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=192, byte encoding
|
|
static inline void ldpc192_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<12; i2++) {
|
|
c2=&csimd[i2];
|
|
d2=&dsimd[i2];
|
|
|
|
//row: 0
|
|
d2[0]=simde_mm_xor_si128(c2[1591],simde_mm_xor_si128(c2[6340],simde_mm_xor_si128(c2[6864],simde_mm_xor_si128(c2[2117],simde_mm_xor_si128(c2[1609],simde_mm_xor_si128(c2[5307],simde_mm_xor_si128(c2[3725],simde_mm_xor_si128(c2[1107],simde_mm_xor_si128(c2[6388],simde_mm_xor_si128(c2[4280],simde_mm_xor_si128(c2[611],simde_mm_xor_si128(c2[4828],simde_mm_xor_si128(c2[72],simde_mm_xor_si128(c2[102],simde_mm_xor_si128(c2[6440],simde_mm_xor_si128(c2[1685],simde_mm_xor_si128(c2[2771],simde_mm_xor_si128(c2[129],simde_mm_xor_si128(c2[124],simde_mm_xor_si128(c2[4369],simde_mm_xor_si128(c2[682],simde_mm_xor_si128(c2[3840],simde_mm_xor_si128(c2[5984],simde_mm_xor_si128(c2[5980],simde_mm_xor_si128(c2[4929],simde_mm_xor_si128(c2[6008],simde_mm_xor_si128(c2[202],simde_mm_xor_si128(c2[4419],simde_mm_xor_si128(c2[7087],simde_mm_xor_si128(c2[1283],simde_mm_xor_si128(c2[8139],simde_mm_xor_si128(c2[246],simde_mm_xor_si128(c2[776],simde_mm_xor_si128(c2[2360],simde_mm_xor_si128(c2[7134],simde_mm_xor_si128(c2[3966],simde_mm_xor_si128(c2[799],simde_mm_xor_si128(c2[817],simde_mm_xor_si128(c2[3465],simde_mm_xor_si128(c2[3986],simde_mm_xor_si128(c2[2962],simde_mm_xor_si128(c2[4536],simde_mm_xor_si128(c2[2961],simde_mm_xor_si128(c2[5089],simde_mm_xor_si128(c2[4565],simde_mm_xor_si128(c2[2449],simde_mm_xor_si128(c2[4057],simde_mm_xor_si128(c2[1950],simde_mm_xor_si128(c2[3003],simde_mm_xor_si128(c2[5670],simde_mm_xor_si128(c2[1450],simde_mm_xor_si128(c2[391],simde_mm_xor_si128(c2[1471],simde_mm_xor_si128(c2[2000],simde_mm_xor_si128(c2[6750],simde_mm_xor_si128(c2[1491],simde_mm_xor_si128(c2[435],simde_mm_xor_si128(c2[8356],simde_mm_xor_si128(c2[2579],simde_mm_xor_si128(c2[6272],simde_mm_xor_si128(c2[7320],simde_mm_xor_si128(c2[5768],simde_mm_xor_si128(c2[7344],simde_mm_xor_si128(c2[4178],simde_mm_xor_si128(c2[5793],simde_mm_xor_si128(c2[511],c2[3154]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 1
|
|
d2[12]=simde_mm_xor_si128(c2[1591],simde_mm_xor_si128(c2[2119],simde_mm_xor_si128(c2[6868],simde_mm_xor_si128(c2[7392],simde_mm_xor_si128(c2[2645],simde_mm_xor_si128(c2[1609],simde_mm_xor_si128(c2[2137],simde_mm_xor_si128(c2[5835],simde_mm_xor_si128(c2[4253],simde_mm_xor_si128(c2[1107],simde_mm_xor_si128(c2[1635],simde_mm_xor_si128(c2[6916],simde_mm_xor_si128(c2[4808],simde_mm_xor_si128(c2[611],simde_mm_xor_si128(c2[1139],simde_mm_xor_si128(c2[5356],simde_mm_xor_si128(c2[600],simde_mm_xor_si128(c2[630],simde_mm_xor_si128(c2[6968],simde_mm_xor_si128(c2[2213],simde_mm_xor_si128(c2[2771],simde_mm_xor_si128(c2[3299],simde_mm_xor_si128(c2[657],simde_mm_xor_si128(c2[652],simde_mm_xor_si128(c2[4369],simde_mm_xor_si128(c2[4897],simde_mm_xor_si128(c2[1210],simde_mm_xor_si128(c2[4368],simde_mm_xor_si128(c2[6512],simde_mm_xor_si128(c2[6508],simde_mm_xor_si128(c2[5457],simde_mm_xor_si128(c2[6536],simde_mm_xor_si128(c2[730],simde_mm_xor_si128(c2[4947],simde_mm_xor_si128(c2[7087],simde_mm_xor_si128(c2[7615],simde_mm_xor_si128(c2[1811],simde_mm_xor_si128(c2[220],simde_mm_xor_si128(c2[246],simde_mm_xor_si128(c2[774],simde_mm_xor_si128(c2[1304],simde_mm_xor_si128(c2[2888],simde_mm_xor_si128(c2[7134],simde_mm_xor_si128(c2[7662],simde_mm_xor_si128(c2[4494],simde_mm_xor_si128(c2[1327],simde_mm_xor_si128(c2[817],simde_mm_xor_si128(c2[1345],simde_mm_xor_si128(c2[3993],simde_mm_xor_si128(c2[4514],simde_mm_xor_si128(c2[2962],simde_mm_xor_si128(c2[3490],simde_mm_xor_si128(c2[5064],simde_mm_xor_si128(c2[3489],simde_mm_xor_si128(c2[5617],simde_mm_xor_si128(c2[5093],simde_mm_xor_si128(c2[2977],simde_mm_xor_si128(c2[4057],simde_mm_xor_si128(c2[4585],simde_mm_xor_si128(c2[2478],simde_mm_xor_si128(c2[3531],simde_mm_xor_si128(c2[5670],simde_mm_xor_si128(c2[6198],simde_mm_xor_si128(c2[1978],simde_mm_xor_si128(c2[919],simde_mm_xor_si128(c2[1999],simde_mm_xor_si128(c2[2528],simde_mm_xor_si128(c2[7278],simde_mm_xor_si128(c2[1491],simde_mm_xor_si128(c2[2019],simde_mm_xor_si128(c2[963],simde_mm_xor_si128(c2[437],simde_mm_xor_si128(c2[2579],simde_mm_xor_si128(c2[3107],simde_mm_xor_si128(c2[6800],simde_mm_xor_si128(c2[7848],simde_mm_xor_si128(c2[5768],simde_mm_xor_si128(c2[6296],simde_mm_xor_si128(c2[7872],simde_mm_xor_si128(c2[4706],simde_mm_xor_si128(c2[5793],simde_mm_xor_si128(c2[6321],simde_mm_xor_si128(c2[1039],c2[3682])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 2
|
|
d2[24]=simde_mm_xor_si128(c2[2119],simde_mm_xor_si128(c2[6868],simde_mm_xor_si128(c2[6864],simde_mm_xor_si128(c2[7392],simde_mm_xor_si128(c2[2117],simde_mm_xor_si128(c2[2645],simde_mm_xor_si128(c2[2137],simde_mm_xor_si128(c2[5307],simde_mm_xor_si128(c2[5835],simde_mm_xor_si128(c2[3725],simde_mm_xor_si128(c2[4253],simde_mm_xor_si128(c2[1635],simde_mm_xor_si128(c2[6916],simde_mm_xor_si128(c2[4280],simde_mm_xor_si128(c2[4808],simde_mm_xor_si128(c2[1139],simde_mm_xor_si128(c2[5356],simde_mm_xor_si128(c2[72],simde_mm_xor_si128(c2[600],simde_mm_xor_si128(c2[630],simde_mm_xor_si128(c2[6440],simde_mm_xor_si128(c2[6968],simde_mm_xor_si128(c2[1685],simde_mm_xor_si128(c2[2213],simde_mm_xor_si128(c2[3299],simde_mm_xor_si128(c2[657],simde_mm_xor_si128(c2[124],simde_mm_xor_si128(c2[652],simde_mm_xor_si128(c2[4897],simde_mm_xor_si128(c2[682],simde_mm_xor_si128(c2[1210],simde_mm_xor_si128(c2[3840],simde_mm_xor_si128(c2[4368],simde_mm_xor_si128(c2[6512],simde_mm_xor_si128(c2[5980],simde_mm_xor_si128(c2[6508],simde_mm_xor_si128(c2[4929],simde_mm_xor_si128(c2[5457],simde_mm_xor_si128(c2[6536],simde_mm_xor_si128(c2[202],simde_mm_xor_si128(c2[730],simde_mm_xor_si128(c2[4419],simde_mm_xor_si128(c2[4947],simde_mm_xor_si128(c2[7615],simde_mm_xor_si128(c2[1811],simde_mm_xor_si128(c2[8139],simde_mm_xor_si128(c2[220],simde_mm_xor_si128(c2[774],simde_mm_xor_si128(c2[776],simde_mm_xor_si128(c2[1304],simde_mm_xor_si128(c2[2360],simde_mm_xor_si128(c2[2888],simde_mm_xor_si128(c2[7662],simde_mm_xor_si128(c2[4494],simde_mm_xor_si128(c2[799],simde_mm_xor_si128(c2[1327],simde_mm_xor_si128(c2[1345],simde_mm_xor_si128(c2[3993],simde_mm_xor_si128(c2[3986],simde_mm_xor_si128(c2[4514],simde_mm_xor_si128(c2[3490],simde_mm_xor_si128(c2[4536],simde_mm_xor_si128(c2[5064],simde_mm_xor_si128(c2[2961],simde_mm_xor_si128(c2[3489],simde_mm_xor_si128(c2[5617],simde_mm_xor_si128(c2[4565],simde_mm_xor_si128(c2[5093],simde_mm_xor_si128(c2[2449],simde_mm_xor_si128(c2[2977],simde_mm_xor_si128(c2[4585],simde_mm_xor_si128(c2[2478],simde_mm_xor_si128(c2[3003],simde_mm_xor_si128(c2[3531],simde_mm_xor_si128(c2[6198],simde_mm_xor_si128(c2[1978],simde_mm_xor_si128(c2[391],simde_mm_xor_si128(c2[919],simde_mm_xor_si128(c2[1999],simde_mm_xor_si128(c2[2000],simde_mm_xor_si128(c2[2528],simde_mm_xor_si128(c2[6750],simde_mm_xor_si128(c2[7278],simde_mm_xor_si128(c2[2019],simde_mm_xor_si128(c2[435],simde_mm_xor_si128(c2[963],simde_mm_xor_si128(c2[8356],simde_mm_xor_si128(c2[437],simde_mm_xor_si128(c2[3107],simde_mm_xor_si128(c2[6800],simde_mm_xor_si128(c2[7320],simde_mm_xor_si128(c2[7848],simde_mm_xor_si128(c2[6296],simde_mm_xor_si128(c2[7344],simde_mm_xor_si128(c2[7872],simde_mm_xor_si128(c2[4178],simde_mm_xor_si128(c2[4706],simde_mm_xor_si128(c2[6321],simde_mm_xor_si128(c2[1039],simde_mm_xor_si128(c2[3154],c2[3682]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 3
|
|
