Raymond Knopp
a7bc282302
Testing on Neoverse N1 (Ampere 3GHz). Tested only on a subset of phy-simulators. - Changes: - use SIMDE consistently - adaptations of LDPC decoder generator for ARMv8 performance - SIMDe modifications of Intel CRC to allow for aarch64 build. optimizations for 128-bit to improve performance of LDPC encode/decode on aarch64 (Neoverse 1) - added BG2 files for 128-bit ldpc encoder (aarch64) - testing on Xeon - testing on x86 - minor changes to build/run on x86 - change in crc.h after returning to aarch64 - removed some warning in ldpc decoder generator for x86_64 - char - Delete irrelevant constants - Correctly declare variables - Define stdbool for all architectures - Remove definition of _MM_SHUFFLE and use SIMDE_MM_SHUFFLE - Remove commented code - Fix CMakeLists.txt - Include SIMDE avx2 functions in tools_defs.h
154 lines
60 KiB
C
154 lines
60 KiB
C
#ifndef __AVX2__
|
|
#include "PHY/sse_intrin.h"
|
|
// generated code for Zc=256, byte encoding
|
|
static inline void ldpc256_byte(uint8_t *c,uint8_t *d) {
|
|
simde__m128i *csimd=(simde__m128i *)c,*dsimd=(simde__m128i *)d;
|
|
|
|
simde__m128i *c2,*d2;
|
|
|
|
int i2;
|
|
for (i2=0; i2<16; i2++) {
|
|
c2=&csimd[i2];
|
|
d2=&dsimd[i2];
|
|
|
|
//row: 0
|
|
d2[0]=simde_mm_xor_si128(c2[7055],simde_mm_xor_si128(c2[1408],simde_mm_xor_si128(c2[7046],simde_mm_xor_si128(c2[6343],simde_mm_xor_si128(c2[3556],simde_mm_xor_si128(c2[10598],simde_mm_xor_si128(c2[6373],simde_mm_xor_si128(c2[1486],simde_mm_xor_si128(c2[10638],simde_mm_xor_si128(c2[6411],simde_mm_xor_si128(c2[10665],simde_mm_xor_si128(c2[3623],simde_mm_xor_si128(c2[2917],simde_mm_xor_si128(c2[8583],simde_mm_xor_si128(c2[10691],simde_mm_xor_si128(c2[2945],simde_mm_xor_si128(c2[2982],simde_mm_xor_si128(c2[5092],simde_mm_xor_si128(c2[3687],simde_mm_xor_si128(c2[7232],simde_mm_xor_si128(c2[9349],simde_mm_xor_si128(c2[4425],simde_mm_xor_si128(c2[10093],simde_mm_xor_si128(c2[3758],simde_mm_xor_si128(c2[2344],simde_mm_xor_si128(c2[5894],simde_mm_xor_si128(c2[971],simde_mm_xor_si128(c2[2383],simde_mm_xor_si128(c2[8035],simde_mm_xor_si128(c2[9450],simde_mm_xor_si128(c2[10853],simde_mm_xor_si128(c2[3854],simde_mm_xor_si128(c2[5250],simde_mm_xor_si128(c2[5960],simde_mm_xor_si128(c2[10214],simde_mm_xor_si128(c2[8813],simde_mm_xor_si128(c2[4581],simde_mm_xor_si128(c2[10955],simde_mm_xor_si128(c2[4614],simde_mm_xor_si128(c2[4623],simde_mm_xor_si128(c2[6752],simde_mm_xor_si128(c2[10280],simde_mm_xor_si128(c2[8173],simde_mm_xor_si128(c2[9606],simde_mm_xor_si128(c2[1166],simde_mm_xor_si128(c2[2573],simde_mm_xor_si128(c2[2604],simde_mm_xor_si128(c2[3304],simde_mm_xor_si128(c2[1198],simde_mm_xor_si128(c2[5441],simde_mm_xor_si128(c2[10376],simde_mm_xor_si128(c2[527],simde_mm_xor_si128(c2[8297],simde_mm_xor_si128(c2[4079],simde_mm_xor_si128(c2[8996],simde_mm_xor_si128(c2[10443],simde_mm_xor_si128(c2[9740],simde_mm_xor_si128(c2[3407],simde_mm_xor_si128(c2[2722],simde_mm_xor_si128(c2[11183],simde_mm_xor_si128(c2[8367],simde_mm_xor_si128(c2[11214],simde_mm_xor_si128(c2[4167],simde_mm_xor_si128(c2[649],simde_mm_xor_si128(c2[11233],simde_mm_xor_si128(c2[9121],c2[9120]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 1
|
|
d2[16]=simde_mm_xor_si128(c2[7055],simde_mm_xor_si128(c2[7759],simde_mm_xor_si128(c2[2112],simde_mm_xor_si128(c2[7750],simde_mm_xor_si128(c2[7047],simde_mm_xor_si128(c2[3556],simde_mm_xor_si128(c2[4260],simde_mm_xor_si128(c2[39],simde_mm_xor_si128(c2[7077],simde_mm_xor_si128(c2[1486],simde_mm_xor_si128(c2[2190],simde_mm_xor_si128(c2[79],simde_mm_xor_si128(c2[7115],simde_mm_xor_si128(c2[10665],simde_mm_xor_si128(c2[106],simde_mm_xor_si128(c2[4327],simde_mm_xor_si128(c2[3621],simde_mm_xor_si128(c2[9287],simde_mm_xor_si128(c2[132],simde_mm_xor_si128(c2[3649],simde_mm_xor_si128(c2[2982],simde_mm_xor_si128(c2[3686],simde_mm_xor_si128(c2[5796],simde_mm_xor_si128(c2[4391],simde_mm_xor_si128(c2[7232],simde_mm_xor_si128(c2[7936],simde_mm_xor_si128(c2[10053],simde_mm_xor_si128(c2[5129],simde_mm_xor_si128(c2[10797],simde_mm_xor_si128(c2[4462],simde_mm_xor_si128(c2[3048],simde_mm_xor_si128(c2[6598],simde_mm_xor_si128(c2[1675],simde_mm_xor_si128(c2[3087],simde_mm_xor_si128(c2[8035],simde_mm_xor_si128(c2[8739],simde_mm_xor_si128(c2[10154],simde_mm_xor_si128(c2[294],simde_mm_xor_si128(c2[3854],simde_mm_xor_si128(c2[4558],simde_mm_xor_si128(c2[5954],simde_mm_xor_si128(c2[6664],simde_mm_xor_si128(c2[10214],simde_mm_xor_si128(c2[10918],simde_mm_xor_si128(c2[9517],simde_mm_xor_si128(c2[5285],simde_mm_xor_si128(c2[10955],simde_mm_xor_si128(c2[396],simde_mm_xor_si128(c2[5318],simde_mm_xor_si128(c2[5327],simde_mm_xor_si128(c2[6752],simde_mm_xor_si128(c2[7456],simde_mm_xor_si128(c2[10984],simde_mm_xor_si128(c2[8877],simde_mm_xor_si128(c2[10310],simde_mm_xor_si128(c2[1870],simde_mm_xor_si128(c2[3277],simde_mm_xor_si128(c2[2604],simde_mm_xor_si128(c2[3308],simde_mm_xor_si128(c2[4008],simde_mm_xor_si128(c2[1902],simde_mm_xor_si128(c2[5441],simde_mm_xor_si128(c2[6145],simde_mm_xor_si128(c2[11080],simde_mm_xor_si128(c2[1231],simde_mm_xor_si128(c2[9001],simde_mm_xor_si128(c2[4783],simde_mm_xor_si128(c2[9700],simde_mm_xor_si128(c2[10443],simde_mm_xor_si128(c2[11147],simde_mm_xor_si128(c2[10444],simde_mm_xor_si128(c2[4111],simde_mm_xor_si128(c2[2722],simde_mm_xor_si128(c2[3426],simde_mm_xor_si128(c2[608],simde_mm_xor_si128(c2[9071],simde_mm_xor_si128(c2[11214],simde_mm_xor_si128(c2[655],simde_mm_xor_si128(c2[4871],simde_mm_xor_si128(c2[1353],simde_mm_xor_si128(c2[11233],simde_mm_xor_si128(c2[674],simde_mm_xor_si128(c2[9825],c2[9824])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 2
|
|
d2[32]=simde_mm_xor_si128(c2[7759],simde_mm_xor_si128(c2[2112],simde_mm_xor_si128(c2[7046],simde_mm_xor_si128(c2[7750],simde_mm_xor_si128(c2[6343],simde_mm_xor_si128(c2[7047],simde_mm_xor_si128(c2[4260],simde_mm_xor_si128(c2[10598],simde_mm_xor_si128(c2[39],simde_mm_xor_si128(c2[6373],simde_mm_xor_si128(c2[7077],simde_mm_xor_si128(c2[2190],simde_mm_xor_si128(c2[79],simde_mm_xor_si128(c2[6411],simde_mm_xor_si128(c2[7115],simde_mm_xor_si128(c2[106],simde_mm_xor_si128(c2[4327],simde_mm_xor_si128(c2[2917],simde_mm_xor_si128(c2[3621],simde_mm_xor_si128(c2[9287],simde_mm_xor_si128(c2[10691],simde_mm_xor_si128(c2[132],simde_mm_xor_si128(c2[2945],simde_mm_xor_si128(c2[3649],simde_mm_xor_si128(c2[3686],simde_mm_xor_si128(c2[5796],simde_mm_xor_si128(c2[3687],simde_mm_xor_si128(c2[4391],simde_mm_xor_si128(c2[7936],simde_mm_xor_si128(c2[9349],simde_mm_xor_si128(c2[10053],simde_mm_xor_si128(c2[4425],simde_mm_xor_si128(c2[5129],simde_mm_xor_si128(c2[10797],simde_mm_xor_si128(c2[3758],simde_mm_xor_si128(c2[4462],simde_mm_xor_si128(c2[2344],simde_mm_xor_si128(c2[3048],simde_mm_xor_si128(c2[6598],simde_mm_xor_si128(c2[971],simde_mm_xor_si128(c2[1675],simde_mm_xor_si128(c2[2383],simde_mm_xor_si128(c2[3087],simde_mm_xor_si128(c2[8739],simde_mm_xor_si128(c2[10154],simde_mm_xor_si128(c2[10853],simde_mm_xor_si128(c2[294],simde_mm_xor_si128(c2[4558],simde_mm_xor_si128(c2[5250],simde_mm_xor_si128(c2[5954],simde_mm_xor_si128(c2[5960],simde_mm_xor_si128(c2[6664],simde_mm_xor_si128(c2[10918],simde_mm_xor_si128(c2[9517],simde_mm_xor_si128(c2[4581],simde_mm_xor_si128(c2[5285],simde_mm_xor_si128(c2[396],simde_mm_xor_si128(c2[5318],simde_mm_xor_si128(c2[4623],simde_mm_xor_si128(c2[5327],simde_mm_xor_si128(c2[7456],simde_mm_xor_si128(c2[10280],simde_mm_xor_si128(c2[10984],simde_mm_xor_si128(c2[8173],simde_mm_xor_si128(c2[8877],simde_mm_xor_si128(c2[10310],simde_mm_xor_si128(c2[1166],simde_mm_xor_si128(c2[1870],simde_mm_xor_si128(c2[2573],simde_mm_xor_si128(c2[3277],simde_mm_xor_si128(c2[3308],simde_mm_xor_si128(c2[4008],simde_mm_xor_si128(c2[1198],simde_mm_xor_si128(c2[1902],simde_mm_xor_si128(c2[6145],simde_mm_xor_si128(c2[11080],simde_mm_xor_si128(c2[527],simde_mm_xor_si128(c2[1231],simde_mm_xor_si128(c2[9001],simde_mm_xor_si128(c2[4079],simde_mm_xor_si128(c2[4783],simde_mm_xor_si128(c2[8996],simde_mm_xor_si128(c2[9700],simde_mm_xor_si128(c2[11147],simde_mm_xor_si128(c2[9740],simde_mm_xor_si128(c2[10444],simde_mm_xor_si128(c2[3407],simde_mm_xor_si128(c2[4111],simde_mm_xor_si128(c2[3426],simde_mm_xor_si128(c2[608],simde_mm_xor_si128(c2[8367],simde_mm_xor_si128(c2[9071],simde_mm_xor_si128(c2[655],simde_mm_xor_si128(c2[4167],simde_mm_xor_si128(c2[4871],simde_mm_xor_si128(c2[649],simde_mm_xor_si128(c2[1353],simde_mm_xor_si128(c2[674],simde_mm_xor_si128(c2[9825],simde_mm_xor_si128(c2[9120],c2[9824]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 3
