⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 atl_cncmmjik_c.c

📁 基于Blas CLapck的.用过的人知道是干啥的
💻 C
📖 第 1 页 / 共 2 页
字号:
         mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),NT),0x0x0_aX_bX);      }   }   else if (TA == AtlasConjTrans)   {      if (TB == AtlasNoTrans)      {         ai0 = ar0 = ATL_rnone; ai1 = ATL_rone;         r0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_aX_b0);         r0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_aX_bX);         r1mm    = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_a1_bX);         i0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_aX_b0);         i0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_aX_bX);         i1mm    = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_a1_b1);         mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),TN),0x0x0_aX_bX);         mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),TN),0x0x0_aX_bX);      }      else if (TB == AtlasConjTrans)      {         ar0 = ATL_rone; ai1 = ai0 = ATL_rnone;         r0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_b0);         r0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_bX);         r1mm    = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_bX);         i0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_b0);         i0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_bX);         i1mm    = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_bX);         mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),TT),0x0x0_aX_bX);         mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),TT),0x0x0_aX_bX);      }      else  /* TA == AtlasConjTrans, TB == AtlasTrans */      {         ai0 = ar0 = ATL_rnone; ai1 = ATL_rone;         r0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_b0);         r0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_bX);         r1mm    = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_bX);         i0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_b0);         i0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_bX);         i1mm    = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_b1);         mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),TT),0x0x0_aX_bX);         mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),TT),0x0x0_aX_bX);      }   }   else   {      if (TB == AtlasNoTrans)      {         ar0 = ai0 = ai1 = ATL_rone;         i0mm_bX = r0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_a1_b0);         i1mm = i0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_a1_b1);         r0mm_b1 = r1mm = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_a1_bX);         mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),TN),0x0x0_aX_bX);         mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),TN),0x0x0_aX_bX);      }      else if (TB == AtlasConjTrans)      {         ai1 = ar0 = ATL_rnone; ai0 = ATL_rone;         r0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_b0);         r0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_bX);         r1mm    = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_bX);         i0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_b0);         i0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_b1);         i1mm    = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_bX);         mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),TT),0x0x0_aX_bX);         mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),TT),0x0x0_aX_bX);      }      else      {         ar0 = ai0 = ai1 = ATL_rone;         i0mm_bX = r0mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_b0);         i1mm = i0mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_b1);         r0mm_b1 = r1mm = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_bX);         mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),TT),0x0x0_aX_bX);         mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),TT),0x0x0_aX_bX);      }   }   if (TA == AtlasNoTrans)   {      incAk = lda * (KB<<1);      incAm = (MB<<1) - Kb * incAk;      incAn = -Mb * (MB<<1);   }   else   {      incAk = KB<<1;      incAm = (lda*MB - Kb*KB)<<1;      incAn = -lda*(MB<<1)*Mb;   }   if (TB == AtlasNoTrans)   {      incBk = KB<<1;      incBm = -(KB<<1)*Kb;      