d2[36]=simde_mm_xor_si128(c2[2119],simde_mm_xor_si128(c2[6868],simde_mm_xor_si128(c2[7392],simde_mm_xor_si128(c2[2117],simde_mm_xor_si128(c2[2645],simde_mm_xor_si128(c2[2137],simde_mm_xor_si128(c2[5835],simde_mm_xor_si128(c2[3725],simde_mm_xor_si128(c2[4253],simde_mm_xor_si128(c2[1635],simde_mm_xor_si128(c2[6916],simde_mm_xor_si128(c2[4808],simde_mm_xor_si128(c2[1139],simde_mm_xor_si128(c2[5356],simde_mm_xor_si128(c2[72],simde_mm_xor_si128(c2[600],simde_mm_xor_si128(c2[630],simde_mm_xor_si128(c2[6968],simde_mm_xor_si128(c2[1685],simde_mm_xor_si128(c2[2213],simde_mm_xor_si128(c2[3299],simde_mm_xor_si128(c2[657],simde_mm_xor_si128(c2[652],simde_mm_xor_si128(c2[4897],simde_mm_xor_si128(c2[1210],simde_mm_xor_si128(c2[3840],simde_mm_xor_si128(c2[4368],simde_mm_xor_si128(c2[6512],simde_mm_xor_si128(c2[6508],simde_mm_xor_si128(c2[4929],simde_mm_xor_si128(c2[5457],simde_mm_xor_si128(c2[6536],simde_mm_xor_si128(c2[730],simde_mm_xor_si128(c2[4419],simde_mm_xor_si128(c2[4947],simde_mm_xor_si128(c2[7615],simde_mm_xor_si128(c2[1811],simde_mm_xor_si128(c2[220],simde_mm_xor_si128(c2[774],simde_mm_xor_si128(c2[1304],simde_mm_xor_si128(c2[2360],simde_mm_xor_si128(c2[2888],simde_mm_xor_si128(c2[7662],simde_mm_xor_si128(c2[4494],simde_mm_xor_si128(c2[799],simde_mm_xor_si128(c2[1327],simde_mm_xor_si128(c2[1345],simde_mm_xor_si128(c2[3993],simde_mm_xor_si128(c2[3986],simde_mm_xor_si128(c2[4514],simde_mm_xor_si128(c2[3490],simde_mm_xor_si128(c2[5064],simde_mm_xor_si128(c2[2961],simde_mm_xor_si128(c2[3489],simde_mm_xor_si128(c2[5617],simde_mm_xor_si128(c2[5093],simde_mm_xor_si128(c2[2449],simde_mm_xor_si128(c2[2977],simde_mm_xor_si128(c2[4585],simde_mm_xor_si128(c2[2478],simde_mm_xor_si128(c2[3531],simde_mm_xor_si128(c2[6198],simde_mm_xor_si128(c2[1978],simde_mm_xor_si128(c2[391],simde_mm_xor_si128(c2[919],simde_mm_xor_si128(c2[1999],simde_mm_xor_si128(c2[2528],simde_mm_xor_si128(c2[6750],simde_mm_xor_si128(c2[7278],simde_mm_xor_si128(c2[2019],simde_mm_xor_si128(c2[963],simde_mm_xor_si128(c2[8356],simde_mm_xor_si128(c2[437],simde_mm_xor_si128(c2[3107],simde_mm_xor_si128(c2[6800],simde_mm_xor_si128(c2[7848],simde_mm_xor_si128(c2[6296],simde_mm_xor_si128(c2[7872],simde_mm_xor_si128(c2[4178],simde_mm_xor_si128(c2[4706],simde_mm_xor_si128(c2[6321],simde_mm_xor_si128(c2[1039],simde_mm_xor_si128(c2[3154],c2[3682])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 4
|
|
d2[48]=simde_mm_xor_si128(c2[6344],c2[2675]);
|
|
|
|
//row: 5
|
|
d2[60]=simde_mm_xor_si128(c2[5],simde_mm_xor_si128(c2[4754],simde_mm_xor_si128(c2[5290],simde_mm_xor_si128(c2[531],simde_mm_xor_si128(c2[1584],simde_mm_xor_si128(c2[35],simde_mm_xor_si128(c2[3721],simde_mm_xor_si128(c2[2139],simde_mm_xor_si128(c2[7416],simde_mm_xor_si128(c2[7968],simde_mm_xor_si128(c2[4802],simde_mm_xor_si128(c2[2694],simde_mm_xor_si128(c2[7472],simde_mm_xor_si128(c2[3242],simde_mm_xor_si128(c2[6945],simde_mm_xor_si128(c2[1663],simde_mm_xor_si128(c2[6963],simde_mm_xor_si128(c2[4854],simde_mm_xor_si128(c2[99],simde_mm_xor_si128(c2[1185],simde_mm_xor_si128(c2[6990],simde_mm_xor_si128(c2[6985],simde_mm_xor_si128(c2[2795],simde_mm_xor_si128(c2[7543],simde_mm_xor_si128(c2[2266],simde_mm_xor_si128(c2[4398],simde_mm_xor_si128(c2[4394],simde_mm_xor_si128(c2[3343],simde_mm_xor_si128(c2[4422],simde_mm_xor_si128(c2[7063],simde_mm_xor_si128(c2[2833],simde_mm_xor_si128(c2[5501],simde_mm_xor_si128(c2[8144],simde_mm_xor_si128(c2[6553],simde_mm_xor_si128(c2[7107],simde_mm_xor_si128(c2[7637],simde_mm_xor_si128(c2[774],simde_mm_xor_si128(c2[5548],simde_mm_xor_si128(c2[2380],simde_mm_xor_si128(c2[7660],simde_mm_xor_si128(c2[7690],simde_mm_xor_si128(c2[1879],simde_mm_xor_si128(c2[2400],simde_mm_xor_si128(c2[3466],simde_mm_xor_si128(c2[1376],simde_mm_xor_si128(c2[2962],simde_mm_xor_si128(c2[1375],simde_mm_xor_si128(c2[3515],simde_mm_xor_si128(c2[2979],simde_mm_xor_si128(c2[875],simde_mm_xor_si128(c2[2483],simde_mm_xor_si128(c2[364],simde_mm_xor_si128(c2[1417],simde_mm_xor_si128(c2[4084],simde_mm_xor_si128(c2[8311],simde_mm_xor_si128(c2[7252],simde_mm_xor_si128(c2[915],simde_mm_xor_si128(c2[8332],simde_mm_xor_si128(c2[414],simde_mm_xor_si128(c2[5164],simde_mm_xor_si128(c2[8352],simde_mm_xor_si128(c2[7296],simde_mm_xor_si128(c2[6770],simde_mm_xor_si128(c2[993],simde_mm_xor_si128(c2[4686],simde_mm_xor_si128(c2[5746],simde_mm_xor_si128(c2[4182],simde_mm_xor_si128(c2[5770],simde_mm_xor_si128(c2[2592],simde_mm_xor_si128(c2[4207],simde_mm_xor_si128(c2[7372],simde_mm_xor_si128(c2[1568],c2[2091]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 6
|
|
d2[72]=simde_mm_xor_si128(c2[3173],simde_mm_xor_si128(c2[676],simde_mm_xor_si128(c2[768],simde_mm_xor_si128(c2[8193],simde_mm_xor_si128(c2[6653],simde_mm_xor_si128(c2[7275],simde_mm_xor_si128(c2[1489],c2[1010])))))));
|
|
|
|
//row: 7
|
|
d2[84]=simde_mm_xor_si128(c2[4752],simde_mm_xor_si128(c2[7419],simde_mm_xor_si128(c2[6439],simde_mm_xor_si128(c2[7040],simde_mm_xor_si128(c2[1254],c2[1399])))));
|
|
|
|
//row: 8
|
|
d2[96]=simde_mm_xor_si128(c2[3179],simde_mm_xor_si128(c2[7397],simde_mm_xor_si128(c2[7928],simde_mm_xor_si128(c2[3699],simde_mm_xor_si128(c2[5],simde_mm_xor_si128(c2[3707],simde_mm_xor_si128(c2[4235],simde_mm_xor_si128(c2[3705],simde_mm_xor_si128(c2[7395],simde_mm_xor_si128(c2[7923],simde_mm_xor_si128(c2[1591],simde_mm_xor_si128(c2[3197],simde_mm_xor_si128(c2[7427],simde_mm_xor_si128(c2[6895],simde_mm_xor_si128(c2[2138],simde_mm_xor_si128(c2[2666],simde_mm_xor_si128(c2[5313],simde_mm_xor_si128(c2[556],simde_mm_xor_si128(c2[1084],simde_mm_xor_si128(c2[1619],simde_mm_xor_si128(c2[2695],simde_mm_xor_si128(c2[6913],simde_mm_xor_si128(c2[7976],simde_mm_xor_si128(c2[3747],simde_mm_xor_si128(c2[5856],simde_mm_xor_si128(c2[1111],simde_mm_xor_si128(c2[1639],simde_mm_xor_si128(c2[2187],simde_mm_xor_si128(c2[6417],simde_mm_xor_si128(c2[6416],simde_mm_xor_si128(c2[2187],simde_mm_xor_si128(c2[1660],simde_mm_xor_si128(c2[5362],simde_mm_xor_si128(c2[5890],simde_mm_xor_si128(c2[2722],simde_mm_xor_si128(c2[1690],simde_mm_xor_si128(c2[5908],simde_mm_xor_si128(c2[8016],simde_mm_xor_si128(c2[3271],simde_mm_xor_si128(c2[3799],simde_mm_xor_si128(c2[3273],simde_mm_xor_si128(c2[6963],simde_mm_xor_si128(c2[7491],simde_mm_xor_si128(c2[4347],simde_mm_xor_si128(c2[130],simde_mm_xor_si128(c2[1705],simde_mm_xor_si128(c2[5935],simde_mm_xor_si128(c2[1712],simde_mm_xor_si128(c2[5402],simde_mm_xor_si128(c2[5930],simde_mm_xor_si128(c2[5957],simde_mm_xor_si128(c2[1728],simde_mm_xor_si128(c2[2258],simde_mm_xor_si128(c2[5960],simde_mm_xor_si128(c2[6488],simde_mm_xor_si128(c2[5428],simde_mm_xor_si128(c2[683],simde_mm_xor_si128(c2[1211],simde_mm_xor_si128(c2[7560],simde_mm_xor_si128(c2[3343],simde_mm_xor_si128(c2[7568],simde_mm_xor_si128(c2[2811],simde_mm_xor_si128(c2[3339],simde_mm_xor_si128(c2[6505],simde_mm_xor_si128(c2[1760],simde_mm_xor_si128(c2[2288],simde_mm_xor_si128(c2[7584],simde_mm_xor_si128(c2[3367],simde_mm_xor_si128(c2[1778],simde_mm_xor_si128(c2[5480],simde_mm_xor_si128(c2[6008],simde_mm_xor_si128(c2[6007],simde_mm_xor_si128(c2[1250],simde_mm_xor_si128(c2[1778],simde_mm_xor_si128(c2[216],simde_mm_xor_si128(c2[4446],simde_mm_xor_si128(c2[2859],simde_mm_xor_si128(c2[7089],simde_mm_xor_si128(c2[1280],simde_mm_xor_si128(c2[4970],simde_mm_xor_si128(c2[5498],simde_mm_xor_si128(c2[1834],simde_mm_xor_si128(c2[6052],simde_mm_xor_si128(c2[2352],simde_mm_xor_si128(c2[6054],simde_mm_xor_si128(c2[6582],simde_mm_xor_si128(c2[3936],simde_mm_xor_si128(c2[7638],simde_mm_xor_si128(c2[8166],simde_mm_xor_si128(c2[275],simde_mm_xor_si128(c2[4493],simde_mm_xor_si128(c2[5554],simde_mm_xor_si128(c2[1325],simde_mm_xor_si128(c2[2387],simde_mm_xor_si128(c2[6077],simde_mm_xor_si128(c2[6605],simde_mm_xor_si128(c2[2405],simde_mm_xor_si128(c2[6635],simde_mm_xor_si128(c2[5041],simde_