "AVX512BW"],
4214
- "_mm256_cmplt_epi8_mask"=>["AVX512VL", "AVX512BW"],
4215
- "_mm256_cmpneq_epi8_mask"=>["AVX512VL", "AVX512BW"],
4216
- "_mm256_mask_cmp_epi8_mask"=>["AVX512VL", "AVX512BW"],
4217
- "_mm256_mask_cmpeq_epi8_mask"=>["AVX512VL", "AVX512BW"],
4218
- "_mm256_mask_cmpge_epi8_mask"=>["AVX512VL", "AVX512BW"],
4219
- "_mm256_mask_cmpgt_epi8_mask"=>["AVX512VL", "AVX512BW"],
4220
- "_mm256_mask_cmple_epi8_mask"=>["AVX512VL", "AVX512BW"],
4221
- "_mm256_mask_cmplt_epi8_mask"=>["AVX512VL", "AVX512BW"],
4222
- "_mm256_mask_cmpneq_epi8_mask"=>["AVX512VL", "AVX512BW"],
4205
+ "_mm_mask_broadcastw_epi16"=>[["AVX512VL", "AVX512BW"]],
4206
+ "_mm_mask_set1_epi16"=>[["AVX512VL", "AVX512BW"]],
4207
+ "_mm_maskz_broadcastw_epi16"=>[["AVX512VL", "AVX512BW"]],
4208
+ "_mm_maskz_set1_epi16"=>[["AVX512VL", "AVX512BW"]],
4209
+ "_mm256_cmp_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4210
+ "_mm256_cmpeq_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4211
+ "_mm256_cmpge_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4212
+ "_mm256_cmpgt_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4213
+ "_mm256_cmple_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4214
+ "_mm256_cmplt_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4215
+ "_mm256_cmpneq_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4216
+ "_mm256_mask_cmp_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4217
+ "_mm256_mask_cmpeq_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4218
+ "_mm256_mask_cmpge_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4219
+ "_mm256_mask_cmpgt_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4220
+ "_mm256_mask_cmple_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4221
+ "_mm256_mask_cmplt_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4222
+ "_mm256_mask_cmpneq_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4223
4223
  "_mm512_cmp_epi8_mask"=>["AVX512BW"],
4224
4224
  "_mm512_cmpeq_epi8_mask"=>["AVX512BW"],
4225
4225
  "_mm512_cmpge_epi8_mask"=>["AVX512BW"],
@@ -4234,90 +4234,90 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
4234
4234
  "_mm512_mask_cmple_epi8_mask"=>["AVX512BW"],
4235
4235
  "_mm512_mask_cmplt_epi8_mask"=>["AVX512BW"],
4236
4236
  "_mm512_mask_cmpneq_epi8_mask"=>["AVX512BW"],
4237
- "_mm_cmp_epi8_mask"=>["AVX512VL", "AVX512BW"],
4238
- "_mm_cmpeq_epi8_mask"=>["AVX512VL", "AVX512BW"],
4239
- "_mm_cmpge_epi8_mask"=>["AVX512VL", "AVX512BW"],
4240
- "_mm_cmpgt_epi8_mask"=>["AVX512VL", "AVX512BW"],
4241
- "_mm_cmple_epi8_mask"=>["AVX512VL", "AVX512BW"],
4242
- "_mm_cmplt_epi8_mask"=>["AVX512VL", "AVX512BW"],
4243
- "_mm_cmpneq_epi8_mask"=>["AVX512VL", "AVX512BW"],
4244
- "_mm_mask_cmp_epi8_mask"=>["AVX512VL", "AVX512BW"],
4245
- "_mm_mask_cmpeq_epi8_mask"=>["AVX512VL", "AVX512BW"],
4246
- "_mm_mask_cmpge_epi8_mask"=>["AVX512VL", "AVX512BW"],
4247
- "_mm_mask_cmpgt_epi8_mask"=>["AVX512VL", "AVX512BW"],
4248
- "_mm_mask_cmple_epi8_mask"=>["AVX512VL", "AVX512BW"],
4249
- "_mm_mask_cmplt_epi8_mask"=>["AVX512VL", "AVX512BW"],
4250
- "_mm_mask_cmpneq_epi8_mask"=>["AVX512VL", "AVX512BW"],
4251
- "_mm256_cmp_epi32_mask"=>["AVX512VL", "AVX512F"],
4252
- "_mm256_cmpeq_epi32_mask"=>["AVX512VL", "AVX512F"],
4253
- "_mm256_cmpge_epi32_mask"=>["AVX512VL", "AVX512F"],
4254
- "_mm256_cmpgt_epi32_mask"=>["AVX512VL", "AVX512F"],
4255
- "_mm256_cmple_epi32_mask"=>["AVX512VL", "AVX512F"],
4256
- "_mm256_cmplt_epi32_mask"=>["AVX512VL", "AVX512F"],
4257
- "_mm256_cmpneq_epi32_mask"=>["AVX512VL", "AVX512F"],
4258
- "_mm256_mask_cmp_epi32_mask"=>["AVX512VL", "AVX512F"],
4259
- "_mm256_mask_cmpeq_epi32_mask"=>["AVX512VL", "AVX512F"],
4260
- "_mm256_mask_cmpge_epi32_mask"=>["AVX512VL", "AVX512F"],
4261
- "_mm256_mask_cmpgt_epi32_mask"=>["AVX512VL", "AVX512F"],
4262
- "_mm256_mask_cmple_epi32_mask"=>["AVX512VL", "AVX512F"],
4263
- "_mm256_mask_cmplt_epi32_mask"=>["AVX512VL", "AVX512F"],
4264
- "_mm256_mask_cmpneq_epi32_mask"=>["AVX512VL", "AVX512F"],
4265
- "_mm_cmp_epi32_mask"=>["AVX512VL", "AVX512F"],
4266
- "_mm_cmpeq_epi32_mask"=>["AVX512VL", "AVX512F"],
4267
- "_mm_cmpge_epi32_mask"=>["AVX512VL", "AVX512F"],
4268
- "_mm_cmpgt_epi32_mask"=>["AVX512VL", "AVX512F"],
4269
- "_mm_cmple_epi32_mask"=>["AVX512VL", "AVX512F"],
4270
- "_mm_cmplt_epi32_mask"=>["AVX512VL", "AVX512F"],
4271
- "_mm_cmpneq_epi32_mask"=>["AVX512VL", "AVX512F"],
4272
- "_mm_mask_cmp_epi32_mask"=>["AVX512VL", "AVX512F"],
4273
- "_mm_mask_cmpeq_epi32_mask"=>["AVX512VL", "AVX512F"],
4274
- "_mm_mask_cmpge_epi32_mask"=>["AVX512VL", "AVX512F"],
4275
- "_mm_mask_cmpgt_epi32_mask"=>["AVX512VL", "AVX512F"],
4276
- "_mm_mask_cmple_epi32_mask"=>["AVX512VL", "AVX512F"],
4277
- "_mm_mask_cmplt_epi32_mask"=>["AVX512VL", "AVX512F"],
4278
- "_mm_mask_cmpneq_epi32_mask"=>["AVX512VL", "AVX512F"],
4279
- "_mm256_cmp_epi64_mask"=>["AVX512VL", "AVX512F"],
4280
- "_mm256_cmpeq_epi64_mask"=>["AVX512VL", "AVX512F"],
4281
- "_mm256_cmpge_epi64_mask"=>["AVX512VL", "AVX512F"],
4282
- "_mm256_cmpgt_epi64_mask"=>["AVX512VL", "AVX512F"],
4283
- "_mm256_cmple_epi64_mask"=>["AVX512VL", "AVX512F"],
4284
- "_mm256_cmplt_epi64_mask"=>["AVX512VL", "AVX512F"],
4285
- "_mm256_cmpneq_epi64_mask"=>["AVX512VL", "AVX512F"],
4286
- "_mm256_mask_cmp_epi64_mask"=>["AVX512VL", "AVX512F"],
4287
- "_mm256_mask_cmpeq_epi64_mask"=>["AVX512VL", "AVX512F"],
4288
- "_mm256_mask_cmpge_epi64_mask"=>["AVX512VL", "AVX512F"],
4289
- "_mm256_mask_cmpgt_epi64_mask"=>["AVX512VL", "AVX512F"],
4290
- "_mm256_mask_cmple_epi64_mask"=>["AVX512VL", "AVX512F"],
4291
- "_mm256_mask_cmplt_epi64_mask"=>["AVX512VL", "AVX512F"],
4292
- "_mm256_mask_cmpneq_epi64_mask"=>["AVX512VL", "AVX512F"],
4293
- "_mm_cmp_epi64_mask"=>["AVX512VL", "AVX512F"],
4294
- "_mm_cmpeq_epi64_mask"=>["AVX512VL", "AVX512F"],
4295
- "_mm_cmpge_epi64_mask"=>["AVX512VL", "AVX512F"],
4296
- "_mm_cmpgt_epi64_mask"=>["AVX512VL", "AVX512F"],
4297
- "_mm_cmple_epi64_mask"=>["AVX512VL", "AVX512F"],
4298
- "_mm_cmplt_epi64_mask"=>["AVX512VL", "AVX512F"],
4299
- "_mm_cmpneq_epi64_mask"=>["AVX512VL", "AVX512F"],
4300
- "_mm_mask_cmp_epi64_mask"=>["AVX512VL", "AVX512F"],
4301
- "_mm_mask_cmpeq_epi64_mask"=>["AVX512VL", "AVX512F"],
4302
- "_mm_mask_cmpge_epi64_mask"=>["AVX512VL", "AVX512F"],
4303
- "_mm_mask_cmpgt_epi64_mask"=>["AVX512VL", "AVX512F"],
4304
- "_mm_mask_cmple_epi64_mask"=>["AVX512VL", "AVX512F"],
4305
- "_mm_mask_cmplt_epi64_mask"=>["AVX512VL", "AVX512F"],
4306
- "_mm_mask_cmpneq_epi64_mask"=>["AVX512VL", "AVX512F"],
4307
- "_mm256_cmp_epu8_mask"=>["AVX512VL", "AVX512BW"],
4308
- "_mm256_cmpeq_epu8_mask"=>["AVX512VL", "AVX512BW"],
4309
- "_mm256_cmpge_epu8_mask"=>["AVX512VL", "AVX512BW"],
4310
- "_mm256_cmpgt_epu8_mask"=>["AVX512VL", "AVX512BW"],
4311
- "_mm256_cmple_epu8_mask"=>["AVX512VL", "AVX512BW"],
4312
- "_mm256_cmplt_epu8_mask"=>["AVX512VL", "AVX512BW"],
4313
- "_mm256_cmpneq_epu8_mask"=>["AVX512VL", "AVX512BW"],
4314
- "_mm256_mask_cmp_epu8_mask"=>["AVX512VL", "AVX512BW"],
4315
- "_mm256_mask_cmpeq_epu8_mask"=>["AVX512VL", "AVX512BW"],
4316
- "_mm256_mask_cmpge_epu8_mask"=>["AVX512VL", "AVX512BW"],
4317
- "_mm256_mask_cmpgt_epu8_mask"=>["AVX512VL", "AVX512BW"],
4318
- "_mm256_mask_cmple_epu8_mask"=>["AVX512VL", "AVX512BW"],
4319
- "_mm256_mask_cmplt_epu8_mask"=>["AVX512VL", "AVX512BW"],
4320
- "_mm256_mask_cmpneq_epu8_mask"=>["AVX512VL", "AVX512BW"],
4237
+ "_mm_cmp_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4238
+ "_mm_cmpeq_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4239
+ "_mm_cmpge_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4240
+ "_mm_cmpgt_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4241
+ "_mm_cmple_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4242
+ "_mm_cmplt_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4243
+ "_mm_cmpneq_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4244
+ "_mm_mask_cmp_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4245
+ "_mm_mask_cmpeq_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4246
+ "_mm_mask_cmpge_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4247
+ "_mm_mask_cmpgt_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4248
+ "_mm_mask_cmple_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4249
+ "_mm_mask_cmplt_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4250
+ "_mm_mask_cmpneq_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
4251
+ "_mm256_cmp_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4252
+ "_mm256_cmpeq_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4253
+ "_mm256_cmpge_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4254
+ "_mm256_cmpgt_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4255
+ "_mm256_cmple_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4256
+ "_mm256_cmplt_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4257
+ "_mm256_cmpneq_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4258
+ "_mm256_mask_cmp_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4259
+ "_mm256_mask_cmpeq_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4260
+ "_mm256_mask_cmpge_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4261
+ "_mm256_mask_cmpgt_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4262
+ "_mm256_mask_cmple_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4263
+ "_mm256_mask_cmplt_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4264
+ "_mm256_mask_cmpneq_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4265
+ "_mm_cmp_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4266
+ "_mm_cmpeq_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4267
+ "_mm_cmpge_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4268
+ "_mm_cmpgt_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4269
+ "_mm_cmple_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4270
+ "_mm_cmplt_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4271
+ "_mm_cmpneq_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4272
+ "_mm_mask_cmp_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4273
+ "_mm_mask_cmpeq_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4274
+ "_mm_mask_cmpge_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4275
+ "_mm_mask_cmpgt_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4276
+ "_mm_mask_cmple_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4277
+ "_mm_mask_cmplt_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4278
+ "_mm_mask_cmpneq_epi32_mask"=>[["AVX512VL", "AVX512F"]],
4279
+ "_mm256_cmp_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4280
+ "_mm256_cmpeq_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4281
+ "_mm256_cmpge_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4282
+ "_mm256_cmpgt_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4283
+ "_mm256_cmple_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4284
+ "_mm256_cmplt_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4285
+ "_mm256_cmpneq_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4286
+ "_mm256_mask_cmp_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4287
+ "_mm256_mask_cmpeq_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4288
+ "_mm256_mask_cmpge_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4289
+ "_mm256_mask_cmpgt_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4290
+ "_mm256_mask_cmple_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4291
+ "_mm256_mask_cmplt_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4292
+ "_mm256_mask_cmpneq_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4293
+ "_mm_cmp_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4294
+ "_mm_cmpeq_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4295
+ "_mm_cmpge_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4296
+ "_mm_cmpgt_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4297
+ "_mm_cmple_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4298
+ "_mm_cmplt_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4299
+ "_mm_cmpneq_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4300
+ "_mm_mask_cmp_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4301
+ "_mm_mask_cmpeq_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4302
+ "_mm_mask_cmpge_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4303
+ "_mm_mask_cmpgt_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4304
+ "_mm_mask_cmple_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4305
+ "_mm_mask_cmplt_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4306
+ "_mm_mask_cmpneq_epi64_mask"=>[["AVX512VL", "AVX512F"]],
4307
+ "_mm256_cmp_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4308
+ "_mm256_cmpeq_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4309
+ "_mm256_cmpge_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4310
+ "_mm256_cmpgt_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4311
+ "_mm256_cmple_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4312
+ "_mm256_cmplt_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4313
+ "_mm256_cmpneq_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4314
+ "_mm256_mask_cmp_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4315
+ "_mm256_mask_cmpeq_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4316
+ "_mm256_mask_cmpge_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4317
+ "_mm256_mask_cmpgt_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4318
+ "_mm256_mask_cmple_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4319
+ "_mm256_mask_cmplt_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4320
+ "_mm256_mask_cmpneq_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4321
4321
  "_mm512_cmp_epu8_mask"=>["AVX512BW"],
4322
4322
  "_mm512_cmpeq_epu8_mask"=>["AVX512BW"],
4323
4323
  "_mm512_cmpge_epu8_mask"=>["AVX512BW"],
@@ -4332,90 +4332,90 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
4332
4332
  "_mm512_mask_cmple_epu8_mask"=>["AVX512BW"],
4333
4333
  "_mm512_mask_cmplt_epu8_mask"=>["AVX512BW"],
4334
4334
  "_mm512_mask_cmpneq_epu8_mask"=>["AVX512BW"],
4335
- "_mm_cmp_epu8_mask"=>["AVX512VL", "AVX512BW"],
4336
- "_mm_cmpeq_epu8_mask"=>["AVX512VL", "AVX512BW"],
4337
- "_mm_cmpge_epu8_mask"=>["AVX512VL", "AVX512BW"],
4338
- "_mm_cmpgt_epu8_mask"=>["AVX512VL", "AVX512BW"],
4339
- "_mm_cmple_epu8_mask"=>["AVX512VL", "AVX512BW"],
4340
- "_mm_cmplt_epu8_mask"=>["AVX512VL", "AVX512BW"],
4341
- "_mm_cmpneq_epu8_mask"=>["AVX512VL", "AVX512BW"],
4342
- "_mm_mask_cmp_epu8_mask"=>["AVX512VL", "AVX512BW"],
4343
- "_mm_mask_cmpeq_epu8_mask"=>["AVX512VL", "AVX512BW"],
4344
- "_mm_mask_cmpge_epu8_mask"=>["AVX512VL", "AVX512BW"],
4345
- "_mm_mask_cmpgt_epu8_mask"=>["AVX512VL", "AVX512BW"],
4346
- "_mm_mask_cmple_epu8_mask"=>["AVX512VL", "AVX512BW"],
4347
- "_mm_mask_cmplt_epu8_mask"=>["AVX512VL", "AVX512BW"],
4348
- "_mm_mask_cmpneq_epu8_mask"=>["AVX512VL", "AVX512BW"],
4349
- "_mm256_cmp_epu32_mask"=>["AVX512VL", "AVX512F"],
4350
- "_mm256_cmpeq_epu32_mask"=>["AVX512VL", "AVX512F"],
4351
- "_mm256_cmpge_epu32_mask"=>["AVX512VL", "AVX512F"],
4352
- "_mm256_cmpgt_epu32_mask"=>["AVX512VL", "AVX512F"],
4353
- "_mm256_cmple_epu32_mask"=>["AVX512VL", "AVX512F"],
4354
- "_mm256_cmplt_epu32_mask"=>["AVX512VL", "AVX512F"],
4355
- "_mm256_cmpneq_epu32_mask"=>["AVX512VL", "AVX512F"],
4356
- "_mm256_mask_cmp_epu32_mask"=>["AVX512VL", "AVX512F"],
4357
- "_mm256_mask_cmpeq_epu32_mask"=>["AVX512VL", "AVX512F"],
4358
- "_mm256_mask_cmpge_epu32_mask"=>["AVX512VL", "AVX512F"],
4359
- "_mm256_mask_cmpgt_epu32_mask"=>["AVX512VL", "AVX512F"],
4360
- "_mm256_mask_cmple_epu32_mask"=>["AVX512VL", "AVX512F"],
4361
- "_mm256_mask_cmplt_epu32_mask"=>["AVX512VL", "AVX512F"],
4362
- "_mm256_mask_cmpneq_epu32_mask"=>["AVX512VL", "AVX512F"],
4363
- "_mm_cmp_epu32_mask"=>["AVX512VL", "AVX512F"],
4364
- "_mm_cmpeq_epu32_mask"=>["AVX512VL", "AVX512F"],
4365
- "_mm_cmpge_epu32_mask"=>["AVX512VL", "AVX512F"],
4366
- "_mm_cmpgt_epu32_mask"=>["AVX512VL", "AVX512F"],
4367
- "_mm_cmple_epu32_mask"=>["AVX512VL", "AVX512F"],
4368
- "_mm_cmplt_epu32_mask"=>["AVX512VL", "AVX512F"],
4369
- "_mm_cmpneq_epu32_mask"=>["AVX512VL", "AVX512F"],
4370
- "_mm_mask_cmp_epu32_mask"=>["AVX512VL", "AVX512F"],
4371
- "_mm_mask_cmpeq_epu32_mask"=>["AVX512VL", "AVX512F"],
4372
- "_mm_mask_cmpge_epu32_mask"=>["AVX512VL", "AVX512F"],
4373
- "_mm_mask_cmpgt_epu32_mask"=>["AVX512VL", "AVX512F"],
4374
- "_mm_mask_cmple_epu32_mask"=>["AVX512VL", "AVX512F"],
4375
- "_mm_mask_cmplt_epu32_mask"=>["AVX512VL", "AVX512F"],
4376
- "_mm_mask_cmpneq_epu32_mask"=>["AVX512VL", "AVX512F"],
4377
- "_mm256_cmp_epu64_mask"=>["AVX512VL", "AVX512F"],
4378
- "_mm256_cmpeq_epu64_mask"=>["AVX512VL", "AVX512F"],
4379
- "_mm256_cmpge_epu64_mask"=>["AVX512VL", "AVX512F"],
4380
- "_mm256_cmpgt_epu64_mask"=>["AVX512VL", "AVX512F"],
4381
- "_mm256_cmple_epu64_mask"=>["AVX512VL", "AVX512F"],
4382
- "_mm256_cmplt_epu64_mask"=>["AVX512VL", "AVX512F"],
4383
- "_mm256_cmpneq_epu64_mask"=>["AVX512VL", "AVX512F"],
4384
- "_mm256_mask_cmp_epu64_mask"=>["AVX512VL", "AVX512F"],
4385
- "_mm256_mask_cmpeq_epu64_mask"=>["AVX512VL", "AVX512F"],
4386
- "_mm256_mask_cmpge_epu64_mask"=>["AVX512VL", "AVX512F"],
4387
- "_mm256_mask_cmpgt_epu64_mask"=>["AVX512VL", "AVX512F"],
4388
- "_mm256_mask_cmple_epu64_mask"=>["AVX512VL", "AVX512F"],
4389
- "_mm256_mask_cmplt_epu64_mask"=>["AVX512VL", "AVX512F"],
4390
- "_mm256_mask_cmpneq_epu64_mask"=>["AVX512VL", "AVX512F"],
4391
- "_mm_cmp_epu64_mask"=>["AVX512VL", "AVX512F"],
4392
- "_mm_cmpeq_epu64_mask"=>["AVX512VL", "AVX512F"],
4393
- "_mm_cmpge_epu64_mask"=>["AVX512VL", "AVX512F"],
4394
- "_mm_cmpgt_epu64_mask"=>["AVX512VL", "AVX512F"],
4395
- "_mm_cmple_epu64_mask"=>["AVX512VL", "AVX512F"],
4396
- "_mm_cmplt_epu64_mask"=>["AVX512VL", "AVX512F"],
4397
- "_mm_cmpneq_epu64_mask"=>["AVX512VL", "AVX512F"],
4398
- "_mm_mask_cmp_epu64_mask"=>["AVX512VL", "AVX512F"],
4399
- "_mm_mask_cmpeq_epu64_mask"=>["AVX512VL", "AVX512F"],
4400
- "_mm_mask_cmpge_epu64_mask"=>["AVX512VL", "AVX512F"],
4401
- "_mm_mask_cmpgt_epu64_mask"=>["AVX512VL", "AVX512F"],
4402
- "_mm_mask_cmple_epu64_mask"=>["AVX512VL", "AVX512F"],
4403
- "_mm_mask_cmplt_epu64_mask"=>["AVX512VL", "AVX512F"],
4404
- "_mm_mask_cmpneq_epu64_mask"=>["AVX512VL", "AVX512F"],
4405
- "_mm256_cmp_epu16_mask"=>["AVX512VL", "AVX512BW"],
4406
- "_mm256_cmpeq_epu16_mask"=>["AVX512VL", "AVX512BW"],
4407
- "_mm256_cmpge_epu16_mask"=>["AVX512VL", "AVX512BW"],
4408
- "_mm256_cmpgt_epu16_mask"=>["AVX512VL", "AVX512BW"],
4409
- "_mm256_cmple_epu16_mask"=>["AVX512VL", "AVX512BW"],
4410
- "_mm256_cmplt_epu16_mask"=>["AVX512VL", "AVX512BW"],
4411
- "_mm256_cmpneq_epu16_mask"=>["AVX512VL", "AVX512BW"],
4412
- "_mm256_mask_cmp_epu16_mask"=>["AVX512VL", "AVX512BW"],
4413
- "_mm256_mask_cmpeq_epu16_mask"=>["AVX512VL", "AVX512BW"],
4414
- "_mm256_mask_cmpge_epu16_mask"=>["AVX512VL", "AVX512BW"],
4415
- "_mm256_mask_cmpgt_epu16_mask"=>["AVX512VL", "AVX512BW"],
4416
- "_mm256_mask_cmple_epu16_mask"=>["AVX512VL", "AVX512BW"],
4417
- "_mm256_mask_cmplt_epu16_mask"=>["AVX512VL", "AVX512BW"],
4418
- "_mm256_mask_cmpneq_epu16_mask"=>["AVX512VL", "AVX512BW"],
4335
+ "_mm_cmp_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4336
+ "_mm_cmpeq_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4337
+ "_mm_cmpge_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4338
+ "_mm_cmpgt_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4339
+ "_mm_cmple_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4340
+ "_mm_cmplt_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4341
+ "_mm_cmpneq_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4342
+ "_mm_mask_cmp_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4343
+ "_mm_mask_cmpeq_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4344
+ "_mm_mask_cmpge_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4345
+ "_mm_mask_cmpgt_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4346
+ "_mm_mask_cmple_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4347
+ "_mm_mask_cmplt_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4348
+ "_mm_mask_cmpneq_epu8_mask"=>[["AVX512VL", "AVX512BW"]],
4349
