atl_mm4x4x8_bpfab.c

来自「基于Blas CLapck的.用过的人知道是干啥的」· C语言 代码 · 共 744 行 · 第 1/3 页

            rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[8];
            rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+8];
            rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+8];
            rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+8];

            rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+8];
            rC1_3 += m1; m1 = rA1 * rB0; pA0 += 8;
            rC2_3 += m2; m2 = rA2 * rB0; pB0 += 8;
            rC3_3 += m3; m3 = rA3 * rB0;
         }
         rC0_0 += m0; m0 = rA0 * rB1; rB0 = pB0[1];     ATL_pfl1R(pA0+KB4);
         rC1_0 += m1; m1 = rA1 * rB1;                   ATL_pfl1R(pA0+KB5);
         rC2_0 += m2; m2 = rA2 * rB1;                   ATL_pfl1R(pA0+KB6);
         rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+1];  ATL_pfl1R(pA0+KB7);
         rC0_1 += m0; m0 = rA0 * rB2;
         rC1_1 += m1; m1 = rA1 * rB2;
         rC2_1 += m2; m2 = rA2 * rB2;
         rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+1];
         rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[1];
         rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+1];
         rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+1];
         rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+1];

         rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+1];
         rC1_3 += m1; m1 = rA1 * rB0;
         rC2_3 += m2; m2 = rA2 * rB0;
         rC3_3 += m3; m3 = rA3 * rB0; rB0 = pB0[2];
         rC0_0 += m0; m0 = rA0 * rB1;
         rC1_0 += m1; m1 = rA1 * rB1;
         rC2_0 += m2; m2 = rA2 * rB1;
         rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+2];
         rC0_1 += m0; m0 = rA0 * rB2;
         rC1_1 += m1; m1 = rA1 * rB2;
         rC2_1 += m2; m2 = rA2 * rB2;
         rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+2];
         rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[2];
         rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+2];
         rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+2];
         rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+2];

         rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+2];
         rC1_3 += m1; m1 = rA1 * rB0;
         rC2_3 += m2; m2 = rA2 * rB0;
         rC3_3 += m3; m3 = rA3 * rB0; rB0 = pB0[3];
         rC0_0 += m0; m0 = rA0 * rB1;
         rC1_0 += m1; m1 = rA1 * rB1;
         rC2_0 += m2; m2 = rA2 * rB1;
         rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+3];
         rC0_1 += m0; m0 = rA0 * rB2;
         rC1_1 += m1; m1 = rA1 * rB2;
         rC2_1 += m2; m2 = rA2 * rB2;
         rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+3];
         rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[3];
         rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+3];
         rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+3];
         rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+3];

         rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+3];
         rC1_3 += m1; m1 = rA1 * rB0;
         rC2_3 += m2; m2 = rA2 * rB0;
         rC3_3 += m3; m3 = rA3 * rB0; rB0 = pB0[4];
         rC0_0 += m0; m0 = rA0 * rB1;
         rC1_0 += m1; m1 = rA1 * rB1;
         rC2_0 += m2; m2 = rA2 * rB1;
         rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+4];
         rC0_1 += m0; m0 = rA0 * rB2;
         rC1_1 += m1; m1 = rA1 * rB2;
         rC2_1 += m2; m2 = rA2 * rB2;
         rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+4];
         rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[4];
         rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+4];
         rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+4];
         rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+4];

         rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+4];
         rC1_3 += m1; m1 = rA1 * rB0;
         rC2_3 += m2; m2 = rA2 * rB0;
         rC3_3 += m3; m3 = rA3 * rB0; rB0 = pB0[5];
         rC0_0 += m0; m0 = rA0 * rB1;
         rC1_0 += m1; m1 = rA1 * rB1;
         rC2_0 += m2; m2 = rA2 * rB1;
         rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+5];
         rC0_1 += m0; m0 = rA0 * rB2;
         rC1_1 += m1; m1 = rA1 * rB2;
         rC2_1 += m2; m2 = rA2 * rB2;
         rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+5];
         rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[5];
         rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+5];
         rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+5];
         rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+5];

         rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+5];
         rC1_3 += m1; m1 = rA1 * rB0;
         rC2_3 += m2; m2 = rA2 * rB0;
         rC3_3 += m3; m3 = rA3 * rB0; rB0 = pB0[6];
         rC0_0 += m0; m0 = rA0 * rB1;
         rC1_0 += m1; m1 = rA1 * rB1;
         rC2_0 += m2; m2 = rA2 * rB1;
         rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+6];
         rC0_1 += m0; m0 = rA0 * rB2;
         rC1_1 += m1; m1 = rA1 * rB2;
         rC2_1 += m2; m2 = rA2 * rB2;
         rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+6];
         rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[6];
         rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+6];
         rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+6];
         rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+6];

         rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+6];
         rC1_3 += m1; m1 = rA1 * rB0;
         rC2_3 += m2; m2 = rA2 * rB0;
         rC3_3 += m3; m3 = rA3 * rB0; rB0 = pB0[7];
         rC0_0 += m0; m0 = rA0 * rB1;
         rC1_0 += m1; m1 = rA1 * rB1;
         rC2_0 += m2; m2 = rA2 * rB1;
         rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+7];
         rC0_1 += m0; m0 = rA0 * rB2;
         rC1_1 += m1; m1 = rA1 * rB2;
         rC2_1 += m2; m2 = rA2 * rB2;
         rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+7];
         rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[7];
         rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+7];
         rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+7];
         rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+7];

         rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+7];
         rC1_3 += m1; m1 = rA1 * rB0;                 pB0 += incBm;
         rC2_3 += m2; m2 = rA2 * rB0;                 pA0 += incAm;
         rC3_3 += m3; m3 = rA3 * rB0;
/*
 *       Drain pipe on last iteration of K-loop
 */
         rC0_0 += m0; m0 = rA0 * rB1;
         rC1_0 += m1; m1 = rA1 * rB1;
         rC2_0 += m2; m2 = rA2 * rB1;
         rC3_0 += m3; m3 = rA3 * rB1;
         rC0_1 += m0; m0 = rA0 * rB2;
         rC1_1 += m1; m1 = rA1 * rB2;
         rC2_1 += m2; m2 = rA2 * rB2;
         rC3_1 += m3; m3 = rA3 * rB2;
         rC0_2 += m0; m0 = rA0 * rB3;
         rC1_2 += m1; m1 = rA1 * rB3;
         rC2_2 += m2; m2 = rA2 * rB3;
         rC3_2 += m3; m3 = rA3 * rB3;
         rC0_3 += m0;
         rC1_3 += m1;
         rC2_3 += m2;
         rC3_3 += m3;
         #ifdef TREAL
         *pC0 = rC0_0; pC0[1] = rC1_0; pC0[2] = rC2_0; pC0[3] = rC3_0; pC0 += 4;
         *pC1 = rC0_1; pC1[1] = rC1_1; pC1[2] = rC2_1; pC1[3] = rC3_1; pC1 += 4;
         *pC2 = rC0_2; pC2[1] = rC1_2; pC2[2] = rC2_2; pC2[3] = rC3_2; pC2 += 4;
         *pC3 = rC0_3; pC3[1] = rC1_3; pC3[2] = rC2_3; pC3[3] = rC3_3; pC3 += 4;
         #else
         *pC0 = rC0_0; pC0[2] = rC1_0; pC0[4] = rC2_0; pC0[6] = rC3_0; pC0 += 8;
         *pC1 = rC0_1; pC1[2] = rC1_1; pC1[4] = rC2_1; pC1[6] = rC3_1; pC1 += 8;
         *pC2 = rC0_2; pC2[2] = rC1_2; pC2[4] = rC2_2; pC2[6] = rC3_2; pC2 += 8;
         *pC3 = rC0_3; pC3[2] = rC1_3; pC3[4] = rC2_3; pC3[6] = rC3_3; pC3 += 8;
         #endif
      }
      while(pA0 != stM);
      if (M >= 8)
      {
         #ifdef BETA0
            rC0_0 = rC1_0 = rC2_0 = rC3_0 =
            rC0_1 = rC1_1 = rC2_1 = rC3_1 =
            rC0_2 = rC1_2 = rC2_2 = rC3_2 =
            rC0_3 = rC1_3 = rC2_3 = rC3_3 = ATL_rzero;
         #else
            #ifdef TREAL
               rC0_0 = *pC0; rC0_1 = *pC1; rC0_2 = *pC2; rC0_3 = *pC3;
               rC1_0 = pC0[1]; rC1_1 = pC1[1]; rC1_2 = pC2[1]; rC1_3 = pC3[1];
               rC2_0 = pC0[2]; rC2_1 = pC1[2]; rC2_2 = pC2[2]; rC2_3 = pC3[2];
               rC3_0 = pC0[3]; rC3_1 = pC1[3]; rC3_2 = pC2[3]; rC3_3 = pC3[3];
            #else
               rC0_0 = *pC0; rC0_1 = *pC1; rC0_2 = *pC2; rC0_3 = *pC3;
               rC1_0 = pC0[2]; rC1_1 = pC1[2]; rC1_2 = pC2[2]; rC1_3 = pC3[2];
               rC2_0 = pC0[4]; rC2_1 = pC1[4]; rC2_2 = pC2[4]; rC2_3 = pC3[4];
               rC3_0 = pC0[6]; rC3_1 = pC1[6]; rC3_2 = pC2[6]; rC3_3 = pC3[6];
            #endif
            #ifdef BETAX
               rB3 = beta;
               rC0_0 *= rB3; rC0_1 *= rB3; rC0_2 *= rB3; rC0_3 *= rB3;
               rC1_0 *= rB3; rC1_1 *= rB3; rC1_2 *= rB3; rC1_3 *= rB3;
               rC2_0 *= rB3; rC2_1 *= rB3; rC2_2 *= rB3; rC2_3 *= rB3;
               rC3_0 *= rB3; rC3_1 *= rB3; rC3_2 *= rB3; rC3_3 *= rB3;
            #else
            #endif
         #endif
/*
 *       Start pipeline
 */
         rA0 = *pA0; rB0 = *pB0;
         rA1 = pA0[KB]; rA2 = pA0[KB2];
         m0 = rA0 * rB0; rA3 = pA0[KB3];
         m1 = rA1 * rB0; rB1 = pB0[KB];
         m2 = rA2 * rB0; rB2 = pB0[KB2];
         m3 = rA3 * rB0; rB3 = pB0[KB3];