mm_xor_si128(c2[824],simde_mm_xor_si128(c2[5574],simde_mm_xor_si128(c2[817],simde_mm_xor_si128(c2[1345],simde_mm_xor_si128(c2[1345],simde_mm_xor_si128(c2[4538],simde_mm_xor_si128(c2[321],simde_mm_xor_si128(c2[6124],simde_mm_xor_si128(c2[1379],simde_mm_xor_si128(c2[1907],simde_mm_xor_si128(c2[4537],simde_mm_xor_si128(c2[8239],simde_mm_xor_si128(c2[320],simde_mm_xor_si128(c2[6677],simde_mm_xor_si128(c2[2448],simde_mm_xor_si128(c2[6153],simde_mm_xor_si128(c2[1396],simde_mm_xor_si128(c2[1924],simde_mm_xor_si128(c2[4037],simde_mm_xor_si128(c2[7739],simde_mm_xor_si128(c2[8267],simde_mm_xor_si128(c2[5645],simde_mm_xor_si128(c2[1416],simde_mm_xor_si128(c2[3538],simde_mm_xor_si128(c2[7756],simde_mm_xor_si128(c2[4591],simde_mm_xor_si128(c2[8281],simde_mm_xor_si128(c2[362],simde_mm_xor_si128(c2[7258],simde_mm_xor_si128(c2[3029],simde_mm_xor_si128(c2[3026],simde_mm_xor_si128(c2[7256],simde_mm_xor_si128(c2[1979],simde_mm_xor_si128(c2[5669],simde_mm_xor_si128(c2[6197],simde_mm_xor_si128(c2[4082],simde_mm_xor_si128(c2[3059],simde_mm_xor_si128(c2[7277],simde_mm_xor_si128(c2[3576],simde_mm_xor_si128(c2[7278],simde_mm_xor_si128(c2[7806],simde_mm_xor_si128(c2[8338],simde_mm_xor_si128(c2[3581],simde_mm_xor_si128(c2[4109],simde_mm_xor_si128(c2[3079],simde_mm_xor_si128(c2[7297],simde_mm_xor_si128(c2[2023],simde_mm_xor_si128(c2[5713],simde_mm_xor_si128(c2[6241],simde_mm_xor_si128(c2[1497],simde_mm_xor_si128(c2[5187],simde_mm_xor_si128(c2[5715],simde_mm_xor_si128(c2[4155],simde_mm_xor_si128(c2[8385],simde_mm_xor_si128(c2[7848],simde_mm_xor_si128(c2[3631],simde_mm_xor_si128(c2[461],simde_mm_xor_si128(c2[4163],simde_mm_xor_si128(c2[4691],simde_mm_xor_si128(c2[458],simde_mm_xor_si128(c2[7344],simde_mm_xor_si128(c2[3127],simde_mm_xor_si128(c2[485],simde_mm_xor_si128(c2[4187],simde_mm_xor_si128(c2[4715],simde_mm_xor_si128(c2[5766],simde_mm_xor_si128(c2[1009],simde_mm_xor_si128(c2[1537],simde_mm_xor_si128(c2[7369],simde_mm_xor_si128(c2[3152],simde_mm_xor_si128(c2[2099],simde_mm_xor_si128(c2[6317],simde_mm_xor_si128(c2[4730],simde_mm_xor_si128(c2[8432],simde_mm_xor_si128(c2[513],c2[515]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 9
|
|
d2[108]=simde_mm_xor_si128(c2[7402],simde_mm_xor_si128(c2[4250],simde_mm_xor_si128(c2[776],simde_mm_xor_si128(c2[2912],simde_mm_xor_si128(c2[5067],simde_mm_xor_si128(c2[8334],simde_mm_xor_si128(c2[8355],c2[1541])))))));
|
|
|
|
//row: 10
|
|
d2[120]=simde_mm_xor_si128(c2[2670],simde_mm_xor_si128(c2[1641],simde_mm_xor_si128(c2[1157],simde_mm_xor_si128(c2[8094],simde_mm_xor_si128(c2[8123],c2[1402])))));
|
|
|
|
//row: 11
|
|
d2[132]=simde_mm_xor_si128(c2[532],simde_mm_xor_si128(c2[3170],simde_mm_xor_si128(c2[3698],simde_mm_xor_si128(c2[5281],simde_mm_xor_si128(c2[0],simde_mm_xor_si128(c2[5817],simde_mm_xor_si128(c2[536],simde_mm_xor_si128(c2[1058],simde_mm_xor_si128(c2[4224],simde_mm_xor_si128(c2[3],simde_mm_xor_si128(c2[562],simde_mm_xor_si128(c2[3200],simde_mm_xor_si128(c2[3728],simde_mm_xor_si128(c2[4248],simde_mm_xor_si128(c2[7426],simde_mm_xor_si128(c2[2666],simde_mm_xor_si128(c2[5832],simde_mm_xor_si128(c2[3198],simde_mm_xor_si128(c2[48],simde_mm_xor_si128(c2[2698],simde_mm_xor_si128(c2[3226],simde_mm_xor_si128(c2[5329],simde_mm_xor_si128(c2[48],simde_mm_xor_si128(c2[3221],simde_mm_xor_si128(c2[6387],simde_mm_xor_si128(c2[7999],simde_mm_xor_si128(c2[2190],simde_mm_xor_si128(c2[2718],simde_mm_xor_si128(c2[3769],simde_mm_xor_si128(c2[6947],simde_mm_xor_si128(c2[7472],simde_mm_xor_si128(c2[2191],simde_mm_xor_si128(c2[7490],simde_mm_xor_si128(c2[2209],simde_mm_xor_si128(c2[5381],simde_mm_xor_si128(c2[100],simde_mm_xor_si128(c2[626],simde_mm_xor_si128(c2[3792],simde_mm_xor_si128(c2[1712],simde_mm_xor_si128(c2[4350],simde_mm_xor_si128(c2[4878],simde_mm_xor_si128(c2[7517],simde_mm_xor_si128(c2[2236],simde_mm_xor_si128(c2[7512],simde_mm_xor_si128(c2[2243],simde_mm_xor_si128(c2[3322],simde_mm_xor_si128(c2[5960],simde_mm_xor_si128(c2[6488],simde_mm_xor_si128(c2[8070],simde_mm_xor_si128(c2[2789],simde_mm_xor_si128(c2[2793],simde_mm_xor_si128(c2[5959],simde_mm_xor_si128(c2[4925],simde_mm_xor_si128(c2[8091],simde_mm_xor_si128(c2[4921],simde_mm_xor_si128(c2[8099],simde_mm_xor_si128(c2[3870],simde_mm_xor_si128(c2[7036],simde_mm_xor_si128(c2[4949],simde_mm_xor_si128(c2[8115],simde_mm_xor_si128(c2[7590],simde_mm_xor_si128(c2[2309],simde_mm_xor_si128(c2[3360],simde_mm_xor_si128(c2[6538],simde_mm_xor_si128(c2[6028],simde_mm_xor_si128(c2[219],simde_mm_xor_si128(c2[747],simde_mm_xor_si128(c2[224],simde_mm_xor_si128(c2[3390],simde_mm_xor_si128(c2[7080],simde_mm_xor_si128(c2[1811],simde_mm_xor_si128(c2[7634],simde_mm_xor_si128(c2[1825],simde_mm_xor_si128(c2[2353],simde_mm_xor_si128(c2[8164],simde_mm_xor_si128(c2[2883],simde_mm_xor_si128(c2[1301],simde_mm_xor_si128(c2[4467],simde_mm_xor_si128(c2[6075],simde_mm_xor_si128(c2[266],simde_mm_xor_si128(c2[794],simde_mm_xor_si128(c2[2907],simde_mm_xor_si128(c2[6073],simde_mm_xor_si128(c2[8187],simde_mm_xor_si128(c2[2906],simde_mm_xor_si128(c2[8217],simde_mm_xor_si128(c2[2408],simde_mm_xor_si128(c2[2936],simde_mm_xor_si128(c2[2406],simde_mm_xor_si128(c2[5572],simde_mm_xor_si128(c2[2939],simde_mm_xor_si128(c2[6105],simde_mm_xor_si128(c2[4512],simde_mm_xor_si128(c2[1903],simde_mm_xor_si128(c2[4541],simde_mm_xor_si128(c2[5069],simde_mm_xor_si128(c2[3489],simde_mm_xor_si128(c2[6655],simde_mm_xor_si128(c2[1902],simde_mm_xor_si128(c2[5068],simde_mm_xor_si128(c2[4042],simde_mm_xor_si128(c2[7208],simde_mm_xor_si128(c2[3506],simde_mm_xor_si128(c2[6672],simde_mm_xor_si128(c2[1402],simde_mm_xor_si128(c2[4568],simde_mm_xor_si128(c2[3010],simde_mm_xor_si128(c2[5648],simde_mm_xor_si128(c2[6176],simde_mm_xor_si128(c2[891],simde_mm_xor_si128(c2[4057],simde_mm_xor_si128(c2[1944],simde_mm_xor_si128(c2[5122],simde_mm_xor_si128(c2[4611],simde_mm_xor_si128(c2[7249],simde_mm_xor_si128(c2[7777],simde_mm_xor_si128(c2[391],simde_mm_xor_si128(c2[3557],simde_mm_xor_si128(c2[7779],simde_mm_xor_si128(c2[2498],simde_mm_xor_si128(c2[8306],simde_mm_xor_si128(c2[412],simde_mm_xor_si128(c2[3578],simde_mm_xor_si128(c2[941],simde_mm_xor_si128(c2[4107],simde_mm_xor_si128(c2[5691],simde_mm_xor_si128(c2[410],simde_mm_xor_si128(c2[432],simde_mm_xor_si128(c2[3082],simde_mm_xor_si128(c2[3610],simde_mm_xor_si128(c2[7835],simde_mm_xor_si128(c2[2554],simde_mm_xor_si128(c2[7297],simde_mm_xor_si128(c2[2016],simde_mm_xor_si128(c2[1520],simde_mm_xor_si128(c2[4158],simde_mm_xor_si128(c2[4686],simde_mm_xor_si128(c2[5213],simde_mm_xor_si128(c2[8379],simde_mm_xor_si128(c2[6273],simde_mm_xor_si128(c2[992],simde_mm_xor_si128(c2[4709],simde_mm_xor_si128(c2[7347],simde_mm_xor_si128(c2[7875],simde_mm_xor_si128(c2[6297],simde_mm_xor_si128(c2[1016],simde_mm_xor_si128(c2[3131],simde_mm_xor_si128(c2[6297],simde_mm_xor_si128(c2[4734],simde_mm_xor_si128(c2[7372],simde_mm_xor_si128(c2[7900],simde_mm_xor_si128(c2[7899],simde_mm_xor_si128(c2[2618],simde_mm_xor_si128(c2[2095],simde_mm_xor_si128(c2[5261],c2[6851])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 12
|
|
d2[144]=simde_mm_xor_si128(c2[6868],simde_mm_xor_si128(c2[5315],simde_mm_xor_si128(c2[7642],simde_mm_xor_si128(c2[4490],simde_mm_xor_si128(c2[1371],c2[5716])))));
|
|
|
|
//row: 13
|
|