|
|
d2[48]=simde_mm_xor_si128(c2[7759],simde_mm_xor_si128(c2[2112],simde_mm_xor_si128(c2[7750],simde_mm_xor_si128(c2[6343],simde_mm_xor_si128(c2[7047],simde_mm_xor_si128(c2[4260],simde_mm_xor_si128(c2[39],simde_mm_xor_si128(c2[6373],simde_mm_xor_si128(c2[7077],simde_mm_xor_si128(c2[2190],simde_mm_xor_si128(c2[79],simde_mm_xor_si128(c2[7115],simde_mm_xor_si128(c2[106],simde_mm_xor_si128(c2[4327],simde_mm_xor_si128(c2[2917],simde_mm_xor_si128(c2[3621],simde_mm_xor_si128(c2[9287],simde_mm_xor_si128(c2[132],simde_mm_xor_si128(c2[2945],simde_mm_xor_si128(c2[3649],simde_mm_xor_si128(c2[3686],simde_mm_xor_si128(c2[5796],simde_mm_xor_si128(c2[4391],simde_mm_xor_si128(c2[7936],simde_mm_xor_si128(c2[10053],simde_mm_xor_si128(c2[4425],simde_mm_xor_si128(c2[5129],simde_mm_xor_si128(c2[10797],simde_mm_xor_si128(c2[4462],simde_mm_xor_si128(c2[2344],simde_mm_xor_si128(c2[3048],simde_mm_xor_si128(c2[6598],simde_mm_xor_si128(c2[1675],simde_mm_xor_si128(c2[2383],simde_mm_xor_si128(c2[3087],simde_mm_xor_si128(c2[8739],simde_mm_xor_si128(c2[10154],simde_mm_xor_si128(c2[294],simde_mm_xor_si128(c2[4558],simde_mm_xor_si128(c2[5954],simde_mm_xor_si128(c2[5960],simde_mm_xor_si128(c2[6664],simde_mm_xor_si128(c2[10918],simde_mm_xor_si128(c2[9517],simde_mm_xor_si128(c2[4581],simde_mm_xor_si128(c2[5285],simde_mm_xor_si128(c2[396],simde_mm_xor_si128(c2[5318],simde_mm_xor_si128(c2[4623],simde_mm_xor_si128(c2[5327],simde_mm_xor_si128(c2[7456],simde_mm_xor_si128(c2[10984],simde_mm_xor_si128(c2[8173],simde_mm_xor_si128(c2[8877],simde_mm_xor_si128(c2[10310],simde_mm_xor_si128(c2[1870],simde_mm_xor_si128(c2[2573],simde_mm_xor_si128(c2[3277],simde_mm_xor_si128(c2[3308],simde_mm_xor_si128(c2[4008],simde_mm_xor_si128(c2[1902],simde_mm_xor_si128(c2[6145],simde_mm_xor_si128(c2[11080],simde_mm_xor_si128(c2[527],simde_mm_xor_si128(c2[1231],simde_mm_xor_si128(c2[9001],simde_mm_xor_si128(c2[4783],simde_mm_xor_si128(c2[8996],simde_mm_xor_si128(c2[9700],simde_mm_xor_si128(c2[11147],simde_mm_xor_si128(c2[10444],simde_mm_xor_si128(c2[3407],simde_mm_xor_si128(c2[4111],simde_mm_xor_si128(c2[3426],simde_mm_xor_si128(c2[608],simde_mm_xor_si128(c2[9071],simde_mm_xor_si128(c2[655],simde_mm_xor_si128(c2[4871],simde_mm_xor_si128(c2[649],simde_mm_xor_si128(c2[1353],simde_mm_xor_si128(c2[674],simde_mm_xor_si128(c2[9825],simde_mm_xor_si128(c2[9120],c2[9824])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 4
|
|
d2[64]=simde_mm_xor_si128(c2[9161],c2[4262]);
|
|
|
|
//row: 5
|
|
d2[80]=simde_mm_xor_si128(c2[9158],simde_mm_xor_si128(c2[3527],simde_mm_xor_si128(c2[9165],simde_mm_xor_si128(c2[8462],simde_mm_xor_si128(c2[9164],simde_mm_xor_si128(c2[5675],simde_mm_xor_si128(c2[1454],simde_mm_xor_si128(c2[8492],simde_mm_xor_si128(c2[8494],simde_mm_xor_si128(c2[3589],simde_mm_xor_si128(c2[1478],simde_mm_xor_si128(c2[8514],simde_mm_xor_si128(c2[1505],simde_mm_xor_si128(c2[5742],simde_mm_xor_si128(c2[5036],simde_mm_xor_si128(c2[1516],simde_mm_xor_si128(c2[10702],simde_mm_xor_si128(c2[1547],simde_mm_xor_si128(c2[5064],simde_mm_xor_si128(c2[5101],simde_mm_xor_si128(c2[7211],simde_mm_xor_si128(c2[5806],simde_mm_xor_si128(c2[9351],simde_mm_xor_si128(c2[205],simde_mm_xor_si128(c2[6528],simde_mm_xor_si128(c2[933],simde_mm_xor_si128(c2[5861],simde_mm_xor_si128(c2[4463],simde_mm_xor_si128(c2[8013],simde_mm_xor_si128(c2[3074],simde_mm_xor_si128(c2[4486],simde_mm_xor_si128(c2[10154],simde_mm_xor_si128(c2[290],simde_mm_xor_si128(c2[1709],simde_mm_xor_si128(c2[5957],simde_mm_xor_si128(c2[7369],simde_mm_xor_si128(c2[8079],simde_mm_xor_si128(c2[1070],simde_mm_xor_si128(c2[10916],simde_mm_xor_si128(c2[6700],simde_mm_xor_si128(c2[1795],simde_mm_xor_si128(c2[6733],simde_mm_xor_si128(c2[6726],simde_mm_xor_si128(c2[5326],simde_mm_xor_si128(c2[8871],simde_mm_xor_si128(c2[1120],simde_mm_xor_si128(c2[10276],simde_mm_xor_si128(c2[462],simde_mm_xor_si128(c2[3269],simde_mm_xor_si128(c2[4676],simde_mm_xor_si128(c2[4707],simde_mm_xor_si128(c2[5423],simde_mm_xor_si128(c2[3301],simde_mm_xor_si128(c2[7560],simde_mm_xor_si128(c2[1216],simde_mm_xor_si128(c2[2630],simde_mm_xor_si128(c2[8961],simde_mm_xor_si128(c2[10400],simde_mm_xor_si128(c2[6182],simde_mm_xor_si128(c2[11115],simde_mm_xor_si128(c2[1283],simde_mm_xor_si128(c2[580],simde_mm_xor_si128(c2[5510],simde_mm_xor_si128(c2[4841],simde_mm_xor_si128(c2[2023],simde_mm_xor_si128(c2[10470],simde_mm_xor_si128(c2[2054],simde_mm_xor_si128(c2[6286],simde_mm_xor_si128(c2[2752],simde_mm_xor_si128(c2[2089],simde_mm_xor_si128(c2[11240],simde_mm_xor_si128(c2[11239],c2[8423]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 6
|
|
d2[96]=simde_mm_xor_si128(c2[4939],simde_mm_xor_si128(c2[4417],simde_mm_xor_si128(c2[8769],simde_mm_xor_si128(c2[2468],simde_mm_xor_si128(c2[3247],simde_mm_xor_si128(c2[8288],simde_mm_xor_si128(c2[9737],c2[2765])))))));
|
|
|
|
//row: 7
|
|
d2[112]=simde_mm_xor_si128(c2[8461],simde_mm_xor_si128(c2[8482],simde_mm_xor_si128(c2[10697],simde_mm_xor_si128(c2[10785],simde_mm_xor_si128(c2[5194],c2[6086])))));
|
|
|
|
//row: 8
|
|
d2[128]=simde_mm_xor_si128(c2[7758],simde_mm_xor_si128(c2[3525],simde_mm_xor_si128(c2[2127],simde_mm_xor_si128(c2[9157],simde_mm_xor_si128(c2[7749],simde_mm_xor_si128(c2[2828],simde_mm_xor_si128(c2[3532],simde_mm_xor_si128(c2[7046],simde_mm_xor_si128(c2[2125],simde_mm_xor_si128(c2[2829],simde_mm_xor_si128(c2[7],simde_mm_xor_si128(c2[4259],simde_mm_xor_si128(c2[42],simde_mm_xor_si128(c2[38],simde_mm_xor_si128(c2[6380],simde_mm_xor_si128(c2[7084],simde_mm_xor_si128(c2[7076],simde_mm_xor_si128(c2[2155],simde_mm_xor_si128(c2[2859],simde_mm_xor_si128(c2[2848],simde_mm_xor_si128(c2[2189],simde_mm_xor_si128(c2[9219],simde_mm_xor_si128(c2[78],simde_mm_xor_si128(c2[7108],simde_mm_xor_si128(c2[7114],simde_mm_xor_si128(c2[2177],simde_mm_xor_si128(c2[2881],simde_mm_xor_si128(c2[105],simde_mm_xor_si128(c2[7151],simde_mm_xor_si128(c2[4326],simde_mm_xor_si128(c2[109],simde_mm_xor_si128(c2[3620],simde_mm_xor_si128(c2[9962],simde_mm_xor_si128(c2[10666],simde_mm_xor_si128(c2[5024],simde_mm_xor_si128(c2[9286],simde_mm_xor_si128(c2[5069],simde_mm_xor_si128(c2[131],simde_mm_xor_si128(c2[6473],simde_mm_xor_si128(c2[7177],simde_mm_xor_si128(c2[3648],simde_mm_xor_si128(c2[9990],simde_mm_xor_si128(c2[10694],simde_mm_xor_si128(c2[3685],simde_mm_xor_si128(c2[10731],simde_mm_xor_si128(c2[5795],simde_mm_xor_si128(c2[1578],simde_mm_xor_si128(c2[4390],simde_mm_xor_si128(c2[10732],simde_mm_xor_si128(c2[173],simde_mm_xor_si128(c2[7951],simde_mm_xor_si128(c2[3718],simde_mm_xor_si128(c2[10052],simde_mm_xor_si128(c2[5131],simde_mm_xor_si128(c2[5835],simde_mm_xor_si128(c2[5128],simde_mm_xor_si128(c2[207],simde_mm_xor_si128(c2[911],simde_mm_xor_si128(c2[10796],simde_mm_xor_si128(c2[6563],simde_mm_xor_si128(c2[4461],simde_mm_xor_si128(c2[10787],simde_mm_xor_si128(c2[228],simde_mm_xor_si128(c2[3047],simde_mm_xor_si128(c2[9389],simde_mm_xor_si128(c2[10093],simde_mm_xor_si128(c2[6597],simde_mm_xor_si128(c2[2380],simde_mm_xor_si128(c2[1674],simde_mm_xor_si128(c2[8000],simde_mm_xor_si128(c2[8704],simde_mm_xor_si128(c2[3086],simde_mm_xor_si128(c2[9412],simde_mm_xor_si128(c2[10116],simde_mm_xor_si128(c2[8738],simde_mm_xor_si128(c2[4521],simde_mm_xor_si128(c2[10153],simde_mm_xor_si128(c2[5920],simde_mm_xor_si128(c2[293],simde_mm_xor_si128(c2[6635],simde_mm_xor_si128(c2[7339],simde_mm_xor_si128(c2[4557],simde_mm_xor_si128(c2[324],simde_mm_xor_si128(c2[5953],simde_mm_xor_si128(c2[1032],simde_mm_xor_si128(c2[1736],simde_mm_xor_si128(c2[6663],simde_mm_xor_si128(c2[1742],simde_mm_xor_si128(c2[2446],simde_mm_xor_si128(c2[10917],simde_mm_xor_si128(c2[6700],simde_mm_xor_si128(c2[9516],simde_mm_xor_si128(c2[5283],simde_mm_xor_si128(c2[5284],simde_mm_xor_si128(c2[363],simde_mm_xor_si128(c2[1067],simde_mm_xor_si128(c2[395],simde_mm_xor_si128(c2[7425],simde_mm_xor_si128(c2[5317],simde_mm_xor_si128(c2[1100],simde_mm_xor_si128(c2[5326],simde_mm_xor_si128(c2[389],simde_mm_xor_si128(c2[1093],simde_mm_xor_si128(c2[2509],simde_mm_xor_si128(c2[7471],simde_mm_xor_si128(c2[3238],simde_mm_xor_si128(c2[10983],simde_mm_xor_si128(c2[6062],simde_mm_xor_si128(c2[6766],simde_mm_xor_si128(c2[8876],simde_mm_xor_si128(c2[3939],simde_mm_xor_si128(c2[4643],simde_mm_xor_si128(c2[10309],simde_mm_xor_si128(c2[6092],simde_mm_xor_si128(c2[1869],simde_mm_xor_si128(c2[8195],simde_mm_xor_si128(c2[8899],simde_mm_xor_si128(c2[3276],simde_mm_xor_si128(c2[9602],simde_mm_xor_si128(c2[10306],simde_mm_xor_si128(c2[3307],simde_mm_xor_si128(c2[10337],simde_mm_xor_si128(c2[4007],simde_mm_xor_si128(c2[11053],simde_mm_xor_si128(c2[1901],simde_mm_xor_si128(c2[8227],simde_mm_xor_si128(c2[8931],simde_mm_xor_si128(c2[6144],simde_mm_xor_si128(c2[1927],simde_mm_xor_si128(c2[11079],simde_mm_xor_si128(c2[6862],simde_mm_xor_si128(c2[1230],simde_mm_xor_si128(c2[7556],simde_mm_xor_si128(c2[8260],simde_mm_xor_si128(c2[4742],simde_mm_xor_si128(c2[9000],simde_mm_xor_si128(c2[4783],simde_mm_xor_si128(c2[4782],simde_mm_xor_si128(c2[11108],simde_mm_xor_si128(c2[549],simde_mm_xor_si128(c2[9699],simde_mm_xor_si128(c2[4778],simde_mm_xor_si128(c2[5482],simde_mm_xor_si128(c2[11146],simde_mm_xor_si128(c2[6913],simde_mm_xor_si128(c2[10443],simde_mm_xor_si128(c2[5506],simde_mm_xor_si128(c2[6210],simde_mm_xor_si128(c2[4110],simde_mm_xor_si128(c2[10436],simde_mm_xor_si128(c2[11140],simde_mm_xor_si128(c2[3425],simde_mm_xor_si128(c2[10471],simde_mm_xor_si128(c2[623],simde_mm_xor_si128(c2[7653],simde_mm_xor_si128(c2[9070],simde_mm_xor_si128(c2[4133],simde_mm_xor_si128(c2[4837],simde_mm_xor_si128(c2[3434],simde_mm_xor_si128(c2[654],simde_mm_xor_si128(c2[7684],simde_mm_xor_si128(c2[4870],simde_mm_xor_si128(c2[11212],simde_mm_xor_si128(c2[653],simde_mm_xor_si128(c2[1352],simde_mm_xor_si128(c2[7694],simde_mm_xor_si128(c2[8398],simde_mm_xor_si128(c2[673],simde_mm_xor_si128(c2[7719],simde_mm_xor_si128(c2[9824],simde_mm_xor_si128(c2[5607],simde_mm_xor_si128(c2[9839],simde_mm_xor_si128(c2[4902],simde_mm_xor_si128(c2[5606],c2[9830]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 9
|
|
d2[144]=simde_mm_xor_si128(c2[4934],simde_mm_xor_si128(c2[4267],simde_mm_xor_si128(c2[9478],simde_mm_xor_si128(c2[3873],simde_mm_xor_si128(c2[10280],simde_mm_xor_si128(c2[10400],simde_mm_xor_si128(c2[9731],c2[6285])))))));
|
|
|
|
//row: 10
|
|
d2[160]=simde_mm_xor_si128(c2[1446],simde_mm_xor_si128(c2[3593],simde_mm_xor_si128(c2[5066],simde_mm_xor_si128(c2[234],simde_mm_xor_si128(c2[963],c2[7491])))));
|
|
|
|
//row: 11
|
|
d2[176]=simde_mm_xor_si128(c2[4239],simde_mm_xor_si128(c2[1],simde_mm_xor_si128(c2[705],simde_mm_xor_si128(c2[9871],simde_mm_xor_si128(c2[6337],simde_mm_xor_si128(c2[4230],simde_mm_xor_si128(c2[712],simde_mm_xor_si128(c2[3527],simde_mm_xor_si128(c2[9],simde_mm_xor_si128(c2[9156],simde_mm_xor_si128(c2[740],simde_mm_xor_si128(c2[7781],simde_mm_xor_si128(c2[8485],simde_mm_xor_si128(c2[7782],simde_mm_xor_si128(c2[4264],simde_mm_xor_si128(c2[3557],simde_mm_xor_si128(c2[39],simde_mm_xor_si128(c2[6370],simde_mm_xor_si128(c2[9933],simde_mm_xor_si128(c2[5711],simde_mm_xor_si128(c2[6415],simde_mm_xor_si128(c2[7822],simde_mm_xor_si128(c2[4288],simde_mm_xor_si128(c2[3595],simde_mm_xor_si128(c2[77],simde_mm_xor_si128(c2[7849],simde_mm_xor_si128(c2[3627],simde_mm_xor_si128(c2[4331],simde_mm_xor_si128(c2[807],simde_mm_xor_si128(c2[8552],simde_mm_xor_si128(c2[101],simde_mm_xor_si128(c2[7846],simde_mm_xor_si128(c2[5767],simde_mm_xor_si128(c2[2249],simde_mm_xor_si128(c2[7875],simde_mm_xor_si128(c2[4357],simde_mm_xor_si128(c2[129],simde_mm_xor_si128(c2[7874],simde_mm_xor_si128(c2[166],simde_mm_xor_si128(c2[7207],simde_mm_xor_si128(c2[7911],simde_mm_xor_si128(c2[2276],simde_mm_xor_si128(c2[10021],simde_mm_xor_si128(c2[871],simde_mm_xor_si128(c2[8616],simde_mm_xor_si128(c2[4416],simde_mm_xor_si128(c2[194],simde_mm_xor_si128(c2[898],simde_mm_xor_si128(c2[6533],simde_mm_xor_si128(c2[3015],simde_mm_xor_si128(c2[1609],simde_mm_xor_si128(c2[9354],simde_mm_xor_si128(c2[7277],simde_mm_xor_si128(c2[3759],simde_mm_xor_si128(c2[942],simde_mm_xor_si128(c2[8687],simde_mm_xor_si128(c2[10791],simde_mm_xor_si128(c2[7273],simde_mm_xor_si128(c2[3078],simde_mm_xor_si128(c2[10823],simde_mm_xor_si128(c2[9418],simde_mm_xor_si128(c2[5900],simde_mm_xor_si128(c2[10830],simde_mm_xor_si128(c2[7296],simde_mm_xor_si128(c2[5219],simde_mm_xor_si128(c2[997],simde_mm_xor_si128(c2[1701],simde_mm_xor_si128(c2[6634],simde_mm_xor_si128(c2[3116],simde_mm_xor_si128(c2[8037],simde_mm_xor_si128(c2[4519],simde_mm_xor_si128(c2[1038],simde_mm_xor_si128(c2[8079],simde_mm_xor_si128(c2[8783],simde_mm_xor_si128(c2[2434],simde_mm_xor_si128(c2[10179],simde_mm_xor_si128(c2[3144],simde_mm_xor_si128(c2[10889],simde_mm_xor_si128(c2[7398],simde_mm_xor_si128(c2[3176],simde_mm_xor_si128(c2[3880],simde_mm_xor_si128(c2[5997],simde_mm_xor_si128(c2[2479],simde_mm_xor_si128(c2[1765],simde_mm_xor_si128(c2[9510],simde_mm_xor_si128(c2[8139],simde_mm_xor_si128(c2[3917],simde_mm_xor_si128(c2[4621],simde_mm_xor_si128(c2[1798],simde_mm_xor_si128(c2[9543],simde_mm_xor_si128(c2[1807],simde_mm_xor_si128(c2[9536],simde_mm_xor_si128(c2[2501],simde_mm_xor_si128(c2[3936],simde_mm_xor_si128(c2[10977],simde_mm_xor_si128(c2[418],simde_mm_xor_si128(c2[7464],simde_mm_xor_si128(c2[3946],simde_mm_xor_si128(c2[5357],simde_mm_xor_si128(c2[1839],simde_mm_xor_si128(c2[6790],simde_mm_xor_si128(c2[3272],simde_mm_xor_si128(c2[9613],simde_mm_xor_si128(c2[6095],simde_mm_xor_si128(c2[11020],simde_mm_xor_si128(c2[7502],simde_mm_xor_si128(c2[11051],simde_mm_xor_si128(c2[6829],simde_mm_xor_si128(c2[7533],simde_mm_xor_si128(c2[488],simde_mm_xor_si128(c2[8233],simde_mm_xor_si128(c2[9645],simde_mm_xor_si128(c2[6127],simde_mm_xor_si128(c2[2625],simde_mm_xor_si128(c2[9666],simde_mm_xor_si128(c2[10370],simde_mm_xor_si128(c2[7560],simde_mm_xor_si128(c2[4042],simde_mm_xor_si128(c2[8974],simde_mm_xor_si128(c2[5440],simde_mm_xor_si128(c2[4747],simde_mm_xor_si128(c2[5481],simde_mm_xor_si128(c2[1963],simde_mm_xor_si128(c2[1263],simde_mm_xor_si128(c2[8992],simde_mm_xor_si128(c2[6180],simde_mm_xor_si128(c2[2662],simde_mm_xor_si128(c2[7627],simde_mm_xor_si128(c2[3405],simde_mm_xor_si128(c2[4109],simde_mm_xor_si128(c2[6924],simde_mm_xor_si128(c2[3406],simde_mm_xor_si128(c2[591],simde_mm_xor_si128(c2[8320],simde_mm_xor_si128(c2[11169],simde_mm_xor_si128(c2[6947],simde_mm_xor_si128(c2[7651],simde_mm_xor_si128(c2[8367],simde_mm_xor_si128(c2[4833],simde_mm_xor_si128(c2[5551],simde_mm_xor_si128(c2[2017],simde_mm_xor_si128(c2[8398],simde_mm_xor_si128(c2[4160],simde_mm_xor_si128(c2[4864],simde_mm_xor_si128(c2[1351],simde_mm_xor_si128(c2[9096],simde_mm_xor_si128(c2[9096],simde_mm_xor_si128(c2[5578],simde_mm_xor_si128(c2[8417],simde_mm_xor_si128(c2[4195],simde_mm_xor_si128(c2[4899],simde_mm_xor_si128(c2[6305],simde_mm_xor_si128(c2[2787],simde_mm_xor_si128(c2[6304],simde_mm_xor_si128(c2[2786],c2[10532])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 12