+ "_mm256_cmp_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4350
+ "_mm256_cmpeq_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4351
+ "_mm256_cmpge_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4352
+ "_mm256_cmpgt_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4353
+ "_mm256_cmple_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4354
+ "_mm256_cmplt_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4355
+ "_mm256_cmpneq_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4356
+ "_mm256_mask_cmp_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4357
+ "_mm256_mask_cmpeq_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4358
+ "_mm256_mask_cmpge_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4359
+ "_mm256_mask_cmpgt_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4360
+ "_mm256_mask_cmple_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4361
+ "_mm256_mask_cmplt_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4362
+ "_mm256_mask_cmpneq_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4363
+ "_mm_cmp_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4364
+ "_mm_cmpeq_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4365
+ "_mm_cmpge_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4366
+ "_mm_cmpgt_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4367
+ "_mm_cmple_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4368
+ "_mm_cmplt_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4369
+ "_mm_cmpneq_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4370
+ "_mm_mask_cmp_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4371
+ "_mm_mask_cmpeq_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4372
+ "_mm_mask_cmpge_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4373
+ "_mm_mask_cmpgt_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4374
+ "_mm_mask_cmple_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4375
+ "_mm_mask_cmplt_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4376
+ "_mm_mask_cmpneq_epu32_mask"=>[["AVX512VL", "AVX512F"]],
4377
+ "_mm256_cmp_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4378
+ "_mm256_cmpeq_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4379
+ "_mm256_cmpge_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4380
+ "_mm256_cmpgt_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4381
+ "_mm256_cmple_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4382
+ "_mm256_cmplt_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4383
+ "_mm256_cmpneq_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4384
+ "_mm256_mask_cmp_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4385
+ "_mm256_mask_cmpeq_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4386
+ "_mm256_mask_cmpge_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4387
+ "_mm256_mask_cmpgt_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4388
+ "_mm256_mask_cmple_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4389
+ "_mm256_mask_cmplt_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4390
+ "_mm256_mask_cmpneq_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4391
+ "_mm_cmp_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4392
+ "_mm_cmpeq_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4393
+ "_mm_cmpge_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4394
+ "_mm_cmpgt_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4395
+ "_mm_cmple_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4396
+ "_mm_cmplt_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4397
+ "_mm_cmpneq_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4398
+ "_mm_mask_cmp_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4399
+ "_mm_mask_cmpeq_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4400
+ "_mm_mask_cmpge_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4401
+ "_mm_mask_cmpgt_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4402
+ "_mm_mask_cmple_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4403
+ "_mm_mask_cmplt_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4404
+ "_mm_mask_cmpneq_epu64_mask"=>[["AVX512VL", "AVX512F"]],
4405
+ "_mm256_cmp_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4406
+ "_mm256_cmpeq_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4407
+ "_mm256_cmpge_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4408
+ "_mm256_cmpgt_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4409
+ "_mm256_cmple_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4410
+ "_mm256_cmplt_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4411
+ "_mm256_cmpneq_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4412
+ "_mm256_mask_cmp_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4413
+ "_mm256_mask_cmpeq_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4414
+ "_mm256_mask_cmpge_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4415
+ "_mm256_mask_cmpgt_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4416
+ "_mm256_mask_cmple_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4417
+ "_mm256_mask_cmplt_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4418
+ "_mm256_mask_cmpneq_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4419
4419
  "_mm512_cmp_epu16_mask"=>["AVX512BW"],
4420
4420
  "_mm512_cmpeq_epu16_mask"=>["AVX512BW"],
4421
4421
  "_mm512_cmpge_epu16_mask"=>["AVX512BW"],
@@ -4430,34 +4430,34 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
4430
4430
  "_mm512_mask_cmple_epu16_mask"=>["AVX512BW"],
4431
4431
  "_mm512_mask_cmplt_epu16_mask"=>["AVX512BW"],
4432
4432
  "_mm512_mask_cmpneq_epu16_mask"=>["AVX512BW"],
4433
- "_mm_cmp_epu16_mask"=>["AVX512VL", "AVX512BW"],
4434
- "_mm_cmpeq_epu16_mask"=>["AVX512VL", "AVX512BW"],
4435
- "_mm_cmpge_epu16_mask"=>["AVX512VL", "AVX512BW"],
4436
- "_mm_cmpgt_epu16_mask"=>["AVX512VL", "AVX512BW"],
4437
- "_mm_cmple_epu16_mask"=>["AVX512VL", "AVX512BW"],
4438
- "_mm_cmplt_epu16_mask"=>["AVX512VL", "AVX512BW"],
4439
- "_mm_cmpneq_epu16_mask"=>["AVX512VL", "AVX512BW"],
4440
- "_mm_mask_cmp_epu16_mask"=>["AVX512VL", "AVX512BW"],
4441
- "_mm_mask_cmpeq_epu16_mask"=>["AVX512VL", "AVX512BW"],
4442
- "_mm_mask_cmpge_epu16_mask"=>["AVX512VL", "AVX512BW"],
4443
- "_mm_mask_cmpgt_epu16_mask"=>["AVX512VL", "AVX512BW"],
4444
- "_mm_mask_cmple_epu16_mask"=>["AVX512VL", "AVX512BW"],
4445
- "_mm_mask_cmplt_epu16_mask"=>["AVX512VL", "AVX512BW"],
4446
- "_mm_mask_cmpneq_epu16_mask"=>["AVX512VL", "AVX512BW"],
4447
- "_mm256_cmp_epi16_mask"=>["AVX512VL", "AVX512BW"],
4448
- "_mm256_cmpeq_epi16_mask"=>["AVX512VL", "AVX512BW"],
4449
- "_mm256_cmpge_epi16_mask"=>["AVX512VL", "AVX512BW"],
4450
- "_mm256_cmpgt_epi16_mask"=>["AVX512VL", "AVX512BW"],
4451
- "_mm256_cmple_epi16_mask"=>["AVX512VL", "AVX512BW"],
4452
- "_mm256_cmplt_epi16_mask"=>["AVX512VL", "AVX512BW"],
4453
- "_mm256_cmpneq_epi16_mask"=>["AVX512VL", "AVX512BW"],
4454
- "_mm256_mask_cmp_epi16_mask"=>["AVX512VL", "AVX512BW"],
4455
- "_mm256_mask_cmpeq_epi16_mask"=>["AVX512VL", "AVX512BW"],
4456
- "_mm256_mask_cmpge_epi16_mask"=>["AVX512VL", "AVX512BW"],
4457
- "_mm256_mask_cmpgt_epi16_mask"=>["AVX512VL", "AVX512BW"],
4458
- "_mm256_mask_cmple_epi16_mask"=>["AVX512VL", "AVX512BW"],
4459
- "_mm256_mask_cmplt_epi16_mask"=>["AVX512VL", "AVX512BW"],
4460
- "_mm256_mask_cmpneq_epi16_mask"=>["AVX512VL", "AVX512BW"],
4433
+ "_mm_cmp_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4434
+ "_mm_cmpeq_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4435
+ "_mm_cmpge_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4436
+ "_mm_cmpgt_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4437
+ "_mm_cmple_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4438
+ "_mm_cmplt_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4439
+ "_mm_cmpneq_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4440
+ "_mm_mask_cmp_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4441
+ "_mm_mask_cmpeq_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4442
+ "_mm_mask_cmpge_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4443
+ "_mm_mask_cmpgt_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4444
+ "_mm_mask_cmple_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4445
+ "_mm_mask_cmplt_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4446
+ "_mm_mask_cmpneq_epu16_mask"=>[["AVX512VL", "AVX512BW"]],
4447
+ "_mm256_cmp_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4448
+ "_mm256_cmpeq_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4449
+ "_mm256_cmpge_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4450
+ "_mm256_cmpgt_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4451
+ "_mm256_cmple_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4452
+ "_mm256_cmplt_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4453
+ "_mm256_cmpneq_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4454
+ "_mm256_mask_cmp_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4455
+ "_mm256_mask_cmpeq_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4456
+ "_mm256_mask_cmpge_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4457
+ "_mm256_mask_cmpgt_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4458
+ "_mm256_mask_cmple_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4459
+ "_mm256_mask_cmplt_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4460
+ "_mm256_mask_cmpneq_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4461
4461
  "_mm512_cmp_epi16_mask"=>["AVX512BW"],
4462
4462
  "_mm512_cmpeq_epi16_mask"=>["AVX512BW"],
4463
4463
  "_mm512_cmpge_epi16_mask"=>["AVX512BW"],
@@ -4472,986 +4472,986 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
4472
4472
  "_mm512_mask_cmple_epi16_mask"=>["AVX512BW"],
4473
4473
  "_mm512_mask_cmplt_epi16_mask"=>["AVX512BW"],
4474
4474
  "_mm512_mask_cmpneq_epi16_mask"=>["AVX512BW"],
4475
- "_mm_cmp_epi16_mask"=>["AVX512VL", "AVX512BW"],
4476
- "_mm_cmpeq_epi16_mask"=>["AVX512VL", "AVX512BW"],
4477
- "_mm_cmpge_epi16_mask"=>["AVX512VL", "AVX512BW"],
4478
- "_mm_cmpgt_epi16_mask"=>["AVX512VL", "AVX512BW"],
4479
- "_mm_cmple_epi16_mask"=>["AVX512VL", "AVX512BW"],
4480
- "_mm_cmplt_epi16_mask"=>["AVX512VL", "AVX512BW"],
4481
- "_mm_cmpneq_epi16_mask"=>["AVX512VL", "AVX512BW"],
4482
- "_mm_mask_cmp_epi16_mask"=>["AVX512VL", "AVX512BW"],
4483
- "_mm_mask_cmpeq_epi16_mask"=>["AVX512VL", "AVX512BW"],
4484
- "_mm_mask_cmpge_epi16_mask"=>["AVX512VL", "AVX512BW"],
4485
- "_mm_mask_cmpgt_epi16_mask"=>["AVX512VL", "AVX512BW"],
4486
- "_mm_mask_cmple_epi16_mask"=>["AVX512VL", "AVX512BW"],
4487
- "_mm_mask_cmplt_epi16_mask"=>["AVX512VL", "AVX512BW"],
4488
- "_mm_mask_cmpneq_epi16_mask"=>["AVX512VL", "AVX512BW"],
4489
- "_mm256_mask_compress_epi32"=>["AVX512VL", "AVX512F"],
4490
- "_mm256_mask_compressstoreu_epi32"=>["AVX512VL", "AVX512F"],
4491
- "_mm256_maskz_compress_epi32"=>["AVX512VL", "AVX512F"],
4492
- "_mm_mask_compress_epi32"=>["AVX512VL", "AVX512F"],
4493
- "_mm_mask_compressstoreu_epi32"=>["AVX512VL", "AVX512F"],
4494
- "_mm_maskz_compress_epi32"=>["AVX512VL", "AVX512F"],
4495
- "_mm256_mask_compress_epi64"=>["AVX512VL", "AVX512F"],
4496
- "_mm256_mask_compressstoreu_epi64"=>["AVX512VL", "AVX512F"],
4497
- "_mm256_maskz_compress_epi64"=>["AVX512VL", "AVX512F"],
4498
- "_mm_mask_compress_epi64"=>["AVX512VL", "AVX512F"],
4499
- "_mm_mask_compressstoreu_epi64"=>["AVX512VL", "AVX512F"],
4500
- "_mm_maskz_compress_epi64"=>["AVX512VL", "AVX512F"],
4501
- "_mm256_conflict_epi32"=>["AVX512VL", "AVX512CD"],
4502
- "_mm256_mask_conflict_epi32"=>["AVX512VL", "AVX512CD"],
4503
- "_mm256_maskz_conflict_epi32"=>["AVX512VL", "AVX512CD"],
4504
- "_mm_conflict_epi32"=>["AVX512VL", "AVX512CD"],
4505
- "_mm_mask_conflict_epi32"=>["AVX512VL", "AVX512CD"],
4506
- "_mm_maskz_conflict_epi32"=>["AVX512VL", "AVX512CD"],
4507
- "_mm256_conflict_epi64"=>["AVX512VL", "AVX512CD"],
4508
- "_mm256_mask_conflict_epi64"=>["AVX512VL", "AVX512CD"],
4509
- "_mm256_maskz_conflict_epi64"=>["AVX512VL", "AVX512CD"],
4510
- "_mm_conflict_epi64"=>["AVX512VL", "AVX512CD"],
4511
- "_mm_mask_conflict_epi64"=>["AVX512VL", "AVX512CD"],
4512
- "_mm_maskz_conflict_epi64"=>["AVX512VL", "AVX512CD"],
4513
- "_mm256_mask_permutexvar_epi32"=>["AVX512VL", "AVX512F"],
4514
- "_mm256_maskz_permutexvar_epi32"=>["AVX512VL", "AVX512F"],
4515
- "_mm256_permutexvar_epi32"=>["AVX512VL", "AVX512F"],
4516
- "_mm256_mask2_permutex2var_epi32"=>["AVX512VL", "AVX512F"],
4517
- "_mm256_mask_permutex2var_epi32"=>["AVX512VL", "AVX512F"],
4518
- "_mm256_maskz_permutex2var_epi32"=>["AVX512VL", "AVX512F"],
4519
- "_mm256_permutex2var_epi32"=>["AVX512VL", "AVX512F"],
4520
- "_mm_mask2_permutex2var_epi32"=>["AVX512VL", "AVX512F"],
4521
- "_mm_mask_permutex2var_epi32"=>["AVX512VL", "AVX512F"],
4522
- "_mm_maskz_permutex2var_epi32"=>["AVX512VL", "AVX512F"],
4523
- "_mm_permutex2var_epi32"=>["AVX512VL", "AVX512F"],
4524
- "_mm256_mask2_permutex2var_pd"=>["AVX512VL", "AVX512F"],
4525
- "_mm256_mask_permutex2var_pd"=>["AVX512VL", "AVX512F"],
4526
- "_mm256_maskz_permutex2var_pd"=>["AVX512VL", "AVX512F"],
4527
- "_mm256_permutex2var_pd"=>["AVX512VL", "AVX512F"],
4528
- "_mm_mask2_permutex2var_pd"=>["AVX512VL", "AVX512F"],
4529
- "_mm_mask_permutex2var_pd"=>["AVX512VL", "AVX512F"],
4530
- "_mm_maskz_permutex2var_pd"=>["AVX512VL", "AVX512F"],
4531
- "_mm_permutex2var_pd"=>["AVX512VL", "AVX512F"],
4532
- "_mm256_mask2_permutex2var_ps"=>["AVX512VL", "AVX512F"],
4533
- "_mm256_mask_permutex2var_ps"=>["AVX512VL", "AVX512F"],
4534
- "_mm256_maskz_permutex2var_ps"=>["AVX512VL", "AVX512F"],
4535
- "_mm256_permutex2var_ps"=>["AVX512VL", "AVX512F"],
4536
- "_mm_mask2_permutex2var_ps"=>["AVX512VL", "AVX512F"],
4537
- "_mm_mask_permutex2var_ps"=>["AVX512VL", "AVX512F"],
4538
- "_mm_maskz_permutex2var_ps"=>["AVX512VL", "AVX512F"],
4539
- "_mm_permutex2var_ps"=>["AVX512VL", "AVX512F"],
4540
- "_mm256_mask2_permutex2var_epi64"=>["AVX512VL", "AVX512F"],
4541
- "_mm256_mask_permutex2var_epi64"=>["AVX512VL", "AVX512F"],
4542
- "_mm256_maskz_permutex2var_epi64"=>["AVX512VL", "AVX512F"],
4543
- "_mm256_permutex2var_epi64"=>["AVX512VL", "AVX512F"],
4544
- "_mm_mask2_permutex2var_epi64"=>["AVX512VL", "AVX512F"],
4545
- "_mm_mask_permutex2var_epi64"=>["AVX512VL", "AVX512F"],
4546
- "_mm_maskz_permutex2var_epi64"=>["AVX512VL", "AVX512F"],
4547
- "_mm_permutex2var_epi64"=>["AVX512VL", "AVX512F"],
4548
- "_mm256_mask2_permutex2var_epi16"=>["AVX512VL", "AVX512BW"],
4549
- "_mm256_mask_permutex2var_epi16"=>["AVX512VL", "AVX512BW"],
4550
- "_mm256_maskz_permutex2var_epi16"=>["AVX512VL", "AVX512BW"],
4551
- "_mm256_permutex2var_epi16"=>["AVX512VL", "AVX512BW"],
4475
+ "_mm_cmp_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4476
+ "_mm_cmpeq_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4477
+ "_mm_cmpge_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4478
+ "_mm_cmpgt_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4479
+ "_mm_cmple_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4480
+ "_mm_cmplt_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4481
+ "_mm_cmpneq_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4482
+ "_mm_mask_cmp_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4483
+ "_mm_mask_cmpeq_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4484
+ "_mm_mask_cmpge_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4485
+ "_mm_mask_cmpgt_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4486
+ "_mm_mask_cmple_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4487
+ "_mm_mask_cmplt_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4488
+ "_mm_mask_cmpneq_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
4489
+ "_mm256_mask_compress_epi32"=>[["AVX512VL", "AVX512F"]],
4490
+ "_mm256_mask_compressstoreu_epi32"=>[["AVX512VL", "AVX512F"]],
4491
+ "_mm256_maskz_compress_epi32"=>[["AVX512VL", "AVX512F"]],
4492
+ "_mm_mask_compress_epi32"=>[["AVX512VL", "AVX512F"]],
4493
+ "_mm_mask_compressstoreu_epi32"=>[["AVX512VL", "AVX512F"]],
4494
+ "_mm_maskz_compress_epi32"=>[["AVX512VL", "AVX512F"]],
4495
+ "_mm256_mask_compress_epi64"=>[["AVX512VL", "AVX512F"]],
4496
+ "_mm256_mask_compressstoreu_epi64"=>[["AVX512VL", "AVX512F"]],
4497
+ "_mm256_maskz_compress_epi64"=>[["AVX512VL", "AVX512F"]],
4498
+ "_mm_mask_compress_epi64"=>[["AVX512VL", "AVX512F"]],
4499
+ "_mm_mask_compressstoreu_epi64"=>[["AVX512VL", "AVX512F"]],
4500
+ "_mm_maskz_compress_epi64"=>[["AVX512VL", "AVX512F"]],
4501
+ "_mm256_conflict_epi32"=>[["AVX512VL", "AVX512CD"]],
4502
+ "_mm256_mask_conflict_epi32"=>[["AVX512VL", "AVX512CD"]],
4503
+ "_mm256_maskz_conflict_epi32"=>[["AVX512VL", "AVX512CD"]],
4504
+ "_mm_conflict_epi32"=>[["AVX512VL", "AVX512CD"]],
4505
+ "_mm_mask_conflict_epi32"=>[["AVX512VL", "AVX512CD"]],
4506
+ "_mm_maskz_conflict_epi32"=>[["AVX512VL", "AVX512CD"]],
4507
+ "_mm256_conflict_epi64"=>[["AVX512VL", "AVX512CD"]],
4508
+ "_mm256_mask_conflict_epi64"=>[["AVX512VL", "AVX512CD"]],
4509
+ "_mm256_maskz_conflict_epi64"=>[["AVX512VL", "AVX512CD"]],
4510
+ "_mm_conflict_epi64"=>[["AVX512VL", "AVX512CD"]],
4511
+ "_mm_mask_conflict_epi64"=>[["AVX512VL", "AVX512CD"]],
4512
+ "_mm_maskz_conflict_epi64"=>[["AVX512VL", "AVX512CD"]],
4513
+ "_mm256_mask_permutexvar_epi32"=>[["AVX512VL", "AVX512F"]],
4514
+ "_mm256_maskz_permutexvar_epi32"=>[["AVX512VL", "AVX512F"]],
4515
+ "_mm256_permutexvar_epi32"=>[["AVX512VL", "AVX512F"]],
4516
+ "_mm256_mask2_permutex2var_epi32"=>[["AVX512VL", "AVX512F"]],
4517
+ "_mm256_mask_permutex2var_epi32"=>[["AVX512VL", "AVX512F"]],
4518
+ "_mm256_maskz_permutex2var_epi32"=>[["AVX512VL", "AVX512F"]],
4519
+ "_mm256_permutex2var_epi32"=>[["AVX512VL", "AVX512F"]],
4520
+ "_mm_mask2_permutex2var_epi32"=>[["AVX512VL", "AVX512F"]],
4521
+ "_mm_mask_permutex2var_epi32"=>[["AVX512VL", "AVX512F"]],
4522
+ "_mm_maskz_permutex2var_epi32"=>[["AVX512VL", "AVX512F"]],
4523
+ "_mm_permutex2var_epi32"=>[["AVX512VL", "AVX512F"]],
4524
+ "_mm256_mask2_permutex2var_pd"=>[["AVX512VL", "AVX512F"]],
4525
+ "_mm256_mask_permutex2var_pd"=>[["AVX512VL", "AVX512F"]],
4526
+ "_mm256_maskz_permutex2var_pd"=>[["AVX512VL", "AVX512F"]],
4527
+ "_mm256_permutex2var_pd"=>[["AVX512VL", "AVX512F"]],
4528
+ "_mm_mask2_permutex2var_pd"=>[["AVX512VL", "AVX512F"]],
4529
+ "_mm_mask_permutex2var_pd"=>[["AVX512VL", "AVX512F"]],
4530
+ "_mm_maskz_permutex2var_pd"=>[["AVX512VL", "AVX512F"]],
4531
+ "_mm_permutex2var_pd"=>[["AVX512VL", "AVX512F"]],
4532
+ "_mm256_mask2_permutex2var_ps"=>[["AVX512VL", "AVX512F"]],
4533
+ "_mm256_mask_permutex2var_ps"=>[["AVX512VL", "AVX512F"]],
4534
+ "_mm256_maskz_permutex2var_ps"=>[["AVX512VL", "AVX512F"]],
4535
+ "_mm256_permutex2var_ps"=>[["AVX512VL", "AVX512F"]],
4536
+ "_mm_mask2_permutex2var_ps"=>[["AVX512VL", "AVX512F"]],
4537
+ "_mm_mask_permutex2var_ps"=>[["AVX512VL", "AVX512F"]],
4538
+ "_mm_maskz_permutex2var_ps"=>[["AVX512VL", "AVX512F"]],
4539
+ "_mm_permutex2var_ps"=>[["AVX512VL", "AVX512F"]],
4540
+ "_mm256_mask2_permutex2var_epi64"=>[["AVX512VL", "AVX512F"]],
4541
+ "_mm256_mask_permutex2var_epi64"=>[["AVX512VL", "AVX512F"]],
4542
+ "_mm256_maskz_permutex2var_epi64"=>[["AVX512VL", "AVX512F"]],
4543
+ "_mm256_permutex2var_epi64"=>[["AVX512VL", "AVX512F"]],
4544
+ "_mm_mask2_permutex2var_epi64"=>[["AVX512VL", "AVX512F"]],
4545
+ "_mm_mask_permutex2var_epi64"=>[["AVX512VL", "AVX512F"]],
4546
+ "_mm_maskz_permutex2var_epi64"=>[["AVX512VL", "AVX512F"]],
4547
+ "_mm_permutex2var_epi64"=>[["AVX512VL", "AVX512F"]],
4548
+ "_mm256_mask2_permutex2var_epi16"=>[["AVX512VL", "AVX512BW"]],
4549
+ "_mm256_mask_permutex2var_epi16"=>[["AVX512VL", "AVX512BW"]],
4550
+ "_mm256_maskz_permutex2var_epi16"=>[["AVX512VL", "AVX512BW"]],
4551
+ "_mm256_permutex2var_epi16"=>[["AVX512VL", "AVX512BW"]],
4552
4552
  "_mm512_mask2_permutex2var_epi16"=>["AVX512BW"],
4553
4553
  "_mm512_mask_permutex2var_epi16"=>["AVX512BW"],
4554
4554
  "_mm512_maskz_permutex2var_epi16"=>["AVX512BW"],
4555
4555
  "_mm512_permutex2var_epi16"=>["AVX512BW"],
4556
- "_mm_mask2_permutex2var_epi16"=>["AVX512VL", "AVX512BW"],
4557
- "_mm_mask_permutex2var_epi16"=>["AVX512VL", "AVX512BW"],
4558
- "_mm_maskz_permutex2var_epi16"=>["AVX512VL", "AVX512BW"],