d2[156]=simde_mm_xor_si128(c2[2646],simde_mm_xor_si128(c2[3174],simde_mm_xor_si128(c2[7923],simde_mm_xor_si128(c2[0],simde_mm_xor_si128(c2[3700],simde_mm_xor_si128(c2[4759],simde_mm_xor_si128(c2[2664],simde_mm_xor_si128(c2[3192],simde_mm_xor_si128(c2[6890],simde_mm_xor_si128(c2[5308],simde_mm_xor_si128(c2[2162],simde_mm_xor_si128(c2[2690],simde_mm_xor_si128(c2[7971],simde_mm_xor_si128(c2[5863],simde_mm_xor_si128(c2[1666],simde_mm_xor_si128(c2[2194],simde_mm_xor_si128(c2[6411],simde_mm_xor_si128(c2[1667],simde_mm_xor_si128(c2[611],simde_mm_xor_si128(c2[1685],simde_mm_xor_si128(c2[8023],simde_mm_xor_si128(c2[3268],simde_mm_xor_si128(c2[3826],simde_mm_xor_si128(c2[4354],simde_mm_xor_si128(c2[1712],simde_mm_xor_si128(c2[1707],simde_mm_xor_si128(c2[5424],simde_mm_xor_si128(c2[5952],simde_mm_xor_si128(c2[2265],simde_mm_xor_si128(c2[5435],simde_mm_xor_si128(c2[7567],simde_mm_xor_si128(c2[7563],simde_mm_xor_si128(c2[6512],simde_mm_xor_si128(c2[5452],simde_mm_xor_si128(c2[7591],simde_mm_xor_si128(c2[1785],simde_mm_xor_si128(c2[6002],simde_mm_xor_si128(c2[8142],simde_mm_xor_si128(c2[223],simde_mm_xor_si128(c2[2866],simde_mm_xor_si128(c2[1275],simde_mm_xor_si128(c2[1301],simde_mm_xor_si128(c2[1829],simde_mm_xor_si128(c2[2359],simde_mm_xor_si128(c2[3943],simde_mm_xor_si128(c2[8189],simde_mm_xor_si128(c2[270],simde_mm_xor_si128(c2[5549],simde_mm_xor_si128(c2[2382],simde_mm_xor_si128(c2[1872],simde_mm_xor_si128(c2[2400],simde_mm_xor_si128(c2[5048],simde_mm_xor_si128(c2[5569],simde_mm_xor_si128(c2[4017],simde_mm_xor_si128(c2[4545],simde_mm_xor_si128(c2[6131],simde_mm_xor_si128(c2[4544],simde_mm_xor_si128(c2[6672],simde_mm_xor_si128(c2[6148],simde_mm_xor_si128(c2[4032],simde_mm_xor_si128(c2[5112],simde_mm_xor_si128(c2[5640],simde_mm_xor_si128(c2[3533],simde_mm_xor_si128(c2[4586],simde_mm_xor_si128(c2[6725],simde_mm_xor_si128(c2[7253],simde_mm_xor_si128(c2[3033],simde_mm_xor_si128(c2[1974],simde_mm_xor_si128(c2[3054],simde_mm_xor_si128(c2[3583],simde_mm_xor_si128(c2[8333],simde_mm_xor_si128(c2[2546],simde_mm_xor_si128(c2[3074],simde_mm_xor_si128(c2[2018],simde_mm_xor_si128(c2[1492],simde_mm_xor_si128(c2[3634],simde_mm_xor_si128(c2[4162],simde_mm_xor_si128(c2[7855],simde_mm_xor_si128(c2[456],simde_mm_xor_si128(c2[6823],simde_mm_xor_si128(c2[7351],simde_mm_xor_si128(c2[480],simde_mm_xor_si128(c2[5761],simde_mm_xor_si128(c2[2071],simde_mm_xor_si128(c2[6848],simde_mm_xor_si128(c2[7376],simde_mm_xor_si128(c2[2094],c2[4737])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 14
|
|
d2[168]=simde_mm_xor_si128(c2[7400],simde_mm_xor_si128(c2[4515],simde_mm_xor_si128(c2[5120],simde_mm_xor_si128(c2[5141],simde_mm_xor_si128(c2[6225],c2[6840])))));
|
|
|
|
//row: 15
|
|
d2[180]=simde_mm_xor_si128(c2[528],simde_mm_xor_si128(c2[5289],simde_mm_xor_si128(c2[5813],simde_mm_xor_si128(c2[538],simde_mm_xor_si128(c2[1066],simde_mm_xor_si128(c2[531],simde_mm_xor_si128(c2[558],simde_mm_xor_si128(c2[4256],simde_mm_xor_si128(c2[2146],simde_mm_xor_si128(c2[2674],simde_mm_xor_si128(c2[1080],simde_mm_xor_si128(c2[56],simde_mm_xor_si128(c2[5337],simde_mm_xor_si128(c2[3217],simde_mm_xor_si128(c2[7995],simde_mm_xor_si128(c2[3777],simde_mm_xor_si128(c2[6940],simde_mm_xor_si128(c2[7468],simde_mm_xor_si128(c2[7498],simde_mm_xor_si128(c2[5377],simde_mm_xor_si128(c2[106],simde_mm_xor_si128(c2[634],simde_mm_xor_si128(c2[1708],simde_mm_xor_si128(c2[7513],simde_mm_xor_si128(c2[7520],simde_mm_xor_si128(c2[3318],simde_mm_xor_si128(c2[8066],simde_mm_xor_si128(c2[2261],simde_mm_xor_si128(c2[2789],simde_mm_xor_si128(c2[4921],simde_mm_xor_si128(c2[4929],simde_mm_xor_si128(c2[3338],simde_mm_xor_si128(c2[3866],simde_mm_xor_si128(c2[4945],simde_mm_xor_si128(c2[7586],simde_mm_xor_si128(c2[2840],simde_mm_xor_si128(c2[3368],simde_mm_xor_si128(c2[6024],simde_mm_xor_si128(c2[220],simde_mm_xor_si128(c2[7088],simde_mm_xor_si128(c2[7642],simde_mm_xor_si128(c2[8160],simde_mm_xor_si128(c2[769],simde_mm_xor_si128(c2[1297],simde_mm_xor_si128(c2[1297],simde_mm_xor_si128(c2[6083],simde_mm_xor_si128(c2[2915],simde_mm_xor_si128(c2[7667],simde_mm_xor_si128(c2[8195],simde_mm_xor_si128(c2[8213],simde_mm_xor_si128(c2[2402],simde_mm_xor_si128(c2[2407],simde_mm_xor_si128(c2[2935],simde_mm_xor_si128(c2[1899],simde_mm_xor_si128(c2[3485],simde_mm_xor_si128(c2[1370],simde_mm_xor_si128(c2[1898],simde_mm_xor_si128(c2[7711],simde_mm_xor_si128(c2[4038],simde_mm_xor_si128(c2[3514],simde_mm_xor_si128(c2[870],simde_mm_xor_si128(c2[1398],simde_mm_xor_si128(c2[3006],simde_mm_xor_si128(c2[899],simde_mm_xor_si128(c2[1952],simde_mm_xor_si128(c2[4619],simde_mm_xor_si128(c2[387],simde_mm_xor_si128(c2[7259],simde_mm_xor_si128(c2[7787],simde_mm_xor_si128(c2[408],simde_mm_xor_si128(c2[937],simde_mm_xor_si128(c2[5171],simde_mm_xor_si128(c2[5699],simde_mm_xor_si128(c2[440],simde_mm_xor_si128(c2[7831],simde_mm_xor_si128(c2[6777],simde_mm_xor_si128(c2[7305],simde_mm_xor_si128(c2[4131],simde_mm_xor_si128(c2[1516],simde_mm_xor_si128(c2[5209],simde_mm_xor_si128(c2[6269],simde_mm_xor_si128(c2[4705],simde_mm_xor_si128(c2[6293],simde_mm_xor_si128(c2[2599],simde_mm_xor_si128(c2[3127],simde_mm_xor_si128(c2[4730],simde_mm_xor_si128(c2[7907],simde_mm_xor_si128(c2[1563],c2[2091]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 16
|
|
d2[192]=simde_mm_xor_si128(c2[6346],simde_mm_xor_si128(c2[2648],simde_mm_xor_si128(c2[3172],simde_mm_xor_si128(c2[6872],simde_mm_xor_si128(c2[6364],simde_mm_xor_si128(c2[1615],simde_mm_xor_si128(c2[33],simde_mm_xor_si128(c2[6888],simde_mm_xor_si128(c2[5862],simde_mm_xor_si128(c2[2696],simde_mm_xor_si128(c2[576],simde_mm_xor_si128(c2[5354],simde_mm_xor_si128(c2[1136],simde_mm_xor_si128(c2[4827],simde_mm_xor_si128(c2[1137],simde_mm_xor_si128(c2[4857],simde_mm_xor_si128(c2[2736],simde_mm_xor_si128(c2[6440],simde_mm_xor_si128(c2[7514],simde_mm_xor_si128(c2[4872],simde_mm_xor_si128(c2[4879],simde_mm_xor_si128(c2[677],simde_mm_xor_si128(c2[5425],simde_mm_xor_si128(c2[148],simde_mm_xor_si128(c2[2280],simde_mm_xor_si128(c2[2288],simde_mm_xor_si128(c2[1225],simde_mm_xor_si128(c2[2304],simde_mm_xor_si128(c2[4945],simde_mm_xor_si128(c2[727],simde_mm_xor_si128(c2[3395],simde_mm_xor_si128(c2[6026],simde_mm_xor_si128(c2[4447],simde_mm_xor_si128(c2[5001],simde_mm_xor_si128(c2[5531],simde_mm_xor_si128(c2[7115],simde_mm_xor_si128(c2[3442],simde_mm_xor_si128(c2[274],simde_mm_xor_si128(c2[5554],simde_mm_xor_si128(c2[5019],simde_mm_xor_si128(c2[5572],simde_mm_xor_si128(c2[8208],simde_mm_xor_si128(c2[294],simde_mm_xor_si128(c2[7705],simde_mm_xor_si128(c2[844],simde_mm_xor_si128(c2[7704],simde_mm_xor_si128(c2[1397],simde_mm_xor_si128(c2[873],simde_mm_xor_si128(c2[7204],simde_mm_xor_si128(c2[365],simde_mm_xor_si128(c2[6705],simde_mm_xor_si128(c2[7758],simde_mm_xor_si128(c2[1978],simde_mm_xor_si128(c2[6193],simde_mm_xor_si128(c2[5146],simde_mm_xor_si128(c2[6226],simde_mm_xor_si128(c2[6755],simde_mm_xor_si128(c2[3058],simde_mm_xor_si128(c2[6246],simde_mm_xor_si128(c2[5190],simde_mm_xor_si128(c2[4664],simde_mm_xor_si128(c2[7322],simde_mm_xor_si128(c2[2568],simde_mm_xor_si128(c2[3628],simde_mm_xor_si128(c2[2064],simde_mm_xor_si128(c2[3652],simde_mm_xor_si128(c2[486],simde_mm_xor_si128(c2[1014],simde_mm_xor_si128(c2[2089],simde_mm_xor_si128(c2[5266],c2[7897]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 17
|
|
d2[204]=simde_mm_xor_si128(c2[2116],simde_mm_xor_si128(c2[8262],simde_mm_xor_si128(c2[917],simde_mm_xor_si128(c2[3586],c2[4211]))));
|
|
|
|
//row: 18
|
|
d2[216]=simde_mm_xor_si128(c2[1088],simde_mm_xor_si128(c2[1882],simde_mm_xor_si128(c2[4541],simde_mm_xor_si128(c2[2552],c2[2568]))));
|
|
|
|
//row: 19
|
|
d2[228]=simde_mm_xor_si128(c2[537],simde_mm_xor_si128(c2[2665],simde_mm_xor_si128(c2[4401],simde_mm_xor_si128(c2[1251],c2[2880]))));
|
|
|
|
//row: 20
|
|