|
|
d2[192]=simde_mm_xor_si128(c2[10],simde_mm_xor_si128(c2[7074],simde_mm_xor_si128(c2[3841],simde_mm_xor_si128(c2[354],simde_mm_xor_si128(c2[7470],c2[5504])))));
|
|
|
|
//row: 13
|
|
d2[208]=simde_mm_xor_si128(c2[5635],simde_mm_xor_si128(c2[6339],simde_mm_xor_si128(c2[708],simde_mm_xor_si128(c2[6346],simde_mm_xor_si128(c2[5643],simde_mm_xor_si128(c2[715],simde_mm_xor_si128(c2[2152],simde_mm_xor_si128(c2[2856],simde_mm_xor_si128(c2[9898],simde_mm_xor_si128(c2[5673],simde_mm_xor_si128(c2[66],simde_mm_xor_si128(c2[770],simde_mm_xor_si128(c2[9922],simde_mm_xor_si128(c2[5711],simde_mm_xor_si128(c2[9261],simde_mm_xor_si128(c2[9965],simde_mm_xor_si128(c2[2923],simde_mm_xor_si128(c2[2217],simde_mm_xor_si128(c2[5743],simde_mm_xor_si128(c2[7883],simde_mm_xor_si128(c2[9991],simde_mm_xor_si128(c2[2245],simde_mm_xor_si128(c2[1578],simde_mm_xor_si128(c2[2282],simde_mm_xor_si128(c2[4392],simde_mm_xor_si128(c2[2987],simde_mm_xor_si128(c2[5828],simde_mm_xor_si128(c2[6532],simde_mm_xor_si128(c2[8649],simde_mm_xor_si128(c2[3725],simde_mm_xor_si128(c2[9377],simde_mm_xor_si128(c2[3042],simde_mm_xor_si128(c2[1644],simde_mm_xor_si128(c2[5161],simde_mm_xor_si128(c2[5194],simde_mm_xor_si128(c2[271],simde_mm_xor_si128(c2[1667],simde_mm_xor_si128(c2[6631],simde_mm_xor_si128(c2[7335],simde_mm_xor_si128(c2[8750],simde_mm_xor_si128(c2[10153],simde_mm_xor_si128(c2[2434],simde_mm_xor_si128(c2[3138],simde_mm_xor_si128(c2[4550],simde_mm_xor_si128(c2[5260],simde_mm_xor_si128(c2[8810],simde_mm_xor_si128(c2[9514],simde_mm_xor_si128(c2[8097],simde_mm_xor_si128(c2[3881],simde_mm_xor_si128(c2[9551],simde_mm_xor_si128(c2[10255],simde_mm_xor_si128(c2[3914],simde_mm_xor_si128(c2[3907],simde_mm_xor_si128(c2[5348],simde_mm_xor_si128(c2[6052],simde_mm_xor_si128(c2[9580],simde_mm_xor_si128(c2[7457],simde_mm_xor_si128(c2[8906],simde_mm_xor_si128(c2[450],simde_mm_xor_si128(c2[1857],simde_mm_xor_si128(c2[1184],simde_mm_xor_si128(c2[1888],simde_mm_xor_si128(c2[2604],simde_mm_xor_si128(c2[482],simde_mm_xor_si128(c2[4037],simde_mm_xor_si128(c2[4741],simde_mm_xor_si128(c2[9676],simde_mm_xor_si128(c2[11074],simde_mm_xor_si128(c2[7597],simde_mm_xor_si128(c2[3363],simde_mm_xor_si128(c2[8296],simde_mm_xor_si128(c2[9039],simde_mm_xor_si128(c2[9743],simde_mm_xor_si128(c2[9024],simde_mm_xor_si128(c2[2691],simde_mm_xor_si128(c2[1318],simde_mm_xor_si128(c2[2022],simde_mm_xor_si128(c2[10467],simde_mm_xor_si128(c2[7651],simde_mm_xor_si128(c2[9794],simde_mm_xor_si128(c2[10498],simde_mm_xor_si128(c2[3467],simde_mm_xor_si128(c2[11212],simde_mm_xor_si128(c2[6987],simde_mm_xor_si128(c2[9829],simde_mm_xor_si128(c2[10533],simde_mm_xor_si128(c2[8421],c2[8420])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 14
|
|
d2[224]=simde_mm_xor_si128(c2[9868],simde_mm_xor_si128(c2[5315],simde_mm_xor_si128(c2[10348],simde_mm_xor_si128(c2[11079],simde_mm_xor_si128(c2[545],c2[4206])))));
|
|
|
|
//row: 15
|
|
d2[240]=simde_mm_xor_si128(c2[9866],simde_mm_xor_si128(c2[4235],simde_mm_xor_si128(c2[9857],simde_mm_xor_si128(c2[8450],simde_mm_xor_si128(c2[9154],simde_mm_xor_si128(c2[5634],simde_mm_xor_si128(c2[6383],simde_mm_xor_si128(c2[2146],simde_mm_xor_si128(c2[8480],simde_mm_xor_si128(c2[9184],simde_mm_xor_si128(c2[38],simde_mm_xor_si128(c2[4297],simde_mm_xor_si128(c2[2186],simde_mm_xor_si128(c2[9222],simde_mm_xor_si128(c2[2213],simde_mm_xor_si128(c2[6434],simde_mm_xor_si128(c2[5024],simde_mm_xor_si128(c2[5728],simde_mm_xor_si128(c2[131],simde_mm_xor_si128(c2[2255],simde_mm_xor_si128(c2[5068],simde_mm_xor_si128(c2[5772],simde_mm_xor_si128(c2[5793],simde_mm_xor_si128(c2[7919],simde_mm_xor_si128(c2[6498],simde_mm_xor_si128(c2[10059],simde_mm_xor_si128(c2[897],simde_mm_xor_si128(c2[6532],simde_mm_xor_si128(c2[7236],simde_mm_xor_si128(c2[1641],simde_mm_xor_si128(c2[6569],simde_mm_xor_si128(c2[4451],simde_mm_xor_si128(c2[5155],simde_mm_xor_si128(c2[8705],simde_mm_xor_si128(c2[3782],simde_mm_xor_si128(c2[4490],simde_mm_xor_si128(c2[5194],simde_mm_xor_si128(c2[10862],simde_mm_xor_si128(c2[998],simde_mm_xor_si128(c2[2401],simde_mm_xor_si128(c2[6665],simde_mm_xor_si128(c2[8077],simde_mm_xor_si128(c2[8067],simde_mm_xor_si128(c2[8771],simde_mm_xor_si128(c2[1028],simde_mm_xor_si128(c2[1762],simde_mm_xor_si128(c2[361],simde_mm_xor_si128(c2[6688],simde_mm_xor_si128(c2[7392],simde_mm_xor_si128(c2[2503],simde_mm_xor_si128(c2[7425],simde_mm_xor_si128(c2[6730],simde_mm_xor_si128(c2[7434],simde_mm_xor_si128(c2[9579],simde_mm_xor_si128(c2[1828],simde_mm_xor_si128(c2[10280],simde_mm_xor_si128(c2[10984],simde_mm_xor_si128(c2[10979],simde_mm_xor_si128(c2[1154],simde_mm_xor_si128(c2[3977],simde_mm_xor_si128(c2[4680],simde_mm_xor_si128(c2[5384],simde_mm_xor_si128(c2[5415],simde_mm_xor_si128(c2[6115],simde_mm_xor_si128(c2[4009],simde_mm_xor_si128(c2[8268],simde_mm_xor_si128(c2[1924],simde_mm_xor_si128(c2[2634],simde_mm_xor_si128(c2[3338],simde_mm_xor_si128(c2[11108],simde_mm_xor_si128(c2[6890],simde_mm_xor_si128(c2[11119],simde_mm_xor_si128(c2[544],simde_mm_xor_si128(c2[1991],simde_mm_xor_si128(c2[1288],simde_mm_xor_si128(c2[5514],simde_mm_xor_si128(c2[6218],simde_mm_xor_si128(c2[8324],simde_mm_xor_si128(c2[5549],simde_mm_xor_si128(c2[2731],simde_mm_xor_si128(c2[11178],simde_mm_xor_si128(c2[2762],simde_mm_xor_si128(c2[6978],simde_mm_xor_si128(c2[2756],simde_mm_xor_si128(c2[3460],simde_mm_xor_si128(c2[2797],simde_mm_xor_si128(c2[685],simde_mm_xor_si128(c2[11243],c2[684]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 16
|
|
d2[256]=simde_mm_xor_si128(c2[2825],simde_mm_xor_si128(c2[8457],simde_mm_xor_si128(c2[2816],simde_mm_xor_si128(c2[2113],simde_mm_xor_si128(c2[10605],simde_mm_xor_si128(c2[6368],simde_mm_xor_si128(c2[2159],simde_mm_xor_si128(c2[36],simde_mm_xor_si128(c2[8519],simde_mm_xor_si128(c2[6408],simde_mm_xor_si128(c2[2181],simde_mm_xor_si128(c2[6435],simde_mm_xor_si128(c2[10656],simde_mm_xor_si128(c2[9966],simde_mm_xor_si128(c2[803],simde_mm_xor_si128(c2[4353],simde_mm_xor_si128(c2[6477],simde_mm_xor_si128(c2[9994],simde_mm_xor_si128(c2[10031],simde_mm_xor_si128(c2[878],simde_mm_xor_si128(c2[10720],simde_mm_xor_si128(c2[3018],simde_mm_xor_si128(c2[5135],simde_mm_xor_si128(c2[195],simde_mm_xor_si128(c2[5863],simde_mm_xor_si128(c2[10791],simde_mm_xor_si128(c2[9377],simde_mm_xor_si128(c2[1664],simde_mm_xor_si128(c2[8004],simde_mm_xor_si128(c2[9416],simde_mm_xor_si128(c2[3821],simde_mm_xor_si128(c2[5220],simde_mm_xor_si128(c2[6639],simde_mm_xor_si128(c2[10887],simde_mm_xor_si128(c2[1036],simde_mm_xor_si128(c2[1730],simde_mm_xor_si128(c2[5984],simde_mm_xor_si128(c2[4583],simde_mm_xor_si128(c2[367],simde_mm_xor_si128(c2[1059],simde_mm_xor_si128(c2[6725],simde_mm_xor_si128(c2[384],simde_mm_xor_si128(c2[393],simde_mm_xor_si128(c2[2538],simde_mm_xor_si128(c2[6050],simde_mm_xor_si128(c2[3943],simde_mm_xor_si128(c2[5376],simde_mm_xor_si128(c2[8199],simde_mm_xor_si128(c2[9606],simde_mm_xor_si128(c2[9637],simde_mm_xor_si128(c2[10337],simde_mm_xor_si128(c2[8231],simde_mm_xor_si128(c2[1227],simde_mm_xor_si128(c2[6146],simde_mm_xor_si128(c2[7560],simde_mm_xor_si128(c2[4067],simde_mm_xor_si128(c2[11112],simde_mm_xor_si128(c2[4782],simde_mm_xor_si128(c2[6213],simde_mm_xor_si128(c2[5510],simde_mm_xor_si128(c2[10440],simde_mm_xor_si128(c2[9771],simde_mm_xor_si128(c2[6953],simde_mm_xor_si128(c2[4137],simde_mm_xor_si128(c2[6984],simde_mm_xor_si128(c2[11200],simde_mm_xor_si128(c2[7682],simde_mm_xor_si128(c2[2755],simde_mm_xor_si128(c2[7019],simde_mm_xor_si128(c2[4907],c2[4906]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 17
|
|
d2[272]=simde_mm_xor_si128(c2[4928],simde_mm_xor_si128(c2[3274],simde_mm_xor_si128(c2[8259],simde_mm_xor_si128(c2[1248],c2[681]))));
|
|
|
|
//row: 18
|
|