4559
- "_mm_permutex2var_epi16"=>["AVX512VL", "AVX512BW"],
4560
- "_mm256_mask_permute_pd"=>["AVX512VL", "AVX512F"],
4561
- "_mm256_mask_permutevar_pd"=>["AVX512VL", "AVX512F"],
4562
- "_mm256_maskz_permute_pd"=>["AVX512VL", "AVX512F"],
4563
- "_mm256_maskz_permutevar_pd"=>["AVX512VL", "AVX512F"],
4564
- "_mm_mask_permute_pd"=>["AVX512VL", "AVX512F"],
4565
- "_mm_mask_permutevar_pd"=>["AVX512VL", "AVX512F"],
4566
- "_mm_maskz_permute_pd"=>["AVX512VL", "AVX512F"],
4567
- "_mm_maskz_permutevar_pd"=>["AVX512VL", "AVX512F"],
4568
- "_mm256_mask_permute_ps"=>["AVX512VL", "AVX512F"],
4569
- "_mm256_mask_permutevar_ps"=>["AVX512VL", "AVX512F"],
4570
- "_mm256_maskz_permute_ps"=>["AVX512VL", "AVX512F"],
4571
- "_mm256_maskz_permutevar_ps"=>["AVX512VL", "AVX512F"],
4572
- "_mm_mask_permute_ps"=>["AVX512VL", "AVX512F"],
4573
- "_mm_mask_permutevar_ps"=>["AVX512VL", "AVX512F"],
4574
- "_mm_maskz_permute_ps"=>["AVX512VL", "AVX512F"],
4575
- "_mm_maskz_permutevar_ps"=>["AVX512VL", "AVX512F"],
4576
- "_mm256_mask_permutex_pd"=>["AVX512VL", "AVX512F"],
4577
- "_mm256_mask_permutexvar_pd"=>["AVX512VL", "AVX512F"],
4578
- "_mm256_maskz_permutex_pd"=>["AVX512VL", "AVX512F"],
4579
- "_mm256_maskz_permutexvar_pd"=>["AVX512VL", "AVX512F"],
4580
- "_mm256_permutex_pd"=>["AVX512VL", "AVX512F"],
4581
- "_mm256_permutexvar_pd"=>["AVX512VL", "AVX512F"],
4582
- "_mm256_mask_permutexvar_ps"=>["AVX512VL", "AVX512F"],
4583
- "_mm256_maskz_permutexvar_ps"=>["AVX512VL", "AVX512F"],
4584
- "_mm256_permutexvar_ps"=>["AVX512VL", "AVX512F"],
4585
- "_mm256_mask_permutex_epi64"=>["AVX512VL", "AVX512F"],
4586
- "_mm256_mask_permutexvar_epi64"=>["AVX512VL", "AVX512F"],
4587
- "_mm256_maskz_permutex_epi64"=>["AVX512VL", "AVX512F"],
4588
- "_mm256_maskz_permutexvar_epi64"=>["AVX512VL", "AVX512F"],
4589
- "_mm256_permutex_epi64"=>["AVX512VL", "AVX512F"],
4590
- "_mm256_permutexvar_epi64"=>["AVX512VL", "AVX512F"],
4591
- "_mm256_mask_permutexvar_epi16"=>["AVX512VL", "AVX512BW"],
4592
- "_mm256_maskz_permutexvar_epi16"=>["AVX512VL", "AVX512BW"],
4593
- "_mm256_permutexvar_epi16"=>["AVX512VL", "AVX512BW"],
4556
+ "_mm_mask2_permutex2var_epi16"=>[["AVX512VL", "AVX512BW"]],
4557
+ "_mm_mask_permutex2var_epi16"=>[["AVX512VL", "AVX512BW"]],
4558
+ "_mm_maskz_permutex2var_epi16"=>[["AVX512VL", "AVX512BW"]],
4559
+ "_mm_permutex2var_epi16"=>[["AVX512VL", "AVX512BW"]],
4560
+ "_mm256_mask_permute_pd"=>[["AVX512VL", "AVX512F"]],
4561
+ "_mm256_mask_permutevar_pd"=>[["AVX512VL", "AVX512F"]],
4562
+ "_mm256_maskz_permute_pd"=>[["AVX512VL", "AVX512F"]],
4563
+ "_mm256_maskz_permutevar_pd"=>[["AVX512VL", "AVX512F"]],
4564
+ "_mm_mask_permute_pd"=>[["AVX512VL", "AVX512F"]],
4565
+ "_mm_mask_permutevar_pd"=>[["AVX512VL", "AVX512F"]],
4566
+ "_mm_maskz_permute_pd"=>[["AVX512VL", "AVX512F"]],
4567
+ "_mm_maskz_permutevar_pd"=>[["AVX512VL", "AVX512F"]],
4568
+ "_mm256_mask_permute_ps"=>[["AVX512VL", "AVX512F"]],
4569
+ "_mm256_mask_permutevar_ps"=>[["AVX512VL", "AVX512F"]],
4570
+ "_mm256_maskz_permute_ps"=>[["AVX512VL", "AVX512F"]],
4571
+ "_mm256_maskz_permutevar_ps"=>[["AVX512VL", "AVX512F"]],
4572
+ "_mm_mask_permute_ps"=>[["AVX512VL", "AVX512F"]],
4573
+ "_mm_mask_permutevar_ps"=>[["AVX512VL", "AVX512F"]],
4574
+ "_mm_maskz_permute_ps"=>[["AVX512VL", "AVX512F"]],
4575
+ "_mm_maskz_permutevar_ps"=>[["AVX512VL", "AVX512F"]],
4576
+ "_mm256_mask_permutex_pd"=>[["AVX512VL", "AVX512F"]],
4577
+ "_mm256_mask_permutexvar_pd"=>[["AVX512VL", "AVX512F"]],
4578
+ "_mm256_maskz_permutex_pd"=>[["AVX512VL", "AVX512F"]],
4579
+ "_mm256_maskz_permutexvar_pd"=>[["AVX512VL", "AVX512F"]],
4580
+ "_mm256_permutex_pd"=>[["AVX512VL", "AVX512F"]],
4581
+ "_mm256_permutexvar_pd"=>[["AVX512VL", "AVX512F"]],
4582
+ "_mm256_mask_permutexvar_ps"=>[["AVX512VL", "AVX512F"]],
4583
+ "_mm256_maskz_permutexvar_ps"=>[["AVX512VL", "AVX512F"]],
4584
+ "_mm256_permutexvar_ps"=>[["AVX512VL", "AVX512F"]],
4585
+ "_mm256_mask_permutex_epi64"=>[["AVX512VL", "AVX512F"]],
4586
+ "_mm256_mask_permutexvar_epi64"=>[["AVX512VL", "AVX512F"]],
4587
+ "_mm256_maskz_permutex_epi64"=>[["AVX512VL", "AVX512F"]],
4588
+ "_mm256_maskz_permutexvar_epi64"=>[["AVX512VL", "AVX512F"]],
4589
+ "_mm256_permutex_epi64"=>[["AVX512VL", "AVX512F"]],
4590
+ "_mm256_permutexvar_epi64"=>[["AVX512VL", "AVX512F"]],
4591
+ "_mm256_mask_permutexvar_epi16"=>[["AVX512VL", "AVX512BW"]],
4592
+ "_mm256_maskz_permutexvar_epi16"=>[["AVX512VL", "AVX512BW"]],
4593
+ "_mm256_permutexvar_epi16"=>[["AVX512VL", "AVX512BW"]],
4594
4594
  "_mm512_mask_permutexvar_epi16"=>["AVX512BW"],
4595
4595
  "_mm512_maskz_permutexvar_epi16"=>["AVX512BW"],
4596
4596
  "_mm512_permutexvar_epi16"=>["AVX512BW"],
4597
- "_mm_mask_permutexvar_epi16"=>["AVX512VL", "AVX512BW"],
4598
- "_mm_maskz_permutexvar_epi16"=>["AVX512VL", "AVX512BW"],
4599
- "_mm_permutexvar_epi16"=>["AVX512VL", "AVX512BW"],
4600
- "_mm256_mask_expand_epi32"=>["AVX512VL", "AVX512F"],
4601
- "_mm256_mask_expandloadu_epi32"=>["AVX512VL", "AVX512F"],
4602
- "_mm256_maskz_expand_epi32"=>["AVX512VL", "AVX512F"],
4603
- "_mm256_maskz_expandloadu_epi32"=>["AVX512VL", "AVX512F"],
4604
- "_mm_mask_expand_epi32"=>["AVX512VL", "AVX512F"],
4605
- "_mm_mask_expandloadu_epi32"=>["AVX512VL", "AVX512F"],
4606
- "_mm_maskz_expand_epi32"=>["AVX512VL", "AVX512F"],
4607
- "_mm_maskz_expandloadu_epi32"=>["AVX512VL", "AVX512F"],
4608
- "_mm256_mask_expand_epi64"=>["AVX512VL", "AVX512F"],
4609
- "_mm256_mask_expandloadu_epi64"=>["AVX512VL", "AVX512F"],
4610
- "_mm256_maskz_expand_epi64"=>["AVX512VL", "AVX512F"],
4611
- "_mm256_maskz_expandloadu_epi64"=>["AVX512VL", "AVX512F"],
4612
- "_mm_mask_expand_epi64"=>["AVX512VL", "AVX512F"],
4613
- "_mm_mask_expandloadu_epi64"=>["AVX512VL", "AVX512F"],
4614
- "_mm_maskz_expand_epi64"=>["AVX512VL", "AVX512F"],
4615
- "_mm_maskz_expandloadu_epi64"=>["AVX512VL", "AVX512F"],
4616
- "_mm256_mmask_i32gather_epi32"=>["AVX512VL", "AVX512F"],
4617
- "_mm_mmask_i32gather_epi32"=>["AVX512VL", "AVX512F"],
4618
- "_mm256_mmask_i32gather_epi64"=>["AVX512VL", "AVX512F"],
4619
- "_mm_mmask_i32gather_epi64"=>["AVX512VL", "AVX512F"],
4620
- "_mm256_mmask_i64gather_epi32"=>["AVX512VL", "AVX512F"],
4621
- "_mm_mmask_i64gather_epi32"=>["AVX512VL", "AVX512F"],
4622
- "_mm256_mmask_i64gather_epi64"=>["AVX512VL", "AVX512F"],
4623
- "_mm_mmask_i64gather_epi64"=>["AVX512VL", "AVX512F"],
4624
- "_mm256_lzcnt_epi32"=>["AVX512VL", "AVX512CD"],
4625
- "_mm256_mask_lzcnt_epi32"=>["AVX512VL", "AVX512CD"],
4626
- "_mm256_maskz_lzcnt_epi32"=>["AVX512VL", "AVX512CD"],
4627
- "_mm_lzcnt_epi32"=>["AVX512VL", "AVX512CD"],
4628
- "_mm_mask_lzcnt_epi32"=>["AVX512VL", "AVX512CD"],
4629
- "_mm_maskz_lzcnt_epi32"=>["AVX512VL", "AVX512CD"],
4630
- "_mm256_lzcnt_epi64"=>["AVX512VL", "AVX512CD"],
4631
- "_mm256_mask_lzcnt_epi64"=>["AVX512VL", "AVX512CD"],
4632
- "_mm256_maskz_lzcnt_epi64"=>["AVX512VL", "AVX512CD"],
4633
- "_mm_lzcnt_epi64"=>["AVX512VL", "AVX512CD"],
4634
- "_mm_mask_lzcnt_epi64"=>["AVX512VL", "AVX512CD"],
4635
- "_mm_maskz_lzcnt_epi64"=>["AVX512VL", "AVX512CD"],
4636
- "_mm256_mask_maddubs_epi16"=>["AVX512VL", "AVX512BW"],
4637
- "_mm256_maskz_maddubs_epi16"=>["AVX512VL", "AVX512BW"],
4597
+ "_mm_mask_permutexvar_epi16"=>[["AVX512VL", "AVX512BW"]],
4598
+ "_mm_maskz_permutexvar_epi16"=>[["AVX512VL", "AVX512BW"]],
4599
+ "_mm_permutexvar_epi16"=>[["AVX512VL", "AVX512BW"]],
4600
+ "_mm256_mask_expand_epi32"=>[["AVX512VL", "AVX512F"]],
4601
+ "_mm256_mask_expandloadu_epi32"=>[["AVX512VL", "AVX512F"]],
4602
+ "_mm256_maskz_expand_epi32"=>[["AVX512VL", "AVX512F"]],
4603
+ "_mm256_maskz_expandloadu_epi32"=>[["AVX512VL", "AVX512F"]],
4604
+ "_mm_mask_expand_epi32"=>[["AVX512VL", "AVX512F"]],
4605
+ "_mm_mask_expandloadu_epi32"=>[["AVX512VL", "AVX512F"]],
4606
+ "_mm_maskz_expand_epi32"=>[["AVX512VL", "AVX512F"]],
4607
+ "_mm_maskz_expandloadu_epi32"=>[["AVX512VL", "AVX512F"]],
4608
+ "_mm256_mask_expand_epi64"=>[["AVX512VL", "AVX512F"]],
4609
+ "_mm256_mask_expandloadu_epi64"=>[["AVX512VL", "AVX512F"]],
4610
+ "_mm256_maskz_expand_epi64"=>[["AVX512VL", "AVX512F"]],
4611
+ "_mm256_maskz_expandloadu_epi64"=>[["AVX512VL", "AVX512F"]],
4612
+ "_mm_mask_expand_epi64"=>[["AVX512VL", "AVX512F"]],
4613
+ "_mm_mask_expandloadu_epi64"=>[["AVX512VL", "AVX512F"]],
4614
+ "_mm_maskz_expand_epi64"=>[["AVX512VL", "AVX512F"]],
4615
+ "_mm_maskz_expandloadu_epi64"=>[["AVX512VL", "AVX512F"]],
4616
+ "_mm256_mmask_i32gather_epi32"=>[["AVX512VL", "AVX512F"]],
4617
+ "_mm_mmask_i32gather_epi32"=>[["AVX512VL", "AVX512F"]],
4618
+ "_mm256_mmask_i32gather_epi64"=>[["AVX512VL", "AVX512F"]],
4619
+ "_mm_mmask_i32gather_epi64"=>[["AVX512VL", "AVX512F"]],
4620
+ "_mm256_mmask_i64gather_epi32"=>[["AVX512VL", "AVX512F"]],
4621
+ "_mm_mmask_i64gather_epi32"=>[["AVX512VL", "AVX512F"]],
4622
+ "_mm256_mmask_i64gather_epi64"=>[["AVX512VL", "AVX512F"]],
4623
+ "_mm_mmask_i64gather_epi64"=>[["AVX512VL", "AVX512F"]],
4624
+ "_mm256_lzcnt_epi32"=>[["AVX512VL", "AVX512CD"]],
4625
+ "_mm256_mask_lzcnt_epi32"=>[["AVX512VL", "AVX512CD"]],
4626
+ "_mm256_maskz_lzcnt_epi32"=>[["AVX512VL", "AVX512CD"]],
4627
+ "_mm_lzcnt_epi32"=>[["AVX512VL", "AVX512CD"]],
4628
+ "_mm_mask_lzcnt_epi32"=>[["AVX512VL", "AVX512CD"]],
4629
+ "_mm_maskz_lzcnt_epi32"=>[["AVX512VL", "AVX512CD"]],
4630
+ "_mm256_lzcnt_epi64"=>[["AVX512VL", "AVX512CD"]],
4631
+ "_mm256_mask_lzcnt_epi64"=>[["AVX512VL", "AVX512CD"]],
4632
+ "_mm256_maskz_lzcnt_epi64"=>[["AVX512VL", "AVX512CD"]],
4633
+ "_mm_lzcnt_epi64"=>[["AVX512VL", "AVX512CD"]],
4634
+ "_mm_mask_lzcnt_epi64"=>[["AVX512VL", "AVX512CD"]],
4635
+ "_mm_maskz_lzcnt_epi64"=>[["AVX512VL", "AVX512CD"]],
4636
+ "_mm256_mask_maddubs_epi16"=>[["AVX512VL", "AVX512BW"]],
4637
+ "_mm256_maskz_maddubs_epi16"=>[["AVX512VL", "AVX512BW"]],
4638
4638
  "_mm512_maddubs_epi16"=>["AVX512BW"],
4639
4639
  "_mm512_mask_maddubs_epi16"=>["AVX512BW"],
4640
4640
  "_mm512_maskz_maddubs_epi16"=>["AVX512BW"],
4641
- "_mm_mask_maddubs_epi16"=>["AVX512VL", "AVX512BW"],
4642
- "_mm_maskz_maddubs_epi16"=>["AVX512VL", "AVX512BW"],
4643
- "_mm256_mask_madd_epi16"=>["AVX512VL", "AVX512BW"],
4644
- "_mm256_maskz_madd_epi16"=>["AVX512VL", "AVX512BW"],
4641
+ "_mm_mask_maddubs_epi16"=>[["AVX512VL", "AVX512BW"]],
4642
+ "_mm_maskz_maddubs_epi16"=>[["AVX512VL", "AVX512BW"]],
4643
+ "_mm256_mask_madd_epi16"=>[["AVX512VL", "AVX512BW"]],
4644
+ "_mm256_maskz_madd_epi16"=>[["AVX512VL", "AVX512BW"]],
4645
4645
  "_mm512_madd_epi16"=>["AVX512BW"],
4646
4646
  "_mm512_mask_madd_epi16"=>["AVX512BW"],
4647
4647
  "_mm512_maskz_madd_epi16"=>["AVX512BW"],
4648
- "_mm_mask_madd_epi16"=>["AVX512VL", "AVX512BW"],
4649
- "_mm_maskz_madd_epi16"=>["AVX512VL", "AVX512BW"],
4650
- "_mm256_mask_max_epi8"=>["AVX512VL", "AVX512BW"],
4651
- "_mm256_maskz_max_epi8"=>["AVX512VL", "AVX512BW"],
4648
+ "_mm_mask_madd_epi16"=>[["AVX512VL", "AVX512BW"]],
4649
+ "_mm_maskz_madd_epi16"=>[["AVX512VL", "AVX512BW"]],
4650
+ "_mm256_mask_max_epi8"=>[["AVX512VL", "AVX512BW"]],
4651
+ "_mm256_maskz_max_epi8"=>[["AVX512VL", "AVX512BW"]],
4652
4652
  "_mm512_mask_max_epi8"=>["AVX512BW"],
4653
4653
  "_mm512_maskz_max_epi8"=>["AVX512BW"],
4654
4654
  "_mm512_max_epi8"=>["AVX512BW"],
4655
- "_mm_mask_max_epi8"=>["AVX512VL", "AVX512BW"],
4656
- "_mm_maskz_max_epi8"=>["AVX512VL", "AVX512BW"],
4657
- "_mm256_mask_max_epi32"=>["AVX512VL", "AVX512F"],
4658
- "_mm256_maskz_max_epi32"=>["AVX512VL", "AVX512F"],
4659
- "_mm_mask_max_epi32"=>["AVX512VL", "AVX512F"],
4660
- "_mm_maskz_max_epi32"=>["AVX512VL", "AVX512F"],
4661
- "_mm256_mask_max_epi64"=>["AVX512VL", "AVX512F"],
4662
- "_mm256_maskz_max_epi64"=>["AVX512VL", "AVX512F"],
4663
- "_mm256_max_epi64"=>["AVX512VL", "AVX512F"],
4664
- "_mm_mask_max_epi64"=>["AVX512VL", "AVX512F"],
4665
- "_mm_maskz_max_epi64"=>["AVX512VL", "AVX512F"],
4666
- "_mm_max_epi64"=>["AVX512VL", "AVX512F"],
4667
- "_mm256_mask_max_epi16"=>["AVX512VL", "AVX512BW"],
4668
- "_mm256_maskz_max_epi16"=>["AVX512VL", "AVX512BW"],
4655
+ "_mm_mask_max_epi8"=>[["AVX512VL", "AVX512BW"]],
4656
+ "_mm_maskz_max_epi8"=>[["AVX512VL", "AVX512BW"]],
4657
+ "_mm256_mask_max_epi32"=>[["AVX512VL", "AVX512F"]],
4658
+ "_mm256_maskz_max_epi32"=>[["AVX512VL", "AVX512F"]],
4659
+ "_mm_mask_max_epi32"=>[["AVX512VL", "AVX512F"]],
4660
+ "_mm_maskz_max_epi32"=>[["AVX512VL", "AVX512F"]],
4661
+ "_mm256_mask_max_epi64"=>[["AVX512VL", "AVX512F"]],
4662
+ "_mm256_maskz_max_epi64"=>[["AVX512VL", "AVX512F"]],
4663
+ "_mm256_max_epi64"=>[["AVX512VL", "AVX512F"]],
4664
+ "_mm_mask_max_epi64"=>[["AVX512VL", "AVX512F"]],
4665
+ "_mm_maskz_max_epi64"=>[["AVX512VL", "AVX512F"]],
4666
+ "_mm_max_epi64"=>[["AVX512VL", "AVX512F"]],
4667
+ "_mm256_mask_max_epi16"=>[["AVX512VL", "AVX512BW"]],
4668
+ "_mm256_maskz_max_epi16"=>[["AVX512VL", "AVX512BW"]],
4669
4669
  "_mm512_mask_max_epi16"=>["AVX512BW"],
4670
4670
  "_mm512_maskz_max_epi16"=>["AVX512BW"],
4671
4671
  "_mm512_max_epi16"=>["AVX512BW"],
4672
- "_mm_mask_max_epi16"=>["AVX512VL", "AVX512BW"],
4673
- "_mm_maskz_max_epi16"=>["AVX512VL", "AVX512BW"],
4674
- "_mm256_mask_max_epu8"=>["AVX512VL", "AVX512BW"],
4675
- "_mm256_maskz_max_epu8"=>["AVX512VL", "AVX512BW"],
4672
+ "_mm_mask_max_epi16"=>[["AVX512VL", "AVX512BW"]],
4673
+ "_mm_maskz_max_epi16"=>[["AVX512VL", "AVX512BW"]],
4674
+ "_mm256_mask_max_epu8"=>[["AVX512VL", "AVX512BW"]],
4675
+ "_mm256_maskz_max_epu8"=>[["AVX512VL", "AVX512BW"]],
4676
4676
  "_mm512_mask_max_epu8"=>["AVX512BW"],
4677
4677
  "_mm512_maskz_max_epu8"=>["AVX512BW"],
4678
4678
  "_mm512_max_epu8"=>["AVX512BW"],
4679
- "_mm_mask_max_epu8"=>["AVX512VL", "AVX512BW"],
4680
- "_mm_maskz_max_epu8"=>["AVX512VL", "AVX512BW"],
4681
- "_mm256_mask_max_epu32"=>["AVX512VL", "AVX512F"],
4682
- "_mm256_maskz_max_epu32"=>["AVX512VL", "AVX512F"],
4683
- "_mm_mask_max_epu32"=>["AVX512VL", "AVX512F"],
4684
- "_mm_maskz_max_epu32"=>["AVX512VL", "AVX512F"],
4685
- "_mm256_mask_max_epu64"=>["AVX512VL", "AVX512F"],
4686
- "_mm256_maskz_max_epu64"=>["AVX512VL", "AVX512F"],
4687
- "_mm256_max_epu64"=>["AVX512VL", "AVX512F"],
4688
- "_mm_mask_max_epu64"=>["AVX512VL", "AVX512F"],
4689
- "_mm_maskz_max_epu64"=>["AVX512VL", "AVX512F"],
4690
- "_mm_max_epu64"=>["AVX512VL", "AVX512F"],
4691
- "_mm256_mask_max_epu16"=>["AVX512VL", "AVX512BW"],
4692
- "_mm256_maskz_max_epu16"=>["AVX512VL", "AVX512BW"],
4679
+ "_mm_mask_max_epu8"=>[["AVX512VL", "AVX512BW"]],
4680
+ "_mm_maskz_max_epu8"=>[["AVX512VL", "AVX512BW"]],
4681
+ "_mm256_mask_max_epu32"=>[["AVX512VL", "AVX512F"]],
4682
+ "_mm256_maskz_max_epu32"=>[["AVX512VL", "AVX512F"]],
4683
+ "_mm_mask_max_epu32"=>[["AVX512VL", "AVX512F"]],
4684
+ "_mm_maskz_max_epu32"=>[["AVX512VL", "AVX512F"]],
4685
+ "_mm256_mask_max_epu64"=>[["AVX512VL", "AVX512F"]],
4686
+ "_mm256_maskz_max_epu64"=>[["AVX512VL", "AVX512F"]],
4687
+ "_mm256_max_epu64"=>[["AVX512VL", "AVX512F"]],
4688
+ "_mm_mask_max_epu64"=>[["AVX512VL", "AVX512F"]],
4689
+ "_mm_maskz_max_epu64"=>[["AVX512VL", "AVX512F"]],
4690
+ "_mm_max_epu64"=>[["AVX512VL", "AVX512F"]],
4691
+ "_mm256_mask_max_epu16"=>[["AVX512VL", "AVX512BW"]],
4692
+ "_mm256_maskz_max_epu16"=>[["AVX512VL", "AVX512BW"]],
4693
4693
  "_mm512_mask_max_epu16"=>["AVX512BW"],
4694
4694
  "_mm512_maskz_max_epu16"=>["AVX512BW"],
4695
4695
  "_mm512_max_epu16"=>["AVX512BW"],
4696
- "_mm_mask_max_epu16"=>["AVX512VL", "AVX512BW"],
4697
- "_mm_maskz_max_epu16"=>["AVX512VL", "AVX512BW"],
4698
- "_mm256_mask_min_epi8"=>["AVX512VL", "AVX512BW"],
4699
- "_mm256_maskz_min_epi8"=>["AVX512VL", "AVX512BW"],
4696
+ "_mm_mask_max_epu16"=>[["AVX512VL", "AVX512BW"]],
4697
+ "_mm_maskz_max_epu16"=>[["AVX512VL", "AVX512BW"]],
4698
+ "_mm256_mask_min_epi8"=>[["AVX512VL", "AVX512BW"]],
4699
+ "_mm256_maskz_min_epi8"=>[["AVX512VL", "AVX512BW"]],
4700
4700
  "_mm512_mask_min_epi8"=>["AVX512BW"],
4701
4701
  "_mm512_maskz_min_epi8"=>["AVX512BW"],
4702
4702
  "_mm512_min_epi8"=>["AVX512BW"],
4703
- "_mm_mask_min_epi8"=>["AVX512VL", "AVX512BW"],
4704
- "_mm_maskz_min_epi8"=>["AVX512VL", "AVX512BW"],
4705
- "_mm256_mask_min_epi32"=>["AVX512VL", "AVX512F"],
4706
- "_mm256_maskz_min_epi32"=>["AVX512VL", "AVX512F"],
4707
- "_mm_mask_min_epi32"=>["AVX512VL", "AVX512F"],
4708
- "_mm_maskz_min_epi32"=>["AVX512VL", "AVX512F"],
4709
- "_mm256_mask_min_epi64"=>["AVX512VL", "AVX512F"],
4710
- "_mm256_maskz_min_epi64"=>["AVX512VL", "AVX512F"],
4711
- "_mm256_min_epi64"=>["AVX512VL", "AVX512F"],
4712
- "_mm_mask_min_epi64"=>["AVX512VL", "AVX512F"],
4713
- "_mm_maskz_min_epi64"=>["AVX512VL", "AVX512F"],
4714
- "_mm_min_epi64"=>["AVX512VL", "AVX512F"],
4715
- "_mm256_mask_min_epi16"=>["AVX512VL", "AVX512BW"],
4716
- "_mm256_maskz_min_epi16"=>["AVX512VL", "AVX512BW"],
4703
+ "_mm_mask_min_epi8"=>[["AVX512VL", "AVX512BW"]],
4704
+ "_mm_maskz_min_epi8"=>[["AVX512VL", "AVX512BW"]],
4705
+ "_mm256_mask_min_epi32"=>[["AVX512VL", "AVX512F"]],
4706
+ "_mm256_maskz_min_epi32"=>[["AVX512VL", "AVX512F"]],
4707
+ "_mm_mask_min_epi32"=>[["AVX512VL", "AVX512F"]],
4708
+ "_mm_maskz_min_epi32"=>[["AVX512VL", "AVX512F"]],
4709
+ "_mm256_mask_min_epi64"=>[["AVX512VL", "AVX512F"]],
4710
+ "_mm256_maskz_min_epi64"=>[["AVX512VL", "AVX512F"]],
4711
+ "_mm256_min_epi64"=>[["AVX512VL", "AVX512F"]],
4712
+ "_mm_mask_min_epi64"=>[["AVX512VL", "AVX512F"]],
4713
+ "_mm_maskz_min_epi64"=>[["AVX512VL", "AVX512F"]],
4714
+ "_mm_min_epi64"=>[["AVX512VL", "AVX512F"]],
4715
+ "_mm256_mask_min_epi16"=>[["AVX512VL", "AVX512BW"]],
4716
+ "_mm256_maskz_min_epi16"=>[["AVX512VL", "AVX512BW"]],
4717
4717
  "_mm512_mask_min_epi16"=>["AVX512BW"],
4718
4718
  "_mm512_maskz_min_epi16"=>["AVX512BW"],
4719
4719
  "_mm512_min_epi16"=>["AVX512BW"],
4720
- "_mm_mask_min_epi16"=>["AVX512VL", "AVX512BW"],
4721
- "_mm_maskz_min_epi16"=>["AVX512VL", "AVX512BW"],
4722
- "_mm256_mask_min_epu8"=>["AVX512VL", "AVX512BW"],
4723
- "_mm256_maskz_min_epu8"=>["AVX512VL", "AVX512BW"],
4720
+ "_mm_mask_min_epi16"=>[["AVX512VL", "AVX512BW"]],
4721
+ "_mm_maskz_min_epi16"=>[["AVX512VL", "AVX512BW"]],
4722
+ "_mm256_mask_min_epu8"=>[["AVX512VL", "AVX512BW"]],
4723
+ "_mm256_maskz_min_epu8"=>[["AVX512VL", "AVX512BW"]],
4724
4724
  "_mm512_mask_min_epu8"=>["AVX512BW"],
4725
4725
  "_mm512_maskz_min_epu8"=>["AVX512BW"],
4726
4726
  "_mm512_min_epu8"=>["AVX512BW"],
4727
- "_mm_mask_min_epu8"=>["AVX512VL", "AVX512BW"],
4728
- "_mm_maskz_min_epu8"=>["AVX512VL", "AVX512BW"],
4729
- "_mm256_mask_min_epu32"=>["AVX512VL", "AVX512F"],
4730
- "_mm256_maskz_min_epu32"=>["AVX512VL", "AVX512F"],
4731
- "_mm_mask_min_epu32"=>["AVX512VL", "AVX512F"],
4732
- "_mm_maskz_min_epu32"=>["AVX512VL", "AVX512F"],
4733
- "_mm256_mask_min_epu64"=>["AVX512VL", "AVX512F"],
4734
- "_mm256_maskz_min_epu64"=>["AVX512VL", "AVX512F"],
4735
- "_mm256_min_epu64"=>["AVX512VL", "AVX512F"],
4736
- "_mm_mask_min_epu64"=>["AVX512VL", "AVX512F"],
4737
- "_mm_maskz_min_epu64"=>["AVX512VL", "AVX512F"],
4738
- "_mm_min_epu64"=>["AVX512VL", "AVX512F"],
4739
- "_mm256_mask_min_epu16"=>["AVX512VL", "AVX512BW"],
4740
- "_mm256_maskz_min_epu16"=>["AVX512VL", "AVX512BW"],
4727
+ "_mm_mask_min_epu8"=>[["AVX512VL", "AVX512BW"]],
4728
+ "_mm_maskz_min_epu8"=>[["AVX512VL", "AVX512BW"]],
4729
+ "_mm256_mask_min_epu32"=>[["AVX512VL", "AVX512F"]],
4730
+ "_mm256_maskz_min_epu32"=>[["AVX512VL", "AVX512F"]],
4731
+ "_mm_mask_min_epu32"=>[["AVX512VL", "AVX512F"]],
4732
+ "_mm_maskz_min_epu32"=>[["AVX512VL", "AVX512F"]],
4733
+ "_mm256_mask_min_epu64"=>[["AVX512VL", "AVX512F"]],
4734
+ "_mm256_maskz_min_epu64"=>[["AVX512VL", "AVX512F"]],
4735
+ "_mm256_min_epu64"=>[["AVX512VL", "AVX512F"]],
4736
+ "_mm_mask_min_epu64"=>[["AVX512VL", "AVX512F"]],
4737
+ "_mm_maskz_min_epu64"=>[["AVX512VL", "AVX512F"]],
4738
+ "_mm_min_epu64"=>[["AVX512VL", "AVX512F"]],
4739
+ "_mm256_mask_min_epu16"=>[["AVX512VL", "AVX512BW"]],
4740
+ "_mm256_maskz_min_epu16"=>[["AVX512VL", "AVX512BW"]],
4741
4741
  "_mm512_mask_min_epu16"=>["AVX512BW"],
4742
4742
  "_mm512_maskz_min_epu16"=>["AVX512BW"],
4743
4743
  "_mm512_min_epu16"=>["AVX512BW"],
4744
- "_mm_mask_min_epu16"=>["AVX512VL", "AVX512BW"],
4745
- "_mm_maskz_min_epu16"=>["AVX512VL", "AVX512BW"],
4746
- "_mm256_movepi8_mask"=>["AVX512VL", "AVX512BW"],
4744
+ "_mm_mask_min_epu16"=>[["AVX512VL", "AVX512BW"]],