d2[240]=simde_mm_xor_si128(c2[7402],simde_mm_xor_si128(c2[3704],simde_mm_xor_si128(c2[4228],simde_mm_xor_si128(c2[7928],simde_mm_xor_si128(c2[5819],simde_mm_xor_si128(c2[7420],simde_mm_xor_si128(c2[2671],simde_mm_xor_si128(c2[1089],simde_mm_xor_si128(c2[6918],simde_mm_xor_si128(c2[3752],simde_mm_xor_si128(c2[1632],simde_mm_xor_si128(c2[6410],simde_mm_xor_si128(c2[2192],simde_mm_xor_si128(c2[5883],simde_mm_xor_si128(c2[7464],simde_mm_xor_si128(c2[5913],simde_mm_xor_si128(c2[3792],simde_mm_xor_si128(c2[7496],simde_mm_xor_si128(c2[123],simde_mm_xor_si128(c2[5928],simde_mm_xor_si128(c2[5935],simde_mm_xor_si128(c2[1733],simde_mm_xor_si128(c2[6481],simde_mm_xor_si128(c2[1204],simde_mm_xor_si128(c2[3336],simde_mm_xor_si128(c2[3344],simde_mm_xor_si128(c2[2281],simde_mm_xor_si128(c2[3360],simde_mm_xor_si128(c2[6001],simde_mm_xor_si128(c2[1783],simde_mm_xor_si128(c2[4451],simde_mm_xor_si128(c2[7082],simde_mm_xor_si128(c2[5503],simde_mm_xor_si128(c2[4444],simde_mm_xor_si128(c2[6057],simde_mm_xor_si128(c2[6587],simde_mm_xor_si128(c2[8171],simde_mm_xor_si128(c2[4498],simde_mm_xor_si128(c2[1330],simde_mm_xor_si128(c2[6610],simde_mm_xor_si128(c2[2913],simde_mm_xor_si128(c2[6628],simde_mm_xor_si128(c2[817],simde_mm_xor_si128(c2[1350],simde_mm_xor_si128(c2[314],simde_mm_xor_si128(c2[1900],simde_mm_xor_si128(c2[313],simde_mm_xor_si128(c2[2453],simde_mm_xor_si128(c2[1929],simde_mm_xor_si128(c2[8260],simde_mm_xor_si128(c2[1421],simde_mm_xor_si128(c2[7761],simde_mm_xor_si128(c2[367],simde_mm_xor_si128(c2[3034],simde_mm_xor_si128(c2[7249],simde_mm_xor_si128(c2[6202],simde_mm_xor_si128(c2[7282],simde_mm_xor_si128(c2[7811],simde_mm_xor_si128(c2[4114],simde_mm_xor_si128(c2[7302],simde_mm_xor_si128(c2[6246],simde_mm_xor_si128(c2[5720],simde_mm_xor_si128(c2[8378],simde_mm_xor_si128(c2[3624],simde_mm_xor_si128(c2[4684],simde_mm_xor_si128(c2[3120],simde_mm_xor_si128(c2[4708],simde_mm_xor_si128(c2[1542],simde_mm_xor_si128(c2[3145],simde_mm_xor_si128(c2[6322],c2[506]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 21
|
|
d2[252]=simde_mm_xor_si128(c2[6888],simde_mm_xor_si128(c2[3294],simde_mm_xor_si128(c2[4616],simde_mm_xor_si128(c2[3121],c2[1038]))));
|
|
|
|
//row: 22
|
|
d2[264]=simde_mm_xor_si128(c2[7393],simde_mm_xor_si128(c2[6096],simde_mm_xor_si128(c2[5066],c2[3577])));
|
|
|
|
//row: 23
|
|
d2[276]=simde_mm_xor_si128(c2[4249],simde_mm_xor_si128(c2[4805],simde_mm_xor_si128(c2[7107],c2[6241])));
|
|
|
|
//row: 24
|
|
d2[288]=simde_mm_xor_si128(c2[6873],simde_mm_xor_si128(c2[3175],simde_mm_xor_si128(c2[3699],simde_mm_xor_si128(c2[7399],simde_mm_xor_si128(c2[5286],simde_mm_xor_si128(c2[6891],simde_mm_xor_si128(c2[2142],simde_mm_xor_si128(c2[560],simde_mm_xor_si128(c2[6389],simde_mm_xor_si128(c2[3223],simde_mm_xor_si128(c2[1115],simde_mm_xor_si128(c2[5881],simde_mm_xor_si128(c2[1663],simde_mm_xor_si128(c2[5354],simde_mm_xor_si128(c2[7473],simde_mm_xor_si128(c2[5384],simde_mm_xor_si128(c2[3275],simde_mm_xor_si128(c2[6967],simde_mm_xor_si128(c2[5906],simde_mm_xor_si128(c2[8041],simde_mm_xor_si128(c2[5411],simde_mm_xor_si128(c2[5406],simde_mm_xor_si128(c2[1204],simde_mm_xor_si128(c2[5952],simde_mm_xor_si128(c2[675],simde_mm_xor_si128(c2[2819],simde_mm_xor_si128(c2[2815],simde_mm_xor_si128(c2[1752],simde_mm_xor_si128(c2[2843],simde_mm_xor_si128(c2[5472],simde_mm_xor_si128(c2[1254],simde_mm_xor_si128(c2[3922],simde_mm_xor_si128(c2[6553],simde_mm_xor_si128(c2[4974],simde_mm_xor_si128(c2[5528],simde_mm_xor_si128(c2[6058],simde_mm_xor_si128(c2[7642],simde_mm_xor_si128(c2[3969],simde_mm_xor_si128(c2[801],simde_mm_xor_si128(c2[6081],simde_mm_xor_si128(c2[1857],simde_mm_xor_si128(c2[6099],simde_mm_xor_si128(c2[288],simde_mm_xor_si128(c2[821],simde_mm_xor_si128(c2[8232],simde_mm_xor_si128(c2[1371],simde_mm_xor_si128(c2[8243],simde_mm_xor_si128(c2[1924],simde_mm_xor_si128(c2[1400],simde_mm_xor_si128(c2[7731],simde_mm_xor_si128(c2[892],simde_mm_xor_si128(c2[7232],simde_mm_xor_si128(c2[8285],simde_mm_xor_si128(c2[2505],simde_mm_xor_si128(c2[6720],simde_mm_xor_si128(c2[5673],simde_mm_xor_si128(c2[6753],simde_mm_xor_si128(c2[7282],simde_mm_xor_si128(c2[3585],simde_mm_xor_si128(c2[6773],simde_mm_xor_si128(c2[5717],simde_mm_xor_si128(c2[5191],simde_mm_xor_si128(c2[7849],simde_mm_xor_si128(c2[3107],simde_mm_xor_si128(c2[4155],simde_mm_xor_si128(c2[2603],simde_mm_xor_si128(c2[4179],simde_mm_xor_si128(c2[1013],simde_mm_xor_si128(c2[2616],simde_mm_xor_si128(c2[5793],c2[8424]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 25
|
|
d2[300]=simde_mm_xor_si128(c2[4252],simde_mm_xor_si128(c2[673],simde_mm_xor_si128(c2[8099],c2[4567])));
|
|
|
|
//row: 26
|
|
d2[312]=simde_mm_xor_si128(c2[3700],simde_mm_xor_si128(c2[581],simde_mm_xor_si128(c2[6436],c2[4592])));
|
|
|
|
//row: 27
|
|
d2[324]=simde_mm_xor_si128(c2[1080],simde_mm_xor_si128(c2[1200],c2[2838]));
|
|
|
|
//row: 28
|
|
d2[336]=simde_mm_xor_si128(c2[7393],simde_mm_xor_si128(c2[1681],simde_mm_xor_si128(c2[2571],c2[1565])));
|
|
|
|
//row: 29
|
|
d2[348]=simde_mm_xor_si128(c2[1056],simde_mm_xor_si128(c2[5817],simde_mm_xor_si128(c2[6341],simde_mm_xor_si128(c2[1066],simde_mm_xor_si128(c2[1594],simde_mm_xor_si128(c2[1086],simde_mm_xor_si128(c2[4784],simde_mm_xor_si128(c2[2674],simde_mm_xor_si128(c2[3202],simde_mm_xor_si128(c2[6363],simde_mm_xor_si128(c2[584],simde_mm_xor_si128(c2[5865],simde_mm_xor_si128(c2[3745],simde_mm_xor_si128(c2[76],simde_mm_xor_si128(c2[4305],simde_mm_xor_si128(c2[7468],simde_mm_xor_si128(c2[7996],simde_mm_xor_si128(c2[8026],simde_mm_xor_si128(c2[5905],simde_mm_xor_si128(c2[634],simde_mm_xor_si128(c2[1162],simde_mm_xor_si128(c2[2236],simde_mm_xor_si128(c2[8041],simde_mm_xor_si128(c2[8048],simde_mm_xor_si128(c2[3846],simde_mm_xor_si128(c2[147],simde_mm_xor_si128(c2[2789],simde_mm_xor_si128(c2[3317],simde_mm_xor_si128(c2[5449],simde_mm_xor_si128(c2[5457],simde_mm_xor_si128(c2[3866],simde_mm_xor_si128(c2[4394],simde_mm_xor_si128(c2[5473],simde_mm_xor_si128(c2[8114],simde_mm_xor_si128(c2[3368],simde_mm_xor_si128(c2[3896],simde_mm_xor_si128(c2[6552],simde_mm_xor_si128(c2[748],simde_mm_xor_si128(c2[7616],simde_mm_xor_si128(c2[8170],simde_mm_xor_si128(c2[241],simde_mm_xor_si128(c2[1297],simde_mm_xor_si128(c2[1825],simde_mm_xor_si128(c2[6611],simde_mm_xor_si128(c2[3443],simde_mm_xor_si128(c2[8195],simde_mm_xor_si128(c2[264],simde_mm_xor_si128(c2[294],simde_mm_xor_si128(c2[2930],simde_mm_xor_si128(c2[2935],simde_mm_xor_si128(c2[3463],simde_mm_xor_si128(c2[2427],simde_mm_xor_si128(c2[4013],simde_mm_xor_si128(c2[1898],simde_mm_xor_si128(c2[2426],simde_mm_xor_si128(c2[4566],simde_mm_xor_si128(c2[4042],simde_mm_xor_si128(c2[1398],simde_mm_xor_si128(c2[1926],simde_mm_xor_si128(c2[2976],simde_mm_xor_si128(c2[3534],simde_mm_xor_si128(c2[1427],simde_mm_xor_si128(c2[2480],simde_mm_xor_si128(c2[5147],simde_mm_xor_si128(c2[915],simde_mm_xor_si128(c2[7787],simde_mm_xor_si128(c2[8315],simde_mm_xor_si128(c2[936],simde_mm_xor_si128(c2[1465],simde_mm_xor_si128(c2[5699],simde_mm_xor_si128(c2[6227],simde_mm_xor_si128(c2[968],simde_mm_xor_si128(c2[8359],simde_mm_xor_si128(c2[7305],simde_mm_xor_si128(c2[7833],simde_mm_xor_si128(c2[2025],simde_mm_xor_si128(c2[2044],simde_mm_xor_si128(c2[5737],simde_mm_xor_si128(c2[6797],simde_mm_xor_si128(c2[5233],simde_mm_xor_si128(c2[6821],simde_mm_xor_si128(c2[3127],simde_mm_xor_si128(c2[3655],simde_mm_xor_si128(c2[5258],simde_mm_xor_si128(c2[8435],simde_mm_xor_si128(c2[2091],c2[2619]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 30
|
|
