#ifndef ATL_RESTRICT #if defined(__STDC_VERSION__) && (__STDC_VERSION__/100 >= 1999) #define ATL_RESTRICT restrict #else #define ATL_RESTRICT #endif #endif #include "atlas_prefetch.h" #ifndef ATL_RESTRICT #if defined(__STDC_VERSION__) && (__STDC_VERSION__/100 >= 1999) #define ATL_RESTRICT restrict #else #define ATL_RESTRICT #endif #endif #include "atlas_prefetch.h" static void ATL_dJIK8x0x56TN8x1x56_a1_b1 (const int M, const int N, const int K, const double alpha, const double * ATL_RESTRICT A, const int lda, const double * ATL_RESTRICT B, const int ldb, const double beta, double * ATL_RESTRICT C, const int ldc) /* * matmul with TA=T, TB=N, MB=8, NB=0, KB=56, * lda=56, ldb=56, ldc=0, mu=8, nu=1, ku=56, pf=1 * Generated by ATLAS/tune/blas/gemm/emit_mm.c (3.8.0) */ { #define Nb N const double *stM = A + 448; const double *stN = B + (56*(Nb)); const double *pfA = stM; const int incPFA0 = (((int)(stM - A))*8*1)/(8*N*sizeof(double)); const int incPFA = (1 > incPFA0) ? 1 : incPFA0; #define incAk 56 const int incAm = 392, incAn = -448; #define incBk 56 const int incBm = -56, incBn = 56; const int incAk0 = ((incAk) / 56), incBk0 = ((incBk) / 56); #define incCm 8 const int incCn = (ldc) - 8; double *pC0=C; const double *pA0=A; const double *pB0=B; register int k; register double rA0, rA1, rA2, rA3, rA4, rA5, rA6, rA7; register double rB0; register double m0; register double rC0_0, rC1_0, rC2_0, rC3_0, rC4_0, rC5_0, rC6_0, rC7_0; do /* N-loop */ { ATL_pfl1R(pfA); pfA += incPFA; ATL_pfl1W(pC0); /* * Start pipeline */ rA0 = *pA0; rB0 = *pB0; rA1 = pA0[56]; rA2 = pA0[112]; rA3 = pA0[168]; rA4 = pA0[224]; rA5 = pA0[280]; rA6 = pA0[336]; rA7 = pA0[392]; m0 = rA0 * rB0; /* * Completely unrolled K-loop */ rC0_0 = m0; m0 = rA1 * rB0; rC1_0 = m0; m0 = rA2 * rB0; rC2_0 = m0; m0 = rA3 * rB0; rC3_0 = m0; m0 = rA4 * rB0; rC4_0 = m0; m0 = rA5 * rB0; rC5_0 = m0; m0 = rA6 * rB0; rC6_0 = m0; m0 = rA7 * rB0; rA0 = pA0[1]; rB0 = pB0[1]; rA1 = pA0[57]; rA2 = pA0[113]; rA3 = pA0[169]; rA4 = pA0[225]; rA5 = pA0[281]; rA6 = pA0[337]; rA7 = pA0[393]; rC7_0 = m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[2]; rB0 = pB0[2]; rA1 = pA0[58]; rA2 = pA0[114]; rA3 = pA0[170]; rA4 = pA0[226]; rA5 = pA0[282]; rA6 = pA0[338]; rA7 = pA0[394]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[3]; rB0 = pB0[3]; rA1 = pA0[59]; rA2 = pA0[115]; rA3 = pA0[171]; rA4 = pA0[227]; rA5 = pA0[283]; rA6 = pA0[339]; rA7 = pA0[395]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[4]; rB0 = pB0[4]; rA1 = pA0[60]; rA2 = pA0[116]; rA3 = pA0[172]; rA4 = pA0[228]; rA5 = pA0[284]; rA6 = pA0[340]; rA7 = pA0[396]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[5]; rB0 = pB0[5]; rA1 = pA0[61]; rA2 = pA0[117]; rA3 = pA0[173]; rA4 = pA0[229]; rA5 = pA0[285]; rA6 = pA0[341]; rA7 = pA0[397]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[6]; rB0 = pB0[6]; rA1 = pA0[62]; rA2 = pA0[118]; rA3 = pA0[174]; rA4 = pA0[230]; rA5 = pA0[286]; rA6 = pA0[342]; rA7 = pA0[398]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[7]; rB0 = pB0[7]; rA1 = pA0[63]; rA2 = pA0[119]; rA3 = pA0[175]; rA4 = pA0[231]; rA5 = pA0[287]; rA6 = pA0[343]; rA7 = pA0[399]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[8]; rB0 = pB0[8]; rA1 = pA0[64]; rA2 = pA0[120]; rA3 = pA0[176]; rA4 = pA0[232]; rA5 = pA0[288]; rA6 = pA0[344]; rA7 = pA0[400]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[9]; rB0 = pB0[9]; rA1 = pA0[65]; rA2 = pA0[121]; rA3 = pA0[177]; rA4 = pA0[233]; rA5 = pA0[289]; rA6 = pA0[345]; rA7 = pA0[401]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[10]; rB0 = pB0[10]; rA1 = pA0[66]; rA2 = pA0[122]; rA3 = pA0[178]; rA4 = pA0[234]; rA5 = pA0[290]; rA6 = pA0[346]; rA7 = pA0[402]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[11]; rB0 = pB0[11]; rA1 = pA0[67]; rA2 = pA0[123]; rA3 = pA0[179]; rA4 = pA0[235]; rA5 = pA0[291]; rA6 = pA0[347]; rA7 = pA0[403]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[12]; rB0 = pB0[12]; rA1 = pA0[68]; rA2 = pA0[124]; rA3 = pA0[180]; rA4 = pA0[236]; rA5 = pA0[292]; rA6 = pA0[348]; rA7 = pA0[404]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[13]; rB0 = pB0[13]; rA1 = pA0[69]; rA2 = pA0[125]; rA3 = pA0[181]; rA4 = pA0[237]; rA5 = pA0[293]; rA6 = pA0[349]; rA7 = pA0[405]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[14]; rB0 = pB0[14]; rA1 = pA0[70]; rA2 = pA0[126]; rA3 = pA0[182]; rA4 = pA0[238]; rA5 = pA0[294]; rA6 = pA0[350]; rA7 = pA0[406]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[15]; rB0 = pB0[15]; rA1 = pA0[71]; rA2 = pA0[127]; rA3 = pA0[183]; rA4 = pA0[239]; rA5 = pA0[295]; rA6 = pA0[351]; rA7 = pA0[407]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[16]; rB0 = pB0[16]; rA1 = pA0[72]; rA2 = pA0[128]; rA3 = pA0[184]; rA4 = pA0[240]; rA5 = pA0[296]; rA6 = pA0[352]; rA7 = pA0[408]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[17]; rB0 = pB0[17]; rA1 = pA0[73]; rA2 = pA0[129]; rA3 = pA0[185]; rA4 = pA0[241]; rA5 = pA0[297]; rA6 = pA0[353]; rA7 = pA0[409]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[18]; rB0 = pB0[18]; rA1 = pA0[74]; rA2 = pA0[130]; rA3 = pA0[186]; rA4 = pA0[242]; rA5 = pA0[298]; rA6 = pA0[354]; rA7 = pA0[410]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[19]; rB0 = pB0[19]; rA1 = pA0[75]; rA2 = pA0[131]; rA3 = pA0[187]; rA4 = pA0[243]; rA5 = pA0[299]; rA6 = pA0[355]; rA7 = pA0[411]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[20]; rB0 = pB0[20]; rA1 = pA0[76]; rA2 = pA0[132]; rA3 = pA0[188]; rA4 = pA0[244]; rA5 = pA0[300]; rA6 = pA0[356]; rA7 = pA0[412]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[21]; rB0 = pB0[21]; rA1 = pA0[77]; rA2 = pA0[133]; rA3 = pA0[189]; rA4 = pA0[245]; rA5 = pA0[301]; rA6 = pA0[357]; rA7 = pA0[413]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[22]; rB0 = pB0[22]; rA1 = pA0[78]; rA2 = pA0[134]; rA3 = pA0[190]; rA4 = pA0[246]; rA5 = pA0[302]; rA6 = pA0[358]; rA7 = pA0[414]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[23]; rB0 = pB0[23]; rA1 = pA0[79]; rA2 = pA0[135]; rA3 = pA0[191]; rA4 = pA0[247]; rA5 = pA0[303]; rA6 = pA0[359]; rA7 = pA0[415]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[24]; rB0 = pB0[24]; rA1 = pA0[80]; rA2 = pA0[136]; rA3 = pA0[192]; rA4 = pA0[248]; rA5 = pA0[304]; rA6 = pA0[360]; rA7 = pA0[416]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[25]; rB0 = pB0[25]; rA1 = pA0[81]; rA2 = pA0[137]; rA3 = pA0[193]; rA4 = pA0[249]; rA5 = pA0[305]; rA6 = pA0[361]; rA7 = pA0[417]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[26]; rB0 = pB0[26]; rA1 = pA0[82]; rA2 = pA0[138]; rA3 = pA0[194]; rA4 = pA0[250]; rA5 = pA0[306]; rA6 = pA0[362]; rA7 = pA0[418]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[27]; rB0 = pB0[27]; rA1 = pA0[83]; rA2 = pA0[139]; rA3 = pA0[195]; rA4 = pA0[251]; rA5 = pA0[307]; rA6 = pA0[363]; rA7 = pA0[419]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[28]; rB0 = pB0[28]; rA1 = pA0[84]; rA2 = pA0[140]; rA3 = pA0[196]; rA4 = pA0[252]; rA5 = pA0[308]; rA6 = pA0[364]; rA7 = pA0[420]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[29]; rB0 = pB0[29]; rA1 = pA0[85]; rA2 = pA0[141]; rA3 = pA0[197]; rA4 = pA0[253]; rA5 = pA0[309]; rA6 = pA0[365]; rA7 = pA0[421]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[30]; rB0 = pB0[30]; rA1 = pA0[86]; rA2 = pA0[142]; rA3 = pA0[198]; rA4 = pA0[254]; rA5 = pA0[310]; rA6 = pA0[366]; rA7 = pA0[422]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[31]; rB0 = pB0[31]; rA1 = pA0[87]; rA2 = pA0[143]; rA3 = pA0[199]; rA4 = pA0[255]; rA5 = pA0[311]; rA6 = pA0[367]; rA7 = pA0[423]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[32]; rB0 = pB0[32]; rA1 = pA0[88]; rA2 = pA0[144]; rA3 = pA0[200]; rA4 = pA0[256]; rA5 = pA0[312]; rA6 = pA0[368]; rA7 = pA0[424]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[33]; rB0 = pB0[33]; rA1 = pA0[89]; rA2 = pA0[145]; rA3 = pA0[201]; rA4 = pA0[257]; rA5 = pA0[313]; rA6 = pA0[369]; rA7 = pA0[425]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[34]; rB0 = pB0[34]; rA1 = pA0[90]; rA2 = pA0[146]; rA3 = pA0[202]; rA4 = pA0[258]; rA5 = pA0[314]; rA6 = pA0[370]; rA7 = pA0[426]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[35]; rB0 = pB0[35]; rA1 = pA0[91]; rA2 = pA0[147]; rA3 = pA0[203]; rA4 = pA0[259]; rA5 = pA0[315]; rA6 = pA0[371]; rA7 = pA0[427]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[36]; rB0 = pB0[36]; rA1 = pA0[92]; rA2 = pA0[148]; rA3 = pA0[204]; rA4 = pA0[260]; rA5 = pA0[316]; rA6 = pA0[372]; rA7 = pA0[428]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[37]; rB0 = pB0[37]; rA1 = pA0[93]; rA2 = pA0[149]; rA3 = pA0[205]; rA4 = pA0[261]; rA5 = pA0[317]; rA6 = pA0[373]; rA7 = pA0[429]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[38]; rB0 = pB0[38]; rA1 = pA0[94]; rA2 = pA0[150]; rA3 = pA0[206]; rA4 = pA0[262]; rA5 = pA0[318]; rA6 = pA0[374]; rA7 = pA0[430]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[39]; rB0 = pB0[39]; rA1 = pA0[95]; rA2 = pA0[151]; rA3 = pA0[207]; rA4 = pA0[263]; rA5 = pA0[319]; rA6 = pA0[375]; rA7 = pA0[431]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[40]; rB0 = pB0[40]; rA1 = pA0[96]; rA2 = pA0[152]; rA3 = pA0[208]; rA4 = pA0[264]; rA5 = pA0[320]; rA6 = pA0[376]; rA7 = pA0[432]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[41]; rB0 = pB0[41]; rA1 = pA0[97]; rA2 = pA0[153]; rA3 = pA0[209]; rA4 = pA0[265]; rA5 = pA0[321]; rA6 = pA0[377]; rA7 = pA0[433]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[42]; rB0 = pB0[42]; rA1 = pA0[98]; rA2 = pA0[154]; rA3 = pA0[210]; rA4 = pA0[266]; rA5 = pA0[322]; rA6 = pA0[378]; rA7 = pA0[434]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[43]; rB0 = pB0[43]; rA1 = pA0[99]; rA2 = pA0[155]; rA3 = pA0[211]; rA4 = pA0[267]; rA5 = pA0[323]; rA6 = pA0[379]; rA7 = pA0[435]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[44]; rB0 = pB0[44]; rA1 = pA0[100]; rA2 = pA0[156]; rA3 = pA0[212]; rA4 = pA0[268]; rA5 = pA0[324]; rA6 = pA0[380]; rA7 = pA0[436]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[45]; rB0 = pB0[45]; rA1 = pA0[101]; rA2 = pA0[157]; rA3 = pA0[213]; rA4 = pA0[269]; rA5 = pA0[325]; rA6 = pA0[381]; rA7 = pA0[437]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[46]; rB0 = pB0[46]; rA1 = pA0[102]; rA2 = pA0[158]; rA3 = pA0[214]; rA4 = pA0[270]; rA5 = pA0[326]; rA6 = pA0[382]; rA7 = pA0[438]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[47]; rB0 = pB0[47]; rA1 = pA0[103]; rA2 = pA0[159]; rA3 = pA0[215]; rA4 = pA0[271]; rA5 = pA0[327]; rA6 = pA0[383]; rA7 = pA0[439]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[48]; rB0 = pB0[48]; rA1 = pA0[104]; rA2 = pA0[160]; rA3 = pA0[216]; rA4 = pA0[272]; rA5 = pA0[328]; rA6 = pA0[384]; rA7 = pA0[440]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[49]; rB0 = pB0[49]; rA1 = pA0[105]; rA2 = pA0[161]; rA3 = pA0[217]; rA4 = pA0[273]; rA5 = pA0[329]; rA6 = pA0[385]; rA7 = pA0[441]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[50]; rB0 = pB0[50]; rA1 = pA0[106]; rA2 = pA0[162]; rA3 = pA0[218]; rA4 = pA0[274]; rA5 = pA0[330]; rA6 = pA0[386]; rA7 = pA0[442]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[51]; rB0 = pB0[51]; rA1 = pA0[107]; rA2 = pA0[163]; rA3 = pA0[219]; rA4 = pA0[275]; rA5 = pA0[331]; rA6 = pA0[387]; rA7 = pA0[443]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[52]; rB0 = pB0[52]; rA1 = pA0[108]; rA2 = pA0[164]; rA3 = pA0[220]; rA4 = pA0[276]; rA5 = pA0[332]; rA6 = pA0[388]; rA7 = pA0[444]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[53]; rB0 = pB0[53]; rA1 = pA0[109]; rA2 = pA0[165]; rA3 = pA0[221]; rA4 = pA0[277]; rA5 = pA0[333]; rA6 = pA0[389]; rA7 = pA0[445]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[54]; rB0 = pB0[54]; rA1 = pA0[110]; rA2 = pA0[166]; rA3 = pA0[222]; rA4 = pA0[278]; rA5 = pA0[334]; rA6 = pA0[390]; rA7 = pA0[446]; rC7_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rA0 = pA0[55]; rB0 = pB0[55]; rA1 = pA0[111]; rA2 = pA0[167]; rA3 = pA0[223]; rA4 = pA0[279]; rA5 = pA0[335]; rA6 = pA0[391]; rA7 = pA0[447]; rC7_0 += m0; m0 = rA0 * rB0; /* * Drain pipe on last iteration of K-loop */ rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; pA0 += incAk; pB0 += incBk; *pC0 += rC0_0; pC0[1] += rC1_0; pC0[2] += rC2_0; pC0[3] += rC3_0; pC0[4] += rC4_0; pC0[5] += rC5_0; pC0[6] += rC6_0; pC0[7] += rC7_0; pC0 += incCm; pA0 += incAm; pB0 += incBm; pC0 += incCn; pA0 += incAn; pB0 += incBn; } while(pB0 != stN); } #ifdef incAm #undef incAm #endif #ifdef incAn #undef incAn #endif #ifdef incAk #undef incAk #endif #ifdef incBm #undef incBm #endif #ifdef incBn #undef incBn #endif #ifdef incBk #undef incBk #endif #ifdef incCm #undef incCm #endif #ifdef incCn #undef incCn #endif #ifdef incCk #undef incCk #endif #ifdef Mb #undef Mb #endif #ifdef Nb #undef Nb #endif #ifdef Kb #undef Kb #endif void ATL_dJIK56x0x56TN56x56x0_a1_b1 (const int M, const int N, const int K, const double alpha, const double * ATL_RESTRICT A, const int lda, const double * ATL_RESTRICT B, const int ldb, const double beta, double * ATL_RESTRICT C, const int ldc) /* * matmul with TA=T, TB=N, MB=56, NB=0, KB=56, * lda=56, ldb=56, ldc=0, mu=12, nu=1, ku=56, pf=1 * Generated by ATLAS/tune/blas/gemm/emit_mm.c (3.8.0) */ { #define Nb N const double *ca=A, *cb=B; double *cc=C; const double *stM = A + 2688; const double *stN = B + (56*(Nb)); const double *pfA = stM; const int incPFA0 = (((int)(stM - A))*12*1)/(56*N*sizeof(double)); const int incPFA = (1 > incPFA0) ? 1 : incPFA0; #define incAk 56 const int incAm = 616, incAn = -2688; #define incBk 56 const int incBm = -56, incBn = 56; const int incAk0 = ((incAk) / 56), incBk0 = ((incBk) / 56); #define incCm 12 const int incCn = (ldc) - 48; double *pC0=C; const double *pA0=A; const double *pB0=B; register int k; register double rA0, rA1, rA2, rA3, rA4, rA5, rA6, rA7, rA8, rA9, rA10, rA11; register double rB0; register double m0; register double rC0_0, rC1_0, rC2_0, rC3_0, rC4_0, rC5_0, rC6_0, rC7_0, rC8_0, rC9_0, rC10_0, rC11_0; do /* N-loop */ { do /* M-loop */ { ATL_pfl1R(pfA); pfA += incPFA; ATL_pfl1W(pC0); /* * Start pipeline */ rA0 = *pA0; rB0 = *pB0; rA1 = pA0[56]; rA2 = pA0[112]; rA3 = pA0[168]; rA4 = pA0[224]; rA5 = pA0[280]; rA6 = pA0[336]; rA7 = pA0[392]; rA8 = pA0[448]; rA9 = pA0[504]; rA10 = pA0[560]; m0 = rA0 * rB0; rA11 = pA0[616]; /* * Completely unrolled K-loop */ rC0_0 = m0; m0 = rA1 * rB0; rC1_0 = m0; m0 = rA2 * rB0; rC2_0 = m0; m0 = rA3 * rB0; rC3_0 = m0; m0 = rA4 * rB0; rC4_0 = m0; m0 = rA5 * rB0; rC5_0 = m0; m0 = rA6 * rB0; rC6_0 = m0; m0 = rA7 * rB0; rC7_0 = m0; m0 = rA8 * rB0; rC8_0 = m0; m0 = rA9 * rB0; rC9_0 = m0; m0 = rA10 * rB0; rC10_0 = m0; m0 = rA11 * rB0; rA0 = pA0[1]; rB0 = pB0[1]; rA1 = pA0[57]; rA2 = pA0[113]; rA3 = pA0[169]; rA4 = pA0[225]; rA5 = pA0[281]; rA6 = pA0[337]; rA7 = pA0[393]; rA8 = pA0[449]; rA9 = pA0[505]; rA10 = pA0[561]; rA11 = pA0[617]; rC11_0 = m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[2]; rB0 = pB0[2]; rA1 = pA0[58]; rA2 = pA0[114]; rA3 = pA0[170]; rA4 = pA0[226]; rA5 = pA0[282]; rA6 = pA0[338]; rA7 = pA0[394]; rA8 = pA0[450]; rA9 = pA0[506]; rA10 = pA0[562]; rA11 = pA0[618]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[3]; rB0 = pB0[3]; rA1 = pA0[59]; rA2 = pA0[115]; rA3 = pA0[171]; rA4 = pA0[227]; rA5 = pA0[283]; rA6 = pA0[339]; rA7 = pA0[395]; rA8 = pA0[451]; rA9 = pA0[507]; rA10 = pA0[563]; rA11 = pA0[619]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[4]; rB0 = pB0[4]; rA1 = pA0[60]; rA2 = pA0[116]; rA3 = pA0[172]; rA4 = pA0[228]; rA5 = pA0[284]; rA6 = pA0[340]; rA7 = pA0[396]; rA8 = pA0[452]; rA9 = pA0[508]; rA10 = pA0[564]; rA11 = pA0[620]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[5]; rB0 = pB0[5]; rA1 = pA0[61]; rA2 = pA0[117]; rA3 = pA0[173]; rA4 = pA0[229]; rA5 = pA0[285]; rA6 = pA0[341]; rA7 = pA0[397]; rA8 = pA0[453]; rA9 = pA0[509]; rA10 = pA0[565]; rA11 = pA0[621]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[6]; rB0 = pB0[6]; rA1 = pA0[62]; rA2 = pA0[118]; rA3 = pA0[174]; rA4 = pA0[230]; rA5 = pA0[286]; rA6 = pA0[342]; rA7 = pA0[398]; rA8 = pA0[454]; rA9 = pA0[510]; rA10 = pA0[566]; rA11 = pA0[622]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[7]; rB0 = pB0[7]; rA1 = pA0[63]; rA2 = pA0[119]; rA3 = pA0[175]; rA4 = pA0[231]; rA5 = pA0[287]; rA6 = pA0[343]; rA7 = pA0[399]; rA8 = pA0[455]; rA9 = pA0[511]; rA10 = pA0[567]; rA11 = pA0[623]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[8]; rB0 = pB0[8]; rA1 = pA0[64]; rA2 = pA0[120]; rA3 = pA0[176]; rA4 = pA0[232]; rA5 = pA0[288]; rA6 = pA0[344]; rA7 = pA0[400]; rA8 = pA0[456]; rA9 = pA0[512]; rA10 = pA0[568]; rA11 = pA0[624]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[9]; rB0 = pB0[9]; rA1 = pA0[65]; rA2 = pA0[121]; rA3 = pA0[177]; rA4 = pA0[233]; rA5 = pA0[289]; rA6 = pA0[345]; rA7 = pA0[401]; rA8 = pA0[457]; rA9 = pA0[513]; rA10 = pA0[569]; rA11 = pA0[625]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[10]; rB0 = pB0[10]; rA1 = pA0[66]; rA2 = pA0[122]; rA3 = pA0[178]; rA4 = pA0[234]; rA5 = pA0[290]; rA6 = pA0[346]; rA7 = pA0[402]; rA8 = pA0[458]; rA9 = pA0[514]; rA10 = pA0[570]; rA11 = pA0[626]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[11]; rB0 = pB0[11]; rA1 = pA0[67]; rA2 = pA0[123]; rA3 = pA0[179]; rA4 = pA0[235]; rA5 = pA0[291]; rA6 = pA0[347]; rA7 = pA0[403]; rA8 = pA0[459]; rA9 = pA0[515]; rA10 = pA0[571]; rA11 = pA0[627]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[12]; rB0 = pB0[12]; rA1 = pA0[68]; rA2 = pA0[124]; rA3 = pA0[180]; rA4 = pA0[236]; rA5 = pA0[292]; rA6 = pA0[348]; rA7 = pA0[404]; rA8 = pA0[460]; rA9 = pA0[516]; rA10 = pA0[572]; rA11 = pA0[628]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[13]; rB0 = pB0[13]; rA1 = pA0[69]; rA2 = pA0[125]; rA3 = pA0[181]; rA4 = pA0[237]; rA5 = pA0[293]; rA6 = pA0[349]; rA7 = pA0[405]; rA8 = pA0[461]; rA9 = pA0[517]; rA10 = pA0[573]; rA11 = pA0[629]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[14]; rB0 = pB0[14]; rA1 = pA0[70]; rA2 = pA0[126]; rA3 = pA0[182]; rA4 = pA0[238]; rA5 = pA0[294]; rA6 = pA0[350]; rA7 = pA0[406]; rA8 = pA0[462]; rA9 = pA0[518]; rA10 = pA0[574]; rA11 = pA0[630]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[15]; rB0 = pB0[15]; rA1 = pA0[71]; rA2 = pA0[127]; rA3 = pA0[183]; rA4 = pA0[239]; rA5 = pA0[295]; rA6 = pA0[351]; rA7 = pA0[407]; rA8 = pA0[463]; rA9 = pA0[519]; rA10 = pA0[575]; rA11 = pA0[631]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[16]; rB0 = pB0[16]; rA1 = pA0[72]; rA2 = pA0[128]; rA3 = pA0[184]; rA4 = pA0[240]; rA5 = pA0[296]; rA6 = pA0[352]; rA7 = pA0[408]; rA8 = pA0[464]; rA9 = pA0[520]; rA10 = pA0[576]; rA11 = pA0[632]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[17]; rB0 = pB0[17]; rA1 = pA0[73]; rA2 = pA0[129]; rA3 = pA0[185]; rA4 = pA0[241]; rA5 = pA0[297]; rA6 = pA0[353]; rA7 = pA0[409]; rA8 = pA0[465]; rA9 = pA0[521]; rA10 = pA0[577]; rA11 = pA0[633]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[18]; rB0 = pB0[18]; rA1 = pA0[74]; rA2 = pA0[130]; rA3 = pA0[186]; rA4 = pA0[242]; rA5 = pA0[298]; rA6 = pA0[354]; rA7 = pA0[410]; rA8 = pA0[466]; rA9 = pA0[522]; rA10 = pA0[578]; rA11 = pA0[634]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[19]; rB0 = pB0[19]; rA1 = pA0[75]; rA2 = pA0[131]; rA3 = pA0[187]; rA4 = pA0[243]; rA5 = pA0[299]; rA6 = pA0[355]; rA7 = pA0[411]; rA8 = pA0[467]; rA9 = pA0[523]; rA10 = pA0[579]; rA11 = pA0[635]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[20]; rB0 = pB0[20]; rA1 = pA0[76]; rA2 = pA0[132]; rA3 = pA0[188]; rA4 = pA0[244]; rA5 = pA0[300]; rA6 = pA0[356]; rA7 = pA0[412]; rA8 = pA0[468]; rA9 = pA0[524]; rA10 = pA0[580]; rA11 = pA0[636]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[21]; rB0 = pB0[21]; rA1 = pA0[77]; rA2 = pA0[133]; rA3 = pA0[189]; rA4 = pA0[245]; rA5 = pA0[301]; rA6 = pA0[357]; rA7 = pA0[413]; rA8 = pA0[469]; rA9 = pA0[525]; rA10 = pA0[581]; rA11 = pA0[637]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[22]; rB0 = pB0[22]; rA1 = pA0[78]; rA2 = pA0[134]; rA3 = pA0[190]; rA4 = pA0[246]; rA5 = pA0[302]; rA6 = pA0[358]; rA7 = pA0[414]; rA8 = pA0[470]; rA9 = pA0[526]; rA10 = pA0[582]; rA11 = pA0[638]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[23]; rB0 = pB0[23]; rA1 = pA0[79]; rA2 = pA0[135]; rA3 = pA0[191]; rA4 = pA0[247]; rA5 = pA0[303]; rA6 = pA0[359]; rA7 = pA0[415]; rA8 = pA0[471]; rA9 = pA0[527]; rA10 = pA0[583]; rA11 = pA0[639]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[24]; rB0 = pB0[24]; rA1 = pA0[80]; rA2 = pA0[136]; rA3 = pA0[192]; rA4 = pA0[248]; rA5 = pA0[304]; rA6 = pA0[360]; rA7 = pA0[416]; rA8 = pA0[472]; rA9 = pA0[528]; rA10 = pA0[584]; rA11 = pA0[640]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[25]; rB0 = pB0[25]; rA1 = pA0[81]; rA2 = pA0[137]; rA3 = pA0[193]; rA4 = pA0[249]; rA5 = pA0[305]; rA6 = pA0[361]; rA7 = pA0[417]; rA8 = pA0[473]; rA9 = pA0[529]; rA10 = pA0[585]; rA11 = pA0[641]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[26]; rB0 = pB0[26]; rA1 = pA0[82]; rA2 = pA0[138]; rA3 = pA0[194]; rA4 = pA0[250]; rA5 = pA0[306]; rA6 = pA0[362]; rA7 = pA0[418]; rA8 = pA0[474]; rA9 = pA0[530]; rA10 = pA0[586]; rA11 = pA0[642]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[27]; rB0 = pB0[27]; rA1 = pA0[83]; rA2 = pA0[139]; rA3 = pA0[195]; rA4 = pA0[251]; rA5 = pA0[307]; rA6 = pA0[363]; rA7 = pA0[419]; rA8 = pA0[475]; rA9 = pA0[531]; rA10 = pA0[587]; rA11 = pA0[643]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[28]; rB0 = pB0[28]; rA1 = pA0[84]; rA2 = pA0[140]; rA3 = pA0[196]; rA4 = pA0[252]; rA5 = pA0[308]; rA6 = pA0[364]; rA7 = pA0[420]; rA8 = pA0[476]; rA9 = pA0[532]; rA10 = pA0[588]; rA11 = pA0[644]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[29]; rB0 = pB0[29]; rA1 = pA0[85]; rA2 = pA0[141]; rA3 = pA0[197]; rA4 = pA0[253]; rA5 = pA0[309]; rA6 = pA0[365]; rA7 = pA0[421]; rA8 = pA0[477]; rA9 = pA0[533]; rA10 = pA0[589]; rA11 = pA0[645]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[30]; rB0 = pB0[30]; rA1 = pA0[86]; rA2 = pA0[142]; rA3 = pA0[198]; rA4 = pA0[254]; rA5 = pA0[310]; rA6 = pA0[366]; rA7 = pA0[422]; rA8 = pA0[478]; rA9 = pA0[534]; rA10 = pA0[590]; rA11 = pA0[646]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[31]; rB0 = pB0[31]; rA1 = pA0[87]; rA2 = pA0[143]; rA3 = pA0[199]; rA4 = pA0[255]; rA5 = pA0[311]; rA6 = pA0[367]; rA7 = pA0[423]; rA8 = pA0[479]; rA9 = pA0[535]; rA10 = pA0[591]; rA11 = pA0[647]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[32]; rB0 = pB0[32]; rA1 = pA0[88]; rA2 = pA0[144]; rA3 = pA0[200]; rA4 = pA0[256]; rA5 = pA0[312]; rA6 = pA0[368]; rA7 = pA0[424]; rA8 = pA0[480]; rA9 = pA0[536]; rA10 = pA0[592]; rA11 = pA0[648]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[33]; rB0 = pB0[33]; rA1 = pA0[89]; rA2 = pA0[145]; rA3 = pA0[201]; rA4 = pA0[257]; rA5 = pA0[313]; rA6 = pA0[369]; rA7 = pA0[425]; rA8 = pA0[481]; rA9 = pA0[537]; rA10 = pA0[593]; rA11 = pA0[649]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[34]; rB0 = pB0[34]; rA1 = pA0[90]; rA2 = pA0[146]; rA3 = pA0[202]; rA4 = pA0[258]; rA5 = pA0[314]; rA6 = pA0[370]; rA7 = pA0[426]; rA8 = pA0[482]; rA9 = pA0[538]; rA10 = pA0[594]; rA11 = pA0[650]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[35]; rB0 = pB0[35]; rA1 = pA0[91]; rA2 = pA0[147]; rA3 = pA0[203]; rA4 = pA0[259]; rA5 = pA0[315]; rA6 = pA0[371]; rA7 = pA0[427]; rA8 = pA0[483]; rA9 = pA0[539]; rA10 = pA0[595]; rA11 = pA0[651]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[36]; rB0 = pB0[36]; rA1 = pA0[92]; rA2 = pA0[148]; rA3 = pA0[204]; rA4 = pA0[260]; rA5 = pA0[316]; rA6 = pA0[372]; rA7 = pA0[428]; rA8 = pA0[484]; rA9 = pA0[540]; rA10 = pA0[596]; rA11 = pA0[652]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[37]; rB0 = pB0[37]; rA1 = pA0[93]; rA2 = pA0[149]; rA3 = pA0[205]; rA4 = pA0[261]; rA5 = pA0[317]; rA6 = pA0[373]; rA7 = pA0[429]; rA8 = pA0[485]; rA9 = pA0[541]; rA10 = pA0[597]; rA11 = pA0[653]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[38]; rB0 = pB0[38]; rA1 = pA0[94]; rA2 = pA0[150]; rA3 = pA0[206]; rA4 = pA0[262]; rA5 = pA0[318]; rA6 = pA0[374]; rA7 = pA0[430]; rA8 = pA0[486]; rA9 = pA0[542]; rA10 = pA0[598]; rA11 = pA0[654]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[39]; rB0 = pB0[39]; rA1 = pA0[95]; rA2 = pA0[151]; rA3 = pA0[207]; rA4 = pA0[263]; rA5 = pA0[319]; rA6 = pA0[375]; rA7 = pA0[431]; rA8 = pA0[487]; rA9 = pA0[543]; rA10 = pA0[599]; rA11 = pA0[655]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[40]; rB0 = pB0[40]; rA1 = pA0[96]; rA2 = pA0[152]; rA3 = pA0[208]; rA4 = pA0[264]; rA5 = pA0[320]; rA6 = pA0[376]; rA7 = pA0[432]; rA8 = pA0[488]; rA9 = pA0[544]; rA10 = pA0[600]; rA11 = pA0[656]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[41]; rB0 = pB0[41]; rA1 = pA0[97]; rA2 = pA0[153]; rA3 = pA0[209]; rA4 = pA0[265]; rA5 = pA0[321]; rA6 = pA0[377]; rA7 = pA0[433]; rA8 = pA0[489]; rA9 = pA0[545]; rA10 = pA0[601]; rA11 = pA0[657]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[42]; rB0 = pB0[42]; rA1 = pA0[98]; rA2 = pA0[154]; rA3 = pA0[210]; rA4 = pA0[266]; rA5 = pA0[322]; rA6 = pA0[378]; rA7 = pA0[434]; rA8 = pA0[490]; rA9 = pA0[546]; rA10 = pA0[602]; rA11 = pA0[658]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[43]; rB0 = pB0[43]; rA1 = pA0[99]; rA2 = pA0[155]; rA3 = pA0[211]; rA4 = pA0[267]; rA5 = pA0[323]; rA6 = pA0[379]; rA7 = pA0[435]; rA8 = pA0[491]; rA9 = pA0[547]; rA10 = pA0[603]; rA11 = pA0[659]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[44]; rB0 = pB0[44]; rA1 = pA0[100]; rA2 = pA0[156]; rA3 = pA0[212]; rA4 = pA0[268]; rA5 = pA0[324]; rA6 = pA0[380]; rA7 = pA0[436]; rA8 = pA0[492]; rA9 = pA0[548]; rA10 = pA0[604]; rA11 = pA0[660]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[45]; rB0 = pB0[45]; rA1 = pA0[101]; rA2 = pA0[157]; rA3 = pA0[213]; rA4 = pA0[269]; rA5 = pA0[325]; rA6 = pA0[381]; rA7 = pA0[437]; rA8 = pA0[493]; rA9 = pA0[549]; rA10 = pA0[605]; rA11 = pA0[661]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[46]; rB0 = pB0[46]; rA1 = pA0[102]; rA2 = pA0[158]; rA3 = pA0[214]; rA4 = pA0[270]; rA5 = pA0[326]; rA6 = pA0[382]; rA7 = pA0[438]; rA8 = pA0[494]; rA9 = pA0[550]; rA10 = pA0[606]; rA11 = pA0[662]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[47]; rB0 = pB0[47]; rA1 = pA0[103]; rA2 = pA0[159]; rA3 = pA0[215]; rA4 = pA0[271]; rA5 = pA0[327]; rA6 = pA0[383]; rA7 = pA0[439]; rA8 = pA0[495]; rA9 = pA0[551]; rA10 = pA0[607]; rA11 = pA0[663]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[48]; rB0 = pB0[48]; rA1 = pA0[104]; rA2 = pA0[160]; rA3 = pA0[216]; rA4 = pA0[272]; rA5 = pA0[328]; rA6 = pA0[384]; rA7 = pA0[440]; rA8 = pA0[496]; rA9 = pA0[552]; rA10 = pA0[608]; rA11 = pA0[664]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[49]; rB0 = pB0[49]; rA1 = pA0[105]; rA2 = pA0[161]; rA3 = pA0[217]; rA4 = pA0[273]; rA5 = pA0[329]; rA6 = pA0[385]; rA7 = pA0[441]; rA8 = pA0[497]; rA9 = pA0[553]; rA10 = pA0[609]; rA11 = pA0[665]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[50]; rB0 = pB0[50]; rA1 = pA0[106]; rA2 = pA0[162]; rA3 = pA0[218]; rA4 = pA0[274]; rA5 = pA0[330]; rA6 = pA0[386]; rA7 = pA0[442]; rA8 = pA0[498]; rA9 = pA0[554]; rA10 = pA0[610]; rA11 = pA0[666]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[51]; rB0 = pB0[51]; rA1 = pA0[107]; rA2 = pA0[163]; rA3 = pA0[219]; rA4 = pA0[275]; rA5 = pA0[331]; rA6 = pA0[387]; rA7 = pA0[443]; rA8 = pA0[499]; rA9 = pA0[555]; rA10 = pA0[611]; rA11 = pA0[667]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[52]; rB0 = pB0[52]; rA1 = pA0[108]; rA2 = pA0[164]; rA3 = pA0[220]; rA4 = pA0[276]; rA5 = pA0[332]; rA6 = pA0[388]; rA7 = pA0[444]; rA8 = pA0[500]; rA9 = pA0[556]; rA10 = pA0[612]; rA11 = pA0[668]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[53]; rB0 = pB0[53]; rA1 = pA0[109]; rA2 = pA0[165]; rA3 = pA0[221]; rA4 = pA0[277]; rA5 = pA0[333]; rA6 = pA0[389]; rA7 = pA0[445]; rA8 = pA0[501]; rA9 = pA0[557]; rA10 = pA0[613]; rA11 = pA0[669]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[54]; rB0 = pB0[54]; rA1 = pA0[110]; rA2 = pA0[166]; rA3 = pA0[222]; rA4 = pA0[278]; rA5 = pA0[334]; rA6 = pA0[390]; rA7 = pA0[446]; rA8 = pA0[502]; rA9 = pA0[558]; rA10 = pA0[614]; rA11 = pA0[670]; rC11_0 += m0; m0 = rA0 * rB0; rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rA0 = pA0[55]; rB0 = pB0[55]; rA1 = pA0[111]; rA2 = pA0[167]; rA3 = pA0[223]; rA4 = pA0[279]; rA5 = pA0[335]; rA6 = pA0[391]; rA7 = pA0[447]; rA8 = pA0[503]; rA9 = pA0[559]; rA10 = pA0[615]; rA11 = pA0[671]; rC11_0 += m0; m0 = rA0 * rB0; /* * Drain pipe on last iteration of K-loop */ rC0_0 += m0; m0 = rA1 * rB0; rC1_0 += m0; m0 = rA2 * rB0; rC2_0 += m0; m0 = rA3 * rB0; rC3_0 += m0; m0 = rA4 * rB0; rC4_0 += m0; m0 = rA5 * rB0; rC5_0 += m0; m0 = rA6 * rB0; rC6_0 += m0; m0 = rA7 * rB0; rC7_0 += m0; m0 = rA8 * rB0; rC8_0 += m0; m0 = rA9 * rB0; rC9_0 += m0; m0 = rA10 * rB0; rC10_0 += m0; m0 = rA11 * rB0; rC11_0 += m0; pA0 += incAk; pB0 += incBk; *pC0 += rC0_0; pC0[1] += rC1_0; pC0[2] += rC2_0; pC0[3] += rC3_0; pC0[4] += rC4_0; pC0[5] += rC5_0; pC0[6] += rC6_0; pC0[7] += rC7_0; pC0[8] += rC8_0; pC0[9] += rC9_0; pC0[10] += rC10_0; pC0[11] += rC11_0; pC0 += incCm; pA0 += incAm; pB0 += incBm; } while(pA0 != stM); pC0 += incCn; pA0 += incAn; pB0 += incBn; } while(pB0 != stN); ATL_dJIK8x0x56TN8x1x56_a1_b1(8, N, 56, alpha, ca + 2688, lda, cb, ldb, beta, cc + 48, ldc); } #ifdef incAm #undef incAm #endif #ifdef incAn #undef incAn #endif #ifdef incAk #undef incAk #endif #ifdef incBm #undef incBm #endif #ifdef incBn #undef incBn #endif #ifdef incBk #undef incBk #endif #ifdef incCm #undef incCm #endif #ifdef incCn #undef incCn #endif #ifdef incCk #undef incCk #endif #ifdef Mb #undef Mb #endif #ifdef Nb #undef Nb #endif #ifdef Kb #undef Kb #endif