4745
+ "_mm_maskz_min_epu16"=>[["AVX512VL", "AVX512BW"]],
4746
+ "_mm256_movepi8_mask"=>[["AVX512VL", "AVX512BW"]],
4747
4747
  "_mm512_movepi8_mask"=>["AVX512BW"],
4748
- "_mm_movepi8_mask"=>["AVX512VL", "AVX512BW"],
4749
- "_mm256_movepi32_mask"=>["AVX512VL", "AVX512DQ"],
4748
+ "_mm_movepi8_mask"=>[["AVX512VL", "AVX512BW"]],
4749
+ "_mm256_movepi32_mask"=>[["AVX512VL", "AVX512DQ"]],
4750
4750
  "_mm512_movepi32_mask"=>["AVX512DQ"],
4751
- "_mm_movepi32_mask"=>["AVX512VL", "AVX512DQ"],
4752
- "_mm256_cvtepi32_epi8"=>["AVX512VL", "AVX512F"],
4753
- "_mm256_mask_cvtepi32_epi8"=>["AVX512VL", "AVX512F"],
4754
- "_mm256_mask_cvtepi32_storeu_epi8"=>["AVX512VL", "AVX512F"],
4755
- "_mm256_maskz_cvtepi32_epi8"=>["AVX512VL", "AVX512F"],
4756
- "_mm_cvtepi32_epi8"=>["AVX512VL", "AVX512F"],
4757
- "_mm_mask_cvtepi32_epi8"=>["AVX512VL", "AVX512F"],
4758
- "_mm_mask_cvtepi32_storeu_epi8"=>["AVX512VL", "AVX512F"],
4759
- "_mm_maskz_cvtepi32_epi8"=>["AVX512VL", "AVX512F"],
4760
- "_mm256_cvtepi32_epi16"=>["AVX512VL", "AVX512F"],
4761
- "_mm256_mask_cvtepi32_epi16"=>["AVX512VL", "AVX512F"],
4762
- "_mm256_mask_cvtepi32_storeu_epi16"=>["AVX512VL", "AVX512F"],
4763
- "_mm256_maskz_cvtepi32_epi16"=>["AVX512VL", "AVX512F"],
4764
- "_mm_cvtepi32_epi16"=>["AVX512VL", "AVX512F"],
4765
- "_mm_mask_cvtepi32_epi16"=>["AVX512VL", "AVX512F"],
4766
- "_mm_mask_cvtepi32_storeu_epi16"=>["AVX512VL", "AVX512F"],
4767
- "_mm_maskz_cvtepi32_epi16"=>["AVX512VL", "AVX512F"],
4768
- "_mm256_movm_epi8"=>["AVX512VL", "AVX512BW"],
4751
+ "_mm_movepi32_mask"=>[["AVX512VL", "AVX512DQ"]],
4752
+ "_mm256_cvtepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4753
+ "_mm256_mask_cvtepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4754
+ "_mm256_mask_cvtepi32_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4755
+ "_mm256_maskz_cvtepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4756
+ "_mm_cvtepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4757
+ "_mm_mask_cvtepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4758
+ "_mm_mask_cvtepi32_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4759
+ "_mm_maskz_cvtepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4760
+ "_mm256_cvtepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4761
+ "_mm256_mask_cvtepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4762
+ "_mm256_mask_cvtepi32_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4763
+ "_mm256_maskz_cvtepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4764
+ "_mm_cvtepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4765
+ "_mm_mask_cvtepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4766
+ "_mm_mask_cvtepi32_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4767
+ "_mm_maskz_cvtepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4768
+ "_mm256_movm_epi8"=>[["AVX512VL", "AVX512BW"]],
4769
4769
  "_mm512_movm_epi8"=>["AVX512BW"],
4770
- "_mm_movm_epi8"=>["AVX512BW", "AVX512VL"],
4771
- "_mm256_movm_epi32"=>["AVX512VL", "AVX512DQ"],
4770
+ "_mm_movm_epi8"=>[["AVX512BW", "AVX512VL"]],
4771
+ "_mm256_movm_epi32"=>[["AVX512VL", "AVX512DQ"]],
4772
4772
  "_mm512_movm_epi32"=>["AVX512DQ"],
4773
- "_mm_movm_epi32"=>["AVX512VL", "AVX512DQ"],
4774
- "_mm256_movm_epi64"=>["AVX512VL", "AVX512DQ"],
4773
+ "_mm_movm_epi32"=>[["AVX512VL", "AVX512DQ"]],
4774
+ "_mm256_movm_epi64"=>[["AVX512VL", "AVX512DQ"]],
4775
4775
  "_mm512_movm_epi64"=>["AVX512DQ"],
4776
- "_mm_movm_epi64"=>["AVX512VL", "AVX512DQ"],
4777
- "_mm256_movm_epi16"=>["AVX512VL", "AVX512BW"],
4776
+ "_mm_movm_epi64"=>[["AVX512VL", "AVX512DQ"]],
4777
+ "_mm256_movm_epi16"=>[["AVX512VL", "AVX512BW"]],
4778
4778
  "_mm512_movm_epi16"=>["AVX512BW"],
4779
- "_mm_movm_epi16"=>["AVX512VL", "AVX512BW"],
4780
- "_mm256_movepi64_mask"=>["AVX512VL", "AVX512DQ"],
4779
+ "_mm_movm_epi16"=>[["AVX512VL", "AVX512BW"]],
4780
+ "_mm256_movepi64_mask"=>[["AVX512VL", "AVX512DQ"]],
4781
4781
  "_mm512_movepi64_mask"=>["AVX512DQ"],
4782
- "_mm_movepi64_mask"=>["AVX512VL", "AVX512DQ"],
4783
- "_mm256_cvtepi64_epi8"=>["AVX512VL", "AVX512F"],
4784
- "_mm256_mask_cvtepi64_epi8"=>["AVX512VL", "AVX512F"],
4785
- "_mm256_mask_cvtepi64_storeu_epi8"=>["AVX512VL", "AVX512F"],
4786
- "_mm256_maskz_cvtepi64_epi8"=>["AVX512VL", "AVX512F"],
4787
- "_mm_cvtepi64_epi8"=>["AVX512VL", "AVX512F"],
4788
- "_mm_mask_cvtepi64_epi8"=>["AVX512VL", "AVX512F"],
4789
- "_mm_mask_cvtepi64_storeu_epi8"=>["AVX512VL", "AVX512F"],
4790
- "_mm_maskz_cvtepi64_epi8"=>["AVX512VL", "AVX512F"],
4791
- "_mm256_cvtepi64_epi32"=>["AVX512VL", "AVX512F"],
4792
- "_mm256_mask_cvtepi64_epi32"=>["AVX512VL", "AVX512F"],
4793
- "_mm256_mask_cvtepi64_storeu_epi32"=>["AVX512VL", "AVX512F"],
4794
- "_mm256_maskz_cvtepi64_epi32"=>["AVX512VL", "AVX512F"],
4795
- "_mm_cvtepi64_epi32"=>["AVX512VL", "AVX512F"],
4796
- "_mm_mask_cvtepi64_epi32"=>["AVX512VL", "AVX512F"],
4797
- "_mm_mask_cvtepi64_storeu_epi32"=>["AVX512VL", "AVX512F"],
4798
- "_mm_maskz_cvtepi64_epi32"=>["AVX512VL", "AVX512F"],
4799
- "_mm256_cvtepi64_epi16"=>["AVX512VL", "AVX512F"],
4800
- "_mm256_mask_cvtepi64_epi16"=>["AVX512VL", "AVX512F"],
4801
- "_mm256_mask_cvtepi64_storeu_epi16"=>["AVX512VL", "AVX512F"],
4802
- "_mm256_maskz_cvtepi64_epi16"=>["AVX512VL", "AVX512F"],
4803
- "_mm_cvtepi64_epi16"=>["AVX512VL", "AVX512F"],
4804
- "_mm_mask_cvtepi64_epi16"=>["AVX512VL", "AVX512F"],
4805
- "_mm_mask_cvtepi64_storeu_epi16"=>["AVX512VL", "AVX512F"],
4806
- "_mm_maskz_cvtepi64_epi16"=>["AVX512VL", "AVX512F"],
4807
- "_mm256_cvtsepi32_epi8"=>["AVX512VL", "AVX512F"],
4808
- "_mm256_mask_cvtsepi32_epi8"=>["AVX512VL", "AVX512F"],
4809
- "_mm256_mask_cvtsepi32_storeu_epi8"=>["AVX512VL", "AVX512F"],
4810
- "_mm256_maskz_cvtsepi32_epi8"=>["AVX512VL", "AVX512F"],
4811
- "_mm_cvtsepi32_epi8"=>["AVX512VL", "AVX512F"],
4812
- "_mm_mask_cvtsepi32_epi8"=>["AVX512VL", "AVX512F"],
4813
- "_mm_mask_cvtsepi32_storeu_epi8"=>["AVX512VL", "AVX512F"],
4814
- "_mm_maskz_cvtsepi32_epi8"=>["AVX512VL", "AVX512F"],
4815
- "_mm256_cvtsepi32_epi16"=>["AVX512VL", "AVX512F"],
4816
- "_mm256_mask_cvtsepi32_epi16"=>["AVX512VL", "AVX512F"],
4817
- "_mm256_mask_cvtsepi32_storeu_epi16"=>["AVX512VL", "AVX512F"],
4818
- "_mm256_maskz_cvtsepi32_epi16"=>["AVX512VL", "AVX512F"],
4819
- "_mm_cvtsepi32_epi16"=>["AVX512VL", "AVX512F"],
4820
- "_mm_mask_cvtsepi32_epi16"=>["AVX512VL", "AVX512F"],
4821
- "_mm_mask_cvtsepi32_storeu_epi16"=>["AVX512VL", "AVX512F"],
4822
- "_mm_maskz_cvtsepi32_epi16"=>["AVX512VL", "AVX512F"],
4823
- "_mm256_cvtsepi64_epi8"=>["AVX512VL", "AVX512F"],
4824
- "_mm256_mask_cvtsepi64_epi8"=>["AVX512VL", "AVX512F"],
4825
- "_mm256_mask_cvtsepi64_storeu_epi8"=>["AVX512VL", "AVX512F"],
4826
- "_mm256_maskz_cvtsepi64_epi8"=>["AVX512VL", "AVX512F"],
4827
- "_mm_cvtsepi64_epi8"=>["AVX512VL", "AVX512F"],
4828
- "_mm_mask_cvtsepi64_epi8"=>["AVX512VL", "AVX512F"],
4829
- "_mm_mask_cvtsepi64_storeu_epi8"=>["AVX512VL", "AVX512F"],
4830
- "_mm_maskz_cvtsepi64_epi8"=>["AVX512VL", "AVX512F"],
4831
- "_mm256_cvtsepi64_epi32"=>["AVX512VL", "AVX512F"],
4832
- "_mm256_mask_cvtsepi64_epi32"=>["AVX512VL", "AVX512F"],
4833
- "_mm256_mask_cvtsepi64_storeu_epi32"=>["AVX512VL", "AVX512F"],
4834
- "_mm256_maskz_cvtsepi64_epi32"=>["AVX512VL", "AVX512F"],
4835
- "_mm_cvtsepi64_epi32"=>["AVX512VL", "AVX512F"],
4836
- "_mm_mask_cvtsepi64_epi32"=>["AVX512VL", "AVX512F"],
4837
- "_mm_mask_cvtsepi64_storeu_epi32"=>["AVX512VL", "AVX512F"],
4838
- "_mm_maskz_cvtsepi64_epi32"=>["AVX512VL", "AVX512F"],
4839
- "_mm256_cvtsepi64_epi16"=>["AVX512VL", "AVX512F"],
4840
- "_mm256_mask_cvtsepi64_epi16"=>["AVX512VL", "AVX512F"],
4841
- "_mm256_mask_cvtsepi64_storeu_epi16"=>["AVX512VL", "AVX512F"],
4842
- "_mm256_maskz_cvtsepi64_epi16"=>["AVX512VL", "AVX512F"],
4843
- "_mm_cvtsepi64_epi16"=>["AVX512VL", "AVX512F"],
4844
- "_mm_mask_cvtsepi64_epi16"=>["AVX512VL", "AVX512F"],
4845
- "_mm_mask_cvtsepi64_storeu_epi16"=>["AVX512VL", "AVX512F"],
4846
- "_mm_maskz_cvtsepi64_epi16"=>["AVX512VL", "AVX512F"],
4847
- "_mm256_cvtsepi16_epi8"=>["AVX512VL", "AVX512BW"],
4848
- "_mm256_mask_cvtsepi16_epi8"=>["AVX512VL", "AVX512BW"],
4849
- "_mm256_mask_cvtsepi16_storeu_epi8"=>["AVX512VL", "AVX512BW"],
4850
- "_mm256_maskz_cvtsepi16_epi8"=>["AVX512VL", "AVX512BW"],
4782
+ "_mm_movepi64_mask"=>[["AVX512VL", "AVX512DQ"]],
4783
+ "_mm256_cvtepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4784
+ "_mm256_mask_cvtepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4785
+ "_mm256_mask_cvtepi64_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4786
+ "_mm256_maskz_cvtepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4787
+ "_mm_cvtepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4788
+ "_mm_mask_cvtepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4789
+ "_mm_mask_cvtepi64_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4790
+ "_mm_maskz_cvtepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4791
+ "_mm256_cvtepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4792
+ "_mm256_mask_cvtepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4793
+ "_mm256_mask_cvtepi64_storeu_epi32"=>[["AVX512VL", "AVX512F"]],
4794
+ "_mm256_maskz_cvtepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4795
+ "_mm_cvtepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4796
+ "_mm_mask_cvtepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4797
+ "_mm_mask_cvtepi64_storeu_epi32"=>[["AVX512VL", "AVX512F"]],
4798
+ "_mm_maskz_cvtepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4799
+ "_mm256_cvtepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4800
+ "_mm256_mask_cvtepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4801
+ "_mm256_mask_cvtepi64_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4802
+ "_mm256_maskz_cvtepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4803
+ "_mm_cvtepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4804
+ "_mm_mask_cvtepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4805
+ "_mm_mask_cvtepi64_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4806
+ "_mm_maskz_cvtepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4807
+ "_mm256_cvtsepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4808
+ "_mm256_mask_cvtsepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4809
+ "_mm256_mask_cvtsepi32_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4810
+ "_mm256_maskz_cvtsepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4811
+ "_mm_cvtsepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4812
+ "_mm_mask_cvtsepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4813
+ "_mm_mask_cvtsepi32_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4814
+ "_mm_maskz_cvtsepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4815
+ "_mm256_cvtsepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4816
+ "_mm256_mask_cvtsepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4817
+ "_mm256_mask_cvtsepi32_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4818
+ "_mm256_maskz_cvtsepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4819
+ "_mm_cvtsepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4820
+ "_mm_mask_cvtsepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4821
+ "_mm_mask_cvtsepi32_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4822
+ "_mm_maskz_cvtsepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4823
+ "_mm256_cvtsepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4824
+ "_mm256_mask_cvtsepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4825
+ "_mm256_mask_cvtsepi64_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4826
+ "_mm256_maskz_cvtsepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4827
+ "_mm_cvtsepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4828
+ "_mm_mask_cvtsepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4829
+ "_mm_mask_cvtsepi64_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4830
+ "_mm_maskz_cvtsepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4831
+ "_mm256_cvtsepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4832
+ "_mm256_mask_cvtsepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4833
+ "_mm256_mask_cvtsepi64_storeu_epi32"=>[["AVX512VL", "AVX512F"]],
4834
+ "_mm256_maskz_cvtsepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4835
+ "_mm_cvtsepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4836
+ "_mm_mask_cvtsepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4837
+ "_mm_mask_cvtsepi64_storeu_epi32"=>[["AVX512VL", "AVX512F"]],
4838
+ "_mm_maskz_cvtsepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4839
+ "_mm256_cvtsepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4840
+ "_mm256_mask_cvtsepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4841
+ "_mm256_mask_cvtsepi64_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4842
+ "_mm256_maskz_cvtsepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4843
+ "_mm_cvtsepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4844
+ "_mm_mask_cvtsepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4845
+ "_mm_mask_cvtsepi64_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4846
+ "_mm_maskz_cvtsepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4847
+ "_mm256_cvtsepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4848
+ "_mm256_mask_cvtsepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4849
+ "_mm256_mask_cvtsepi16_storeu_epi8"=>[["AVX512VL", "AVX512BW"]],
4850
+ "_mm256_maskz_cvtsepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4851
4851
  "_mm512_cvtsepi16_epi8"=>["AVX512BW"],
4852
4852
  "_mm512_mask_cvtsepi16_epi8"=>["AVX512BW"],
4853
4853
  "_mm512_mask_cvtsepi16_storeu_epi8"=>["AVX512BW"],
4854
4854
  "_mm512_maskz_cvtsepi16_epi8"=>["AVX512BW"],
4855
- "_mm_cvtsepi16_epi8"=>["AVX512VL", "AVX512BW"],
4856
- "_mm_mask_cvtsepi16_epi8"=>["AVX512VL", "AVX512BW"],
4857
- "_mm_mask_cvtsepi16_storeu_epi8"=>["AVX512VL", "AVX512BW"],
4858
- "_mm_maskz_cvtsepi16_epi8"=>["AVX512VL", "AVX512BW"],
4859
- "_mm256_mask_cvtepi8_epi32"=>["AVX512VL", "AVX512F"],
4860
- "_mm256_maskz_cvtepi8_epi32"=>["AVX512VL", "AVX512F"],
4861
- "_mm_mask_cvtepi8_epi32"=>["AVX512VL", "AVX512F"],
4862
- "_mm_maskz_cvtepi8_epi32"=>["AVX512VL", "AVX512F"],
4863
- "_mm256_mask_cvtepi8_epi64"=>["AVX512VL", "AVX512F"],
4864
- "_mm256_maskz_cvtepi8_epi64"=>["AVX512VL", "AVX512F"],
4865
- "_mm_mask_cvtepi8_epi64"=>["AVX512VL", "AVX512F"],
4866
- "_mm_maskz_cvtepi8_epi64"=>["AVX512VL", "AVX512F"],
4867
- "_mm256_mask_cvtepi8_epi16"=>["AVX512VL", "AVX512BW"],
4868
- "_mm256_maskz_cvtepi8_epi16"=>["AVX512VL", "AVX512BW"],
4855
+ "_mm_cvtsepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4856
+ "_mm_mask_cvtsepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4857
+ "_mm_mask_cvtsepi16_storeu_epi8"=>[["AVX512VL", "AVX512BW"]],
4858
+ "_mm_maskz_cvtsepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4859
+ "_mm256_mask_cvtepi8_epi32"=>[["AVX512VL", "AVX512F"]],
4860
+ "_mm256_maskz_cvtepi8_epi32"=>[["AVX512VL", "AVX512F"]],
4861
+ "_mm_mask_cvtepi8_epi32"=>[["AVX512VL", "AVX512F"]],
4862
+ "_mm_maskz_cvtepi8_epi32"=>[["AVX512VL", "AVX512F"]],
4863
+ "_mm256_mask_cvtepi8_epi64"=>[["AVX512VL", "AVX512F"]],
4864
+ "_mm256_maskz_cvtepi8_epi64"=>[["AVX512VL", "AVX512F"]],
4865
+ "_mm_mask_cvtepi8_epi64"=>[["AVX512VL", "AVX512F"]],
4866
+ "_mm_maskz_cvtepi8_epi64"=>[["AVX512VL", "AVX512F"]],
4867
+ "_mm256_mask_cvtepi8_epi16"=>[["AVX512VL", "AVX512BW"]],
4868
+ "_mm256_maskz_cvtepi8_epi16"=>[["AVX512VL", "AVX512BW"]],
4869
4869
  "_mm512_cvtepi8_epi16"=>["AVX512BW"],
4870
4870
  "_mm512_mask_cvtepi8_epi16"=>["AVX512BW"],
4871
4871
  "_mm512_maskz_cvtepi8_epi16"=>["AVX512BW"],
4872
- "_mm_mask_cvtepi8_epi16"=>["AVX512VL", "AVX512BW"],
4873
- "_mm_maskz_cvtepi8_epi16"=>["AVX512VL", "AVX512BW"],
4874
- "_mm256_mask_cvtepi32_epi64"=>["AVX512VL", "AVX512F"],
4875
- "_mm256_maskz_cvtepi32_epi64"=>["AVX512VL", "AVX512F"],
4876
- "_mm_mask_cvtepi32_epi64"=>["AVX512VL", "AVX512F"],
4877
- "_mm_maskz_cvtepi32_epi64"=>["AVX512VL", "AVX512F"],
4878
- "_mm256_mask_cvtepi16_epi32"=>["AVX512VL", "AVX512F"],
4879
- "_mm256_maskz_cvtepi16_epi32"=>["AVX512VL", "AVX512F"],
4880
- "_mm_mask_cvtepi16_epi32"=>["AVX512VL", "AVX512F"],
4881
- "_mm_maskz_cvtepi16_epi32"=>["AVX512VL", "AVX512F"],
4882
- "_mm256_mask_cvtepi16_epi64"=>["AVX512VL", "AVX512F"],
4883
- "_mm256_maskz_cvtepi16_epi64"=>["AVX512VL", "AVX512F"],
4884
- "_mm_mask_cvtepi16_epi64"=>["AVX512VL", "AVX512F"],
4885
- "_mm_maskz_cvtepi16_epi64"=>["AVX512VL", "AVX512F"],
4886
- "_mm256_cvtusepi32_epi8"=>["AVX512VL", "AVX512F"],
4887
- "_mm256_mask_cvtusepi32_epi8"=>["AVX512VL", "AVX512F"],
4888
- "_mm256_mask_cvtusepi32_storeu_epi8"=>["AVX512VL", "AVX512F"],
4889
- "_mm256_maskz_cvtusepi32_epi8"=>["AVX512VL", "AVX512F"],
4890
- "_mm_cvtusepi32_epi8"=>["AVX512VL", "AVX512F"],
4891
- "_mm_mask_cvtusepi32_epi8"=>["AVX512VL", "AVX512F"],
4892
- "_mm_mask_cvtusepi32_storeu_epi8"=>["AVX512VL", "AVX512F"],
4893
- "_mm_maskz_cvtusepi32_epi8"=>["AVX512VL", "AVX512F"],
4894
- "_mm256_cvtusepi32_epi16"=>["AVX512VL", "AVX512F"],
4895
- "_mm256_mask_cvtusepi32_epi16"=>["AVX512VL", "AVX512F"],
4896
- "_mm256_mask_cvtusepi32_storeu_epi16"=>["AVX512VL", "AVX512F"],
4897
- "_mm256_maskz_cvtusepi32_epi16"=>["AVX512VL", "AVX512F"],
4898
- "_mm_cvtusepi32_epi16"=>["AVX512VL", "AVX512F"],
4899
- "_mm_mask_cvtusepi32_epi16"=>["AVX512VL", "AVX512F"],
4900
- "_mm_mask_cvtusepi32_storeu_epi16"=>["AVX512VL", "AVX512F"],
4901
- "_mm_maskz_cvtusepi32_epi16"=>["AVX512VL", "AVX512F"],
4902
- "_mm256_cvtusepi64_epi8"=>["AVX512VL", "AVX512F"],
4903
- "_mm256_mask_cvtusepi64_epi8"=>["AVX512VL", "AVX512F"],
4904
- "_mm256_mask_cvtusepi64_storeu_epi8"=>["AVX512VL", "AVX512F"],
4905
- "_mm256_maskz_cvtusepi64_epi8"=>["AVX512VL", "AVX512F"],
4906
- "_mm_cvtusepi64_epi8"=>["AVX512VL", "AVX512F"],
4907
- "_mm_mask_cvtusepi64_epi8"=>["AVX512VL", "AVX512F"],
4908
- "_mm_mask_cvtusepi64_storeu_epi8"=>["AVX512VL", "AVX512F"],
4909
- "_mm_maskz_cvtusepi64_epi8"=>["AVX512VL", "AVX512F"],
4910
- "_mm256_cvtusepi64_epi32"=>["AVX512VL", "AVX512F"],
4911
- "_mm256_mask_cvtusepi64_epi32"=>["AVX512VL", "AVX512F"],
4912
- "_mm256_mask_cvtusepi64_storeu_epi32"=>["AVX512VL", "AVX512F"],
4913
- "_mm256_maskz_cvtusepi64_epi32"=>["AVX512VL", "AVX512F"],
4914
- "_mm_cvtusepi64_epi32"=>["AVX512VL", "AVX512F"],
4915
- "_mm_mask_cvtusepi64_epi32"=>["AVX512VL", "AVX512F"],
4916
- "_mm_mask_cvtusepi64_storeu_epi32"=>["AVX512VL", "AVX512F"],
4917
- "_mm_maskz_cvtusepi64_epi32"=>["AVX512VL", "AVX512F"],
4918
- "_mm256_cvtusepi64_epi16"=>["AVX512VL", "AVX512F"],
4919
- "_mm256_mask_cvtusepi64_epi16"=>["AVX512VL", "AVX512F"],
4920
- "_mm256_mask_cvtusepi64_storeu_epi16"=>["AVX512VL", "AVX512F"],
4921
- "_mm256_maskz_cvtusepi64_epi16"=>["AVX512VL", "AVX512F"],
4922
- "_mm_cvtusepi64_epi16"=>["AVX512VL", "AVX512F"],
4923
- "_mm_mask_cvtusepi64_epi16"=>["AVX512VL", "AVX512F"],
4924
- "_mm_mask_cvtusepi64_storeu_epi16"=>["AVX512VL", "AVX512F"],
4925
- "_mm_maskz_cvtusepi64_epi16"=>["AVX512VL", "AVX512F"],
4926
- "_mm256_cvtusepi16_epi8"=>["AVX512VL", "AVX512BW"],
4927
- "_mm256_mask_cvtusepi16_epi8"=>["AVX512VL", "AVX512BW"],
4928
- "_mm256_mask_cvtusepi16_storeu_epi8"=>["AVX512VL", "AVX512BW"],
4929
- "_mm256_maskz_cvtusepi16_epi8"=>["AVX512VL", "AVX512BW"],
4872
+ "_mm_mask_cvtepi8_epi16"=>[["AVX512VL", "AVX512BW"]],
4873
+ "_mm_maskz_cvtepi8_epi16"=>[["AVX512VL", "AVX512BW"]],
4874
+ "_mm256_mask_cvtepi32_epi64"=>[["AVX512VL", "AVX512F"]],
4875
+ "_mm256_maskz_cvtepi32_epi64"=>[["AVX512VL", "AVX512F"]],
4876
+ "_mm_mask_cvtepi32_epi64"=>[["AVX512VL", "AVX512F"]],
4877
+ "_mm_maskz_cvtepi32_epi64"=>[["AVX512VL", "AVX512F"]],
4878
+ "_mm256_mask_cvtepi16_epi32"=>[["AVX512VL", "AVX512F"]],
4879
+ "_mm256_maskz_cvtepi16_epi32"=>[["AVX512VL", "AVX512F"]],
4880
+ "_mm_mask_cvtepi16_epi32"=>[["AVX512VL", "AVX512F"]],
4881
+ "_mm_maskz_cvtepi16_epi32"=>[["AVX512VL", "AVX512F"]],
4882
+ "_mm256_mask_cvtepi16_epi64"=>[["AVX512VL", "AVX512F"]],
4883
+ "_mm256_maskz_cvtepi16_epi64"=>[["AVX512VL", "AVX512F"]],
4884
+ "_mm_mask_cvtepi16_epi64"=>[["AVX512VL", "AVX512F"]],
4885