d2[360]=simde_mm_xor_si128(c2[7392],simde_mm_xor_si128(c2[3706],simde_mm_xor_si128(c2[3702],simde_mm_xor_si128(c2[4230],simde_mm_xor_si128(c2[7402],simde_mm_xor_si128(c2[7930],simde_mm_xor_si128(c2[7929],simde_mm_xor_si128(c2[7422],simde_mm_xor_si128(c2[2145],simde_mm_xor_si128(c2[2673],simde_mm_xor_si128(c2[563],simde_mm_xor_si128(c2[1091],simde_mm_xor_si128(c2[6920],simde_mm_xor_si128(c2[3754],simde_mm_xor_si128(c2[1106],simde_mm_xor_si128(c2[1634],simde_mm_xor_si128(c2[6412],simde_mm_xor_si128(c2[2194],simde_mm_xor_si128(c2[5357],simde_mm_xor_si128(c2[5885],simde_mm_xor_si128(c2[5915],simde_mm_xor_si128(c2[3266],simde_mm_xor_si128(c2[3794],simde_mm_xor_si128(c2[6970],simde_mm_xor_si128(c2[7498],simde_mm_xor_si128(c2[125],simde_mm_xor_si128(c2[5930],simde_mm_xor_si128(c2[5409],simde_mm_xor_si128(c2[5937],simde_mm_xor_si128(c2[1735],simde_mm_xor_si128(c2[5955],simde_mm_xor_si128(c2[6483],simde_mm_xor_si128(c2[678],simde_mm_xor_si128(c2[1206],simde_mm_xor_si128(c2[3338],simde_mm_xor_si128(c2[2818],simde_mm_xor_si128(c2[3346],simde_mm_xor_si128(c2[1755],simde_mm_xor_si128(c2[2283],simde_mm_xor_si128(c2[3362],simde_mm_xor_si128(c2[5475],simde_mm_xor_si128(c2[6003],simde_mm_xor_si128(c2[1257],simde_mm_xor_si128(c2[1785],simde_mm_xor_si128(c2[4441],simde_mm_xor_si128(c2[7084],simde_mm_xor_si128(c2[4977],simde_mm_xor_si128(c2[5505],simde_mm_xor_si128(c2[6059],simde_mm_xor_si128(c2[6049],simde_mm_xor_si128(c2[6577],simde_mm_xor_si128(c2[7633],simde_mm_xor_si128(c2[8161],simde_mm_xor_si128(c2[2882],simde_mm_xor_si128(c2[4488],simde_mm_xor_si128(c2[1320],simde_mm_xor_si128(c2[6072],simde_mm_xor_si128(c2[6600],simde_mm_xor_si128(c2[6630],simde_mm_xor_si128(c2[819],simde_mm_xor_si128(c2[824],simde_mm_xor_si128(c2[1352],simde_mm_xor_si128(c2[316],simde_mm_xor_si128(c2[1374],simde_mm_xor_si128(c2[1902],simde_mm_xor_si128(c2[8234],simde_mm_xor_si128(c2[315],simde_mm_xor_si128(c2[2428],simde_mm_xor_si128(c2[2455],simde_mm_xor_si128(c2[1403],simde_mm_xor_si128(c2[1931],simde_mm_xor_si128(c2[7734],simde_mm_xor_si128(c2[8262],simde_mm_xor_si128(c2[1423],simde_mm_xor_si128(c2[7763],simde_mm_xor_si128(c2[8288],simde_mm_xor_si128(c2[369],simde_mm_xor_si128(c2[3024],simde_mm_xor_si128(c2[7251],simde_mm_xor_si128(c2[5664],simde_mm_xor_si128(c2[6192],simde_mm_xor_si128(c2[7272],simde_mm_xor_si128(c2[7273],simde_mm_xor_si128(c2[7801],simde_mm_xor_si128(c2[3576],simde_mm_xor_si128(c2[4104],simde_mm_xor_si128(c2[7304],simde_mm_xor_si128(c2[5720],simde_mm_xor_si128(c2[6248],simde_mm_xor_si128(c2[5194],simde_mm_xor_si128(c2[5722],simde_mm_xor_si128(c2[8380],simde_mm_xor_si128(c2[3626],simde_mm_xor_si128(c2[4158],simde_mm_xor_si128(c2[4686],simde_mm_xor_si128(c2[3122],simde_mm_xor_si128(c2[4182],simde_mm_xor_si128(c2[4710],simde_mm_xor_si128(c2[1016],simde_mm_xor_si128(c2[1544],simde_mm_xor_si128(c2[3147],simde_mm_xor_si128(c2[6312],simde_mm_xor_si128(c2[8427],c2[508])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 31
|
|
d2[372]=simde_mm_xor_si128(c2[2651],simde_mm_xor_si128(c2[2123],simde_mm_xor_si128(c2[7400],simde_mm_xor_si128(c2[6872],simde_mm_xor_si128(c2[7924],simde_mm_xor_si128(c2[7396],simde_mm_xor_si128(c2[3177],simde_mm_xor_si128(c2[2121],simde_mm_xor_si128(c2[2649],simde_mm_xor_si128(c2[2669],simde_mm_xor_si128(c2[2141],simde_mm_xor_si128(c2[6367],simde_mm_xor_si128(c2[5839],simde_mm_xor_si128(c2[4785],simde_mm_xor_si128(c2[3729],simde_mm_xor_si128(c2[4257],simde_mm_xor_si128(c2[2142],simde_mm_xor_si128(c2[2167],simde_mm_xor_si128(c2[1639],simde_mm_xor_si128(c2[7448],simde_mm_xor_si128(c2[6920],simde_mm_xor_si128(c2[5328],simde_mm_xor_si128(c2[4800],simde_mm_xor_si128(c2[1659],simde_mm_xor_si128(c2[1131],simde_mm_xor_si128(c2[5888],simde_mm_xor_si128(c2[5360],simde_mm_xor_si128(c2[1132],simde_mm_xor_si128(c2[76],simde_mm_xor_si128(c2[604],simde_mm_xor_si128(c2[1162],simde_mm_xor_si128(c2[634],simde_mm_xor_si128(c2[7488],simde_mm_xor_si128(c2[6960],simde_mm_xor_si128(c2[2745],simde_mm_xor_si128(c2[1689],simde_mm_xor_si128(c2[2217],simde_mm_xor_si128(c2[3819],simde_mm_xor_si128(c2[3291],simde_mm_xor_si128(c2[1177],simde_mm_xor_si128(c2[649],simde_mm_xor_si128(c2[1184],simde_mm_xor_si128(c2[656],simde_mm_xor_si128(c2[5429],simde_mm_xor_si128(c2[4901],simde_mm_xor_si128(c2[1730],simde_mm_xor_si128(c2[1202],simde_mm_xor_si128(c2[4900],simde_mm_xor_si128(c2[3844],simde_mm_xor_si128(c2[4372],simde_mm_xor_si128(c2[7032],simde_mm_xor_si128(c2[6504],simde_mm_xor_si128(c2[7040],simde_mm_xor_si128(c2[6512],simde_mm_xor_si128(c2[5977],simde_mm_xor_si128(c2[4921],simde_mm_xor_si128(c2[5449],simde_mm_xor_si128(c2[3865],simde_mm_xor_si128(c2[7056],simde_mm_xor_si128(c2[6528],simde_mm_xor_si128(c2[1250],simde_mm_xor_si128(c2[722],simde_mm_xor_si128(c2[5479],simde_mm_xor_si128(c2[4423],simde_mm_xor_si128(c2[4951],simde_mm_xor_si128(c2[8147],simde_mm_xor_si128(c2[7619],simde_mm_xor_si128(c2[2331],simde_mm_xor_si128(c2[1803],simde_mm_xor_si128(c2[752],simde_mm_xor_si128(c2[224],simde_mm_xor_si128(c2[1306],simde_mm_xor_si128(c2[778],simde_mm_xor_si128(c2[1824],simde_mm_xor_si128(c2[1296],simde_mm_xor_si128(c2[3408],simde_mm_xor_si128(c2[2352],simde_mm_xor_si128(c2[2880],simde_mm_xor_si128(c2[8194],simde_mm_xor_si128(c2[7666],simde_mm_xor_si128(c2[5026],simde_mm_xor_si128(c2[4498],simde_mm_xor_si128(c2[1859],simde_mm_xor_si128(c2[803],simde_mm_xor_si128(c2[1331],simde_mm_xor_si128(c2[1877],simde_mm_xor_si128(c2[1349],simde_mm_xor_si128(c2[4513],simde_mm_xor_si128(c2[3985],simde_mm_xor_si128(c2[5046],simde_mm_xor_si128(c2[3990],simde_mm_xor_si128(c2[4518],simde_mm_xor_si128(c2[4010],simde_mm_xor_si128(c2[3482],simde_mm_xor_si128(c2[5596],simde_mm_xor_si128(c2[5068],simde_mm_xor_si128(c2[4009],simde_mm_xor_si128(c2[2953],simde_mm_xor_si128(c2[3481],simde_mm_xor_si128(c2[6149],simde_mm_xor_si128(c2[5621],simde_mm_xor_si128(c2[5625],simde_mm_xor_si128(c2[5097],simde_mm_xor_si128(c2[3509],simde_mm_xor_si128(c2[2453],simde_mm_xor_si128(c2[2981],simde_mm_xor_si128(c2[5117],simde_mm_xor_si128(c2[4589],simde_mm_xor_si128(c2[3010],simde_mm_xor_si128(c2[2482],simde_mm_xor_si128(c2[4063],simde_mm_xor_si128(c2[3535],simde_mm_xor_si128(c2[6730],simde_mm_xor_si128(c2[6202],simde_mm_xor_si128(c2[2498],simde_mm_xor_si128(c2[1970],simde_mm_xor_si128(c2[1451],simde_mm_xor_si128(c2[395],simde_mm_xor_si128(c2[923],simde_mm_xor_si128(c2[2531],simde_mm_xor_si128(c2[2003],simde_mm_xor_si128(c2[3048],simde_mm_xor_si128(c2[2520],simde_mm_xor_si128(c2[7810],simde_mm_xor_si128(c2[6754],simde_mm_xor_si128(c2[7282],simde_mm_xor_si128(c2[2551],simde_mm_xor_si128(c2[2023],simde_mm_xor_si128(c2[1495],simde_mm_xor_si128(c2[967],simde_mm_xor_si128(c2[969],simde_mm_xor_si128(c2[8360],simde_mm_xor_si128(c2[441],simde_mm_xor_si128(c2[3627],simde_mm_xor_si128(c2[3099],simde_mm_xor_si128(c2[7320],simde_mm_xor_si128(c2[6792],simde_mm_xor_si128(c2[8380],simde_mm_xor_si128(c2[7852],simde_mm_xor_si128(c2[6816],simde_mm_xor_si128(c2[6288],simde_mm_xor_si128(c2[8404],simde_mm_xor_si128(c2[7876],simde_mm_xor_si128(c2[5238],simde_mm_xor_si128(c2[4182],simde_mm_xor_si128(c2[4710],simde_mm_xor_si128(c2[6841],simde_mm_xor_si128(c2[6313],simde_mm_xor_si128(c2[1571],simde_mm_xor_si128(c2[1043],simde_mm_xor_si128(c2[4202],simde_mm_xor_si128(c2[3146],c2[3674]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 32
|
|