         for (k=Kstart; k; k--)
         {
            rC0_0 += m0; m0 = rA0 * rB1; rB0 = pB0[1];     ATL_pfl1R(pB0+KB4);
            rC1_0 += m1; m1 = rA1 * rB1;                   ATL_pfl1R(pB0+KB5);
            rC2_0 += m2; m2 = rA2 * rB1;                   ATL_pfl1R(pB0+KB6);
            rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+1];  ATL_pfl1R(pB0+KB7);
            rC0_1 += m0; m0 = rA0 * rB2;
            rC1_1 += m1; m1 = rA1 * rB2;
            rC2_1 += m2; m2 = rA2 * rB2;
            rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+1];
            rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[1];
            rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+1];
            rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+1];
            rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+1];

            rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+1];
            rC1_3 += m1; m1 = rA1 * rB0;
            rC2_3 += m2; m2 = rA2 * rB0;
            rC3_3 += m3; m3 = rA3 * rB0; rB0 = pB0[2];
            rC0_0 += m0; m0 = rA0 * rB1;
            rC1_0 += m1; m1 = rA1 * rB1;
            rC2_0 += m2; m2 = rA2 * rB1;
            rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+2];
            rC0_1 += m0; m0 = rA0 * rB2;
            rC1_1 += m1; m1 = rA1 * rB2;
            rC2_1 += m2; m2 = rA2 * rB2;
            rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+2];
            rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[2];
            rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+2];
            rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+2];
            rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+2];

            rC0_3 += m0; m0 = rA0 * rB0; rB3 = pB0[KB3+2];
            rC1_3 += m1; m1 = rA1 * rB0;
            rC2_3 += m2; m2 = rA2 * rB0;
            rC3_3 += m3; m3 = rA3 * rB0; rB0 = pB0[3];
            rC0_0 += m0; m0 = rA0 * rB1;
            rC1_0 += m1; m1 = rA1 * rB1;
            rC2_0 += m2; m2 = rA2 * rB1;
            rC3_0 += m3; m3 = rA3 * rB1; rB1 = pB0[KB+3];
            rC0_1 += m0; m0 = rA0 * rB2;
            rC1_1 += m1; m1 = rA1 * rB2;
            rC2_1 += m2; m2 = rA2 * rB2;
            rC3_1 += m3; m3 = rA3 * rB2; rB2 = pB0[KB2+3];
            rC0_2 += m0; m0 = rA0 * rB3; rA0 = pA0[3];
            rC1_2 += m1; m1 = rA1 * rB3; rA1 = pA0[KB+3];
            rC2_2 += m2; m2 = rA2 * rB3; rA2 = pA0[KB2+3];
            rC3_2 += m3; m3 = rA3 * rB3; rA3 = pA0[KB3+3];