+ "_mm_maskz_cvtepi16_epi64"=>[["AVX512VL", "AVX512F"]],
4886
+ "_mm256_cvtusepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4887
+ "_mm256_mask_cvtusepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4888
+ "_mm256_mask_cvtusepi32_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4889
+ "_mm256_maskz_cvtusepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4890
+ "_mm_cvtusepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4891
+ "_mm_mask_cvtusepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4892
+ "_mm_mask_cvtusepi32_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4893
+ "_mm_maskz_cvtusepi32_epi8"=>[["AVX512VL", "AVX512F"]],
4894
+ "_mm256_cvtusepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4895
+ "_mm256_mask_cvtusepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4896
+ "_mm256_mask_cvtusepi32_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4897
+ "_mm256_maskz_cvtusepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4898
+ "_mm_cvtusepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4899
+ "_mm_mask_cvtusepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4900
+ "_mm_mask_cvtusepi32_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4901
+ "_mm_maskz_cvtusepi32_epi16"=>[["AVX512VL", "AVX512F"]],
4902
+ "_mm256_cvtusepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4903
+ "_mm256_mask_cvtusepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4904
+ "_mm256_mask_cvtusepi64_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4905
+ "_mm256_maskz_cvtusepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4906
+ "_mm_cvtusepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4907
+ "_mm_mask_cvtusepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4908
+ "_mm_mask_cvtusepi64_storeu_epi8"=>[["AVX512VL", "AVX512F"]],
4909
+ "_mm_maskz_cvtusepi64_epi8"=>[["AVX512VL", "AVX512F"]],
4910
+ "_mm256_cvtusepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4911
+ "_mm256_mask_cvtusepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4912
+ "_mm256_mask_cvtusepi64_storeu_epi32"=>[["AVX512VL", "AVX512F"]],
4913
+ "_mm256_maskz_cvtusepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4914
+ "_mm_cvtusepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4915
+ "_mm_mask_cvtusepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4916
+ "_mm_mask_cvtusepi64_storeu_epi32"=>[["AVX512VL", "AVX512F"]],
4917
+ "_mm_maskz_cvtusepi64_epi32"=>[["AVX512VL", "AVX512F"]],
4918
+ "_mm256_cvtusepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4919
+ "_mm256_mask_cvtusepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4920
+ "_mm256_mask_cvtusepi64_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4921
+ "_mm256_maskz_cvtusepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4922
+ "_mm_cvtusepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4923
+ "_mm_mask_cvtusepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4924
+ "_mm_mask_cvtusepi64_storeu_epi16"=>[["AVX512VL", "AVX512F"]],
4925
+ "_mm_maskz_cvtusepi64_epi16"=>[["AVX512VL", "AVX512F"]],
4926
+ "_mm256_cvtusepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4927
+ "_mm256_mask_cvtusepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4928
+ "_mm256_mask_cvtusepi16_storeu_epi8"=>[["AVX512VL", "AVX512BW"]],
4929
+ "_mm256_maskz_cvtusepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4930
4930
  "_mm512_cvtusepi16_epi8"=>["AVX512BW"],
4931
4931
  "_mm512_mask_cvtusepi16_epi8"=>["AVX512BW"],
4932
4932
  "_mm512_mask_cvtusepi16_storeu_epi8"=>["AVX512BW"],
4933
4933
  "_mm512_maskz_cvtusepi16_epi8"=>["AVX512BW"],
4934
- "_mm_cvtusepi16_epi8"=>["AVX512VL", "AVX512BW"],
4935
- "_mm_mask_cvtusepi16_epi8"=>["AVX512VL", "AVX512BW"],
4936
- "_mm_mask_cvtusepi16_storeu_epi8"=>["AVX512VL", "AVX512BW"],
4937
- "_mm_maskz_cvtusepi16_epi8"=>["AVX512VL", "AVX512BW"],
4938
- "_mm256_movepi16_mask"=>["AVX512VL", "AVX512BW"],
4934
+ "_mm_cvtusepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4935
+ "_mm_mask_cvtusepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4936
+ "_mm_mask_cvtusepi16_storeu_epi8"=>[["AVX512VL", "AVX512BW"]],
4937
+ "_mm_maskz_cvtusepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4938
+ "_mm256_movepi16_mask"=>[["AVX512VL", "AVX512BW"]],
4939
4939
  "_mm512_movepi16_mask"=>["AVX512BW"],
4940
- "_mm_movepi16_mask"=>["AVX512VL", "AVX512BW"],
4941
- "_mm256_cvtepi16_epi8"=>["AVX512VL", "AVX512BW"],
4942
- "_mm256_mask_cvtepi16_epi8"=>["AVX512VL", "AVX512BW"],
4943
- "_mm256_mask_cvtepi16_storeu_epi8"=>["AVX512VL", "AVX512BW"],
4944
- "_mm256_maskz_cvtepi16_epi8"=>["AVX512VL", "AVX512BW"],
4940
+ "_mm_movepi16_mask"=>[["AVX512VL", "AVX512BW"]],
4941
+ "_mm256_cvtepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4942
+ "_mm256_mask_cvtepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4943
+ "_mm256_mask_cvtepi16_storeu_epi8"=>[["AVX512VL", "AVX512BW"]],
4944
+ "_mm256_maskz_cvtepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4945
4945
  "_mm512_cvtepi16_epi8"=>["AVX512BW"],
4946
4946
  "_mm512_mask_cvtepi16_epi8"=>["AVX512BW"],
4947
4947
  "_mm512_mask_cvtepi16_storeu_epi8"=>["AVX512BW"],
4948
4948
  "_mm512_maskz_cvtepi16_epi8"=>["AVX512BW"],
4949
- "_mm_cvtepi16_epi8"=>["AVX512VL", "AVX512BW"],
4950
- "_mm_mask_cvtepi16_epi8"=>["AVX512VL", "AVX512BW"],
4951
- "_mm_mask_cvtepi16_storeu_epi8"=>["AVX512VL", "AVX512BW"],
4952
- "_mm_maskz_cvtepi16_epi8"=>["AVX512VL", "AVX512BW"],
4953
- "_mm256_mask_cvtepu8_epi32"=>["AVX512VL", "AVX512F"],
4954
- "_mm256_maskz_cvtepu8_epi32"=>["AVX512VL", "AVX512F"],
4955
- "_mm_mask_cvtepu8_epi32"=>["AVX512VL", "AVX512F"],
4956
- "_mm_maskz_cvtepu8_epi32"=>["AVX512VL", "AVX512F"],
4957
- "_mm256_mask_cvtepu8_epi64"=>["AVX512VL", "AVX512F"],
4958
- "_mm256_maskz_cvtepu8_epi64"=>["AVX512VL", "AVX512F"],
4959
- "_mm_mask_cvtepu8_epi64"=>["AVX512VL", "AVX512F"],
4960
- "_mm_maskz_cvtepu8_epi64"=>["AVX512VL", "AVX512F"],
4961
- "_mm256_mask_cvtepu8_epi16"=>["AVX512VL", "AVX512BW"],
4962
- "_mm256_maskz_cvtepu8_epi16"=>["AVX512VL", "AVX512BW"],
4949
+ "_mm_cvtepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4950
+ "_mm_mask_cvtepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4951
+ "_mm_mask_cvtepi16_storeu_epi8"=>[["AVX512VL", "AVX512BW"]],
4952
+ "_mm_maskz_cvtepi16_epi8"=>[["AVX512VL", "AVX512BW"]],
4953
+ "_mm256_mask_cvtepu8_epi32"=>[["AVX512VL", "AVX512F"]],
4954
+ "_mm256_maskz_cvtepu8_epi32"=>[["AVX512VL", "AVX512F"]],
4955
+ "_mm_mask_cvtepu8_epi32"=>[["AVX512VL", "AVX512F"]],
4956
+ "_mm_maskz_cvtepu8_epi32"=>[["AVX512VL", "AVX512F"]],
4957
+ "_mm256_mask_cvtepu8_epi64"=>[["AVX512VL", "AVX512F"]],
4958
+ "_mm256_maskz_cvtepu8_epi64"=>[["AVX512VL", "AVX512F"]],
4959
+ "_mm_mask_cvtepu8_epi64"=>[["AVX512VL", "AVX512F"]],
4960
+ "_mm_maskz_cvtepu8_epi64"=>[["AVX512VL", "AVX512F"]],
4961
+ "_mm256_mask_cvtepu8_epi16"=>[["AVX512VL", "AVX512BW"]],
4962
+ "_mm256_maskz_cvtepu8_epi16"=>[["AVX512VL", "AVX512BW"]],
4963
4963
  "_mm512_cvtepu8_epi16"=>["AVX512BW"],
4964
4964
  "_mm512_mask_cvtepu8_epi16"=>["AVX512BW"],
4965
4965
  "_mm512_maskz_cvtepu8_epi16"=>["AVX512BW"],
4966
- "_mm_mask_cvtepu8_epi16"=>["AVX512VL", "AVX512BW"],
4967
- "_mm_maskz_cvtepu8_epi16"=>["AVX512VL", "AVX512BW"],
4968
- "_mm256_mask_cvtepu32_epi64"=>["AVX512VL", "AVX512F"],
4969
- "_mm256_maskz_cvtepu32_epi64"=>["AVX512VL", "AVX512F"],
4970
- "_mm_mask_cvtepu32_epi64"=>["AVX512VL", "AVX512F"],
4971
- "_mm_maskz_cvtepu32_epi64"=>["AVX512VL", "AVX512F"],
4972
- "_mm256_mask_cvtepu16_epi32"=>["AVX512VL", "AVX512F"],
4973
- "_mm256_maskz_cvtepu16_epi32"=>["AVX512VL", "AVX512F"],
4974
- "_mm_mask_cvtepu16_epi32"=>["AVX512VL", "AVX512F"],
4975
- "_mm_maskz_cvtepu16_epi32"=>["AVX512VL", "AVX512F"],
4976
- "_mm256_mask_cvtepu16_epi64"=>["AVX512VL", "AVX512F"],
4977
- "_mm256_maskz_cvtepu16_epi64"=>["AVX512VL", "AVX512F"],
4978
- "_mm_mask_cvtepu16_epi64"=>["AVX512VL", "AVX512F"],
4979
- "_mm_maskz_cvtepu16_epi64"=>["AVX512VL", "AVX512F"],
4980
- "_mm256_mask_mul_epi32"=>["AVX512VL", "AVX512F"],
4981
- "_mm256_maskz_mul_epi32"=>["AVX512VL", "AVX512F"],
4982
- "_mm_mask_mul_epi32"=>["AVX512VL", "AVX512F"],
4983
- "_mm_maskz_mul_epi32"=>["AVX512VL", "AVX512F"],
4984
- "_mm256_mask_mulhrs_epi16"=>["AVX512VL", "AVX512BW"],
4985
- "_mm256_maskz_mulhrs_epi16"=>["AVX512VL", "AVX512BW"],
4966
+ "_mm_mask_cvtepu8_epi16"=>[["AVX512VL", "AVX512BW"]],
4967
+ "_mm_maskz_cvtepu8_epi16"=>[["AVX512VL", "AVX512BW"]],
4968
+ "_mm256_mask_cvtepu32_epi64"=>[["AVX512VL", "AVX512F"]],
4969
+ "_mm256_maskz_cvtepu32_epi64"=>[["AVX512VL", "AVX512F"]],
4970
+ "_mm_mask_cvtepu32_epi64"=>[["AVX512VL", "AVX512F"]],
4971
+ "_mm_maskz_cvtepu32_epi64"=>[["AVX512VL", "AVX512F"]],
4972
+ "_mm256_mask_cvtepu16_epi32"=>[["AVX512VL", "AVX512F"]],
4973
+ "_mm256_maskz_cvtepu16_epi32"=>[["AVX512VL", "AVX512F"]],
4974
+ "_mm_mask_cvtepu16_epi32"=>[["AVX512VL", "AVX512F"]],
4975
+ "_mm_maskz_cvtepu16_epi32"=>[["AVX512VL", "AVX512F"]],
4976
+ "_mm256_mask_cvtepu16_epi64"=>[["AVX512VL", "AVX512F"]],
4977
+ "_mm256_maskz_cvtepu16_epi64"=>[["AVX512VL", "AVX512F"]],
4978
+ "_mm_mask_cvtepu16_epi64"=>[["AVX512VL", "AVX512F"]],
4979
+ "_mm_maskz_cvtepu16_epi64"=>[["AVX512VL", "AVX512F"]],
4980
+ "_mm256_mask_mul_epi32"=>[["AVX512VL", "AVX512F"]],
4981
+ "_mm256_maskz_mul_epi32"=>[["AVX512VL", "AVX512F"]],
4982
+ "_mm_mask_mul_epi32"=>[["AVX512VL", "AVX512F"]],
4983
+ "_mm_maskz_mul_epi32"=>[["AVX512VL", "AVX512F"]],
4984
+ "_mm256_mask_mulhrs_epi16"=>[["AVX512VL", "AVX512BW"]],
4985
+ "_mm256_maskz_mulhrs_epi16"=>[["AVX512VL", "AVX512BW"]],
4986
4986
  "_mm512_mask_mulhrs_epi16"=>["AVX512BW"],
4987
4987
  "_mm512_maskz_mulhrs_epi16"=>["AVX512BW"],
4988
4988
  "_mm512_mulhrs_epi16"=>["AVX512BW"],
4989
- "_mm_mask_mulhrs_epi16"=>["AVX512VL", "AVX512BW"],
4990
- "_mm_maskz_mulhrs_epi16"=>["AVX512VL", "AVX512BW"],
4991
- "_mm256_mask_mulhi_epu16"=>["AVX512VL", "AVX512BW"],
4992
- "_mm256_maskz_mulhi_epu16"=>["AVX512VL", "AVX512BW"],
4989
+ "_mm_mask_mulhrs_epi16"=>[["AVX512VL", "AVX512BW"]],
4990
+ "_mm_maskz_mulhrs_epi16"=>[["AVX512VL", "AVX512BW"]],
4991
+ "_mm256_mask_mulhi_epu16"=>[["AVX512VL", "AVX512BW"]],
4992
+ "_mm256_maskz_mulhi_epu16"=>[["AVX512VL", "AVX512BW"]],
4993
4993
  "_mm512_mask_mulhi_epu16"=>["AVX512BW"],
4994
4994
  "_mm512_maskz_mulhi_epu16"=>["AVX512BW"],
4995
4995
  "_mm512_mulhi_epu16"=>["AVX512BW"],
4996
- "_mm_mask_mulhi_epu16"=>["AVX512VL", "AVX512BW"],
4997
- "_mm_maskz_mulhi_epu16"=>["AVX512VL", "AVX512BW"],
4998
- "_mm256_mask_mulhi_epi16"=>["AVX512VL", "AVX512BW"],
4999
- "_mm256_maskz_mulhi_epi16"=>["AVX512VL", "AVX512BW"],
4996
+ "_mm_mask_mulhi_epu16"=>[["AVX512VL", "AVX512BW"]],
4997
+ "_mm_maskz_mulhi_epu16"=>[["AVX512VL", "AVX512BW"]],
4998
+ "_mm256_mask_mulhi_epi16"=>[["AVX512VL", "AVX512BW"]],
4999
+ "_mm256_maskz_mulhi_epi16"=>[["AVX512VL", "AVX512BW"]],
5000
5000
  "_mm512_mask_mulhi_epi16"=>["AVX512BW"],
5001
5001
  "_mm512_maskz_mulhi_epi16"=>["AVX512BW"],
5002
5002
  "_mm512_mulhi_epi16"=>["AVX512BW"],
5003
- "_mm_mask_mulhi_epi16"=>["AVX512VL", "AVX512BW"],
5004
- "_mm_maskz_mulhi_epi16"=>["AVX512VL", "AVX512BW"],
5005
- "_mm256_mask_mullo_epi32"=>["AVX512VL", "AVX512F"],
5006
- "_mm256_maskz_mullo_epi32"=>["AVX512VL", "AVX512F"],
5003
+ "_mm_mask_mulhi_epi16"=>[["AVX512VL", "AVX512BW"]],
5004
+ "_mm_maskz_mulhi_epi16"=>[["AVX512VL", "AVX512BW"]],
5005
+ "_mm256_mask_mullo_epi32"=>[["AVX512VL", "AVX512F"]],
5006
+ "_mm256_maskz_mullo_epi32"=>[["AVX512VL", "AVX512F"]],
5007
5007
  "_mm512_maskz_mullo_epi32"=>["AVX512F"],
5008
- "_mm_mask_mullo_epi32"=>["AVX512VL", "AVX512F"],
5009
- "_mm_maskz_mullo_epi32"=>["AVX512VL", "AVX512F"],
5010
- "_mm256_mask_mullo_epi64"=>["AVX512VL", "AVX512DQ"],
5011
- "_mm256_maskz_mullo_epi64"=>["AVX512VL", "AVX512DQ"],
5012
- "_mm256_mullo_epi64"=>["AVX512VL", "AVX512DQ"],
5008
+ "_mm_mask_mullo_epi32"=>[["AVX512VL", "AVX512F"]],
5009
+ "_mm_maskz_mullo_epi32"=>[["AVX512VL", "AVX512F"]],
5010
+ "_mm256_mask_mullo_epi64"=>[["AVX512VL", "AVX512DQ"]],
5011
+ "_mm256_maskz_mullo_epi64"=>[["AVX512VL", "AVX512DQ"]],
5012
+ "_mm256_mullo_epi64"=>[["AVX512VL", "AVX512DQ"]],
5013
5013
  "_mm512_mask_mullo_epi64"=>["AVX512DQ"],
5014
5014
  "_mm512_maskz_mullo_epi64"=>["AVX512DQ"],
5015
5015
  "_mm512_mullo_epi64"=>["AVX512DQ"],
5016
- "_mm_mask_mullo_epi64"=>["AVX512VL", "AVX512DQ"],
5017
- "_mm_maskz_mullo_epi64"=>["AVX512VL", "AVX512DQ"],
5018
- "_mm_mullo_epi64"=>["AVX512VL", "AVX512DQ"],
5019
- "_mm256_mask_mullo_epi16"=>["AVX512VL", "AVX512BW"],
5020
- "_mm256_maskz_mullo_epi16"=>["AVX512VL", "AVX512BW"],
5016
+ "_mm_mask_mullo_epi64"=>[["AVX512VL", "AVX512DQ"]],
5017
+ "_mm_maskz_mullo_epi64"=>[["AVX512VL", "AVX512DQ"]],
5018
+ "_mm_mullo_epi64"=>[["AVX512VL", "AVX512DQ"]],
5019
+ "_mm256_mask_mullo_epi16"=>[["AVX512VL", "AVX512BW"]],
5020
+ "_mm256_maskz_mullo_epi16"=>[["AVX512VL", "AVX512BW"]],
5021
5021
  "_mm512_mask_mullo_epi16"=>["AVX512BW"],
5022
5022
  "_mm512_maskz_mullo_epi16"=>["AVX512BW"],
5023
5023
  "_mm512_mullo_epi16"=>["AVX512BW"],
5024
- "_mm_mask_mullo_epi16"=>["AVX512VL", "AVX512BW"],
5025
- "_mm_maskz_mullo_epi16"=>["AVX512VL", "AVX512BW"],
5026
- "_mm256_mask_mul_epu32"=>["AVX512VL", "AVX512F"],
5027
- "_mm256_maskz_mul_epu32"=>["AVX512VL", "AVX512F"],
5028
- "_mm_mask_mul_epu32"=>["AVX512VL", "AVX512F"],
5029
- "_mm_maskz_mul_epu32"=>["AVX512VL", "AVX512F"],
5030
- "_mm256_mask_or_epi32"=>["AVX512VL", "AVX512F"],
5031
- "_mm256_maskz_or_epi32"=>["AVX512VL", "AVX512F"],
5032
- "_mm_mask_or_epi32"=>["AVX512VL", "AVX512F"],
5033
- "_mm_maskz_or_epi32"=>["AVX512VL", "AVX512F"],
5034
- "_mm256_mask_or_epi64"=>["AVX512VL", "AVX512F"],
5035
- "_mm256_maskz_or_epi64"=>["AVX512VL", "AVX512F"],
5036
- "_mm_mask_or_epi64"=>["AVX512VL", "AVX512F"],
5037
- "_mm_maskz_or_epi64"=>["AVX512VL", "AVX512F"],
5038
- "_mm256_mask_rol_epi32"=>["AVX512VL", "AVX512F"],
5039
- "_mm256_maskz_rol_epi32"=>["AVX512VL", "AVX512F"],
5040
- "_mm256_rol_epi32"=>["AVX512VL", "AVX512F"],
5041
- "_mm_mask_rol_epi32"=>["AVX512VL", "AVX512F"],
5042
- "_mm_maskz_rol_epi32"=>["AVX512VL", "AVX512F"],
5043
- "_mm_rol_epi32"=>["AVX512VL", "AVX512F"],
5044
- "_mm256_mask_rol_epi64"=>["AVX512VL", "AVX512F"],
5045
- "_mm256_maskz_rol_epi64"=>["AVX512VL", "AVX512F"],
5046
- "_mm256_rol_epi64"=>["AVX512VL", "AVX512F"],
5047
- "_mm_mask_rol_epi64"=>["AVX512VL", "AVX512F"],
5048
- "_mm_maskz_rol_epi64"=>["AVX512VL", "AVX512F"],
5049
- "_mm_rol_epi64"=>["AVX512VL", "AVX512F"],
5050
- "_mm256_mask_rolv_epi32"=>["AVX512VL", "AVX512F"],
5051
- "_mm256_maskz_rolv_epi32"=>["AVX512VL", "AVX512F"],
5052
- "_mm256_rolv_epi32"=>["AVX512VL", "AVX512F"],
5053
- "_mm_mask_rolv_epi32"=>["AVX512VL", "AVX512F"],
5054
- "_mm_maskz_rolv_epi32"=>["AVX512VL", "AVX512F"],
5055
- "_mm_rolv_epi32"=>["AVX512VL", "AVX512F"],
5056
- "_mm256_mask_rolv_epi64"=>["AVX512VL", "AVX512F"],
5057
- "_mm256_maskz_rolv_epi64"=>["AVX512VL", "AVX512F"],
5058
- "_mm256_rolv_epi64"=>["AVX512VL", "AVX512F"],
5059
- "_mm_mask_rolv_epi64"=>["AVX512VL", "AVX512F"],
5060
- "_mm_maskz_rolv_epi64"=>["AVX512VL", "AVX512F"],
5061
- "_mm_rolv_epi64"=>["AVX512VL", "AVX512F"],
5062
- "_mm256_mask_ror_epi32"=>["AVX512VL", "AVX512F"],
5063
- "_mm256_maskz_ror_epi32"=>["AVX512VL", "AVX512F"],
5064
- "_mm256_ror_epi32"=>["AVX512VL", "AVX512F"],
5065
- "_mm_mask_ror_epi32"=>["AVX512VL", "AVX512F"],
5066
- "_mm_maskz_ror_epi32"=>["AVX512VL", "AVX512F"],
5067
- "_mm_ror_epi32"=>["AVX512VL", "AVX512F"],
5068
- "_mm256_mask_ror_epi64"=>["AVX512VL", "AVX512F"],
5069
- "_mm256_maskz_ror_epi64"=>["AVX512VL", "AVX512F"],
5070
- "_mm256_ror_epi64"=>["AVX512VL", "AVX512F"],
5071
- "_mm_mask_ror_epi64"=>["AVX512VL", "AVX512F"],
5072
- "_mm_maskz_ror_epi64"=>["AVX512VL", "AVX512F"],
5073
- "_mm_ror_epi64"=>["AVX512VL", "AVX512F"],
5074
- "_mm256_mask_rorv_epi32"=>["AVX512VL", "AVX512F"],
5075
- "_mm256_maskz_rorv_epi32"=>["AVX512VL", "AVX512F"],
5076
- "_mm256_rorv_epi32"=>["AVX512VL", "AVX512F"],
5077
- "_mm_mask_rorv_epi32"=>["AVX512VL", "AVX512F"],
5078
- "_mm_maskz_rorv_epi32"=>["AVX512VL", "AVX512F"],
5079
- "_mm_rorv_epi32"=>["AVX512VL", "AVX512F"],
5080
- "_mm256_mask_rorv_epi64"=>["AVX512VL", "AVX512F"],
5081
- "_mm256_maskz_rorv_epi64"=>["AVX512VL", "AVX512F"],
5082
- "_mm256_rorv_epi64"=>["AVX512VL", "AVX512F"],
5083
- "_mm_mask_rorv_epi64"=>["AVX512VL", "AVX512F"],
5084
- "_mm_maskz_rorv_epi64"=>["AVX512VL", "AVX512F"],
5085
- "_mm_rorv_epi64"=>["AVX512VL", "AVX512F"],
5024
+ "_mm_mask_mullo_epi16"=>[["AVX512VL", "AVX512BW"]],
5025
+ "_mm_maskz_mullo_epi16"=>[["AVX512VL", "AVX512BW"]],
5026
+ "_mm256_mask_mul_epu32"=>[["AVX512VL", "AVX512F"]],
5027
+ "_mm256_maskz_mul_epu32"=>[["AVX512VL", "AVX512F"]],
5028
+ "_mm_mask_mul_epu32"=>[["AVX512VL", "AVX512F"]],
5029
+ "_mm_maskz_mul_epu32"=>[["AVX512VL", "AVX512F"]],
5030
+ "_mm256_mask_or_epi32"=>[["AVX512VL", "AVX512F"]],
5031
+ "_mm256_maskz_or_epi32"=>[["AVX512VL", "AVX512F"]],
5032
+ "_mm_mask_or_epi32"=>[["AVX512VL", "AVX512F"]],
5033
+ "_mm_maskz_or_epi32"=>[["AVX512VL", "AVX512F"]],
5034
+ "_mm256_mask_or_epi64"=>[["AVX512VL", "AVX512F"]],
5035
+ "_mm256_maskz_or_epi64"=>[["AVX512VL", "AVX512F"]],
5036
+ "_mm_mask_or_epi64"=>[["AVX512VL", "AVX512F"]],
5037
+ "_mm_maskz_or_epi64"=>[["AVX512VL", "AVX512F"]],
5038
+ "_mm256_mask_rol_epi32"=>[["AVX512VL", "AVX512F"]],
5039
+ "_mm256_maskz_rol_epi32"=>[["AVX512VL", "AVX512F"]],
5040
+ "_mm256_rol_epi32"=>[["AVX512VL", "AVX512F"]],
5041
+ "_mm_mask_rol_epi32"=>[["AVX512VL", "AVX512F"]],
5042
+ "_mm_maskz_rol_epi32"=>[["AVX512VL", "AVX512F"]],
5043
+ "_mm_rol_epi32"=>[["AVX512VL", "AVX512F"]],
5044
+ "_mm256_mask_rol_epi64"=>[["AVX512VL", "AVX512F"]],
5045
+ "_mm256_maskz_rol_epi64"=>[["AVX512VL", "AVX512F"]],
5046
+ "_mm256_rol_epi64"=>[["AVX512VL", "AVX512F"]],
5047
+ "_mm_mask_rol_epi64"=>[["AVX512VL", "AVX512F"]],
5048
+ "_mm_maskz_rol_epi64"=>[["AVX512VL", "AVX512F"]],
5049
+ "_mm_rol_epi64"=>[["AVX512VL", "AVX512F"]],
5050
+ "_mm256_mask_rolv_epi32"=>[["AVX512VL", "AVX512F"]],
5051
+ "_mm256_maskz_rolv_epi32"=>[["AVX512VL", "AVX512F"]],
5052
+ "_mm256_rolv_epi32"=>[["AVX512VL", "AVX512F"]],
5053
+ "_mm_mask_rolv_epi32"=>[["AVX512VL", "AVX512F"]],
5054
+ "_mm_maskz_rolv_epi32"=>[["AVX512VL", "AVX512F"]],
5055
+ "_mm_rolv_epi32"=>[["AVX512VL", "AVX512F"]],
5056
+ "_mm256_mask_rolv_epi64"=>[["AVX512VL", "AVX512F"]],
5057
+ "_mm256_maskz_rolv_epi64"=>[["AVX512VL", "AVX512F"]],
5058
+ "_mm256_rolv_epi64"=>[["AVX512VL", "AVX512F"]],
5059
+ "_mm_mask_rolv_epi64"=>[["AVX512VL", "AVX512F"]],
5060
+ "_mm_maskz_rolv_epi64"=>[["AVX512VL", "AVX512F"]],
5061
+ "_mm_rolv_epi64"=>[["AVX512VL", "AVX512F"]],
5062
+ "_mm256_mask_ror_epi32"=>[["AVX512VL", "AVX512F"]],
5063
+ "_mm256_maskz_ror_epi32"=>[["AVX512VL", "AVX512F"]],
5064
+ "_mm256_ror_epi32"=>[["AVX512VL", "AVX512F"]],
5065
+ "_mm_mask_ror_epi32"=>[["AVX512VL", "AVX512F"]],
5066
+ "_mm_maskz_ror_epi32"=>[["AVX512VL", "AVX512F"]],
5067
+ "_mm_ror_epi32"=>[["AVX512VL", "AVX512F"]],
5068
+ "_mm256_mask_ror_epi64"=>[["AVX512VL", "AVX512F"]],
5069
+ "_mm256_maskz_ror_epi64"=>[["AVX512VL", "AVX512F"]],
5070
+ "_mm256_ror_epi64"=>[["AVX512VL", "AVX512F"]],
5071
+ "_mm_mask_ror_epi64"=>[["AVX512VL", "AVX512F"]],
5072
+ "_mm_maskz_ror_epi64"=>[["AVX512VL", "AVX512F"]],
5073
+ "_mm_ror_epi64"=>[["AVX512VL", "AVX512F"]],
5074
+ "_mm256_mask_rorv_epi32"=>[["AVX512VL", "AVX512F"]],
5075
+ "_mm256_maskz_rorv_epi32"=>[["AVX512VL", "AVX512F"]],
5076
+ "_mm256_rorv_epi32"=>[["AVX512VL", "AVX512F"]],
5077
+ "_mm_mask_rorv_epi32"=>[["AVX512VL", "AVX512F"]],
5078
+ "_mm_maskz_rorv_epi32"=>[["AVX512VL", "AVX512F"]],
5079
+ "_mm_rorv_epi32"=>[["AVX512VL", "AVX512F"]],
5080
+ "_mm256_mask_rorv_epi64"=>[["AVX512VL", "AVX512F"]],
5081
+ "_mm256_maskz_rorv_epi64"=>[["AVX512VL", "AVX512F"]],
5082
+ "_mm256_rorv_epi64"=>[["AVX512VL", "AVX512F"]],
5083
+ "_mm_mask_rorv_epi64"=>[["AVX512VL", "AVX512F"]],
5084
+ "_mm_maskz_rorv_epi64"=>[["AVX512VL", "AVX512F"]],
5085
+ "_mm_rorv_epi64"=>[["AVX512VL", "AVX512F"]],
5086
5086
  "_mm512_sad_epu8"=>["AVX512BW"],
5087
- "_mm256_i32scatter_epi32"=>["AVX512VL", "AVX512F"],
5088
- "_mm256_mask_i32scatter_epi32"=>["AVX512VL", "AVX512F"],