d2[288]=simde_mm_xor_si128(c2[7074],simde_mm_xor_si128(c2[6734],simde_mm_xor_si128(c2[6048],simde_mm_xor_si128(c2[8329],c2[2729]))));
|
|
|
|
//row: 19
|
|
d2[304]=simde_mm_xor_si128(c2[8451],simde_mm_xor_si128(c2[6372],simde_mm_xor_si128(c2[5860],simde_mm_xor_si128(c2[10823],c2[334]))));
|
|
|
|
//row: 20
|
|
d2[320]=simde_mm_xor_si128(c2[7049],simde_mm_xor_si128(c2[1418],simde_mm_xor_si128(c2[7040],simde_mm_xor_si128(c2[6337],simde_mm_xor_si128(c2[4937],simde_mm_xor_si128(c2[3566],simde_mm_xor_si128(c2[10592],simde_mm_xor_si128(c2[6383],simde_mm_xor_si128(c2[1480],simde_mm_xor_si128(c2[10632],simde_mm_xor_si128(c2[6405],simde_mm_xor_si128(c2[10659],simde_mm_xor_si128(c2[3617],simde_mm_xor_si128(c2[2927],simde_mm_xor_si128(c2[7147],simde_mm_xor_si128(c2[8577],simde_mm_xor_si128(c2[10701],simde_mm_xor_si128(c2[2955],simde_mm_xor_si128(c2[2976],simde_mm_xor_si128(c2[5102],simde_mm_xor_si128(c2[3681],simde_mm_xor_si128(c2[7242],simde_mm_xor_si128(c2[9359],simde_mm_xor_si128(c2[4419],simde_mm_xor_si128(c2[10087],simde_mm_xor_si128(c2[3752],simde_mm_xor_si128(c2[2338],simde_mm_xor_si128(c2[5888],simde_mm_xor_si128(c2[965],simde_mm_xor_si128(c2[2377],simde_mm_xor_si128(c2[8045],simde_mm_xor_si128(c2[9444],simde_mm_xor_si128(c2[10863],simde_mm_xor_si128(c2[6637],simde_mm_xor_si128(c2[3848],simde_mm_xor_si128(c2[5260],simde_mm_xor_si128(c2[5954],simde_mm_xor_si128(c2[10208],simde_mm_xor_si128(c2[8807],simde_mm_xor_si128(c2[4591],simde_mm_xor_si128(c2[10914],simde_mm_xor_si128(c2[10949],simde_mm_xor_si128(c2[4608],simde_mm_xor_si128(c2[4617],simde_mm_xor_si128(c2[6762],simde_mm_xor_si128(c2[10274],simde_mm_xor_si128(c2[8167],simde_mm_xor_si128(c2[9600],simde_mm_xor_si128(c2[1160],simde_mm_xor_si128(c2[2567],simde_mm_xor_si128(c2[2598],simde_mm_xor_si128(c2[3298],simde_mm_xor_si128(c2[1192],simde_mm_xor_si128(c2[5451],simde_mm_xor_si128(c2[10370],simde_mm_xor_si128(c2[521],simde_mm_xor_si128(c2[8291],simde_mm_xor_si128(c2[4073],simde_mm_xor_si128(c2[9006],simde_mm_xor_si128(c2[10437],simde_mm_xor_si128(c2[9734],simde_mm_xor_si128(c2[3401],simde_mm_xor_si128(c2[2732],simde_mm_xor_si128(c2[11177],simde_mm_xor_si128(c2[8361],simde_mm_xor_si128(c2[11208],simde_mm_xor_si128(c2[4161],simde_mm_xor_si128(c2[643],simde_mm_xor_si128(c2[11243],simde_mm_xor_si128(c2[9131],c2[9130]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 21
|
|
d2[336]=simde_mm_xor_si128(c2[6383],simde_mm_xor_si128(c2[6503],simde_mm_xor_si128(c2[9670],simde_mm_xor_si128(c2[2760],c2[8426]))));
|
|
|
|
//row: 22
|
|
d2[352]=simde_mm_xor_si128(c2[4],simde_mm_xor_si128(c2[10248],simde_mm_xor_si128(c2[8875],c2[10409])));
|
|
|
|
//row: 23
|
|
d2[368]=simde_mm_xor_si128(c2[8489],simde_mm_xor_si128(c2[2185],simde_mm_xor_si128(c2[7370],c2[6217])));
|
|
|
|
//row: 24
|
|
d2[384]=simde_mm_xor_si128(c2[5645],simde_mm_xor_si128(c2[14],simde_mm_xor_si128(c2[5636],simde_mm_xor_si128(c2[4933],simde_mm_xor_si128(c2[7],simde_mm_xor_si128(c2[2146],simde_mm_xor_si128(c2[9188],simde_mm_xor_si128(c2[4963],simde_mm_xor_si128(c2[76],simde_mm_xor_si128(c2[9228],simde_mm_xor_si128(c2[5001],simde_mm_xor_si128(c2[9255],simde_mm_xor_si128(c2[2213],simde_mm_xor_si128(c2[1507],simde_mm_xor_si128(c2[4325],simde_mm_xor_si128(c2[7173],simde_mm_xor_si128(c2[9281],simde_mm_xor_si128(c2[1551],simde_mm_xor_si128(c2[8590],simde_mm_xor_si128(c2[1572],simde_mm_xor_si128(c2[3682],simde_mm_xor_si128(c2[2277],simde_mm_xor_si128(c2[5838],simde_mm_xor_si128(c2[7939],simde_mm_xor_si128(c2[3015],simde_mm_xor_si128(c2[8683],simde_mm_xor_si128(c2[2348],simde_mm_xor_si128(c2[934],simde_mm_xor_si128(c2[4484],simde_mm_xor_si128(c2[10824],simde_mm_xor_si128(c2[973],simde_mm_xor_si128(c2[6625],simde_mm_xor_si128(c2[8040],simde_mm_xor_si128(c2[9443],simde_mm_xor_si128(c2[2444],simde_mm_xor_si128(c2[3840],simde_mm_xor_si128(c2[4550],simde_mm_xor_si128(c2[8804],simde_mm_xor_si128(c2[7403],simde_mm_xor_si128(c2[3171],simde_mm_xor_si128(c2[3175],simde_mm_xor_si128(c2[9545],simde_mm_xor_si128(c2[3204],simde_mm_xor_si128(c2[3213],simde_mm_xor_si128(c2[5358],simde_mm_xor_si128(c2[8870],simde_mm_xor_si128(c2[6763],simde_mm_xor_si128(c2[8196],simde_mm_xor_si128(c2[11019],simde_mm_xor_si128(c2[1163],simde_mm_xor_si128(c2[1194],simde_mm_xor_si128(c2[1894],simde_mm_xor_si128(c2[11051],simde_mm_xor_si128(c2[4047],simde_mm_xor_si128(c2[8966],simde_mm_xor_si128(c2[10380],simde_mm_xor_si128(c2[6887],simde_mm_xor_si128(c2[2669],simde_mm_xor_si128(c2[7586],simde_mm_xor_si128(c2[9033],simde_mm_xor_si128(c2[8330],simde_mm_xor_si128(c2[1997],simde_mm_xor_si128(c2[1312],simde_mm_xor_si128(c2[9773],simde_mm_xor_si128(c2[6957],simde_mm_xor_si128(c2[9804],simde_mm_xor_si128(c2[2757],simde_mm_xor_si128(c2[10502],simde_mm_xor_si128(c2[9839],simde_mm_xor_si128(c2[7727],c2[7726]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 25
|
|
d2[400]=simde_mm_xor_si128(c2[4961],simde_mm_xor_si128(c2[5832],simde_mm_xor_si128(c2[3047],c2[4683])));
|
|
|
|
//row: 26
|
|
d2[416]=simde_mm_xor_si128(c2[2124],simde_mm_xor_si128(c2[2191],simde_mm_xor_si128(c2[5069],c2[9635])));
|
|
|
|
//row: 27
|
|
d2[432]=simde_mm_xor_si128(c2[6369],simde_mm_xor_si128(c2[5830],c2[1676]));
|
|
|
|
//row: 28
|
|
d2[448]=simde_mm_xor_si128(c2[8],simde_mm_xor_si128(c2[3658],simde_mm_xor_si128(c2[4139],c2[11235])));
|
|
|
|
//row: 29
|
|
d2[464]=simde_mm_xor_si128(c2[704],simde_mm_xor_si128(c2[6336],simde_mm_xor_si128(c2[711],simde_mm_xor_si128(c2[10567],simde_mm_xor_si128(c2[8],simde_mm_xor_si128(c2[8484],simde_mm_xor_si128(c2[4263],simde_mm_xor_si128(c2[10597],simde_mm_xor_si128(c2[38],simde_mm_xor_si128(c2[4261],simde_mm_xor_si128(c2[6414],simde_mm_xor_si128(c2[4303],simde_mm_xor_si128(c2[76],simde_mm_xor_si128(c2[4330],simde_mm_xor_si128(c2[8551],simde_mm_xor_si128(c2[7141],simde_mm_xor_si128(c2[7845],simde_mm_xor_si128(c2[2248],simde_mm_xor_si128(c2[4356],simde_mm_xor_si128(c2[7169],simde_mm_xor_si128(c2[7873],simde_mm_xor_si128(c2[7910],simde_mm_xor_si128(c2[10020],simde_mm_xor_si128(c2[8615],simde_mm_xor_si128(c2[897],simde_mm_xor_si128(c2[3014],simde_mm_xor_si128(c2[8649],simde_mm_xor_si128(c2[9353],simde_mm_xor_si128(c2[3758],simde_mm_xor_si128(c2[8686],simde_mm_xor_si128(c2[6568],simde_mm_xor_si128(c2[7272],simde_mm_xor_si128(c2[10822],simde_mm_xor_si128(c2[5899],simde_mm_xor_si128(c2[6607],simde_mm_xor_si128(c2[7311],simde_mm_xor_si128(c2[1700],simde_mm_xor_si128(c2[3115],simde_mm_xor_si128(c2[4518],simde_mm_xor_si128(c2[8782],simde_mm_xor_si128(c2[10178],simde_mm_xor_si128(c2[10184],simde_mm_xor_si128(c2[10888],simde_mm_xor_si128(c2[3879],simde_mm_xor_si128(c2[2478],simde_mm_xor_si128(c2[8805],simde_mm_xor_si128(c2[9509],simde_mm_xor_si128(c2[4620],simde_mm_xor_si128(c2[9542],simde_mm_xor_si128(c2[8847],simde_mm_xor_si128(c2[9551],simde_mm_xor_si128(c2[417],simde_mm_xor_si128(c2[3945],simde_mm_xor_si128(c2[1134],simde_mm_xor_si128(c2[1838],simde_mm_xor_si128(c2[3271],simde_mm_xor_si128(c2[6094],simde_mm_xor_si128(c2[6797],simde_mm_xor_si128(c2[7501],simde_mm_xor_si128(c2[8910],simde_mm_xor_si128(c2[7532],simde_mm_xor_si128(c2[8232],simde_mm_xor_si128(c2[6126],simde_mm_xor_si128(c2[10369],simde_mm_xor_si128(c2[4041],simde_mm_xor_si128(c2[4751],simde_mm_xor_si128(c2[5455],simde_mm_xor_si128(c2[1962],simde_mm_xor_si128(c2[9007],simde_mm_xor_si128(c2[1957],simde_mm_xor_si128(c2[2661],simde_mm_xor_si128(c2[4108],simde_mm_xor_si128(c2[3405],simde_mm_xor_si128(c2[7631],simde_mm_xor_si128(c2[8335],simde_mm_xor_si128(c2[3397],simde_mm_xor_si128(c2[7650],simde_mm_xor_si128(c2[4832],simde_mm_xor_si128(c2[2016],simde_mm_xor_si128(c2[4879],simde_mm_xor_si128(c2[9095],simde_mm_xor_si128(c2[4873],simde_mm_xor_si128(c2[5577],simde_mm_xor_si128(c2[4898],simde_mm_xor_si128(c2[2786],simde_mm_xor_si128(c2[2081],c2[2785]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 30