d2[384]=simde_mm_xor_si128(c2[1585],simde_mm_xor_si128(c2[6346],simde_mm_xor_si128(c2[6342],simde_mm_xor_si128(c2[6870],simde_mm_xor_si128(c2[1595],simde_mm_xor_si128(c2[2123],simde_mm_xor_si128(c2[3174],simde_mm_xor_si128(c2[1615],simde_mm_xor_si128(c2[4785],simde_mm_xor_si128(c2[5313],simde_mm_xor_si128(c2[3203],simde_mm_xor_si128(c2[3731],simde_mm_xor_si128(c2[1113],simde_mm_xor_si128(c2[6394],simde_mm_xor_si128(c2[3746],simde_mm_xor_si128(c2[4274],simde_mm_xor_si128(c2[605],simde_mm_xor_si128(c2[4834],simde_mm_xor_si128(c2[7997],simde_mm_xor_si128(c2[78],simde_mm_xor_si128(c2[96],simde_mm_xor_si128(c2[5906],simde_mm_xor_si128(c2[6434],simde_mm_xor_si128(c2[1163],simde_mm_xor_si128(c2[1691],simde_mm_xor_si128(c2[2765],simde_mm_xor_si128(c2[123],simde_mm_xor_si128(c2[8049],simde_mm_xor_si128(c2[130],simde_mm_xor_si128(c2[4375],simde_mm_xor_si128(c2[148],simde_mm_xor_si128(c2[676],simde_mm_xor_si128(c2[3318],simde_mm_xor_si128(c2[3846],simde_mm_xor_si128(c2[5978],simde_mm_xor_si128(c2[5458],simde_mm_xor_si128(c2[5986],simde_mm_xor_si128(c2[4395],simde_mm_xor_si128(c2[4923],simde_mm_xor_si128(c2[6002],simde_mm_xor_si128(c2[8115],simde_mm_xor_si128(c2[196],simde_mm_xor_si128(c2[3897],simde_mm_xor_si128(c2[4425],simde_mm_xor_si128(c2[7081],simde_mm_xor_si128(c2[1277],simde_mm_xor_si128(c2[7617],simde_mm_xor_si128(c2[8145],simde_mm_xor_si128(c2[240],simde_mm_xor_si128(c2[242],simde_mm_xor_si128(c2[770],simde_mm_xor_si128(c2[1826],simde_mm_xor_si128(c2[2354],simde_mm_xor_si128(c2[7128],simde_mm_xor_si128(c2[3960],simde_mm_xor_si128(c2[265],simde_mm_xor_si128(c2[793],simde_mm_xor_si128(c2[823],simde_mm_xor_si128(c2[3459],simde_mm_xor_si128(c2[3464],simde_mm_xor_si128(c2[3992],simde_mm_xor_si128(c2[5040],simde_mm_xor_si128(c2[2956],simde_mm_xor_si128(c2[4014],simde_mm_xor_si128(c2[4542],simde_mm_xor_si128(c2[2427],simde_mm_xor_si128(c2[2955],simde_mm_xor_si128(c2[5095],simde_mm_xor_si128(c2[4043],simde_mm_xor_si128(c2[4571],simde_mm_xor_si128(c2[1927],simde_mm_xor_si128(c2[2455],simde_mm_xor_si128(c2[8266],simde_mm_xor_si128(c2[4063],simde_mm_xor_si128(c2[1944],simde_mm_xor_si128(c2[2481],simde_mm_xor_si128(c2[3009],simde_mm_xor_si128(c2[5664],simde_mm_xor_si128(c2[1444],simde_mm_xor_si128(c2[8304],simde_mm_xor_si128(c2[385],simde_mm_xor_si128(c2[1465],simde_mm_xor_si128(c2[1466],simde_mm_xor_si128(c2[1994],simde_mm_xor_si128(c2[6216],simde_mm_xor_si128(c2[6744],simde_mm_xor_si128(c2[1497],simde_mm_xor_si128(c2[8360],simde_mm_xor_si128(c2[441],simde_mm_xor_si128(c2[7834],simde_mm_xor_si128(c2[8362],simde_mm_xor_si128(c2[2573],simde_mm_xor_si128(c2[6266],simde_mm_xor_si128(c2[6798],simde_mm_xor_si128(c2[7326],simde_mm_xor_si128(c2[5762],simde_mm_xor_si128(c2[6822],simde_mm_xor_si128(c2[7350],simde_mm_xor_si128(c2[3656],simde_mm_xor_si128(c2[4184],simde_mm_xor_si128(c2[5787],simde_mm_xor_si128(c2[505],simde_mm_xor_si128(c2[2620],c2[3148])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 33
|
|
d2[396]=simde_mm_xor_si128(c2[1616],simde_mm_xor_si128(c2[4272],simde_mm_xor_si128(c2[5026],c2[5262])));
|
|
|
|
//row: 34
|
|
d2[408]=simde_mm_xor_si128(c2[3170],simde_mm_xor_si128(c2[2289],simde_mm_xor_si128(c2[5640],c2[4639])));
|
|
|
|
//row: 35
|
|
d2[420]=simde_mm_xor_si128(c2[6875],simde_mm_xor_si128(c2[3177],simde_mm_xor_si128(c2[3701],simde_mm_xor_si128(c2[7401],simde_mm_xor_si128(c2[6893],simde_mm_xor_si128(c2[2144],simde_mm_xor_si128(c2[562],simde_mm_xor_si128(c2[32],simde_mm_xor_si128(c2[6391],simde_mm_xor_si128(c2[3225],simde_mm_xor_si128(c2[1105],simde_mm_xor_si128(c2[5883],simde_mm_xor_si128(c2[1665],simde_mm_xor_si128(c2[5356],simde_mm_xor_si128(c2[5386],simde_mm_xor_si128(c2[3265],simde_mm_xor_si128(c2[6969],simde_mm_xor_si128(c2[8043],simde_mm_xor_si128(c2[5401],simde_mm_xor_si128(c2[5408],simde_mm_xor_si128(c2[1206],simde_mm_xor_si128(c2[5954],simde_mm_xor_si128(c2[677],simde_mm_xor_si128(c2[8072],simde_mm_xor_si128(c2[2809],simde_mm_xor_si128(c2[2817],simde_mm_xor_si128(c2[1754],simde_mm_xor_si128(c2[2833],simde_mm_xor_si128(c2[5474],simde_mm_xor_si128(c2[1256],simde_mm_xor_si128(c2[3912],simde_mm_xor_si128(c2[6555],simde_mm_xor_si128(c2[4976],simde_mm_xor_si128(c2[5530],simde_mm_xor_si128(c2[6048],simde_mm_xor_si128(c2[7632],simde_mm_xor_si128(c2[3971],simde_mm_xor_si128(c2[803],simde_mm_xor_si128(c2[6083],simde_mm_xor_si128(c2[6101],simde_mm_xor_si128(c2[290],simde_mm_xor_si128(c2[823],simde_mm_xor_si128(c2[1344],simde_mm_xor_si128(c2[8234],simde_mm_xor_si128(c2[1373],simde_mm_xor_si128(c2[8233],simde_mm_xor_si128(c2[1926],simde_mm_xor_si128(c2[1402],simde_mm_xor_si128(c2[7733],simde_mm_xor_si128(c2[894],simde_mm_xor_si128(c2[7234],simde_mm_xor_si128(c2[8287],simde_mm_xor_si128(c2[2507],simde_mm_xor_si128(c2[6722],simde_mm_xor_si128(c2[5675],simde_mm_xor_si128(c2[6755],simde_mm_xor_si128(c2[7272],simde_mm_xor_si128(c2[3587],simde_mm_xor_si128(c2[6775],simde_mm_xor_si128(c2[5719],simde_mm_xor_si128(c2[5193],simde_mm_xor_si128(c2[7851],simde_mm_xor_si128(c2[3097],simde_mm_xor_si128(c2[4157],simde_mm_xor_si128(c2[2593],simde_mm_xor_si128(c2[4181],simde_mm_xor_si128(c2[1015],simde_mm_xor_si128(c2[2618],simde_mm_xor_si128(c2[5795],c2[8426])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 36
|
|
d2[432]=simde_mm_xor_si128(c2[1057],simde_mm_xor_si128(c2[5095],simde_mm_xor_si128(c2[5118],c2[3073])));
|
|
|
|
//row: 37
|
|
d2[444]=simde_mm_xor_si128(c2[3170],simde_mm_xor_si128(c2[3698],simde_mm_xor_si128(c2[0],simde_mm_xor_si128(c2[536],simde_mm_xor_si128(c2[4224],simde_mm_xor_si128(c2[3200],simde_mm_xor_si128(c2[3728],simde_mm_xor_si128(c2[7426],simde_mm_xor_si128(c2[5832],simde_mm_xor_si128(c2[6892],simde_mm_xor_si128(c2[2698],simde_mm_xor_si128(c2[3226],simde_mm_xor_si128(c2[48],simde_mm_xor_si128(c2[6387],simde_mm_xor_si128(c2[2190],simde_mm_xor_si128(c2[2718],simde_mm_xor_si128(c2[6947],simde_mm_xor_si128(c2[2191],simde_mm_xor_si128(c2[2209],simde_mm_xor_si128(c2[100],simde_mm_xor_si128(c2[3792],simde_mm_xor_si128(c2[4350],simde_mm_xor_si128(c2[4878],simde_mm_xor_si128(c2[2236],simde_mm_xor_si128(c2[2243],simde_mm_xor_si128(c2[5960],simde_mm_xor_si128(c2[6488],simde_mm_xor_si128(c2[2789],simde_mm_xor_si128(c2[5959],simde_mm_xor_si128(c2[8091],simde_mm_xor_si128(c2[8099],simde_mm_xor_si128(c2[7036],simde_mm_xor_si128(c2[8115],simde_mm_xor_si128(c2[2309],simde_mm_xor_si128(c2[6538],simde_mm_xor_si128(c2[219],simde_mm_xor_si128(c2[747],simde_mm_xor_si128(c2[3390],simde_mm_xor_si128(c2[1811],simde_mm_xor_si128(c2[1825],simde_mm_xor_si128(c2[2353],simde_mm_xor_si128(c2[2883],simde_mm_xor_si128(c2[4467],simde_mm_xor_si128(c2[266],simde_mm_xor_si128(c2[794],simde_mm_xor_si128(c2[6073],simde_mm_xor_si128(c2[2906],simde_mm_xor_si128(c2[2408],simde_mm_xor_si128(c2[2936],simde_mm_xor_si128(c2[5572],simde_mm_xor_si128(c2[6105],simde_mm_xor_si128(c2[4541],simde_mm_xor_si128(c2[5069],simde_mm_xor_si128(c2[6655],simde_mm_xor_si128(c2[5068],simde_mm_xor_si128(c2[1373],simde_mm_xor_si128(c2[7208],simde_mm_xor_si128(c2[6672],simde_mm_xor_si128(c2[4568],simde_mm_xor_si128(c2[5648],simde_mm_xor_si128(c2[6176],simde_mm_xor_si128(c2[4057],simde_mm_xor_si128(c2[5122],simde_mm_xor_si128(c2[7249],simde_mm_xor_si128(c2[7777],simde_mm_xor_si128(c2[3557],simde_mm_xor_si128(c2[2498],simde_mm_xor_si128(c2[3578],simde_mm_xor_si128(c2[4107],simde_mm_xor_si128(c2[410],simde_mm_xor_si128(c2[3082],simde_mm_xor_si128(c2[3610],simde_mm_xor_si128(c2[2554],simde_mm_xor_si128(c2[2016],simde_mm_xor_si128(c2[4158],simde_mm_xor_si128(c2[4686],simde_mm_xor_si128(c2[8379],simde_mm_xor_si128(c2[992],simde_mm_xor_si128(c2[7347],simde_mm_xor_si128(c2[7875],simde_mm_xor_si128(c2[1016],simde_mm_xor_si128(c2[6297],simde_mm_xor_si128(c2[7372],simde_mm_xor_si128(c2[7900],simde_mm_xor_si128(c2[2618],c2[5261])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 38
|
|
d2[456]=simde_mm_xor_si128(c2[4763],simde_mm_xor_si128(c2[755],simde_mm_xor_si128(c2[774],c2[3457])));
|
|
|
|
//row: 39
|
|
d2[468]=simde_mm_xor_si128(c2[1084],simde_mm_xor_si128(c2[6941],simde_mm_xor_si128(c2[5457],c2[5214])));
|
|
|
|
//row: 40
|
|
d2[480]=simde_mm_xor_si128(c2[7930],simde_mm_xor_si128(c2[2834],c2[4639]));
|
|
|
|
//row: 41
|
|
d2[492]=simde_mm_xor_si128(c2[2139],simde_mm_xor_si128(c2[5359],simde_mm_xor_si128(c2[6032],c2[438])));