5089
- "_mm_i32scatter_epi32"=>["AVX512VL", "AVX512F"],
5090
- "_mm_mask_i32scatter_epi32"=>["AVX512VL", "AVX512F"],
5091
- "_mm256_i32scatter_epi64"=>["AVX512VL", "AVX512F"],
5092
- "_mm256_mask_i32scatter_epi64"=>["AVX512VL", "AVX512F"],
5093
- "_mm_i32scatter_epi64"=>["AVX512VL", "AVX512F"],
5094
- "_mm_mask_i32scatter_epi64"=>["AVX512VL", "AVX512F"],
5095
- "_mm256_i64scatter_epi32"=>["AVX512VL", "AVX512F"],
5096
- "_mm256_mask_i64scatter_epi32"=>["AVX512VL", "AVX512F"],
5097
- "_mm_i64scatter_epi32"=>["AVX512VL", "AVX512F"],
5098
- "_mm_mask_i64scatter_epi32"=>["AVX512VL", "AVX512F"],
5099
- "_mm256_i64scatter_epi64"=>["AVX512VL", "AVX512F"],
5100
- "_mm256_mask_i64scatter_epi64"=>["AVX512VL", "AVX512F"],
5101
- "_mm_i64scatter_epi64"=>["AVX512VL", "AVX512F"],
5102
- "_mm_mask_i64scatter_epi64"=>["AVX512VL", "AVX512F"],
5103
- "_mm256_mask_shuffle_epi8"=>["AVX512VL", "AVX512BW"],
5104
- "_mm256_maskz_shuffle_epi8"=>["AVX512VL", "AVX512BW"],
5087
+ "_mm256_i32scatter_epi32"=>[["AVX512VL", "AVX512F"]],
5088
+ "_mm256_mask_i32scatter_epi32"=>[["AVX512VL", "AVX512F"]],
5089
+ "_mm_i32scatter_epi32"=>[["AVX512VL", "AVX512F"]],
5090
+ "_mm_mask_i32scatter_epi32"=>[["AVX512VL", "AVX512F"]],
5091
+ "_mm256_i32scatter_epi64"=>[["AVX512VL", "AVX512F"]],
5092
+ "_mm256_mask_i32scatter_epi64"=>[["AVX512VL", "AVX512F"]],
5093
+ "_mm_i32scatter_epi64"=>[["AVX512VL", "AVX512F"]],
5094
+ "_mm_mask_i32scatter_epi64"=>[["AVX512VL", "AVX512F"]],
5095
+ "_mm256_i64scatter_epi32"=>[["AVX512VL", "AVX512F"]],
5096
+ "_mm256_mask_i64scatter_epi32"=>[["AVX512VL", "AVX512F"]],
5097
+ "_mm_i64scatter_epi32"=>[["AVX512VL", "AVX512F"]],
5098
+ "_mm_mask_i64scatter_epi32"=>[["AVX512VL", "AVX512F"]],
5099
+ "_mm256_i64scatter_epi64"=>[["AVX512VL", "AVX512F"]],
5100
+ "_mm256_mask_i64scatter_epi64"=>[["AVX512VL", "AVX512F"]],
5101
+ "_mm_i64scatter_epi64"=>[["AVX512VL", "AVX512F"]],
5102
+ "_mm_mask_i64scatter_epi64"=>[["AVX512VL", "AVX512F"]],
5103
+ "_mm256_mask_shuffle_epi8"=>[["AVX512VL", "AVX512BW"]],
5104
+ "_mm256_maskz_shuffle_epi8"=>[["AVX512VL", "AVX512BW"]],
5105
5105
  "_mm512_mask_shuffle_epi8"=>["AVX512BW"],
5106
5106
  "_mm512_maskz_shuffle_epi8"=>["AVX512BW"],
5107
5107
  "_mm512_shuffle_epi8"=>["AVX512BW"],
5108
- "_mm_mask_shuffle_epi8"=>["AVX512VL", "AVX512BW"],
5109
- "_mm_maskz_shuffle_epi8"=>["AVX512VL", "AVX512BW"],
5110
- "_mm256_mask_shuffle_epi32"=>["AVX512VL", "AVX512F"],
5111
- "_mm256_maskz_shuffle_epi32"=>["AVX512VL", "AVX512F"],
5112
- "_mm_mask_shuffle_epi32"=>["AVX512VL", "AVX512F"],
5113
- "_mm_maskz_shuffle_epi32"=>["AVX512VL", "AVX512F"],
5114
- "_mm256_mask_shufflehi_epi16"=>["AVX512VL", "AVX512BW"],
5115
- "_mm256_maskz_shufflehi_epi16"=>["AVX512VL", "AVX512BW"],
5108
+ "_mm_mask_shuffle_epi8"=>[["AVX512VL", "AVX512BW"]],
5109
+ "_mm_maskz_shuffle_epi8"=>[["AVX512VL", "AVX512BW"]],
5110
+ "_mm256_mask_shuffle_epi32"=>[["AVX512VL", "AVX512F"]],
5111
+ "_mm256_maskz_shuffle_epi32"=>[["AVX512VL", "AVX512F"]],
5112
+ "_mm_mask_shuffle_epi32"=>[["AVX512VL", "AVX512F"]],
5113
+ "_mm_maskz_shuffle_epi32"=>[["AVX512VL", "AVX512F"]],
5114
+ "_mm256_mask_shufflehi_epi16"=>[["AVX512VL", "AVX512BW"]],
5115
+ "_mm256_maskz_shufflehi_epi16"=>[["AVX512VL", "AVX512BW"]],
5116
5116
  "_mm512_mask_shufflehi_epi16"=>["AVX512BW"],
5117
5117
  "_mm512_maskz_shufflehi_epi16"=>["AVX512BW"],
5118
5118
  "_mm512_shufflehi_epi16"=>["AVX512BW"],
5119
- "_mm_mask_shufflehi_epi16"=>["AVX512VL", "AVX512BW"],
5120
- "_mm_maskz_shufflehi_epi16"=>["AVX512VL", "AVX512BW"],
5121
- "_mm256_mask_shufflelo_epi16"=>["AVX512VL", "AVX512BW"],
5122
- "_mm256_maskz_shufflelo_epi16"=>["AVX512VL", "AVX512BW"],
5119
+ "_mm_mask_shufflehi_epi16"=>[["AVX512VL", "AVX512BW"]],
5120
+ "_mm_maskz_shufflehi_epi16"=>[["AVX512VL", "AVX512BW"]],
5121
+ "_mm256_mask_shufflelo_epi16"=>[["AVX512VL", "AVX512BW"]],
5122
+ "_mm256_maskz_shufflelo_epi16"=>[["AVX512VL", "AVX512BW"]],
5123
5123
  "_mm512_mask_shufflelo_epi16"=>["AVX512BW"],
5124
5124
  "_mm512_maskz_shufflelo_epi16"=>["AVX512BW"],
5125
5125
  "_mm512_shufflelo_epi16"=>["AVX512BW"],
5126
- "_mm_mask_shufflelo_epi16"=>["AVX512VL", "AVX512BW"],
5127
- "_mm_maskz_shufflelo_epi16"=>["AVX512VL", "AVX512BW"],
5128
- "_mm256_mask_sll_epi32"=>["AVX512VL", "AVX512F"],
5129
- "_mm256_mask_slli_epi32"=>["AVX512VL", "AVX512F"],
5130
- "_mm256_maskz_sll_epi32"=>["AVX512VL", "AVX512F"],
5131
- "_mm256_maskz_slli_epi32"=>["AVX512VL", "AVX512F"],
5132
- "_mm_mask_sll_epi32"=>["AVX512VL", "AVX512F"],
5133
- "_mm_mask_slli_epi32"=>["AVX512VL", "AVX512F"],
5134
- "_mm_maskz_sll_epi32"=>["AVX512VL", "AVX512F"],
5135
- "_mm_maskz_slli_epi32"=>["AVX512VL", "AVX512F"],
5126
+ "_mm_mask_shufflelo_epi16"=>[["AVX512VL", "AVX512BW"]],
5127
+ "_mm_maskz_shufflelo_epi16"=>[["AVX512VL", "AVX512BW"]],
5128
+ "_mm256_mask_sll_epi32"=>[["AVX512VL", "AVX512F"]],
5129
+ "_mm256_mask_slli_epi32"=>[["AVX512VL", "AVX512F"]],
5130
+ "_mm256_maskz_sll_epi32"=>[["AVX512VL", "AVX512F"]],
5131
+ "_mm256_maskz_slli_epi32"=>[["AVX512VL", "AVX512F"]],
5132
+ "_mm_mask_sll_epi32"=>[["AVX512VL", "AVX512F"]],
5133
+ "_mm_mask_slli_epi32"=>[["AVX512VL", "AVX512F"]],
5134
+ "_mm_maskz_sll_epi32"=>[["AVX512VL", "AVX512F"]],
5135
+ "_mm_maskz_slli_epi32"=>[["AVX512VL", "AVX512F"]],
5136
5136
  "_mm512_bslli_epi128"=>["AVX512BW"],
5137
- "_mm256_mask_sll_epi64"=>["AVX512VL", "AVX512F"],
5138
- "_mm256_mask_slli_epi64"=>["AVX512VL", "AVX512F"],
5139
- "_mm256_maskz_sll_epi64"=>["AVX512VL", "AVX512F"],
5140
- "_mm256_maskz_slli_epi64"=>["AVX512VL", "AVX512F"],
5141
- "_mm_mask_sll_epi64"=>["AVX512VL", "AVX512F"],
5142
- "_mm_mask_slli_epi64"=>["AVX512VL", "AVX512F"],
5143
- "_mm_maskz_sll_epi64"=>["AVX512VL", "AVX512F"],
5144
- "_mm_maskz_slli_epi64"=>["AVX512VL", "AVX512F"],
5145
- "_mm256_mask_sllv_epi32"=>["AVX512VL", "AVX512F"],
5146
- "_mm256_maskz_sllv_epi32"=>["AVX512VL", "AVX512F"],
5147
- "_mm_mask_sllv_epi32"=>["AVX512VL", "AVX512F"],
5148
- "_mm_maskz_sllv_epi32"=>["AVX512VL", "AVX512F"],
5149
- "_mm256_mask_sllv_epi64"=>["AVX512VL", "AVX512F"],
5150
- "_mm256_maskz_sllv_epi64"=>["AVX512VL", "AVX512F"],
5151
- "_mm_mask_sllv_epi64"=>["AVX512VL", "AVX512F"],
5152
- "_mm_maskz_sllv_epi64"=>["AVX512VL", "AVX512F"],
5153
- "_mm256_mask_sllv_epi16"=>["AVX512VL", "AVX512BW"],
5154
- "_mm256_maskz_sllv_epi16"=>["AVX512VL", "AVX512BW"],
5155
- "_mm256_sllv_epi16"=>["AVX512VL", "AVX512BW"],
5137
+ "_mm256_mask_sll_epi64"=>[["AVX512VL", "AVX512F"]],
5138
+ "_mm256_mask_slli_epi64"=>[["AVX512VL", "AVX512F"]],
5139
+ "_mm256_maskz_sll_epi64"=>[["AVX512VL", "AVX512F"]],
5140
+ "_mm256_maskz_slli_epi64"=>[["AVX512VL", "AVX512F"]],
5141
+ "_mm_mask_sll_epi64"=>[["AVX512VL", "AVX512F"]],
5142
+ "_mm_mask_slli_epi64"=>[["AVX512VL", "AVX512F"]],
5143
+ "_mm_maskz_sll_epi64"=>[["AVX512VL", "AVX512F"]],
5144
+ "_mm_maskz_slli_epi64"=>[["AVX512VL", "AVX512F"]],
5145
+ "_mm256_mask_sllv_epi32"=>[["AVX512VL", "AVX512F"]],
5146
+ "_mm256_maskz_sllv_epi32"=>[["AVX512VL", "AVX512F"]],
5147
+ "_mm_mask_sllv_epi32"=>[["AVX512VL", "AVX512F"]],
5148
+ "_mm_maskz_sllv_epi32"=>[["AVX512VL", "AVX512F"]],
5149
+ "_mm256_mask_sllv_epi64"=>[["AVX512VL", "AVX512F"]],
5150
+ "_mm256_maskz_sllv_epi64"=>[["AVX512VL", "AVX512F"]],
5151
+ "_mm_mask_sllv_epi64"=>[["AVX512VL", "AVX512F"]],
5152
+ "_mm_maskz_sllv_epi64"=>[["AVX512VL", "AVX512F"]],
5153
+ "_mm256_mask_sllv_epi16"=>[["AVX512VL", "AVX512BW"]],
5154
+ "_mm256_maskz_sllv_epi16"=>[["AVX512VL", "AVX512BW"]],
5155
+ "_mm256_sllv_epi16"=>[["AVX512VL", "AVX512BW"]],
5156
5156
  "_mm512_mask_sllv_epi16"=>["AVX512BW"],
5157
5157
  "_mm512_maskz_sllv_epi16"=>["AVX512BW"],
5158
5158
  "_mm512_sllv_epi16"=>["AVX512BW"],
5159
- "_mm_mask_sllv_epi16"=>["AVX512VL", "AVX512BW"],
5160
- "_mm_maskz_sllv_epi16"=>["AVX512VL", "AVX512BW"],
5161
- "_mm_sllv_epi16"=>["AVX512VL", "AVX512BW"],
5162
- "_mm256_mask_sll_epi16"=>["AVX512VL", "AVX512BW"],
5163
- "_mm256_mask_slli_epi16"=>["AVX512VL", "AVX512BW"],
5164
- "_mm256_maskz_sll_epi16"=>["AVX512VL", "AVX512BW"],
5165
- "_mm256_maskz_slli_epi16"=>["AVX512VL", "AVX512BW"],
5159
+ "_mm_mask_sllv_epi16"=>[["AVX512VL", "AVX512BW"]],
5160
+ "_mm_maskz_sllv_epi16"=>[["AVX512VL", "AVX512BW"]],
5161
+ "_mm_sllv_epi16"=>[["AVX512VL", "AVX512BW"]],
5162
+ "_mm256_mask_sll_epi16"=>[["AVX512VL", "AVX512BW"]],
5163
+ "_mm256_mask_slli_epi16"=>[["AVX512VL", "AVX512BW"]],
5164
+ "_mm256_maskz_sll_epi16"=>[["AVX512VL", "AVX512BW"]],
5165
+ "_mm256_maskz_slli_epi16"=>[["AVX512VL", "AVX512BW"]],
5166
5166
  "_mm512_mask_sll_epi16"=>["AVX512BW"],
5167
5167
  "_mm512_mask_slli_epi16"=>["AVX512BW"],
5168
5168
  "_mm512_maskz_sll_epi16"=>["AVX512BW"],
5169
5169
  "_mm512_maskz_slli_epi16"=>["AVX512BW"],
5170
5170
  "_mm512_sll_epi16"=>["AVX512BW"],
5171
5171
  "_mm512_slli_epi16"=>["AVX512BW"],
5172
- "_mm_mask_sll_epi16"=>["AVX512VL", "AVX512BW"],
5173
- "_mm_mask_slli_epi16"=>["AVX512VL", "AVX512BW"],
5174
- "_mm_maskz_sll_epi16"=>["AVX512VL", "AVX512BW"],
5175
- "_mm_maskz_slli_epi16"=>["AVX512VL", "AVX512BW"],
5176
- "_mm256_mask_sra_epi32"=>["AVX512VL", "AVX512F"],
5177
- "_mm256_mask_srai_epi32"=>["AVX512VL", "AVX512F"],
5178
- "_mm256_maskz_sra_epi32"=>["AVX512VL", "AVX512F"],
5179
- "_mm256_maskz_srai_epi32"=>["AVX512VL", "AVX512F"],
5180
- "_mm_mask_sra_epi32"=>["AVX512VL", "AVX512F"],
5181
- "_mm_mask_srai_epi32"=>["AVX512VL", "AVX512F"],
5182
- "_mm_maskz_sra_epi32"=>["AVX512VL", "AVX512F"],
5183
- "_mm_maskz_srai_epi32"=>["AVX512VL", "AVX512F"],
5184
- "_mm256_mask_sra_epi64"=>["AVX512VL", "AVX512F"],
5185
- "_mm256_mask_srai_epi64"=>["AVX512VL", "AVX512F"],
5186
- "_mm256_maskz_sra_epi64"=>["AVX512VL", "AVX512F"],
5187
- "_mm256_maskz_srai_epi64"=>["AVX512VL", "AVX512F"],
5188
- "_mm256_sra_epi64"=>["AVX512VL", "AVX512F"],
5189
- "_mm256_srai_epi64"=>["AVX512VL", "AVX512F"],
5190
- "_mm_mask_sra_epi64"=>["AVX512VL", "AVX512F"],
5191
- "_mm_mask_srai_epi64"=>["AVX512VL", "AVX512F"],
5192
- "_mm_maskz_sra_epi64"=>["AVX512VL", "AVX512F"],
5193
- "_mm_maskz_srai_epi64"=>["AVX512VL", "AVX512F"],
5194
- "_mm_sra_epi64"=>["AVX512VL", "AVX512F"],
5195
- "_mm_srai_epi64"=>["AVX512VL", "AVX512F"],
5196
- "_mm256_mask_srav_epi32"=>["AVX512VL", "AVX512F"],
5197
- "_mm256_maskz_srav_epi32"=>["AVX512VL", "AVX512F"],
5198
- "_mm_mask_srav_epi32"=>["AVX512VL", "AVX512F"],
5199
- "_mm_maskz_srav_epi32"=>["AVX512VL", "AVX512F"],
5200
- "_mm256_mask_srav_epi64"=>["AVX512VL", "AVX512F"],
5201
- "_mm256_maskz_srav_epi64"=>["AVX512VL", "AVX512F"],
5202
- "_mm256_srav_epi64"=>["AVX512VL", "AVX512F"],
5203
- "_mm_mask_srav_epi64"=>["AVX512VL", "AVX512F"],
5204
- "_mm_maskz_srav_epi64"=>["AVX512VL", "AVX512F"],
5205
- "_mm_srav_epi64"=>["AVX512VL", "AVX512F"],
5206
- "_mm256_mask_srav_epi16"=>["AVX512VL", "AVX512BW"],
5207
- "_mm256_maskz_srav_epi16"=>["AVX512VL", "AVX512BW"],
5208
- "_mm256_srav_epi16"=>["AVX512VL", "AVX512BW"],
5172
+ "_mm_mask_sll_epi16"=>[["AVX512VL", "AVX512BW"]],
5173
+ "_mm_mask_slli_epi16"=>[["AVX512VL", "AVX512BW"]],
5174
+ "_mm_maskz_sll_epi16"=>[["AVX512VL", "AVX512BW"]],
5175
+ "_mm_maskz_slli_epi16"=>[["AVX512VL", "AVX512BW"]],
5176
+ "_mm256_mask_sra_epi32"=>[["AVX512VL", "AVX512F"]],
5177
+ "_mm256_mask_srai_epi32"=>[["AVX512VL", "AVX512F"]],
5178
+ "_mm256_maskz_sra_epi32"=>[["AVX512VL", "AVX512F"]],
5179
+ "_mm256_maskz_srai_epi32"=>[["AVX512VL", "AVX512F"]],
5180
+ "_mm_mask_sra_epi32"=>[["AVX512VL", "AVX512F"]],
5181
+ "_mm_mask_srai_epi32"=>[["AVX512VL", "AVX512F"]],
5182
+ "_mm_maskz_sra_epi32"=>[["AVX512VL", "AVX512F"]],
5183
+ "_mm_maskz_srai_epi32"=>[["AVX512VL", "AVX512F"]],
5184
+ "_mm256_mask_sra_epi64"=>[["AVX512VL", "AVX512F"]],
5185
+ "_mm256_mask_srai_epi64"=>[["AVX512VL", "AVX512F"]],
5186
+ "_mm256_maskz_sra_epi64"=>[["AVX512VL", "AVX512F"]],
5187
+ "_mm256_maskz_srai_epi64"=>[["AVX512VL", "AVX512F"]],
5188
+ "_mm256_sra_epi64"=>[["AVX512VL", "AVX512F"]],
5189
+ "_mm256_srai_epi64"=>[["AVX512VL", "AVX512F"]],
5190
+ "_mm_mask_sra_epi64"=>[["AVX512VL", "AVX512F"]],
5191
+ "_mm_mask_srai_epi64"=>[["AVX512VL", "AVX512F"]],
5192
+ "_mm_maskz_sra_epi64"=>[["AVX512VL", "AVX512F"]],
5193
+ "_mm_maskz_srai_epi64"=>[["AVX512VL", "AVX512F"]],
5194
+ "_mm_sra_epi64"=>[["AVX512VL", "AVX512F"]],
5195
+ "_mm_srai_epi64"=>[["AVX512VL", "AVX512F"]],
5196
+ "_mm256_mask_srav_epi32"=>[["AVX512VL", "AVX512F"]],
5197
+ "_mm256_maskz_srav_epi32"=>[["AVX512VL", "AVX512F"]],
5198
+ "_mm_mask_srav_epi32"=>[["AVX512VL", "AVX512F"]],
5199
+ "_mm_maskz_srav_epi32"=>[["AVX512VL", "AVX512F"]],
5200
+ "_mm256_mask_srav_epi64"=>[["AVX512VL", "AVX512F"]],
5201
+ "_mm256_maskz_srav_epi64"=>[["AVX512VL", "AVX512F"]],
5202
+ "_mm256_srav_epi64"=>[["AVX512VL", "AVX512F"]],
5203
+ "_mm_mask_srav_epi64"=>[["AVX512VL", "AVX512F"]],
5204
+ "_mm_maskz_srav_epi64"=>[["AVX512VL", "AVX512F"]],
5205
+ "_mm_srav_epi64"=>[["AVX512VL", "AVX512F"]],
5206
+ "_mm256_mask_srav_epi16"=>[["AVX512VL", "AVX512BW"]],
5207
+ "_mm256_maskz_srav_epi16"=>[["AVX512VL", "AVX512BW"]],
5208
+ "_mm256_srav_epi16"=>[["AVX512VL", "AVX512BW"]],
5209
5209
  "_mm512_mask_srav_epi16"=>["AVX512BW"],
5210
5210
  "_mm512_maskz_srav_epi16"=>["AVX512BW"],
5211
5211
  "_mm512_srav_epi16"=>["AVX512BW"],
5212
- "_mm_mask_srav_epi16"=>["AVX512VL", "AVX512BW"],
5213
- "_mm_maskz_srav_epi16"=>["AVX512VL", "AVX512BW"],
5214
- "_mm_srav_epi16"=>["AVX512VL", "AVX512BW"],
5215
- "_mm256_mask_sra_epi16"=>["AVX512VL", "AVX512BW"],
5216
- "_mm256_mask_srai_epi16"=>["AVX512VL", "AVX512BW"],
5217
- "_mm256_maskz_sra_epi16"=>["AVX512VL", "AVX512BW"],
5218
- "_mm256_maskz_srai_epi16"=>["AVX512VL", "AVX512BW"],
5212
+ "_mm_mask_srav_epi16"=>[["AVX512VL", "AVX512BW"]],
5213
+ "_mm_maskz_srav_epi16"=>[["AVX512VL", "AVX512BW"]],
5214
+ "_mm_srav_epi16"=>[["AVX512VL", "AVX512BW"]],
5215
+ "_mm256_mask_sra_epi16"=>[["AVX512VL", "AVX512BW"]],
5216
+ "_mm256_mask_srai_epi16"=>[["AVX512VL", "AVX512BW"]],
5217
+ "_mm256_maskz_sra_epi16"=>[["AVX512VL", "AVX512BW"]],
5218
+ "_mm256_maskz_srai_epi16"=>[["AVX512VL", "AVX512BW"]],
5219
5219
  "_mm512_mask_sra_epi16"=>["AVX512BW"],
5220
5220
  "_mm512_mask_srai_epi16"=>["AVX512BW"],
5221
5221
  "_mm512_maskz_sra_epi16"=>["AVX512BW"],
5222
5222
  "_mm512_maskz_srai_epi16"=>["AVX512BW"],
5223
5223
  "_mm512_sra_epi16"=>["AVX512BW"],
5224
5224
  "_mm512_srai_epi16"=>["AVX512BW"],
5225
- "_mm_mask_sra_epi16"=>["AVX512VL", "AVX512BW"],
5226
- "_mm_mask_srai_epi16"=>["AVX512VL", "AVX512BW"],
5227
- "_mm_maskz_sra_epi16"=>["AVX512VL", "AVX512BW"],
5228
- "_mm_maskz_srai_epi16"=>["AVX512VL", "AVX512BW"],
5229
- "_mm256_mask_srl_epi32"=>["AVX512VL", "AVX512F"],
5230
- "_mm256_mask_srli_epi32"=>["AVX512VL", "AVX512F"],
5231
- "_mm256_maskz_srl_epi32"=>["AVX512VL", "AVX512F"],
5232
- "_mm256_maskz_srli_epi32"=>["AVX512VL", "AVX512F"],
5233
- "_mm_mask_srl_epi32"=>["AVX512VL", "AVX512F"],
5234
- "_mm_mask_srli_epi32"=>["AVX512VL", "AVX512F"],
5235
- "_mm_maskz_srl_epi32"=>["AVX512VL", "AVX512F"],
5236
- "_mm_maskz_srli_epi32"=>["AVX512VL", "AVX512F"],
5225
+ "_mm_mask_sra_epi16"=>[["AVX512VL", "AVX512BW"]],
5226
+ "_mm_mask_srai_epi16"=>[["AVX512VL", "AVX512BW"]],
5227
+ "_mm_maskz_sra_epi16"=>[["AVX512VL", "AVX512BW"]],
5228
+ "_mm_maskz_srai_epi16"=>[["AVX512VL", "AVX512BW"]],
5229
+ "_mm256_mask_srl_epi32"=>[["AVX512VL", "AVX512F"]],
5230
+ "_mm256_mask_srli_epi32"=>[["AVX512VL", "AVX512F"]],
5231
+ "_mm256_maskz_srl_epi32"=>[["AVX512VL", "AVX512F"]],
5232
+ "_mm256_maskz_srli_epi32"=>[["AVX512VL", "AVX512F"]],
5233
+ "_mm_mask_srl_epi32"=>[["AVX512VL", "AVX512F"]],
5234
+ "_mm_mask_srli_epi32"=>[["AVX512VL", "AVX512F"]],
5235
+ "_mm_maskz_srl_epi32"=>[["AVX512VL", "AVX512F"]],
5236
+ "_mm_maskz_srli_epi32"=>[["AVX512VL", "AVX512F"]],
5237
5237
  "_mm512_bsrli_epi128"=>["AVX512BW"],
5238
- "_mm256_mask_srl_epi64"=>["AVX512VL", "AVX512F"],
5239
- "_mm256_mask_srli_epi64"=>["AVX512VL", "AVX512F"],
5240
- "_mm256_maskz_srl_epi64"=>["AVX512VL", "AVX512F"],
5241
- "_mm256_maskz_srli_epi64"=>["AVX512VL", "AVX512F"],
5242
- "_mm_mask_srl_epi64"=>["AVX512VL", "AVX512F"],
5243
- "_mm_mask_srli_epi64"=>["AVX512VL", "AVX512F"],
5244
- "_mm_maskz_srl_epi64"=>["AVX512VL", "AVX512F"],
5245
- "_mm_maskz_srli_epi64"=>["AVX512VL", "AVX512F"],
5246
- "_mm256_mask_srlv_epi32"=>["AVX512VL", "AVX512F"],
5247
- "_mm256_maskz_srlv_epi32"=>["AVX512VL", "AVX512F"],
5248
- "_mm_mask_srlv_epi32"=>["AVX512VL", "AVX512F"],
5249
- "_mm_maskz_srlv_epi32"=>["AVX512VL", "AVX512F"],
5250
- "_mm256_mask_srlv_epi64"=>["AVX512VL", "AVX512F"],
5251
- "_mm256_maskz_srlv_epi64"=>["AVX512VL", "AVX512F"],
5252
- "_mm_mask_srlv_epi64"=>["AVX512VL", "AVX512F"],
5253
- "_mm_maskz_srlv_epi64"=>["AVX512VL", "AVX512F"],
5254
- "_mm256_mask_srlv_epi16"=>["AVX512VL", "AVX512BW"],
5255
- "_mm256_maskz_srlv_epi16"=>["AVX512VL", "AVX512BW"],
5256
- "_mm256_srlv_epi16"=>["AVX512VL", "AVX512BW"],
5238
+ "_mm256_mask_srl_epi64"=>[["AVX512VL", "AVX512F"]],
5239
+ "_mm256_mask_srli_epi64"=>[["AVX512VL", "AVX512F"]],
5240
+ "_mm256_maskz_srl_epi64"=>[["AVX512VL", "AVX512F"]],
5241
+ "_mm256_maskz_srli_epi64"=>[["AVX512VL", "AVX512F"]],
5242
+ "_mm_mask_srl_epi64"=>[["AVX512VL", "AVX512F"]],
5243
+ "_mm_mask_srli_epi64"=>[["AVX512VL", "AVX512F"]],
5244
+ "_mm_maskz_srl_epi64"=>[["AVX512VL", "AVX512F"]],
5245
+ "_mm_maskz_srli_epi64"=>[["AVX512VL", "AVX512F"]],
5246
+ "_mm256_mask_srlv_epi32"=>[["AVX512VL", "AVX512F"]],
5247
+ "_mm256_maskz_srlv_epi32"=>[["AVX512VL", "AVX512F"]],
5248
+ "_mm_mask_srlv_epi32"=>[["AVX512VL", "AVX512F"]],
5249
+ "_mm_maskz_srlv_epi32"=>[["AVX512VL", "AVX512F"]],
5250
+ "_mm256_mask_srlv_epi64"=>[["AVX512VL", "AVX512F"]],
5251
+ "_mm256_maskz_srlv_epi64"=>[["AVX512VL", "AVX512F"]],
5252
+ "_mm_mask_srlv_epi64"=>[["AVX512VL", "AVX512F"]],
5253
+ "_mm_maskz_srlv_epi64"=>[["AVX512VL", "AVX512F"]],
5254
+ "_mm256_mask_srlv_epi16"=>[["AVX512VL", "AVX512BW"]],
5255
+ "_mm256_maskz_srlv_epi16"=>[["AVX512VL", "AVX512BW"]],
5256
+ "_mm256_srlv_epi16"=>[["AVX512VL", "AVX512BW"]],
5257
5257
  "_mm512_mask_srlv_epi16"=>["AVX512BW"],
5258
5258
  "_mm512_maskz_srlv_epi16"=>["AVX512BW"],
5259
5259
  "_mm512_srlv_epi16"=>["AVX512BW"],
5260
- "_mm_mask_srlv_epi16"=>["AVX512VL", "AVX512BW"],
5261
- "_mm_maskz_srlv_epi16"=>["AVX512VL", "AVX512BW"],
5262
- "_mm_srlv_epi16"=>["AVX512VL", "AVX512BW"],
5263
- "_mm256_mask_srl_epi16"=>["AVX512VL", "AVX512BW"],
5264
- "_mm256_mask_srli_epi16"=>["AVX512VL", "AVX512BW"],
5265
- "_mm256_maskz_srl_epi16"=>["AVX512VL", "AVX512BW"],
5266
- "_mm256_maskz_srli_epi16"=>["AVX512VL", "AVX512BW"],
5260
+ "_mm_mask_srlv_epi16"=>[["AVX512VL", "AVX512BW"]],
5261
+ "_mm_maskz_srlv_epi16"=>[["AVX512VL", "AVX512BW"]],
5262
+ "_mm_srlv_epi16"=>[["AVX512VL", "AVX512BW"]],
5263
+ "_mm256_mask_srl_epi16"=>[["AVX512VL", "AVX512BW"]],
5264
+ "_mm256_mask_srli_epi16"=>[["AVX512VL", "AVX512BW"]],
5265
+ "_mm256_maskz_srl_epi16"=>[["AVX512VL", "AVX512BW"]],
5266
+ "_mm256_maskz_srli_epi16"=>[["AVX512VL", "AVX512BW"]],
5267
5267
  "_mm512_mask_srl_epi16"=>["AVX512BW"],
5268
5268
  "_mm512_mask_srli_epi16"=>["AVX512BW"],
5269
5269
  "_mm512_maskz_srl_epi16"=>["AVX512BW"],
5270
5270
  "_mm512_maskz_srli_epi16"=>["AVX512BW"],
5271
5271
  "_mm512_srl_epi16"=>["AVX512BW"],
5272
5272
  "_mm512_srli_epi16"=>["AVX512BW"],
5273
- "_mm_mask_srl_epi16"=>["AVX512VL", "AVX512BW"],
5274
- "_mm_mask_srli_epi16"=>["AVX512VL", "AVX512BW"],
5275
- "_mm_maskz_srl_epi16"=>["AVX512VL", "AVX512BW"],
5276
- "_mm_maskz_srli_epi16"=>["AVX512VL", "AVX512BW"],
5277
- "_mm256_mask_sub_epi8"=>["AVX512VL", "AVX512BW"],
5278
- "_mm256_maskz_sub_epi8"=>["AVX512VL", "AVX512BW"],
5273
+ "_mm_mask_srl_epi16"=>[["AVX512VL", "AVX512BW"]],
5274
+ "_mm_mask_srli_epi16"=>[["AVX512VL", "AVX512BW"]],
5275
+ "_mm_maskz_srl_epi16"=>[["AVX512VL", "AVX512BW"]],
5276
+ "_mm_maskz_srli_epi16"=>[["AVX512VL", "AVX512BW"]],
5277
+ "_mm256_mask_sub_epi8"=>[["AVX512VL", "AVX512BW"]],
5278
+ "_mm256_maskz_sub_epi8"=>[["AVX512VL", "AVX512BW"]],
5279
5279
  "_mm512_mask_sub_epi8"=>["AVX512BW"],
5280
5280
  "_mm512_maskz_sub_epi8"=>["AVX512BW"],
5281
5281
  "_mm512_sub_epi8"=>["AVX512BW"],
5282
- "_mm_mask_sub_epi8"=>["AVX512VL", "AVX512BW"],
5283