|
|
d2[480]=simde_mm_xor_si128(c2[2816],simde_mm_xor_si128(c2[8448],simde_mm_xor_si128(c2[2119],simde_mm_xor_si128(c2[2823],simde_mm_xor_si128(c2[1416],simde_mm_xor_si128(c2[2120],simde_mm_xor_si128(c2[5645],simde_mm_xor_si128(c2[10596],simde_mm_xor_si128(c2[5671],simde_mm_xor_si128(c2[6375],simde_mm_xor_si128(c2[1446],simde_mm_xor_si128(c2[2150],simde_mm_xor_si128(c2[8526],simde_mm_xor_si128(c2[6415],simde_mm_xor_si128(c2[1484],simde_mm_xor_si128(c2[2188],simde_mm_xor_si128(c2[6442],simde_mm_xor_si128(c2[10663],simde_mm_xor_si128(c2[9253],simde_mm_xor_si128(c2[9957],simde_mm_xor_si128(c2[4360],simde_mm_xor_si128(c2[5764],simde_mm_xor_si128(c2[6468],simde_mm_xor_si128(c2[9281],simde_mm_xor_si128(c2[9985],simde_mm_xor_si128(c2[10022],simde_mm_xor_si128(c2[869],simde_mm_xor_si128(c2[10023],simde_mm_xor_si128(c2[10727],simde_mm_xor_si128(c2[3009],simde_mm_xor_si128(c2[4422],simde_mm_xor_si128(c2[5126],simde_mm_xor_si128(c2[10761],simde_mm_xor_si128(c2[202],simde_mm_xor_si128(c2[5870],simde_mm_xor_si128(c2[10094],simde_mm_xor_si128(c2[10798],simde_mm_xor_si128(c2[8680],simde_mm_xor_si128(c2[9384],simde_mm_xor_si128(c2[1671],simde_mm_xor_si128(c2[7307],simde_mm_xor_si128(c2[8011],simde_mm_xor_si128(c2[8719],simde_mm_xor_si128(c2[9423],simde_mm_xor_si128(c2[3812],simde_mm_xor_si128(c2[5227],simde_mm_xor_si128(c2[5926],simde_mm_xor_si128(c2[6630],simde_mm_xor_si128(c2[10894],simde_mm_xor_si128(c2[323],simde_mm_xor_si128(c2[1027],simde_mm_xor_si128(c2[1033],simde_mm_xor_si128(c2[1737],simde_mm_xor_si128(c2[6660],simde_mm_xor_si128(c2[5991],simde_mm_xor_si128(c2[4590],simde_mm_xor_si128(c2[10917],simde_mm_xor_si128(c2[358],simde_mm_xor_si128(c2[6732],simde_mm_xor_si128(c2[391],simde_mm_xor_si128(c2[10959],simde_mm_xor_si128(c2[384],simde_mm_xor_si128(c2[2529],simde_mm_xor_si128(c2[5353],simde_mm_xor_si128(c2[6057],simde_mm_xor_si128(c2[3246],simde_mm_xor_si128(c2[3950],simde_mm_xor_si128(c2[6055],simde_mm_xor_si128(c2[5383],simde_mm_xor_si128(c2[7502],simde_mm_xor_si128(c2[8206],simde_mm_xor_si128(c2[8909],simde_mm_xor_si128(c2[9613],simde_mm_xor_si128(c2[9644],simde_mm_xor_si128(c2[10344],simde_mm_xor_si128(c2[7534],simde_mm_xor_si128(c2[8238],simde_mm_xor_si128(c2[1218],simde_mm_xor_si128(c2[6153],simde_mm_xor_si128(c2[6863],simde_mm_xor_si128(c2[7567],simde_mm_xor_si128(c2[4074],simde_mm_xor_si128(c2[10415],simde_mm_xor_si128(c2[11119],simde_mm_xor_si128(c2[4069],simde_mm_xor_si128(c2[4773],simde_mm_xor_si128(c2[6220],simde_mm_xor_si128(c2[4813],simde_mm_xor_si128(c2[5517],simde_mm_xor_si128(c2[9743],simde_mm_xor_si128(c2[10447],simde_mm_xor_si128(c2[9762],simde_mm_xor_si128(c2[6944],simde_mm_xor_si128(c2[3424],simde_mm_xor_si128(c2[4128],simde_mm_xor_si128(c2[6991],simde_mm_xor_si128(c2[10503],simde_mm_xor_si128(c2[11207],simde_mm_xor_si128(c2[6985],simde_mm_xor_si128(c2[7689],simde_mm_xor_si128(c2[7010],simde_mm_xor_si128(c2[4898],simde_mm_xor_si128(c2[4193],c2[4897])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 31
|
|
d2[496]=simde_mm_xor_si128(c2[4234],simde_mm_xor_si128(c2[5635],simde_mm_xor_si128(c2[9866],simde_mm_xor_si128(c2[4],simde_mm_xor_si128(c2[4225],simde_mm_xor_si128(c2[5642],simde_mm_xor_si128(c2[3522],simde_mm_xor_si128(c2[4235],simde_mm_xor_si128(c2[4939],simde_mm_xor_si128(c2[751],simde_mm_xor_si128(c2[2152],simde_mm_xor_si128(c2[7777],simde_mm_xor_si128(c2[9194],simde_mm_xor_si128(c2[3552],simde_mm_xor_si128(c2[4265],simde_mm_xor_si128(c2[4969],simde_mm_xor_si128(c2[10597],simde_mm_xor_si128(c2[9928],simde_mm_xor_si128(c2[66],simde_mm_xor_si128(c2[7817],simde_mm_xor_si128(c2[9218],simde_mm_xor_si128(c2[3590],simde_mm_xor_si128(c2[5007],simde_mm_xor_si128(c2[7844],simde_mm_xor_si128(c2[9261],simde_mm_xor_si128(c2[802],simde_mm_xor_si128(c2[2219],simde_mm_xor_si128(c2[96],simde_mm_xor_si128(c2[809],simde_mm_xor_si128(c2[1513],simde_mm_xor_si128(c2[5762],simde_mm_xor_si128(c2[7179],simde_mm_xor_si128(c2[7886],simde_mm_xor_si128(c2[9287],simde_mm_xor_si128(c2[140],simde_mm_xor_si128(c2[837],simde_mm_xor_si128(c2[1541],simde_mm_xor_si128(c2[161],simde_mm_xor_si128(c2[1578],simde_mm_xor_si128(c2[2287],simde_mm_xor_si128(c2[3688],simde_mm_xor_si128(c2[866],simde_mm_xor_si128(c2[2283],simde_mm_xor_si128(c2[4427],simde_mm_xor_si128(c2[5828],simde_mm_xor_si128(c2[6528],simde_mm_xor_si128(c2[7945],simde_mm_xor_si128(c2[1604],simde_mm_xor_si128(c2[2317],simde_mm_xor_si128(c2[3021],simde_mm_xor_si128(c2[7272],simde_mm_xor_si128(c2[8673],simde_mm_xor_si128(c2[937],simde_mm_xor_si128(c2[2338],simde_mm_xor_si128(c2[10786],simde_mm_xor_si128(c2[236],simde_mm_xor_si128(c2[940],simde_mm_xor_si128(c2[939],simde_mm_xor_si128(c2[3073],simde_mm_xor_si128(c2[4490],simde_mm_xor_si128(c2[9413],simde_mm_xor_si128(c2[10830],simde_mm_xor_si128(c2[10825],simde_mm_xor_si128(c2[259],simde_mm_xor_si128(c2[963],simde_mm_xor_si128(c2[5230],simde_mm_xor_si128(c2[6631],simde_mm_xor_si128(c2[6629],simde_mm_xor_si128(c2[8046],simde_mm_xor_si128(c2[8032],simde_mm_xor_si128(c2[9449],simde_mm_xor_si128(c2[1033],simde_mm_xor_si128(c2[2434],simde_mm_xor_si128(c2[2445],simde_mm_xor_si128(c2[3846],simde_mm_xor_si128(c2[3139],simde_mm_xor_si128(c2[3852],simde_mm_xor_si128(c2[4556],simde_mm_xor_si128(c2[7393],simde_mm_xor_si128(c2[8810],simde_mm_xor_si128(c2[5992],simde_mm_xor_si128(c2[7393],simde_mm_xor_si128(c2[1760],simde_mm_xor_si128(c2[2473],simde_mm_xor_si128(c2[3177],simde_mm_xor_si128(c2[8134],simde_mm_xor_si128(c2[9551],simde_mm_xor_si128(c2[1793],simde_mm_xor_si128(c2[3210],simde_mm_xor_si128(c2[1802],simde_mm_xor_si128(c2[2499],simde_mm_xor_si128(c2[3203],simde_mm_xor_si128(c2[3947],simde_mm_xor_si128(c2[5348],simde_mm_xor_si128(c2[7459],simde_mm_xor_si128(c2[8876],simde_mm_xor_si128(c2[5352],simde_mm_xor_si128(c2[6049],simde_mm_xor_si128(c2[6753],simde_mm_xor_si128(c2[6785],simde_mm_xor_si128(c2[8202],simde_mm_xor_si128(c2[9608],simde_mm_xor_si128(c2[11009],simde_mm_xor_si128(c2[11015],simde_mm_xor_si128(c2[449],simde_mm_xor_si128(c2[1153],simde_mm_xor_si128(c2[11046],simde_mm_xor_si128(c2[1184],simde_mm_xor_si128(c2[483],simde_mm_xor_si128(c2[1900],simde_mm_xor_si128(c2[9640],simde_mm_xor_si128(c2[11041],simde_mm_xor_si128(c2[2636],simde_mm_xor_si128(c2[4037],simde_mm_xor_si128(c2[7555],simde_mm_xor_si128(c2[8972],simde_mm_xor_si128(c2[8969],simde_mm_xor_si128(c2[9666],simde_mm_xor_si128(c2[10370],simde_mm_xor_si128(c2[5476],simde_mm_xor_si128(c2[6893],simde_mm_xor_si128(c2[1258],simde_mm_xor_si128(c2[2659],simde_mm_xor_si128(c2[6191],simde_mm_xor_si128(c2[6888],simde_mm_xor_si128(c2[7592],simde_mm_xor_si128(c2[7622],simde_mm_xor_si128(c2[9039],simde_mm_xor_si128(c2[6919],simde_mm_xor_si128(c2[8320],simde_mm_xor_si128(c2[586],simde_mm_xor_si128(c2[1283],simde_mm_xor_si128(c2[1987],simde_mm_xor_si128(c2[11180],simde_mm_xor_si128(c2[1318],simde_mm_xor_si128(c2[8362],simde_mm_xor_si128(c2[9763],simde_mm_xor_si128(c2[5546],simde_mm_xor_si128(c2[6947],simde_mm_xor_si128(c2[8393],simde_mm_xor_si128(c2[9794],simde_mm_xor_si128(c2[1346],simde_mm_xor_si128(c2[2763],simde_mm_xor_si128(c2[9091],simde_mm_xor_si128(c2[9804],simde_mm_xor_si128(c2[10508],simde_mm_xor_si128(c2[8428],simde_mm_xor_si128(c2[9829],simde_mm_xor_si128(c2[6316],simde_mm_xor_si128(c2[7717],simde_mm_xor_si128(c2[6315],simde_mm_xor_si128(c2[7012],c2[7716]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 32
|
|