|
|
|
|
//row: 42
|
|
d2[504]=simde_mm_xor_si128(c2[7400],simde_mm_xor_si128(c2[3702],simde_mm_xor_si128(c2[3698],simde_mm_xor_si128(c2[4226],simde_mm_xor_si128(c2[7398],simde_mm_xor_si128(c2[7926],simde_mm_xor_si128(c2[535],simde_mm_xor_si128(c2[7418],simde_mm_xor_si128(c2[2141],simde_mm_xor_si128(c2[2669],simde_mm_xor_si128(c2[559],simde_mm_xor_si128(c2[1087],simde_mm_xor_si128(c2[6916],simde_mm_xor_si128(c2[3750],simde_mm_xor_si128(c2[1114],simde_mm_xor_si128(c2[1642],simde_mm_xor_si128(c2[6408],simde_mm_xor_si128(c2[2190],simde_mm_xor_si128(c2[5353],simde_mm_xor_si128(c2[5881],simde_mm_xor_si128(c2[5911],simde_mm_xor_si128(c2[3274],simde_mm_xor_si128(c2[3802],simde_mm_xor_si128(c2[6966],simde_mm_xor_si128(c2[7494],simde_mm_xor_si128(c2[7488],simde_mm_xor_si128(c2[121],simde_mm_xor_si128(c2[5938],simde_mm_xor_si128(c2[5405],simde_mm_xor_si128(c2[5933],simde_mm_xor_si128(c2[1731],simde_mm_xor_si128(c2[5963],simde_mm_xor_si128(c2[6491],simde_mm_xor_si128(c2[674],simde_mm_xor_si128(c2[1202],simde_mm_xor_si128(c2[3346],simde_mm_xor_si128(c2[2814],simde_mm_xor_si128(c2[3342],simde_mm_xor_si128(c2[1763],simde_mm_xor_si128(c2[2291],simde_mm_xor_si128(c2[3370],simde_mm_xor_si128(c2[5483],simde_mm_xor_si128(c2[6011],simde_mm_xor_si128(c2[1253],simde_mm_xor_si128(c2[1781],simde_mm_xor_si128(c2[4449],simde_mm_xor_si128(c2[7080],simde_mm_xor_si128(c2[4973],simde_mm_xor_si128(c2[5501],simde_mm_xor_si128(c2[6055],simde_mm_xor_si128(c2[6057],simde_mm_xor_si128(c2[6585],simde_mm_xor_si128(c2[7641],simde_mm_xor_si128(c2[8169],simde_mm_xor_si128(c2[4496],simde_mm_xor_si128(c2[1328],simde_mm_xor_si128(c2[6080],simde_mm_xor_si128(c2[6608],simde_mm_xor_si128(c2[6626],simde_mm_xor_si128(c2[827],simde_mm_xor_si128(c2[820],simde_mm_xor_si128(c2[1348],simde_mm_xor_si128(c2[312],simde_mm_xor_si128(c2[1370],simde_mm_xor_si128(c2[1898],simde_mm_xor_si128(c2[8242],simde_mm_xor_si128(c2[323],simde_mm_xor_si128(c2[2451],simde_mm_xor_si128(c2[1399],simde_mm_xor_si128(c2[1927],simde_mm_xor_si128(c2[7730],simde_mm_xor_si128(c2[8258],simde_mm_xor_si128(c2[1419],simde_mm_xor_si128(c2[7759],simde_mm_xor_si128(c2[8284],simde_mm_xor_si128(c2[365],simde_mm_xor_si128(c2[3032],simde_mm_xor_si128(c2[7259],simde_mm_xor_si128(c2[5672],simde_mm_xor_si128(c2[6200],simde_mm_xor_si128(c2[7280],simde_mm_xor_si128(c2[7281],simde_mm_xor_si128(c2[7809],simde_mm_xor_si128(c2[3584],simde_mm_xor_si128(c2[4112],simde_mm_xor_si128(c2[7300],simde_mm_xor_si128(c2[5716],simde_mm_xor_si128(c2[6244],simde_mm_xor_si128(c2[5190],simde_mm_xor_si128(c2[5718],simde_mm_xor_si128(c2[8376],simde_mm_xor_si128(c2[3634],simde_mm_xor_si128(c2[4154],simde_mm_xor_si128(c2[4682],simde_mm_xor_si128(c2[3130],simde_mm_xor_si128(c2[4178],simde_mm_xor_si128(c2[4706],simde_mm_xor_si128(c2[1012],simde_mm_xor_si128(c2[1540],simde_mm_xor_si128(c2[3155],simde_mm_xor_si128(c2[6320],simde_mm_xor_si128(c2[8435],c2[504]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 43
|
|
d2[516]=simde_mm_xor_si128(c2[6341],simde_mm_xor_si128(c2[2643],simde_mm_xor_si128(c2[3179],simde_mm_xor_si128(c2[6339],simde_mm_xor_si128(c2[6867],simde_mm_xor_si128(c2[6371],simde_mm_xor_si128(c2[1610],simde_mm_xor_si128(c2[7947],simde_mm_xor_si128(c2[28],simde_mm_xor_si128(c2[559],simde_mm_xor_si128(c2[5857],simde_mm_xor_si128(c2[2691],simde_mm_xor_si128(c2[583],simde_mm_xor_si128(c2[5361],simde_mm_xor_si128(c2[1131],simde_mm_xor_si128(c2[4306],simde_mm_xor_si128(c2[4834],simde_mm_xor_si128(c2[4852],simde_mm_xor_si128(c2[2743],simde_mm_xor_si128(c2[5907],simde_mm_xor_si128(c2[6435],simde_mm_xor_si128(c2[7521],simde_mm_xor_si128(c2[4879],simde_mm_xor_si128(c2[4874],simde_mm_xor_si128(c2[672],simde_mm_xor_si128(c2[5432],simde_mm_xor_si128(c2[8074],simde_mm_xor_si128(c2[155],simde_mm_xor_si128(c2[2287],simde_mm_xor_si128(c2[2283],simde_mm_xor_si128(c2[704],simde_mm_xor_si128(c2[1232],simde_mm_xor_si128(c2[2311],simde_mm_xor_si128(c2[4952],simde_mm_xor_si128(c2[194],simde_mm_xor_si128(c2[722],simde_mm_xor_si128(c2[3390],simde_mm_xor_si128(c2[6033],simde_mm_xor_si128(c2[4442],simde_mm_xor_si128(c2[4996],simde_mm_xor_si128(c2[5526],simde_mm_xor_si128(c2[6582],simde_mm_xor_si128(c2[7110],simde_mm_xor_si128(c2[3437],simde_mm_xor_si128(c2[269],simde_mm_xor_si128(c2[5021],simde_mm_xor_si128(c2[5549],simde_mm_xor_si128(c2[5579],simde_mm_xor_si128(c2[8215],simde_mm_xor_si128(c2[8208],simde_mm_xor_si128(c2[289],simde_mm_xor_si128(c2[7712],simde_mm_xor_si128(c2[851],simde_mm_xor_si128(c2[7183],simde_mm_xor_si128(c2[7711],simde_mm_xor_si128(c2[1392],simde_mm_xor_si128(c2[868],simde_mm_xor_si128(c2[6683],simde_mm_xor_si128(c2[7211],simde_mm_xor_si128(c2[360],simde_mm_xor_si128(c2[6700],simde_mm_xor_si128(c2[7753],simde_mm_xor_si128(c2[1973],simde_mm_xor_si128(c2[6200],simde_mm_xor_si128(c2[4613],simde_mm_xor_si128(c2[5141],simde_mm_xor_si128(c2[2504],simde_mm_xor_si128(c2[6221],simde_mm_xor_si128(c2[6750],simde_mm_xor_si128(c2[2525],simde_mm_xor_si128(c2[3053],simde_mm_xor_si128(c2[6241],simde_mm_xor_si128(c2[5185],simde_mm_xor_si128(c2[4131],simde_mm_xor_si128(c2[4659],simde_mm_xor_si128(c2[1495],simde_mm_xor_si128(c2[7329],simde_mm_xor_si128(c2[2575],simde_mm_xor_si128(c2[3635],simde_mm_xor_si128(c2[2071],simde_mm_xor_si128(c2[3659],simde_mm_xor_si128(c2[8400],simde_mm_xor_si128(c2[481],simde_mm_xor_si128(c2[2096],simde_mm_xor_si128(c2[5261],simde_mm_xor_si128(c2[7376],c2[7904]))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 44
|
|
d2[528]=simde_mm_xor_si128(c2[2640],simde_mm_xor_si128(c2[7401],simde_mm_xor_si128(c2[7925],simde_mm_xor_si128(c2[3178],simde_mm_xor_si128(c2[5],simde_mm_xor_si128(c2[2670],simde_mm_xor_si128(c2[6368],simde_mm_xor_si128(c2[4786],simde_mm_xor_si128(c2[2168],simde_mm_xor_si128(c2[7449],simde_mm_xor_si128(c2[5329],simde_mm_xor_si128(c2[1660],simde_mm_xor_si128(c2[5889],simde_mm_xor_si128(c2[1133],simde_mm_xor_si128(c2[1163],simde_mm_xor_si128(c2[7489],simde_mm_xor_si128(c2[2746],simde_mm_xor_si128(c2[3820],simde_mm_xor_si128(c2[1178],simde_mm_xor_si128(c2[1185],simde_mm_xor_si128(c2[5430],simde_mm_xor_si128(c2[1731],simde_mm_xor_si128(c2[4901],simde_mm_xor_si128(c2[7033],simde_mm_xor_si128(c2[7041],simde_mm_xor_si128(c2[5978],simde_mm_xor_si128(c2[7564],simde_mm_xor_si128(c2[7057],simde_mm_xor_si128(c2[1251],simde_mm_xor_si128(c2[5480],simde_mm_xor_si128(c2[8136],simde_mm_xor_si128(c2[2332],simde_mm_xor_si128(c2[753],simde_mm_xor_si128(c2[1810],simde_mm_xor_si128(c2[1307],simde_mm_xor_si128(c2[1825],simde_mm_xor_si128(c2[3409],simde_mm_xor_si128(c2[8195],simde_mm_xor_si128(c2[5027],simde_mm_xor_si128(c2[1848],simde_mm_xor_si128(c2[1878],simde_mm_xor_si128(c2[4514],simde_mm_xor_si128(c2[5047],simde_mm_xor_si128(c2[4011],simde_mm_xor_si128(c2[5597],simde_mm_xor_si128(c2[4010],simde_mm_xor_si128(c2[6150],simde_mm_xor_si128(c2[5626],simde_mm_xor_si128(c2[3510],simde_mm_xor_si128(c2[5118],simde_mm_xor_si128(c2[3011],simde_mm_xor_si128(c2[4064],simde_mm_xor_si128(c2[6731],simde_mm_xor_si128(c2[2499],simde_mm_xor_si128(c2[1440],simde_mm_xor_si128(c2[2520],simde_mm_xor_si128(c2[3049],simde_mm_xor_si128(c2[7811],simde_mm_xor_si128(c2[2552],simde_mm_xor_si128(c2[1496],simde_mm_xor_si128(c2[970],simde_mm_xor_si128(c2[3628],simde_mm_xor_si128(c2[7321],simde_mm_xor_si128(c2[8381],simde_mm_xor_si128(c2[6817],simde_mm_xor_si128(c2[8405],simde_mm_xor_si128(c2[5239],simde_mm_xor_si128(c2[6842],simde_mm_xor_si128(c2[1560],c2[4203])))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))))));
|
|
|
|
//row: 45
|
|
d2[540]=simde_mm_xor_si128(c2[3728],simde_mm_xor_si128(c2[2793],c2[8160]));
|
|
}
|
|
}
|
|
#endif
|