- "_mm_maskz_sub_epi8"=>["AVX512VL", "AVX512BW"],
5284
- "_mm256_mask_sub_epi32"=>["AVX512VL", "AVX512F"],
5285
- "_mm256_maskz_sub_epi32"=>["AVX512VL", "AVX512F"],
5286
- "_mm_mask_sub_epi32"=>["AVX512VL", "AVX512F"],
5287
- "_mm_maskz_sub_epi32"=>["AVX512VL", "AVX512F"],
5288
- "_mm256_mask_sub_epi64"=>["AVX512VL", "AVX512F"],
5289
- "_mm256_maskz_sub_epi64"=>["AVX512VL", "AVX512F"],
5290
- "_mm_mask_sub_epi64"=>["AVX512VL", "AVX512F"],
5291
- "_mm_maskz_sub_epi64"=>["AVX512VL", "AVX512F"],
5292
- "_mm256_mask_subs_epi8"=>["AVX512VL", "AVX512BW"],
5293
- "_mm256_maskz_subs_epi8"=>["AVX512VL", "AVX512BW"],
5282
+ "_mm_mask_sub_epi8"=>[["AVX512VL", "AVX512BW"]],
5283
+ "_mm_maskz_sub_epi8"=>[["AVX512VL", "AVX512BW"]],
5284
+ "_mm256_mask_sub_epi32"=>[["AVX512VL", "AVX512F"]],
5285
+ "_mm256_maskz_sub_epi32"=>[["AVX512VL", "AVX512F"]],
5286
+ "_mm_mask_sub_epi32"=>[["AVX512VL", "AVX512F"]],
5287
+ "_mm_maskz_sub_epi32"=>[["AVX512VL", "AVX512F"]],
5288
+ "_mm256_mask_sub_epi64"=>[["AVX512VL", "AVX512F"]],
5289
+ "_mm256_maskz_sub_epi64"=>[["AVX512VL", "AVX512F"]],
5290
+ "_mm_mask_sub_epi64"=>[["AVX512VL", "AVX512F"]],
5291
+ "_mm_maskz_sub_epi64"=>[["AVX512VL", "AVX512F"]],
5292
+ "_mm256_mask_subs_epi8"=>[["AVX512VL", "AVX512BW"]],
5293
+ "_mm256_maskz_subs_epi8"=>[["AVX512VL", "AVX512BW"]],
5294
5294
  "_mm512_mask_subs_epi8"=>["AVX512BW"],
5295
5295
  "_mm512_maskz_subs_epi8"=>["AVX512BW"],
5296
5296
  "_mm512_subs_epi8"=>["AVX512BW"],
5297
- "_mm_mask_subs_epi8"=>["AVX512VL", "AVX512BW"],
5298
- "_mm_maskz_subs_epi8"=>["AVX512VL", "AVX512BW"],
5299
- "_mm256_mask_subs_epi16"=>["AVX512VL", "AVX512BW"],
5300
- "_mm256_maskz_subs_epi16"=>["AVX512VL", "AVX512BW"],
5297
+ "_mm_mask_subs_epi8"=>[["AVX512VL", "AVX512BW"]],
5298
+ "_mm_maskz_subs_epi8"=>[["AVX512VL", "AVX512BW"]],
5299
+ "_mm256_mask_subs_epi16"=>[["AVX512VL", "AVX512BW"]],
5300
+ "_mm256_maskz_subs_epi16"=>[["AVX512VL", "AVX512BW"]],
5301
5301
  "_mm512_mask_subs_epi16"=>["AVX512BW"],
5302
5302
  "_mm512_maskz_subs_epi16"=>["AVX512BW"],
5303
5303
  "_mm512_subs_epi16"=>["AVX512BW"],
5304
- "_mm_mask_subs_epi16"=>["AVX512VL", "AVX512BW"],
5305
- "_mm_maskz_subs_epi16"=>["AVX512VL", "AVX512BW"],
5306
- "_mm256_mask_subs_epu8"=>["AVX512VL", "AVX512BW"],
5307
- "_mm256_maskz_subs_epu8"=>["AVX512VL", "AVX512BW"],
5304
+ "_mm_mask_subs_epi16"=>[["AVX512VL", "AVX512BW"]],
5305
+ "_mm_maskz_subs_epi16"=>[["AVX512VL", "AVX512BW"]],
5306
+ "_mm256_mask_subs_epu8"=>[["AVX512VL", "AVX512BW"]],
5307
+ "_mm256_maskz_subs_epu8"=>[["AVX512VL", "AVX512BW"]],
5308
5308
  "_mm512_mask_subs_epu8"=>["AVX512BW"],
5309
5309
  "_mm512_maskz_subs_epu8"=>["AVX512BW"],
5310
5310
  "_mm512_subs_epu8"=>["AVX512BW"],
5311
- "_mm_mask_subs_epu8"=>["AVX512VL", "AVX512BW"],
5312
- "_mm_maskz_subs_epu8"=>["AVX512VL", "AVX512BW"],
5313
- "_mm256_mask_subs_epu16"=>["AVX512VL", "AVX512BW"],
5314
- "_mm256_maskz_subs_epu16"=>["AVX512VL", "AVX512BW"],
5311
+ "_mm_mask_subs_epu8"=>[["AVX512VL", "AVX512BW"]],
5312
+ "_mm_maskz_subs_epu8"=>[["AVX512VL", "AVX512BW"]],
5313
+ "_mm256_mask_subs_epu16"=>[["AVX512VL", "AVX512BW"]],
5314
+ "_mm256_maskz_subs_epu16"=>[["AVX512VL", "AVX512BW"]],
5315
5315
  "_mm512_mask_subs_epu16"=>["AVX512BW"],
5316
5316
  "_mm512_maskz_subs_epu16"=>["AVX512BW"],
5317
5317
  "_mm512_subs_epu16"=>["AVX512BW"],
5318
- "_mm_mask_subs_epu16"=>["AVX512VL", "AVX512BW"],
5319
- "_mm_maskz_subs_epu16"=>["AVX512VL", "AVX512BW"],
5320
- "_mm256_mask_sub_epi16"=>["AVX512VL", "AVX512BW"],
5321
- "_mm256_maskz_sub_epi16"=>["AVX512VL", "AVX512BW"],
5318
+ "_mm_mask_subs_epu16"=>[["AVX512VL", "AVX512BW"]],
5319
+ "_mm_maskz_subs_epu16"=>[["AVX512VL", "AVX512BW"]],
5320
+ "_mm256_mask_sub_epi16"=>[["AVX512VL", "AVX512BW"]],
5321
+ "_mm256_maskz_sub_epi16"=>[["AVX512VL", "AVX512BW"]],
5322
5322
  "_mm512_mask_sub_epi16"=>["AVX512BW"],
5323
5323
  "_mm512_maskz_sub_epi16"=>["AVX512BW"],
5324
5324
  "_mm512_sub_epi16"=>["AVX512BW"],
5325
- "_mm_mask_sub_epi16"=>["AVX512VL", "AVX512BW"],
5326
- "_mm_maskz_sub_epi16"=>["AVX512VL", "AVX512BW"],
5327
- "_mm256_mask_ternarylogic_epi32"=>["AVX512VL", "AVX512F"],
5328
- "_mm256_maskz_ternarylogic_epi32"=>["AVX512VL", "AVX512F"],
5329
- "_mm256_ternarylogic_epi32"=>["AVX512VL", "AVX512F"],
5330
- "_mm_mask_ternarylogic_epi32"=>["AVX512VL", "AVX512F"],
5331
- "_mm_maskz_ternarylogic_epi32"=>["AVX512VL", "AVX512F"],
5332
- "_mm_ternarylogic_epi32"=>["AVX512VL", "AVX512F"],
5333
- "_mm256_mask_ternarylogic_epi64"=>["AVX512VL", "AVX512F"],
5334
- "_mm256_maskz_ternarylogic_epi64"=>["AVX512VL", "AVX512F"],
5335
- "_mm256_ternarylogic_epi64"=>["AVX512VL", "AVX512F"],
5336
- "_mm_mask_ternarylogic_epi64"=>["AVX512VL", "AVX512F"],
5337
- "_mm_maskz_ternarylogic_epi64"=>["AVX512VL", "AVX512F"],
5338
- "_mm_ternarylogic_epi64"=>["AVX512VL", "AVX512F"],
5339
- "_mm256_mask_test_epi8_mask"=>["AVX512VL", "AVX512BW"],
5340
- "_mm256_test_epi8_mask"=>["AVX512VL", "AVX512BW"],
5325
+ "_mm_mask_sub_epi16"=>[["AVX512VL", "AVX512BW"]],
5326
+ "_mm_maskz_sub_epi16"=>[["AVX512VL", "AVX512BW"]],
5327
+ "_mm256_mask_ternarylogic_epi32"=>[["AVX512VL", "AVX512F"]],
5328
+ "_mm256_maskz_ternarylogic_epi32"=>[["AVX512VL", "AVX512F"]],
5329
+ "_mm256_ternarylogic_epi32"=>[["AVX512VL", "AVX512F"]],
5330
+ "_mm_mask_ternarylogic_epi32"=>[["AVX512VL", "AVX512F"]],
5331
+ "_mm_maskz_ternarylogic_epi32"=>[["AVX512VL", "AVX512F"]],
5332
+ "_mm_ternarylogic_epi32"=>[["AVX512VL", "AVX512F"]],
5333
+ "_mm256_mask_ternarylogic_epi64"=>[["AVX512VL", "AVX512F"]],
5334
+ "_mm256_maskz_ternarylogic_epi64"=>[["AVX512VL", "AVX512F"]],
5335
+ "_mm256_ternarylogic_epi64"=>[["AVX512VL", "AVX512F"]],
5336
+ "_mm_mask_ternarylogic_epi64"=>[["AVX512VL", "AVX512F"]],
5337
+ "_mm_maskz_ternarylogic_epi64"=>[["AVX512VL", "AVX512F"]],
5338
+ "_mm_ternarylogic_epi64"=>[["AVX512VL", "AVX512F"]],
5339
+ "_mm256_mask_test_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
5340
+ "_mm256_test_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
5341
5341
  "_mm512_mask_test_epi8_mask"=>["AVX512BW"],
5342
5342
  "_mm512_test_epi8_mask"=>["AVX512BW"],
5343
- "_mm_mask_test_epi8_mask"=>["AVX512VL", "AVX512BW"],
5344
- "_mm_test_epi8_mask"=>["AVX512VL", "AVX512BW"],
5345
- "_mm256_mask_test_epi32_mask"=>["AVX512VL", "AVX512F"],
5346
- "_mm256_test_epi32_mask"=>["AVX512VL", "AVX512F"],
5347
- "_mm_mask_test_epi32_mask"=>["AVX512VL", "AVX512F"],
5348
- "_mm_test_epi32_mask"=>["AVX512VL", "AVX512F"],
5349
- "_mm256_mask_test_epi64_mask"=>["AVX512VL", "AVX512F"],
5350
- "_mm256_test_epi64_mask"=>["AVX512VL", "AVX512F"],
5351
- "_mm_mask_test_epi64_mask"=>["AVX512VL", "AVX512F"],
5352
- "_mm_test_epi64_mask"=>["AVX512VL", "AVX512F"],
5353
- "_mm256_mask_test_epi16_mask"=>["AVX512VL", "AVX512BW"],
5354
- "_mm256_test_epi16_mask"=>["AVX512VL", "AVX512BW"],
5343
+ "_mm_mask_test_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
5344
+ "_mm_test_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
5345
+ "_mm256_mask_test_epi32_mask"=>[["AVX512VL", "AVX512F"]],
5346
+ "_mm256_test_epi32_mask"=>[["AVX512VL", "AVX512F"]],
5347
+ "_mm_mask_test_epi32_mask"=>[["AVX512VL", "AVX512F"]],
5348
+ "_mm_test_epi32_mask"=>[["AVX512VL", "AVX512F"]],
5349
+ "_mm256_mask_test_epi64_mask"=>[["AVX512VL", "AVX512F"]],
5350
+ "_mm256_test_epi64_mask"=>[["AVX512VL", "AVX512F"]],
5351
+ "_mm_mask_test_epi64_mask"=>[["AVX512VL", "AVX512F"]],
5352
+ "_mm_test_epi64_mask"=>[["AVX512VL", "AVX512F"]],
5353
+ "_mm256_mask_test_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
5354
+ "_mm256_test_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
5355
5355
  "_mm512_mask_test_epi16_mask"=>["AVX512BW"],
5356
5356
  "_mm512_test_epi16_mask"=>["AVX512BW"],
5357
- "_mm_mask_test_epi16_mask"=>["AVX512VL", "AVX512BW"],
5358
- "_mm_test_epi16_mask"=>["AVX512VL", "AVX512BW"],
5359
- "_mm256_mask_testn_epi8_mask"=>["AVX512VL", "AVX512BW"],
5360
- "_mm256_testn_epi8_mask"=>["AVX512VL", "AVX512BW"],
5357
+ "_mm_mask_test_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
5358
+ "_mm_test_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
5359
+ "_mm256_mask_testn_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
5360
+ "_mm256_testn_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
5361
5361
  "_mm512_mask_testn_epi8_mask"=>["AVX512BW"],
5362
5362
  "_mm512_testn_epi8_mask"=>["AVX512BW"],
5363
- "_mm_mask_testn_epi8_mask"=>["AVX512VL", "AVX512BW"],
5364
- "_mm_testn_epi8_mask"=>["AVX512VL", "AVX512BW"],
5365
- "_mm256_mask_testn_epi32_mask"=>["AVX512VL", "AVX512F"],
5366
- "_mm256_testn_epi32_mask"=>["AVX512VL", "AVX512F"],
5367
- "_mm_mask_testn_epi32_mask"=>["AVX512VL", "AVX512F"],
5368
- "_mm_testn_epi32_mask"=>["AVX512VL", "AVX512F"],
5369
- "_mm256_mask_testn_epi64_mask"=>["AVX512VL", "AVX512F"],
5370
- "_mm256_testn_epi64_mask"=>["AVX512VL", "AVX512F"],
5371
- "_mm_mask_testn_epi64_mask"=>["AVX512VL", "AVX512F"],
5372
- "_mm_testn_epi64_mask"=>["AVX512VL", "AVX512F"],
5373
- "_mm256_mask_testn_epi16_mask"=>["AVX512VL", "AVX512BW"],
5374
- "_mm256_testn_epi16_mask"=>["AVX512VL", "AVX512BW"],
5363
+ "_mm_mask_testn_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
5364
+ "_mm_testn_epi8_mask"=>[["AVX512VL", "AVX512BW"]],
5365
+ "_mm256_mask_testn_epi32_mask"=>[["AVX512VL", "AVX512F"]],
5366
+ "_mm256_testn_epi32_mask"=>[["AVX512VL", "AVX512F"]],
5367
+ "_mm_mask_testn_epi32_mask"=>[["AVX512VL", "AVX512F"]],
5368
+ "_mm_testn_epi32_mask"=>[["AVX512VL", "AVX512F"]],
5369
+ "_mm256_mask_testn_epi64_mask"=>[["AVX512VL", "AVX512F"]],
5370
+ "_mm256_testn_epi64_mask"=>[["AVX512VL", "AVX512F"]],
5371
+ "_mm_mask_testn_epi64_mask"=>[["AVX512VL", "AVX512F"]],
5372
+ "_mm_testn_epi64_mask"=>[["AVX512VL", "AVX512F"]],
5373
+ "_mm256_mask_testn_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
5374
+ "_mm256_testn_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
5375
5375
  "_mm512_mask_testn_epi16_mask"=>["AVX512BW"],
5376
5376
  "_mm512_testn_epi16_mask"=>["AVX512BW"],
5377
- "_mm_mask_testn_epi16_mask"=>["AVX512VL", "AVX512BW"],
5378
- "_mm_testn_epi16_mask"=>["AVX512VL", "AVX512BW"],
5379
- "_mm256_mask_unpackhi_epi8"=>["AVX512VL", "AVX512BW"],
5380
- "_mm256_maskz_unpackhi_epi8"=>["AVX512VL", "AVX512BW"],
5377
+ "_mm_mask_testn_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
5378
+ "_mm_testn_epi16_mask"=>[["AVX512VL", "AVX512BW"]],
5379
+ "_mm256_mask_unpackhi_epi8"=>[["AVX512VL", "AVX512BW"]],
5380
+ "_mm256_maskz_unpackhi_epi8"=>[["AVX512VL", "AVX512BW"]],
5381
5381
  "_mm512_mask_unpackhi_epi8"=>["AVX512BW"],
5382
5382
  "_mm512_maskz_unpackhi_epi8"=>["AVX512BW"],
5383
5383
  "_mm512_unpackhi_epi8"=>["AVX512BW"],
5384
- "_mm_mask_unpackhi_epi8"=>["AVX512VL", "AVX512BW"],
5385
- "_mm_maskz_unpackhi_epi8"=>["AVX512VL", "AVX512BW"],
5386
- "_mm256_mask_unpackhi_epi32"=>["AVX512VL", "AVX512F"],
5387
- "_mm256_maskz_unpackhi_epi32"=>["AVX512VL", "AVX512F"],
5388
- "_mm_mask_unpackhi_epi32"=>["AVX512VL", "AVX512F"],
5389
- "_mm_maskz_unpackhi_epi32"=>["AVX512VL", "AVX512F"],
5390
- "_mm256_mask_unpackhi_epi64"=>["AVX512VL", "AVX512F"],
5391
- "_mm256_maskz_unpackhi_epi64"=>["AVX512VL", "AVX512F"],
5392
- "_mm_mask_unpackhi_epi64"=>["AVX512VL", "AVX512F"],
5393
- "_mm_maskz_unpackhi_epi64"=>["AVX512VL", "AVX512F"],
5394
- "_mm256_mask_unpackhi_epi16"=>["AVX512VL", "AVX512BW"],
5395
- "_mm256_maskz_unpackhi_epi16"=>["AVX512VL", "AVX512BW"],
5384
+ "_mm_mask_unpackhi_epi8"=>[["AVX512VL", "AVX512BW"]],
5385
+ "_mm_maskz_unpackhi_epi8"=>[["AVX512VL", "AVX512BW"]],
5386
+ "_mm256_mask_unpackhi_epi32"=>[["AVX512VL", "AVX512F"]],
5387
+ "_mm256_maskz_unpackhi_epi32"=>[["AVX512VL", "AVX512F"]],
5388
+ "_mm_mask_unpackhi_epi32"=>[["AVX512VL", "AVX512F"]],
5389
+ "_mm_maskz_unpackhi_epi32"=>[["AVX512VL", "AVX512F"]],
5390
+ "_mm256_mask_unpackhi_epi64"=>[["AVX512VL", "AVX512F"]],
5391
+ "_mm256_maskz_unpackhi_epi64"=>[["AVX512VL", "AVX512F"]],
5392
+ "_mm_mask_unpackhi_epi64"=>[["AVX512VL", "AVX512F"]],
5393
+ "_mm_maskz_unpackhi_epi64"=>[["AVX512VL", "AVX512F"]],
5394
+ "_mm256_mask_unpackhi_epi16"=>[["AVX512VL", "AVX512BW"]],
5395
+ "_mm256_maskz_unpackhi_epi16"=>[["AVX512VL", "AVX512BW"]],
5396
5396
  "_mm512_mask_unpackhi_epi16"=>["AVX512BW"],
5397
5397
  "_mm512_maskz_unpackhi_epi16"=>["AVX512BW"],
5398
5398
  "_mm512_unpackhi_epi16"=>["AVX512BW"],
5399
- "_mm_mask_unpackhi_epi16"=>["AVX512VL", "AVX512BW"],
5400
- "_mm_maskz_unpackhi_epi16"=>["AVX512VL", "AVX512BW"],
5401
- "_mm256_mask_unpacklo_epi8"=>["AVX512VL", "AVX512BW"],
5402
- "_mm256_maskz_unpacklo_epi8"=>["AVX512VL", "AVX512BW"],
5399
+ "_mm_mask_unpackhi_epi16"=>[["AVX512VL", "AVX512BW"]],
5400
+ "_mm_maskz_unpackhi_epi16"=>[["AVX512VL", "AVX512BW"]],
5401
+ "_mm256_mask_unpacklo_epi8"=>[["AVX512VL", "AVX512BW"]],
5402
+ "_mm256_maskz_unpacklo_epi8"=>[["AVX512VL", "AVX512BW"]],
5403
5403
  "_mm512_mask_unpacklo_epi8"=>["AVX512BW"],
5404
5404
  "_mm512_maskz_unpacklo_epi8"=>["AVX512BW"],
5405
5405
  "_mm512_unpacklo_epi8"=>["AVX512BW"],
5406
- "_mm_mask_unpacklo_epi8"=>["AVX512VL", "AVX512BW"],
5407
- "_mm_maskz_unpacklo_epi8"=>["AVX512VL", "AVX512BW"],
5408
- "_mm256_mask_unpacklo_epi32"=>["AVX512VL", "AVX512F"],
5409
- "_mm256_maskz_unpacklo_epi32"=>["AVX512VL", "AVX512F"],
5410
- "_mm_mask_unpacklo_epi32"=>["AVX512VL", "AVX512F"],
5411
- "_mm_maskz_unpacklo_epi32"=>["AVX512VL", "AVX512F"],
5412
- "_mm256_mask_unpacklo_epi64"=>["AVX512VL", "AVX512F"],
5413
- "_mm256_maskz_unpacklo_epi64"=>["AVX512VL", "AVX512F"],
5414
- "_mm_mask_unpacklo_epi64"=>["AVX512VL", "AVX512F"],
5415
- "_mm_maskz_unpacklo_epi64"=>["AVX512VL", "AVX512F"],
5416
- "_mm256_mask_unpacklo_epi16"=>["AVX512VL", "AVX512BW"],
5417
- "_mm256_maskz_unpacklo_epi16"=>["AVX512VL", "AVX512BW"],
5406
+ "_mm_mask_unpacklo_epi8"=>[["AVX512VL", "AVX512BW"]],
5407
+ "_mm_maskz_unpacklo_epi8"=>[["AVX512VL", "AVX512BW"]],
5408
+ "_mm256_mask_unpacklo_epi32"=>[["AVX512VL", "AVX512F"]],
5409
+ "_mm256_maskz_unpacklo_epi32"=>[["AVX512VL", "AVX512F"]],
5410
+ "_mm_mask_unpacklo_epi32"=>[["AVX512VL", "AVX512F"]],
5411
+ "_mm_maskz_unpacklo_epi32"=>[["AVX512VL", "AVX512F"]],
5412
+ "_mm256_mask_unpacklo_epi64"=>[["AVX512VL", "AVX512F"]],
5413
+ "_mm256_maskz_unpacklo_epi64"=>[["AVX512VL", "AVX512F"]],
5414
+ "_mm_mask_unpacklo_epi64"=>[["AVX512VL", "AVX512F"]],
5415
+ "_mm_maskz_unpacklo_epi64"=>[["AVX512VL", "AVX512F"]],
5416
+ "_mm256_mask_unpacklo_epi16"=>[["AVX512VL", "AVX512BW"]],
5417
+ "_mm256_maskz_unpacklo_epi16"=>[["AVX512VL", "AVX512BW"]],
5418
5418
  "_mm512_mask_unpacklo_epi16"=>["AVX512BW"],
5419
5419
  "_mm512_maskz_unpacklo_epi16"=>["AVX512BW"],
5420
5420
  "_mm512_unpacklo_epi16"=>["AVX512BW"],
5421
- "_mm_mask_unpacklo_epi16"=>["AVX512VL", "AVX512BW"],
5422
- "_mm_maskz_unpacklo_epi16"=>["AVX512VL", "AVX512BW"],
5423
- "_mm256_mask_xor_epi32"=>["AVX512VL", "AVX512F"],
5424
- "_mm256_maskz_xor_epi32"=>["AVX512VL", "AVX512F"],
5425
- "_mm_mask_xor_epi32"=>["AVX512VL", "AVX512F"],
5426
- "_mm_maskz_xor_epi32"=>["AVX512VL", "AVX512F"],
5427
- "_mm256_mask_xor_epi64"=>["AVX512VL", "AVX512F"],
5428
- "_mm256_maskz_xor_epi64"=>["AVX512VL", "AVX512F"],
5429
- "_mm_mask_xor_epi64"=>["AVX512VL", "AVX512F"],
5430
- "_mm_maskz_xor_epi64"=>["AVX512VL", "AVX512F"],
5431
- "_mm256_mask_range_pd"=>["AVX512VL", "AVX512DQ"],
5432
- "_mm256_maskz_range_pd"=>["AVX512VL", "AVX512DQ"],
5433
- "_mm256_range_pd"=>["AVX512VL", "AVX512DQ"],
5421
+ "_mm_mask_unpacklo_epi16"=>[["AVX512VL", "AVX512BW"]],
5422
+ "_mm_maskz_unpacklo_epi16"=>[["AVX512VL", "AVX512BW"]],
5423
+ "_mm256_mask_xor_epi32"=>[["AVX512VL", "AVX512F"]],
5424
+ "_mm256_maskz_xor_epi32"=>[["AVX512VL", "AVX512F"]],
5425
+ "_mm_mask_xor_epi32"=>[["AVX512VL", "AVX512F"]],
5426
+ "_mm_maskz_xor_epi32"=>[["AVX512VL", "AVX512F"]],
5427
+ "_mm256_mask_xor_epi64"=>[["AVX512VL", "AVX512F"]],
5428
+ "_mm256_maskz_xor_epi64"=>[["AVX512VL", "AVX512F"]],
5429
+ "_mm_mask_xor_epi64"=>[["AVX512VL", "AVX512F"]],
5430
+ "_mm_maskz_xor_epi64"=>[["AVX512VL", "AVX512F"]],
5431
+ "_mm256_mask_range_pd"=>[["AVX512VL", "AVX512DQ"]],
5432
+ "_mm256_maskz_range_pd"=>[["AVX512VL", "AVX512DQ"]],
5433
+ "_mm256_range_pd"=>[["AVX512VL", "AVX512DQ"]],
5434
5434
  "_mm512_mask_range_pd"=>["AVX512DQ"],
5435
5435
  "_mm512_mask_range_round_pd"=>["AVX512DQ"],
5436
5436
  "_mm512_maskz_range_pd"=>["AVX512DQ"],
5437
5437
  "_mm512_maskz_range_round_pd"=>["AVX512DQ"],
5438
5438
  "_mm512_range_pd"=>["AVX512DQ"],
5439
5439
  "_mm512_range_round_pd"=>["AVX512DQ"],
5440
- "_mm_mask_range_pd"=>["AVX512VL", "AVX512DQ"],
5441
- "_mm_maskz_range_pd"=>["AVX512VL", "AVX512DQ"],
5442
- "_mm_range_pd"=>["AVX512VL", "AVX512DQ"],
5443
- "_mm256_mask_range_ps"=>["AVX512VL", "AVX512DQ"],
5444
- "_mm256_maskz_range_ps"=>["AVX512VL", "AVX512DQ"],
5445
- "_mm256_range_ps"=>["AVX512VL", "AVX512DQ"],
5440
+ "_mm_mask_range_pd"=>[["AVX512VL", "AVX512DQ"]],
5441
+ "_mm_maskz_range_pd"=>[["AVX512VL", "AVX512DQ"]],
5442
+ "_mm_range_pd"=>[["AVX512VL", "AVX512DQ"]],
5443
+ "_mm256_mask_range_ps"=>[["AVX512VL", "AVX512DQ"]],
5444
+ "_mm256_maskz_range_ps"=>[["AVX512VL", "AVX512DQ"]],
5445
+ "_mm256_range_ps"=>[["AVX512VL", "AVX512DQ"]],
5446
5446
  "_mm512_mask_range_ps"=>["AVX512DQ"],
5447
5447
  "_mm512_mask_range_round_ps"=>["AVX512DQ"],
5448
5448
  "_mm512_maskz_range_ps"=>["AVX512DQ"],
5449
5449
  "_mm512_maskz_range_round_ps"=>["AVX512DQ"],
5450
5450
  "_mm512_range_ps"=>["AVX512DQ"],
5451
5451
  "_mm512_range_round_ps"=>["AVX512DQ"],
5452
- "_mm_mask_range_ps"=>["AVX512VL", "AVX512DQ"],
5453
- "_mm_maskz_range_ps"=>["AVX512VL", "AVX512DQ"],
5454
- "_mm_range_ps"=>["AVX512VL", "AVX512DQ"],
5452
+ "_mm_mask_range_ps"=>[["AVX512VL", "AVX512DQ"]],
5453
+ "_mm_maskz_range_ps"=>[["AVX512VL", "AVX512DQ"]],
5454
+ "_mm_range_ps"=>[["AVX512VL", "AVX512DQ"]],
5455
5455
  "_mm_mask_range_round_sd"=>["AVX512DQ"],
5456
5456
  "_mm_mask_range_sd"=>["AVX512DQ"],
5457
5457
  "_mm_maskz_range_round_sd"=>["AVX512DQ"],
@@ -5462,42 +5462,42 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
5462
5462
  "_mm_maskz_range_round_ss"=>["AVX512DQ"],
5463
5463
  "_mm_maskz_range_ss"=>["AVX512DQ"],
5464
5464
  "_mm_range_round_ss"=>["AVX512DQ"],
5465
- "_mm256_mask_rcp14_pd"=>["AVX512VL", "AVX512F"],
5466
- "_mm256_maskz_rcp14_pd"=>["AVX512VL", "AVX512F"],
5467
- "_mm256_rcp14_pd"=>["AVX512VL", "AVX512F"],
5468
- "_mm_mask_rcp14_pd"=>["AVX512VL", "AVX512F"],
5469
- "_mm_maskz_rcp14_pd"=>["AVX512VL", "AVX512F"],
5470
- "_mm_rcp14_pd"=>["AVX512VL", "AVX512F"],
5471
- "_mm256_mask_rcp14_ps"=>["AVX512VL", "AVX512F"],
5472
- "_mm256_maskz_rcp14_ps"=>["AVX512VL", "AVX512F"],
5473
- "_mm256_rcp14_ps"=>["AVX512VL", "AVX512F"],
5474
- "_mm_mask_rcp14_ps"=>["AVX512VL", "AVX512F"],
5475
- "_mm_maskz_rcp14_ps"=>["AVX512VL", "AVX512F"],
5476
- "_mm_rcp14_ps"=>["AVX512VL", "AVX512F"],
5477
- "_mm256_mask_reduce_pd"=>["AVX512VL", "AVX512DQ"],
5478
- "_mm256_maskz_reduce_pd"=>["AVX512VL", "AVX512DQ"],
5479
- "_mm256_reduce_pd"=>["AVX512VL", "AVX512DQ"],
5465
+ "_mm256_mask_rcp14_pd"=>[["AVX512VL", "AVX512F"]],
5466
+ "_mm256_maskz_rcp14_pd"=>[["AVX512VL", "AVX512F"]],
5467
+ "_mm256_rcp14_pd"=>[["AVX512VL", "AVX512F"]],
5468
+ "_mm_mask_rcp14_pd"=>[["AVX512VL", "AVX512F"]],
5469
+ "_mm_maskz_rcp14_pd"=>[["AVX512VL", "AVX512F"]],
5470
+ "_mm_rcp14_pd"=>[["AVX512VL", "AVX512F"]],
5471
+ "_mm256_mask_rcp14_ps"=>[["AVX512VL", "AVX512F"]],
5472
+ "_mm256_maskz_rcp14_ps"=>[["AVX512VL", "AVX512F"]],
5473
+ "_mm256_rcp14_ps"=>[["AVX512VL", "AVX512F"]],
5474
+ "_mm_mask_rcp14_ps"=>[["AVX512VL", "AVX512F"]],
5475
+ "_mm_maskz_rcp14_ps"=>[["AVX512VL", "AVX512F"]],
5476
+ "_mm_rcp14_ps"=>[["AVX512VL", "AVX512F"]],
5477
+ "_mm256_mask_reduce_pd"=>[["AVX512VL", "AVX512DQ"]],
5478
+ "_mm256_maskz_reduce_pd"=>[["AVX512VL", "AVX512DQ"]],
5479
+ "_mm256_reduce_pd"=>[["AVX512VL", "AVX512DQ"]],
5480
5480
  "_mm512_mask_reduce_pd"=>["AVX512DQ"],
5481
5481
  "_mm512_mask_reduce_round_pd"=>["AVX512DQ"],
5482
5482
  "_mm512_maskz_reduce_pd"=>["AVX512DQ"],
5483
5483
  "_mm512_maskz_reduce_round_pd"=>["AVX512DQ"],
5484
5484
  "_mm512_reduce_pd"=>["AVX512DQ"],
5485
5485
  "_mm512_reduce_round_pd"=>["AVX512DQ"],
5486
- "_mm_mask_reduce_pd"=>["AVX512VL", "AVX512DQ"],