d2[512]=simde_mm_xor_si128(c2[2823],simde_mm_xor_si128(c2[8455],simde_mm_xor_si128(c2[2126],simde_mm_xor_si128(c2[2830],simde_mm_xor_si128(c2[1423],simde_mm_xor_si128(c2[2127],simde_mm_xor_si128(c2[9165],simde_mm_xor_si128(c2[10603],simde_mm_xor_si128(c2[5678],simde_mm_xor_si128(c2[6382],simde_mm_xor_si128(c2[1453],simde_mm_xor_si128(c2[2157],simde_mm_xor_si128(c2[8517],simde_mm_xor_si128(c2[6406],simde_mm_xor_si128(c2[1475],simde_mm_xor_si128(c2[2179],simde_mm_xor_si128(c2[6433],simde_mm_xor_si128(c2[10670],simde_mm_xor_si128(c2[9260],simde_mm_xor_si128(c2[9964],simde_mm_xor_si128(c2[4367],simde_mm_xor_si128(c2[5771],simde_mm_xor_si128(c2[6475],simde_mm_xor_si128(c2[9288],simde_mm_xor_si128(c2[9992],simde_mm_xor_si128(c2[10029],simde_mm_xor_si128(c2[876],simde_mm_xor_si128(c2[10030],simde_mm_xor_si128(c2[10734],simde_mm_xor_si128(c2[3016],simde_mm_xor_si128(c2[4429],simde_mm_xor_si128(c2[5133],simde_mm_xor_si128(c2[10752],simde_mm_xor_si128(c2[193],simde_mm_xor_si128(c2[5861],simde_mm_xor_si128(c2[10085],simde_mm_xor_si128(c2[10789],simde_mm_xor_si128(c2[8687],simde_mm_xor_si128(c2[9391],simde_mm_xor_si128(c2[1678],simde_mm_xor_si128(c2[7298],simde_mm_xor_si128(c2[8002],simde_mm_xor_si128(c2[8710],simde_mm_xor_si128(c2[9414],simde_mm_xor_si128(c2[3819],simde_mm_xor_si128(c2[5218],simde_mm_xor_si128(c2[5933],simde_mm_xor_si128(c2[6637],simde_mm_xor_si128(c2[10885],simde_mm_xor_si128(c2[330],simde_mm_xor_si128(c2[1034],simde_mm_xor_si128(c2[1024],simde_mm_xor_si128(c2[1728],simde_mm_xor_si128(c2[5998],simde_mm_xor_si128(c2[4581],simde_mm_xor_si128(c2[10924],simde_mm_xor_si128(c2[365],simde_mm_xor_si128(c2[6723],simde_mm_xor_si128(c2[398],simde_mm_xor_si128(c2[10950],simde_mm_xor_si128(c2[391],simde_mm_xor_si128(c2[391],simde_mm_xor_si128(c2[2536],simde_mm_xor_si128(c2[5344],simde_mm_xor_si128(c2[6048],simde_mm_xor_si128(c2[3237],simde_mm_xor_si128(c2[3941],simde_mm_xor_si128(c2[5390],simde_mm_xor_si128(c2[7493],simde_mm_xor_si128(c2[8197],simde_mm_xor_si128(c2[8900],simde_mm_xor_si128(c2[9604],simde_mm_xor_si128(c2[5388],simde_mm_xor_si128(c2[9635],simde_mm_xor_si128(c2[10351],simde_mm_xor_si128(c2[7525],simde_mm_xor_si128(c2[8229],simde_mm_xor_si128(c2[1225],simde_mm_xor_si128(c2[6144],simde_mm_xor_si128(c2[6854],simde_mm_xor_si128(c2[7558],simde_mm_xor_si128(c2[4065],simde_mm_xor_si128(c2[10406],simde_mm_xor_si128(c2[11110],simde_mm_xor_si128(c2[4076],simde_mm_xor_si128(c2[4780],simde_mm_xor_si128(c2[6211],simde_mm_xor_si128(c2[4804],simde_mm_xor_si128(c2[5508],simde_mm_xor_si128(c2[9734],simde_mm_xor_si128(c2[10438],simde_mm_xor_si128(c2[9769],simde_mm_xor_si128(c2[6951],simde_mm_xor_si128(c2[3431],simde_mm_xor_si128(c2[4135],simde_mm_xor_si128(c2[6982],simde_mm_xor_si128(c2[10510],simde_mm_xor_si128(c2[11214],simde_mm_xor_si128(c2[6976],simde_mm_xor_si128(c2[7680],simde_mm_xor_si128(c2[7017],simde_mm_xor_si128(c2[4905],simde_mm_xor_si128(c2[4200],c2[4904])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 33
|
|
d2[528]=simde_mm_xor_si128(c2[1440],simde_mm_xor_si128(c2[7819],simde_mm_xor_si128(c2[6690],c2[2797])));
|
|
|
|
//row: 34
|
|
d2[544]=simde_mm_xor_si128(c2[10567],simde_mm_xor_si128(c2[5162],simde_mm_xor_si128(c2[3306],c2[11113])));
|
|
|
|
//row: 35
|
|
d2[560]=simde_mm_xor_si128(c2[710],simde_mm_xor_si128(c2[6342],simde_mm_xor_si128(c2[717],simde_mm_xor_si128(c2[14],simde_mm_xor_si128(c2[8490],simde_mm_xor_si128(c2[4269],simde_mm_xor_si128(c2[44],simde_mm_xor_si128(c2[746],simde_mm_xor_si128(c2[6404],simde_mm_xor_si128(c2[4293],simde_mm_xor_si128(c2[66],simde_mm_xor_si128(c2[4320],simde_mm_xor_si128(c2[8557],simde_mm_xor_si128(c2[7851],simde_mm_xor_si128(c2[2254],simde_mm_xor_si128(c2[4362],simde_mm_xor_si128(c2[7879],simde_mm_xor_si128(c2[7916],simde_mm_xor_si128(c2[10026],simde_mm_xor_si128(c2[8621],simde_mm_xor_si128(c2[903],simde_mm_xor_si128(c2[3020],simde_mm_xor_si128(c2[9359],simde_mm_xor_si128(c2[3724],simde_mm_xor_si128(c2[3748],simde_mm_xor_si128(c2[8676],simde_mm_xor_si128(c2[7278],simde_mm_xor_si128(c2[10828],simde_mm_xor_si128(c2[5889],simde_mm_xor_si128(c2[7301],simde_mm_xor_si128(c2[1706],simde_mm_xor_si128(c2[3105],simde_mm_xor_si128(c2[4524],simde_mm_xor_si128(c2[8772],simde_mm_xor_si128(c2[10184],simde_mm_xor_si128(c2[10894],simde_mm_xor_si128(c2[3885],simde_mm_xor_si128(c2[2468],simde_mm_xor_si128(c2[9515],simde_mm_xor_si128(c2[4610],simde_mm_xor_si128(c2[9548],simde_mm_xor_si128(c2[9541],simde_mm_xor_si128(c2[10956],simde_mm_xor_si128(c2[423],simde_mm_xor_si128(c2[3951],simde_mm_xor_si128(c2[1828],simde_mm_xor_si128(c2[3277],simde_mm_xor_si128(c2[6084],simde_mm_xor_si128(c2[7491],simde_mm_xor_si128(c2[7522],simde_mm_xor_si128(c2[8238],simde_mm_xor_si128(c2[6116],simde_mm_xor_si128(c2[10375],simde_mm_xor_si128(c2[4047],simde_mm_xor_si128(c2[5445],simde_mm_xor_si128(c2[1952],simde_mm_xor_si128(c2[8997],simde_mm_xor_si128(c2[2667],simde_mm_xor_si128(c2[4098],simde_mm_xor_si128(c2[3395],simde_mm_xor_si128(c2[8325],simde_mm_xor_si128(c2[7656],simde_mm_xor_si128(c2[4838],simde_mm_xor_si128(c2[2022],simde_mm_xor_si128(c2[4869],simde_mm_xor_si128(c2[9101],simde_mm_xor_si128(c2[5583],simde_mm_xor_si128(c2[4904],simde_mm_xor_si128(c2[2792],c2[2791])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 36
|
|
d2[576]=simde_mm_xor_si128(c2[3522],simde_mm_xor_si128(c2[6790],simde_mm_xor_si128(c2[2595],c2[6215])));
|
|
|
|
//row: 37
|
|
d2[592]=simde_mm_xor_si128(c2[2823],simde_mm_xor_si128(c2[3527],simde_mm_xor_si128(c2[9159],simde_mm_xor_si128(c2[3534],simde_mm_xor_si128(c2[2831],simde_mm_xor_si128(c2[10603],simde_mm_xor_si128(c2[44],simde_mm_xor_si128(c2[7086],simde_mm_xor_si128(c2[2861],simde_mm_xor_si128(c2[4268],simde_mm_xor_si128(c2[8517],simde_mm_xor_si128(c2[9221],simde_mm_xor_si128(c2[7110],simde_mm_xor_si128(c2[2883],simde_mm_xor_si128(c2[6433],simde_mm_xor_si128(c2[7137],simde_mm_xor_si128(c2[111],simde_mm_xor_si128(c2[10668],simde_mm_xor_si128(c2[5071],simde_mm_xor_si128(c2[7179],simde_mm_xor_si128(c2[10696],simde_mm_xor_si128(c2[10029],simde_mm_xor_si128(c2[10733],simde_mm_xor_si128(c2[1580],simde_mm_xor_si128(c2[175],simde_mm_xor_si128(c2[3016],simde_mm_xor_si128(c2[3720],simde_mm_xor_si128(c2[5837],simde_mm_xor_si128(c2[897],simde_mm_xor_si128(c2[6565],simde_mm_xor_si128(c2[230],simde_mm_xor_si128(c2[10095],simde_mm_xor_si128(c2[2382],simde_mm_xor_si128(c2[8706],simde_mm_xor_si128(c2[10118],simde_mm_xor_si128(c2[3819],simde_mm_xor_si128(c2[4523],simde_mm_xor_si128(c2[5922],simde_mm_xor_si128(c2[7341],simde_mm_xor_si128(c2[10885],simde_mm_xor_si128(c2[326],simde_mm_xor_si128(c2[1738],simde_mm_xor_si128(c2[2432],simde_mm_xor_si128(c2[5998],simde_mm_xor_si128(c2[6702],simde_mm_xor_si128(c2[5285],simde_mm_xor_si128(c2[1069],simde_mm_xor_si128(c2[6723],simde_mm_xor_si128(c2[7427],simde_mm_xor_si128(c2[1102],simde_mm_xor_si128(c2[1095],simde_mm_xor_si128(c2[2536],simde_mm_xor_si128(c2[3240],simde_mm_xor_si128(c2[6752],simde_mm_xor_si128(c2[4645],simde_mm_xor_si128(c2[8877],simde_mm_xor_si128(c2[6094],simde_mm_xor_si128(c2[8901],simde_mm_xor_si128(c2[10308],simde_mm_xor_si128(c2[9635],simde_mm_xor_si128(c2[10339],simde_mm_xor_si128(c2[11055],simde_mm_xor_si128(c2[8933],simde_mm_xor_si128(c2[1225],simde_mm_xor_si128(c2[1929],simde_mm_xor_si128(c2[6848],simde_mm_xor_si128(c2[8262],simde_mm_xor_si128(c2[4769],simde_mm_xor_si128(c2[551],simde_mm_xor_si128(c2[5484],simde_mm_xor_si128(c2[6211],simde_mm_xor_si128(c2[6915],simde_mm_xor_si128(c2[6212],simde_mm_xor_si128(c2[11142],simde_mm_xor_si128(c2[9769],simde_mm_xor_si128(c2[10473],simde_mm_xor_si128(c2[7655],simde_mm_xor_si128(c2[4839],simde_mm_xor_si128(c2[6982],simde_mm_xor_si128(c2[7686],simde_mm_xor_si128(c2[655],simde_mm_xor_si128(c2[8384],simde_mm_xor_si128(c2[7017],simde_mm_xor_si128(c2[7721],simde_mm_xor_si128(c2[5609],c2[5608])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 38
|
|
d2[608]=simde_mm_xor_si128(c2[4938],simde_mm_xor_si128(c2[5225],simde_mm_xor_si128(c2[9481],c2[2506])));
|
|
|
|
//row: 39
|
|
d2[624]=simde_mm_xor_si128(c2[9194],simde_mm_xor_si128(c2[7848],simde_mm_xor_si128(c2[3753],c2[608])));
|
|
|
|
//row: 40
|
|
d2[640]=simde_mm_xor_si128(c2[9161],simde_mm_xor_si128(c2[6600],c2[4073]));
|
|
|
|
//row: 41
|
|
d2[656]=simde_mm_xor_si128(c2[4970],simde_mm_xor_si128(c2[9258],simde_mm_xor_si128(c2[8040],c2[5513])));
|
|
|
|
//row: 42
|
|
d2[672]=simde_mm_xor_si128(c2[2824],simde_mm_xor_si128(c2[8456],simde_mm_xor_si128(c2[2127],simde_mm_xor_si128(c2[2831],simde_mm_xor_si128(c2[1408],simde_mm_xor_si128(c2[2112],simde_mm_xor_si128(c2[3529],simde_mm_xor_si128(c2[10604],simde_mm_xor_si128(c2[5679],simde_mm_xor_si128(c2[6383],simde_mm_xor_si128(c2[1454],simde_mm_xor_si128(c2[2158],simde_mm_xor_si128(c2[8518],simde_mm_xor_si128(c2[6407],simde_mm_xor_si128(c2[1476],simde_mm_xor_si128(c2[2180],simde_mm_xor_si128(c2[6434],simde_mm_xor_si128(c2[10671],simde_mm_xor_si128(c2[9261],simde_mm_xor_si128(c2[9965],simde_mm_xor_si128(c2[4352],simde_mm_xor_si128(c2[5772],simde_mm_xor_si128(c2[6476],simde_mm_xor_si128(c2[9289],simde_mm_xor_si128(c2[9993],simde_mm_xor_si128(c2[9289],simde_mm_xor_si128(c2[10030],simde_mm_xor_si128(c2[877],simde_mm_xor_si128(c2[10031],simde_mm_xor_si128(c2[10735],simde_mm_xor_si128(c2[3017],simde_mm_xor_si128(c2[4430],simde_mm_xor_si128(c2[5134],simde_mm_xor_si128(c2[10753],simde_mm_xor_si128(c2[194],simde_mm_xor_si128(c2[5862],simde_mm_xor_si128(c2[10086],simde_mm_xor_si128(c2[10790],simde_mm_xor_si128(c2[8672],simde_mm_xor_si128(c2[9376],simde_mm_xor_si128(c2[1679],simde_mm_xor_si128(c2[7299],simde_mm_xor_si128(c2[8003],simde_mm_xor_si128(c2[8711],simde_mm_xor_si128(c2[9415],simde_mm_xor_si128(c2[3820],simde_mm_xor_si128(c2[5219],simde_mm_xor_si128(c2[5934],simde_mm_xor_si128(c2[6638],simde_mm_xor_si128(c2[10886],simde_mm_xor_si128(c2[331],simde_mm_xor_si128(c2[1035],simde_mm_xor_si128(c2[1025],simde_mm_xor_si128(c2[1729],simde_mm_xor_si128(c2[5999],simde_mm_xor_si128(c2[4582],simde_mm_xor_si128(c2[10925],simde_mm_xor_si128(c2[366],simde_mm_xor_si128(c2[6724],simde_mm_xor_si128(c2[399],simde_mm_xor_si128(c2[10951],simde_mm_xor_si128(c2[392],simde_mm_xor_si128(c2[2537],simde_mm_xor_si128(c2[5345],simde_mm_xor_si128(c2[6049],simde_mm_xor_si128(c2[3238],simde_mm_xor_si128(c2[3942],simde_mm_xor_si128(c2[5391],simde_mm_xor_si128(c2[7494],simde_mm_xor_si128(c2[8198],simde_mm_xor_si128(c2[8901],simde_mm_xor_si128(c2[9605],simde_mm_xor_si128(c2[9636],simde_mm_xor_si128(c2[10336],simde_mm_xor_si128(c2[7526],simde_mm_xor_si128(c2[8230],simde_mm_xor_si128(c2[1226],simde_mm_xor_si128(c2[6145],simde_mm_xor_si128(c2[6855],simde_mm_xor_si128(c2[7559],simde_mm_xor_si128(c2[4066],simde_mm_xor_si128(c2[10407],simde_mm_xor_si128(c2[11111],simde_mm_xor_si128(c2[4077],simde_mm_xor_si128(c2[4781],simde_mm_xor_si128(c2[6212],simde_mm_xor_si128(c2[4805],simde_mm_xor_si128(c2[5509],simde_mm_xor_si128(c2[9735],simde_mm_xor_si128(c2[10439],simde_mm_xor_si128(c2[9770],simde_mm_xor_si128(c2[6952],simde_mm_xor_si128(c2[3432],simde_mm_xor_si128(c2[4136],simde_mm_xor_si128(c2[6983],simde_mm_xor_si128(c2[10511],simde_mm_xor_si128(c2[11215],simde_mm_xor_si128(c2[6977],simde_mm_xor_si128(c2[7681],simde_mm_xor_si128(c2[7018],simde_mm_xor_si128(c2[4906],simde_mm_xor_si128(c2[4201],c2[4905]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 43
|
|
d2[688]=simde_mm_xor_si128(c2[4232],simde_mm_xor_si128(c2[9864],simde_mm_xor_si128(c2[4239],simde_mm_xor_si128(c2[2816],simde_mm_xor_si128(c2[3520],simde_mm_xor_si128(c2[749],simde_mm_xor_si128(c2[7791],simde_mm_xor_si128(c2[2862],simde_mm_xor_si128(c2[3566],simde_mm_xor_si128(c2[4969],simde_mm_xor_si128(c2[9926],simde_mm_xor_si128(c2[7815],simde_mm_xor_si128(c2[3588],simde_mm_xor_si128(c2[7842],simde_mm_xor_si128(c2[800],simde_mm_xor_si128(c2[10669],simde_mm_xor_si128(c2[110],simde_mm_xor_si128(c2[5760],simde_mm_xor_si128(c2[7884],simde_mm_xor_si128(c2[10697],simde_mm_xor_si128(c2[138],simde_mm_xor_si128(c2[175],simde_mm_xor_si128(c2[2285],simde_mm_xor_si128(c2[864],simde_mm_xor_si128(c2[4425],simde_mm_xor_si128(c2[6542],simde_mm_xor_si128(c2[898],simde_mm_xor_si128(c2[1602],simde_mm_xor_si128(c2[7270],simde_mm_xor_si128(c2[935],simde_mm_xor_si128(c2[10080],simde_mm_xor_si128(c2[10784],simde_mm_xor_si128(c2[3087],simde_mm_xor_si128(c2[9411],simde_mm_xor_si128(c2[10119],simde_mm_xor_si128(c2[10823],simde_mm_xor_si128(c2[5228],simde_mm_xor_si128(c2[6627],simde_mm_xor_si128(c2[8046],simde_mm_xor_si128(c2[1031],simde_mm_xor_si128(c2[2443],simde_mm_xor_si128(c2[2433],simde_mm_xor_si128(c2[3137],simde_mm_xor_si128(c2[7407],simde_mm_xor_si128(c2[5990],simde_mm_xor_si128(c2[1070],simde_mm_xor_si128(c2[1774],simde_mm_xor_si128(c2[8132],simde_mm_xor_si128(c2[1807],simde_mm_xor_si128(c2[1096],simde_mm_xor_si128(c2[1800],simde_mm_xor_si128(c2[3945],simde_mm_xor_si128(c2[7457],simde_mm_xor_si128(c2[4646],simde_mm_xor_si128(c2[5350],simde_mm_xor_si128(c2[6799],simde_mm_xor_si128(c2[9606],simde_mm_xor_si128(c2[10309],simde_mm_xor_si128(c2[11013],simde_mm_xor_si128(c2[11044],simde_mm_xor_si128(c2[481],simde_mm_xor_si128(c2[9638],simde_mm_xor_si128(c2[2634],simde_mm_xor_si128(c2[7553],simde_mm_xor_si128(c2[8263],simde_mm_xor_si128(c2[8967],simde_mm_xor_si128(c2[2634],simde_mm_xor_si128(c2[5474],simde_mm_xor_si128(c2[1256],simde_mm_xor_si128(c2[5485],simde_mm_xor_si128(c2[6189],simde_mm_xor_si128(c2[7620],simde_mm_xor_si128(c2[6917],simde_mm_xor_si128(c2[11143],simde_mm_xor_si128(c2[584],simde_mm_xor_si128(c2[9738],simde_mm_xor_si128(c2[11178],simde_mm_xor_si128(c2[8360],simde_mm_xor_si128(c2[5544],simde_mm_xor_si128(c2[8391],simde_mm_xor_si128(c2[1344],simde_mm_xor_si128(c2[8385],simde_mm_xor_si128(c2[9089],simde_mm_xor_si128(c2[8426],simde_mm_xor_si128(c2[6314],simde_mm_xor_si128(c2[5609],c2[6313]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 44
|
|
d2[704]=simde_mm_xor_si128(c2[4938],simde_mm_xor_si128(c2[10570],simde_mm_xor_si128(c2[4929],simde_mm_xor_si128(c2[4226],simde_mm_xor_si128(c2[7752],simde_mm_xor_si128(c2[1455],simde_mm_xor_si128(c2[8481],simde_mm_xor_si128(c2[4256],simde_mm_xor_si128(c2[10632],simde_mm_xor_si128(c2[8521],simde_mm_xor_si128(c2[4294],simde_mm_xor_si128(c2[8548],simde_mm_xor_si128(c2[1506],simde_mm_xor_si128(c2[800],simde_mm_xor_si128(c2[6466],simde_mm_xor_si128(c2[8590],simde_mm_xor_si128(c2[844],simde_mm_xor_si128(c2[865],simde_mm_xor_si128(c2[2991],simde_mm_xor_si128(c2[1570],simde_mm_xor_si128(c2[5131],simde_mm_xor_si128(c2[7232],simde_mm_xor_si128(c2[2308],simde_mm_xor_si128(c2[7976],simde_mm_xor_si128(c2[1641],simde_mm_xor_si128(c2[227],simde_mm_xor_si128(c2[9385],simde_mm_xor_si128(c2[3777],simde_mm_xor_si128(c2[10117],simde_mm_xor_si128(c2[266],simde_mm_xor_si128(c2[5934],simde_mm_xor_si128(c2[7333],simde_mm_xor_si128(c2[8736],simde_mm_xor_si128(c2[2410],simde_mm_xor_si128(c2[1737],simde_mm_xor_si128(c2[3149],simde_mm_xor_si128(c2[3843],simde_mm_xor_si128(c2[8097],simde_mm_xor_si128(c2[6696],simde_mm_xor_si128(c2[2464],simde_mm_xor_si128(c2[8838],simde_mm_xor_si128(c2[2497],simde_mm_xor_si128(c2[2506],simde_mm_xor_si128(c2[4651],simde_mm_xor_si128(c2[8163],simde_mm_xor_si128(c2[6056],simde_mm_xor_si128(c2[7489],simde_mm_xor_si128(c2[10312],simde_mm_xor_si128(c2[456],simde_mm_xor_si128(c2[487],simde_mm_xor_si128(c2[1187],simde_mm_xor_si128(c2[10344],simde_mm_xor_si128(c2[3340],simde_mm_xor_si128(c2[8259],simde_mm_xor_si128(c2[9673],simde_mm_xor_si128(c2[6180],simde_mm_xor_si128(c2[1962],simde_mm_xor_si128(c2[6895],simde_mm_xor_si128(c2[8326],simde_mm_xor_si128(c2[7623],simde_mm_xor_si128(c2[1290],simde_mm_xor_si128(c2[621],simde_mm_xor_si128(c2[9066],simde_mm_xor_si128(c2[6250],simde_mm_xor_si128(c2[9097],simde_mm_xor_si128(c2[2050],simde_mm_xor_si128(c2[9795],simde_mm_xor_si128(c2[9132],simde_mm_xor_si128(c2[7020],c2[7019])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 45
|
|
d2[720]=simde_mm_xor_si128(c2[3561],simde_mm_xor_si128(c2[5129],c2[5258]));
|
|
}
|
|
}
|
|
#endif
|