5487
- "_mm_maskz_reduce_pd"=>["AVX512VL", "AVX512DQ"],
5488
- "_mm_reduce_pd"=>["AVX512VL", "AVX512DQ"],
5489
- "_mm256_mask_reduce_ps"=>["AVX512VL", "AVX512DQ"],
5490
- "_mm256_maskz_reduce_ps"=>["AVX512VL", "AVX512DQ"],
5491
- "_mm256_reduce_ps"=>["AVX512VL", "AVX512DQ"],
5486
+ "_mm_mask_reduce_pd"=>[["AVX512VL", "AVX512DQ"]],
5487
+ "_mm_maskz_reduce_pd"=>[["AVX512VL", "AVX512DQ"]],
5488
+ "_mm_reduce_pd"=>[["AVX512VL", "AVX512DQ"]],
5489
+ "_mm256_mask_reduce_ps"=>[["AVX512VL", "AVX512DQ"]],
5490
+ "_mm256_maskz_reduce_ps"=>[["AVX512VL", "AVX512DQ"]],
5491
+ "_mm256_reduce_ps"=>[["AVX512VL", "AVX512DQ"]],
5492
5492
  "_mm512_mask_reduce_ps"=>["AVX512DQ"],
5493
5493
  "_mm512_mask_reduce_round_ps"=>["AVX512DQ"],
5494
5494
  "_mm512_maskz_reduce_ps"=>["AVX512DQ"],
5495
5495
  "_mm512_maskz_reduce_round_ps"=>["AVX512DQ"],
5496
5496
  "_mm512_reduce_ps"=>["AVX512DQ"],
5497
5497
  "_mm512_reduce_round_ps"=>["AVX512DQ"],
5498
- "_mm_mask_reduce_ps"=>["AVX512VL", "AVX512DQ"],
5499
- "_mm_maskz_reduce_ps"=>["AVX512VL", "AVX512DQ"],
5500
- "_mm_reduce_ps"=>["AVX512VL", "AVX512DQ"],
5498
+ "_mm_mask_reduce_ps"=>[["AVX512VL", "AVX512DQ"]],
5499
+ "_mm_maskz_reduce_ps"=>[["AVX512VL", "AVX512DQ"]],
5500
+ "_mm_reduce_ps"=>[["AVX512VL", "AVX512DQ"]],
5501
5501
  "_mm_mask_reduce_sd"=>["AVX512DQ"],
5502
5502
  "_mm_mask_reduce_round_sd"=>["AVX512DQ"],
5503
5503
  "_mm_maskz_reduce_sd"=>["AVX512DQ"],
@@ -5510,167 +5510,167 @@ X86CPUID_by_name ={"_m_from_int64"=>["MMX"],
5510
5510
  "_mm_maskz_reduce_round_ss"=>["AVX512DQ"],
5511
5511
  "_mm_reduce_ss"=>["AVX512DQ"],
5512
5512
  "_mm_reduce_round_ss"=>["AVX512DQ"],
5513
- "_mm256_mask_roundscale_pd"=>["AVX512VL", "AVX512F"],
5514
- "_mm256_maskz_roundscale_pd"=>["AVX512VL", "AVX512F"],
5515
- "_mm256_roundscale_pd"=>["AVX512VL", "AVX512F"],
5516
- "_mm_mask_roundscale_pd"=>["AVX512VL", "AVX512F"],
5517
- "_mm_maskz_roundscale_pd"=>["AVX512VL", "AVX512F"],
5518
- "_mm_roundscale_pd"=>["AVX512VL", "AVX512F"],
5519
- "_mm256_mask_roundscale_ps"=>["AVX512VL", "AVX512F"],
5520
- "_mm256_maskz_roundscale_ps"=>["AVX512VL", "AVX512F"],
5521
- "_mm256_roundscale_ps"=>["AVX512VL", "AVX512F"],
5522
- "_mm_mask_roundscale_ps"=>["AVX512VL", "AVX512F"],
5523
- "_mm_maskz_roundscale_ps"=>["AVX512VL", "AVX512F"],
5524
- "_mm_roundscale_ps"=>["AVX512VL", "AVX512F"],
5525
- "_mm256_mask_rsqrt14_pd"=>["AVX512VL", "AVX512F"],
5526
- "_mm256_maskz_rsqrt14_pd"=>["AVX512VL", "AVX512F"],
5527
- "_mm_mask_rsqrt14_pd"=>["AVX512VL", "AVX512F"],
5528
- "_mm_maskz_rsqrt14_pd"=>["AVX512VL", "AVX512F"],
5529
- "_mm256_mask_rsqrt14_ps"=>["AVX512VL", "AVX512F"],
5530
- "_mm256_maskz_rsqrt14_ps"=>["AVX512VL", "AVX512F"],
5531
- "_mm_mask_rsqrt14_ps"=>["AVX512VL", "AVX512F"],
5532
- "_mm_maskz_rsqrt14_ps"=>["AVX512VL", "AVX512F"],
5533
- "_mm256_mask_scalef_pd"=>["AVX512VL", "AVX512F"],
5534
- "_mm256_maskz_scalef_pd"=>["AVX512VL", "AVX512F"],
5535
- "_mm256_scalef_pd"=>["AVX512VL", "AVX512F"],
5536
- "_mm_mask_scalef_pd"=>["AVX512VL", "AVX512F"],
5537
- "_mm_maskz_scalef_pd"=>["AVX512VL", "AVX512F"],
5538
- "_mm_scalef_pd"=>["AVX512VL", "AVX512F"],
5539
- "_mm256_mask_scalef_ps"=>["AVX512VL", "AVX512F"],
5540
- "_mm256_maskz_scalef_ps"=>["AVX512VL", "AVX512F"],
5541
- "_mm256_scalef_ps"=>["AVX512VL", "AVX512F"],
5542
- "_mm_mask_scalef_ps"=>["AVX512VL", "AVX512F"],
5543
- "_mm_maskz_scalef_ps"=>["AVX512VL", "AVX512F"],
5544
- "_mm_scalef_ps"=>["AVX512VL", "AVX512F"],
5545
- "_mm256_i32scatter_pd"=>["AVX512VL", "AVX512F"],
5546
- "_mm256_mask_i32scatter_pd"=>["AVX512VL", "AVX512F"],
5547
- "_mm_i32scatter_pd"=>["AVX512VL", "AVX512F"],
5548
- "_mm_mask_i32scatter_pd"=>["AVX512VL", "AVX512F"],
5549
- "_mm256_i32scatter_ps"=>["AVX512VL", "AVX512F"],
5550
- "_mm256_mask_i32scatter_ps"=>["AVX512VL", "AVX512F"],
5551
- "_mm_i32scatter_ps"=>["AVX512VL", "AVX512F"],
5552
- "_mm_mask_i32scatter_ps"=>["AVX512VL", "AVX512F"],
5553
- "_mm256_i64scatter_pd"=>["AVX512VL", "AVX512F"],
5554
- "_mm256_mask_i64scatter_pd"=>["AVX512VL", "AVX512F"],
5555
- "_mm_i64scatter_pd"=>["AVX512VL", "AVX512F"],
5556
- "_mm_mask_i64scatter_pd"=>["AVX512VL", "AVX512F"],
5557
- "_mm256_i64scatter_ps"=>["AVX512VL", "AVX512F"],
5558
- "_mm256_mask_i64scatter_ps"=>["AVX512VL", "AVX512F"],
5559
- "_mm_i64scatter_ps"=>["AVX512VL", "AVX512F"],
5560
- "_mm_mask_i64scatter_ps"=>["AVX512VL", "AVX512F"],
5561
- "_mm256_mask_shuffle_f32x4"=>["AVX512VL", "AVX512F"],
5562
- "_mm256_maskz_shuffle_f32x4"=>["AVX512VL", "AVX512F"],
5563
- "_mm256_shuffle_f32x4"=>["AVX512VL", "AVX512F"],
5564
- "_mm256_mask_shuffle_f64x2"=>["AVX512VL", "AVX512F"],
5565
- "_mm256_maskz_shuffle_f64x2"=>["AVX512VL", "AVX512F"],
5566
- "_mm256_shuffle_f64x2"=>["AVX512VL", "AVX512F"],
5567
- "_mm256_mask_shuffle_i32x4"=>["AVX512VL", "AVX512F"],
5568
- "_mm256_maskz_shuffle_i32x4"=>["AVX512VL", "AVX512F"],
5569
- "_mm256_shuffle_i32x4"=>["AVX512VL", "AVX512F"],
5570
- "_mm256_mask_shuffle_i64x2"=>["AVX512VL", "AVX512F"],
5571
- "_mm256_maskz_shuffle_i64x2"=>["AVX512VL", "AVX512F"],
5572
- "_mm256_shuffle_i64x2"=>["AVX512VL", "AVX512F"],
5573
- "_mm256_mask_shuffle_pd"=>["AVX512VL", "AVX512F"],
5574
- "_mm256_maskz_shuffle_pd"=>["AVX512VL", "AVX512F"],
5575
- "_mm_mask_shuffle_pd"=>["AVX512VL", "AVX512F"],
5576
- "_mm_maskz_shuffle_pd"=>["AVX512VL", "AVX512F"],
5577
- "_mm256_mask_shuffle_ps"=>["AVX512VL", "AVX512F"],
5578
- "_mm256_maskz_shuffle_ps"=>["AVX512VL", "AVX512F"],
5579
- "_mm_mask_shuffle_ps"=>["AVX512VL", "AVX512F"],
5580
- "_mm_maskz_shuffle_ps"=>["AVX512VL", "AVX512F"],
5581
- "_mm256_mask_sqrt_pd"=>["AVX512VL", "AVX512F"],
5582
- "_mm256_maskz_sqrt_pd"=>["AVX512VL", "AVX512F"],
5583
- "_mm_mask_sqrt_pd"=>["AVX512VL", "AVX512F"],
5584
- "_mm_maskz_sqrt_pd"=>["AVX512VL", "AVX512F"],
5585
- "_mm256_mask_sqrt_ps"=>["AVX512VL", "AVX512F"],
5586
- "_mm256_maskz_sqrt_ps"=>["AVX512VL", "AVX512F"],
5587
- "_mm_mask_sqrt_ps"=>["AVX512VL", "AVX512F"],
5588
- "_mm_maskz_sqrt_ps"=>["AVX512VL", "AVX512F"],
5589
- "_mm256_mask_sub_pd"=>["AVX512VL", "AVX512F"],
5590
- "_mm256_maskz_sub_pd"=>["AVX512VL", "AVX512F"],
5591
- "_mm_mask_sub_pd"=>["AVX512VL", "AVX512F"],
5592
- "_mm_maskz_sub_pd"=>["AVX512VL", "AVX512F"],
5593
- "_mm256_mask_sub_ps"=>["AVX512VL", "AVX512F"],
5594
- "_mm256_maskz_sub_ps"=>["AVX512VL", "AVX512F"],
5595
- "_mm_mask_sub_ps"=>["AVX512VL", "AVX512F"],
5596
- "_mm_maskz_sub_ps"=>["AVX512VL", "AVX512F"],
5597
- "_mm256_mask_unpackhi_pd"=>["AVX512VL", "AVX512F"],
5598
- "_mm256_maskz_unpackhi_pd"=>["AVX512VL", "AVX512F"],
5599
- "_mm_mask_unpackhi_pd"=>["AVX512VL", "AVX512F"],
5600
- "_mm_maskz_unpackhi_pd"=>["AVX512VL", "AVX512F"],
5601
- "_mm256_mask_unpackhi_ps"=>["AVX512VL", "AVX512F"],
5602
- "_mm256_maskz_unpackhi_ps"=>["AVX512VL", "AVX512F"],
5603
- "_mm_mask_unpackhi_ps"=>["AVX512VL", "AVX512F"],
5604
- "_mm_maskz_unpackhi_ps"=>["AVX512VL", "AVX512F"],
5605
- "_mm256_mask_unpacklo_pd"=>["AVX512VL", "AVX512F"],
5606
- "_mm256_maskz_unpacklo_pd"=>["AVX512VL", "AVX512F"],
5607
- "_mm_mask_unpacklo_pd"=>["AVX512VL", "AVX512F"],
5608
- "_mm_maskz_unpacklo_pd"=>["AVX512VL", "AVX512F"],
5609
- "_mm256_mask_unpacklo_ps"=>["AVX512VL", "AVX512F"],
5610
- "_mm256_maskz_unpacklo_ps"=>["AVX512VL", "AVX512F"],
5611
- "_mm_mask_unpacklo_ps"=>["AVX512VL", "AVX512F"],
5612
- "_mm_maskz_unpacklo_ps"=>["AVX512VL", "AVX512F"],
5613
- "_mm256_mask_xor_pd"=>["AVX512VL", "AVX512DQ"],
5614
- "_mm256_maskz_xor_pd"=>["AVX512VL", "AVX512DQ"],
5513
+ "_mm256_mask_roundscale_pd"=>[["AVX512VL", "AVX512F"]],
5514
+ "_mm256_maskz_roundscale_pd"=>[["AVX512VL", "AVX512F"]],
5515
+ "_mm256_roundscale_pd"=>[["AVX512VL", "AVX512F"]],
5516
+ "_mm_mask_roundscale_pd"=>[["AVX512VL", "AVX512F"]],
5517
+ "_mm_maskz_roundscale_pd"=>[["AVX512VL", "AVX512F"]],
5518
+ "_mm_roundscale_pd"=>[["AVX512VL", "AVX512F"]],
5519
+ "_mm256_mask_roundscale_ps"=>[["AVX512VL", "AVX512F"]],
5520
+ "_mm256_maskz_roundscale_ps"=>[["AVX512VL", "AVX512F"]],
5521
+ "_mm256_roundscale_ps"=>[["AVX512VL", "AVX512F"]],
5522
+ "_mm_mask_roundscale_ps"=>[["AVX512VL", "AVX512F"]],
5523
+ "_mm_maskz_roundscale_ps"=>[["AVX512VL", "AVX512F"]],
5524
+ "_mm_roundscale_ps"=>[["AVX512VL", "AVX512F"]],
5525
+ "_mm256_mask_rsqrt14_pd"=>[["AVX512VL", "AVX512F"]],
5526
+ "_mm256_maskz_rsqrt14_pd"=>[["AVX512VL", "AVX512F"]],
5527
+ "_mm_mask_rsqrt14_pd"=>[["AVX512VL", "AVX512F"]],
5528
+ "_mm_maskz_rsqrt14_pd"=>[["AVX512VL", "AVX512F"]],
5529
+ "_mm256_mask_rsqrt14_ps"=>[["AVX512VL", "AVX512F"]],
5530
+ "_mm256_maskz_rsqrt14_ps"=>[["AVX512VL", "AVX512F"]],
5531
+ "_mm_mask_rsqrt14_ps"=>[["AVX512VL", "AVX512F"]],
5532
+ "_mm_maskz_rsqrt14_ps"=>[["AVX512VL", "AVX512F"]],
5533
+ "_mm256_mask_scalef_pd"=>[["AVX512VL", "AVX512F"]],
5534
+ "_mm256_maskz_scalef_pd"=>[["AVX512VL", "AVX512F"]],
5535
+ "_mm256_scalef_pd"=>[["AVX512VL", "AVX512F"]],
5536
+ "_mm_mask_scalef_pd"=>[["AVX512VL", "AVX512F"]],
5537
+ "_mm_maskz_scalef_pd"=>[["AVX512VL", "AVX512F"]],
5538
+ "_mm_scalef_pd"=>[["AVX512VL", "AVX512F"]],
5539
+ "_mm256_mask_scalef_ps"=>[["AVX512VL", "AVX512F"]],
5540
+ "_mm256_maskz_scalef_ps"=>[["AVX512VL", "AVX512F"]],
5541
+ "_mm256_scalef_ps"=>[["AVX512VL", "AVX512F"]],
5542
+ "_mm_mask_scalef_ps"=>[["AVX512VL", "AVX512F"]],
5543
+ "_mm_maskz_scalef_ps"=>[["AVX512VL", "AVX512F"]],
5544
+ "_mm_scalef_ps"=>[["AVX512VL", "AVX512F"]],
5545
+ "_mm256_i32scatter_pd"=>[["AVX512VL", "AVX512F"]],
5546
+ "_mm256_mask_i32scatter_pd"=>[["AVX512VL", "AVX512F"]],
5547
+ "_mm_i32scatter_pd"=>[["AVX512VL", "AVX512F"]],
5548
+ "_mm_mask_i32scatter_pd"=>[["AVX512VL", "AVX512F"]],
5549
+ "_mm256_i32scatter_ps"=>[["AVX512VL", "AVX512F"]],
5550
+ "_mm256_mask_i32scatter_ps"=>[["AVX512VL", "AVX512F"]],
5551
+ "_mm_i32scatter_ps"=>[["AVX512VL", "AVX512F"]],
5552
+ "_mm_mask_i32scatter_ps"=>[["AVX512VL", "AVX512F"]],
5553
+ "_mm256_i64scatter_pd"=>[["AVX512VL", "AVX512F"]],
5554
+ "_mm256_mask_i64scatter_pd"=>[["AVX512VL", "AVX512F"]],
5555
+ "_mm_i64scatter_pd"=>[["AVX512VL", "AVX512F"]],
5556
+ "_mm_mask_i64scatter_pd"=>[["AVX512VL", "AVX512F"]],
5557
+ "_mm256_i64scatter_ps"=>[["AVX512VL", "AVX512F"]],
5558
+ "_mm256_mask_i64scatter_ps"=>[["AVX512VL", "AVX512F"]],
5559
+ "_mm_i64scatter_ps"=>[["AVX512VL", "AVX512F"]],
5560
+ "_mm_mask_i64scatter_ps"=>[["AVX512VL", "AVX512F"]],
5561
+ "_mm256_mask_shuffle_f32x4"=>[["AVX512VL", "AVX512F"]],
5562
+ "_mm256_maskz_shuffle_f32x4"=>[["AVX512VL", "AVX512F"]],
5563
+ "_mm256_shuffle_f32x4"=>[["AVX512VL", "AVX512F"]],
5564
+ "_mm256_mask_shuffle_f64x2"=>[["AVX512VL", "AVX512F"]],
5565
+ "_mm256_maskz_shuffle_f64x2"=>[["AVX512VL", "AVX512F"]],
5566
+ "_mm256_shuffle_f64x2"=>[["AVX512VL", "AVX512F"]],
5567
+ "_mm256_mask_shuffle_i32x4"=>[["AVX512VL", "AVX512F"]],
5568
+ "_mm256_maskz_shuffle_i32x4"=>[["AVX512VL", "AVX512F"]],
5569
+ "_mm256_shuffle_i32x4"=>[["AVX512VL", "AVX512F"]],
5570
+ "_mm256_mask_shuffle_i64x2"=>[["AVX512VL", "AVX512F"]],
5571
+ "_mm256_maskz_shuffle_i64x2"=>[["AVX512VL", "AVX512F"]],
5572
+ "_mm256_shuffle_i64x2"=>[["AVX512VL", "AVX512F"]],
5573
+ "_mm256_mask_shuffle_pd"=>[["AVX512VL", "AVX512F"]],
5574
+ "_mm256_maskz_shuffle_pd"=>[["AVX512VL", "AVX512F"]],
5575
+ "_mm_mask_shuffle_pd"=>[["AVX512VL", "AVX512F"]],
5576
+ "_mm_maskz_shuffle_pd"=>[["AVX512VL", "AVX512F"]],
5577
+ "_mm256_mask_shuffle_ps"=>[["AVX512VL", "AVX512F"]],
5578
+ "_mm256_maskz_shuffle_ps"=>[["AVX512VL", "AVX512F"]],
5579
+ "_mm_mask_shuffle_ps"=>[["AVX512VL", "AVX512F"]],
5580
+ "_mm_maskz_shuffle_ps"=>[["AVX512VL", "AVX512F"]],
5581
+ "_mm256_mask_sqrt_pd"=>[["AVX512VL", "AVX512F"]],
5582
+ "_mm256_maskz_sqrt_pd"=>[["AVX512VL", "AVX512F"]],
5583
+ "_mm_mask_sqrt_pd"=>[["AVX512VL", "AVX512F"]],
5584
+ "_mm_maskz_sqrt_pd"=>[["AVX512VL", "AVX512F"]],
5585
+ "_mm256_mask_sqrt_ps"=>[["AVX512VL", "AVX512F"]],
5586
+ "_mm256_maskz_sqrt_ps"=>[["AVX512VL", "AVX512F"]],
5587
+ "_mm_mask_sqrt_ps"=>[["AVX512VL", "AVX512F"]],
5588
+ "_mm_maskz_sqrt_ps"=>[["AVX512VL", "AVX512F"]],
5589
+ "_mm256_mask_sub_pd"=>[["AVX512VL", "AVX512F"]],
5590
+ "_mm256_maskz_sub_pd"=>[["AVX512VL", "AVX512F"]],
5591
+ "_mm_mask_sub_pd"=>[["AVX512VL", "AVX512F"]],
5592
+ "_mm_maskz_sub_pd"=>[["AVX512VL", "AVX512F"]],
5593
+ "_mm256_mask_sub_ps"=>[["AVX512VL", "AVX512F"]],
5594
+ "_mm256_maskz_sub_ps"=>[["AVX512VL", "AVX512F"]],
5595
+ "_mm_mask_sub_ps"=>[["AVX512VL", "AVX512F"]],
5596
+ "_mm_maskz_sub_ps"=>[["AVX512VL", "AVX512F"]],
5597
+ "_mm256_mask_unpackhi_pd"=>[["AVX512VL", "AVX512F"]],
5598
+ "_mm256_maskz_unpackhi_pd"=>[["AVX512VL", "AVX512F"]],
5599
+ "_mm_mask_unpackhi_pd"=>[["AVX512VL", "AVX512F"]],
5600
+ "_mm_maskz_unpackhi_pd"=>[["AVX512VL", "AVX512F"]],
5601
+ "_mm256_mask_unpackhi_ps"=>[["AVX512VL", "AVX512F"]],
5602
+ "_mm256_maskz_unpackhi_ps"=>[["AVX512VL", "AVX512F"]],
5603
+ "_mm_mask_unpackhi_ps"=>[["AVX512VL", "AVX512F"]],
5604
+ "_mm_maskz_unpackhi_ps"=>[["AVX512VL", "AVX512F"]],
5605
+ "_mm256_mask_unpacklo_pd"=>[["AVX512VL", "AVX512F"]],
5606
+ "_mm256_maskz_unpacklo_pd"=>[["AVX512VL", "AVX512F"]],
5607
+ "_mm_mask_unpacklo_pd"=>[["AVX512VL", "AVX512F"]],
5608
+ "_mm_maskz_unpacklo_pd"=>[["AVX512VL", "AVX512F"]],
5609
+ "_mm256_mask_unpacklo_ps"=>[["AVX512VL", "AVX512F"]],
5610
+ "_mm256_maskz_unpacklo_ps"=>[["AVX512VL", "AVX512F"]],
5611
+ "_mm_mask_unpacklo_ps"=>[["AVX512VL", "AVX512F"]],
5612
+ "_mm_maskz_unpacklo_ps"=>[["AVX512VL", "AVX512F"]],
5613
+ "_mm256_mask_xor_pd"=>[["AVX512VL", "AVX512DQ"]],
5614
+ "_mm256_maskz_xor_pd"=>[["AVX512VL", "AVX512DQ"]],
5615
5615
  "_mm512_mask_xor_pd"=>["AVX512DQ"],
5616
5616
  "_mm512_maskz_xor_pd"=>["AVX512DQ"],
5617
5617
  "_mm512_xor_pd"=>["AVX512DQ"],
5618
- "_mm_mask_xor_pd"=>["AVX512VL", "AVX512DQ"],
5619
- "_mm_maskz_xor_pd"=>["AVX512VL", "AVX512DQ"],
5620
- "_mm256_mask_xor_ps"=>["AVX512VL", "AVX512DQ"],
5621
- "_mm256_maskz_xor_ps"=>["AVX512VL", "AVX512DQ"],
5618
+ "_mm_mask_xor_pd"=>[["AVX512VL", "AVX512DQ"]],
5619
+ "_mm_maskz_xor_pd"=>[["AVX512VL", "AVX512DQ"]],
5620
+ "_mm256_mask_xor_ps"=>[["AVX512VL", "AVX512DQ"]],
5621
+ "_mm256_maskz_xor_ps"=>[["AVX512VL", "AVX512DQ"]],
5622
5622
  "_mm512_mask_xor_ps"=>["AVX512DQ"],
5623
5623
  "_mm512_maskz_xor_ps"=>["AVX512DQ"],
5624
5624
  "_mm512_xor_ps"=>["AVX512DQ"],
5625
- "_mm_mask_xor_ps"=>["AVX512VL", "AVX512DQ"],
5626
- "_mm_maskz_xor_ps"=>["AVX512VL", "AVX512DQ"],
5625
+ "_mm_mask_xor_ps"=>[["AVX512VL", "AVX512DQ"]],
5626
+ "_mm_maskz_xor_ps"=>[["AVX512VL", "AVX512DQ"]],
5627
5627
  "_mm512_madd52lo_epu64"=>["AVX512IFMA52"],
5628
5628
  "_mm512_mask_madd52lo_epu64"=>["AVX512IFMA52"],
5629
5629
  "_mm512_maskz_madd52lo_epu64"=>["AVX512IFMA52"],
5630
- "_mm256_madd52lo_epu64"=>["AVX512IFMA52", "AVX512VL"],
5631
- "_mm256_mask_madd52lo_epu64"=>["AVX512IFMA52", "AVX512VL"],
5632
- "_mm256_maskz_madd52lo_epu64"=>["AVX512IFMA52", "AVX512VL"],
5633
- "_mm_madd52lo_epu64"=>["AVX512IFMA52", "AVX512VL"],
5634
- "_mm_mask_madd52lo_epu64"=>["AVX512IFMA52", "AVX512VL"],
5635
- "_mm_maskz_madd52lo_epu64"=>["AVX512IFMA52", "AVX512VL"],
5630
+ "_mm256_madd52lo_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5631
+ "_mm256_mask_madd52lo_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5632
+ "_mm256_maskz_madd52lo_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5633
+ "_mm_madd52lo_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5634
+ "_mm_mask_madd52lo_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5635
+ "_mm_maskz_madd52lo_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5636
5636
  "_mm512_madd52hi_epu64"=>["AVX512IFMA52"],
5637
5637
  "_mm512_mask_madd52hi_epu64"=>["AVX512IFMA52"],
5638
5638
  "_mm512_maskz_madd52hi_epu64"=>["AVX512IFMA52"],
5639
- "_mm256_madd52hi_epu64"=>["AVX512IFMA52", "AVX512VL"],
5640
- "_mm256_mask_madd52hi_epu64"=>["AVX512IFMA52", "AVX512VL"],
5641
- "_mm256_maskz_madd52hi_epu64"=>["AVX512IFMA52", "AVX512VL"],
5642
- "_mm_madd52hi_epu64"=>["AVX512IFMA52", "AVX512VL"],
5643
- "_mm_mask_madd52hi_epu64"=>["AVX512IFMA52", "AVX512VL"],
5644
- "_mm_maskz_madd52hi_epu64"=>["AVX512IFMA52", "AVX512VL"],
5639
+ "_mm256_madd52hi_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5640
+ "_mm256_mask_madd52hi_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5641
+ "_mm256_maskz_madd52hi_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5642
+ "_mm_madd52hi_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5643
+ "_mm_mask_madd52hi_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5644
+ "_mm_maskz_madd52hi_epu64"=>[["AVX512IFMA52", "AVX512VL"]],
5645
5645
  "_mm512_multishift_epi64_epi8"=>["AVX512VBMI"],
5646
5646
  "_mm512_mask_multishift_epi64_epi8"=>["AVX512VBMI"],
5647
5647
  "_mm512_maskz_multishift_epi64_epi8"=>["AVX512VBMI"],
5648
- "_mm256_multishift_epi64_epi8"=>["AVX512VBMI", "AVX512VL"],
5649
- "_mm256_mask_multishift_epi64_epi8"=>["AVX512VBMI", "AVX512VL"],
5650
- "_mm256_maskz_multishift_epi64_epi8"=>["AVX512VBMI", "AVX512VL"],
5651
- "_mm_multishift_epi64_epi8"=>["AVX512VBMI", "AVX512VL"],
5652
- "_mm_mask_multishift_epi64_epi8"=>["AVX512VBMI", "AVX512VL"],
5653
- "_mm_maskz_multishift_epi64_epi8"=>["AVX512VBMI", "AVX512VL"],
5648
+ "_mm256_multishift_epi64_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5649
+ "_mm256_mask_multishift_epi64_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5650
+ "_mm256_maskz_multishift_epi64_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5651
+ "_mm_multishift_epi64_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5652
+ "_mm_mask_multishift_epi64_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5653
+ "_mm_maskz_multishift_epi64_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5654
5654
  "_mm512_permutexvar_epi8"=>["AVX512VBMI"],
5655
5655
  "_mm512_mask_permutexvar_epi8"=>["AVX512VBMI"],
5656
5656
  "_mm512_maskz_permutexvar_epi8"=>["AVX512VBMI"],
5657
- "_mm256_permutexvar_epi8"=>["AVX512VBMI", "AVX512VL"],
5658
- "_mm256_mask_permutexvar_epi8"=>["AVX512VBMI", "AVX512VL"],
5659
- "_mm256_maskz_permutexvar_epi8"=>["AVX512VBMI", "AVX512VL"],
5660
- "_mm_permutexvar_epi8"=>["AVX512VBMI", "AVX512VL"],
5661
- "_mm_mask_permutexvar_epi8"=>["AVX512VBMI", "AVX512VL"],
5662
- "_mm_maskz_permutexvar_epi8"=>["AVX512VBMI", "AVX512VL"],
5657
+ "_mm256_permutexvar_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5658
+ "_mm256_mask_permutexvar_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5659
+ "_mm256_maskz_permutexvar_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5660
+ "_mm_permutexvar_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5661
+ "_mm_mask_permutexvar_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5662
+ "_mm_maskz_permutexvar_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5663
5663
  "_mm512_permutex2var_epi8"=>["AVX512VBMI"],
5664
5664
  "_mm512_mask_permutex2var_epi8"=>["AVX512VBMI"],
5665
5665
  "_mm512_mask2_permutex2var_epi8"=>["AVX512VBMI"],
5666
5666
  "_mm512_maskz_permutex2var_epi8"=>["AVX512VBMI"],
5667
- "_mm256_permutex2var_epi8"=>["AVX512VBMI", "AVX512VL"],
5668
- "_mm256_mask_permutex2var_epi8"=>["AVX512VBMI", "AVX512VL"],
5669
- "_mm256_mask2_permutex2var_epi8"=>["AVX512VBMI", "AVX512VL"],
5670
- "_mm256_maskz_permutex2var_epi8"=>["AVX512VBMI", "AVX512VL"],
5671
- "_mm_permutex2var_epi8"=>["AVX512VBMI", "AVX512VL"],
5672
- "_mm_mask_permutex2var_epi8"=>["AVX512VBMI", "AVX512VL"],
5673
- "_mm_mask2_permutex2var_epi8"=>["AVX512VBMI", "AVX512VL"],
5674
- "_mm_maskz_permutex2var_epi8"=>["AVX512VBMI", "AVX512VL"]}
5667
+ "_mm256_permutex2var_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5668
+ "_mm256_mask_permutex2var_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5669
+ "_mm256_mask2_permutex2var_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5670
+ "_mm256_maskz_permutex2var_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5671
+ "_mm_permutex2var_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5672
+ "_mm_mask_permutex2var_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5673
+ "_mm_mask2_permutex2var_epi8"=>[["AVX512VBMI", "AVX512VL"]],
5674
+ "_mm_maskz_permutex2var_epi8"=>[["AVX512VBMI", "AVX512VL"]]}
5675
5675
  private_constant :X86CPUID_by_name
5676
5676
  end