incBn = ldb*(NB<<1);   }   else   {      incBk = (KB<<1)*ldb;      incBm = -Kb * incBk;      incBn = NB<<1;   }   if (AlphaIsOne)   {      if (BetaIsOne) geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_a1),_b1);      else if (BetaIsZero) geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_a1),_b0);      else if (BetaIsReal) geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_a1),_bXi0);      else geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_a1),_bX);   }   else if (AlphaIsReal)   {      if (BetaIsOne) geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_aXi0),_b1);      else if (BetaIsZero) geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_aXi0),_b0);      else if (BetaIsReal) geadd=Mjoin(Mjoin(Mjoin(PATL,geadd),_aXi0),_bXi0);      else geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_aXi0),_bX);   }   else if (BetaIsOne) geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_aX),_b1);   else if (BetaIsZero) geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_aX),_b0);   else if (BetaIsReal) geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_aX),_bXi0);   else geadd = Mjoin(Mjoin(Mjoin(PATL,geadd),_aX),_bX);   vp = malloc(ATL_Cachelen + ATL_MulBySize(MB * NB));   ATL_assert(vp);   cp = ATL_AlignPtr(vp);   if (mr || nr || kr) Mjoin(PATL,gezero)(MB, NB, cp, MB);   for (j=Nb; j; j--, a += incAn, b += incBn, c += incCn)   {      for (i=Mb; i; i--, a += incAm, b += incBm, c += incCm)      {         if (Kb)         {            r0mm_bX(MB, NB, KB, ar0, a+1, lda, b+1, ldb, ATL_rzero, cp, MB);            i0mm_bX(MB, NB, KB, ai0, a+1, lda, b, ldb, ATL_rzero, cp+1, MB);            r1mm(MB, NB, KB, ATL_rone, a, lda, b, ldb, ATL_rnone, cp, MB);            i1mm(MB, NB, KB, ai1, a, lda, b+1, ldb, ATL_rone, cp+1, MB);            a += incAk;  b += incBk;            for (k=Kb-1; k; k--, a += incAk, b += incBk)            {               r0mm_b1(MB, NB, KB, ar0, a+1, lda, b+1, ldb, ATL_rnone, cp, MB);               i0mm_b1(MB, NB, KB, ai0, a+1, lda, b, ldb, ATL_rone, cp+1, MB);               r1mm(MB, NB, KB, ATL_rone, a, lda, b, ldb, ATL_rnone, cp, MB);               i1mm(MB, NB, KB, ai1, a, lda, b+1, ldb, ATL_rone, cp+1, MB);            }            if (kr)            {               mmcu(MB, NB, kr, ar0, a+1, lda, b+1, ldb, ATL_rnone, cp, MB);               mmcu(MB, NB, kr, ai0, a+1, lda, b, ldb, ATL_rone, cp+1, MB);               mmcu(MB, NB, kr, ATL_rone, a, lda, b, ldb, ATL_rnone, cp, MB);               mmcu(MB, NB, kr, ai1, a, lda, b+1, ldb, ATL_rone, cp+1, MB);            }         }         else if (kr)         {            Mjoin(PATL,zero)(MB*NB, cp, 1); /* kill NaN/INF from before */            mmcu(MB, NB, kr, ar0, a+1, lda, b+1, ldb, ATL_rzero, cp, MB);            mmcu(MB, NB, kr, ai0, a+1, lda, b, ldb, ATL_rzero, cp+1, MB);            mmcu(MB, NB, kr, ATL_rone, a, lda, b, ldb, ATL_rnone, cp, MB);            mmcu(MB, NB, kr, ai1, a, lda, b+1, ldb, ATL_rone, cp+1, MB);         }         geadd(MB, NB, alpha, cp, MB, beta, c, ldc);      }   }   if (mr && N != nr)      Mjoin(PATL,NCmmIJK)(TA, TB, mr, N-nr, K, alpha, A+Mb*(incAm+Kb*incAk),                          lda, B, ldb, beta, C+Mb*(MB<<1), ldc);   if (nr)   {      for (i=Mb; i; i--, a += incAm, b += incBm, c += incCm)      {            Mjoin(PATL,zero)(MB*nr, cp, 1); /* kill NaN/INF from before */         if (Kb)         {            mm_fixedKcu(MB, nr, KB, ar0, a+1, lda, b+1, ldb, ATL_rzero,                        cp, MB);            mm_fixedKcu(MB, nr, KB, ai0, a+1, lda, b, ldb, ATL_rzero,                        cp+1, MB);            mm_fixedKcu(MB, nr, KB, ATL_rone, a, lda, b, ldb, ATL_rnone,                        cp, MB);            mm_fixedKcu(MB, nr, KB, ai1, a, lda, b+1, ldb, ATL_rone,                        cp+1, MB);            a += incAk;  b += incBk;            for (k=Kb-1; k; k--, a += incAk, b += incBk)            {               mm_fixedKcu(MB, nr, KB, ar0, a+1, lda, b+1, ldb, ATL_rnone,                           cp, MB);               mm_fixedKcu(MB, nr, KB, ai0, a+1, lda, b, ldb, ATL_rone,                           cp+1, MB);               mm_fixedKcu(MB, nr, KB, ATL_rone, a, lda, b, ldb, ATL_rnone,                           cp, MB);               mm_fixedKcu(MB, nr, KB, ai1, a, lda, b+1, ldb, ATL_rone,                           cp+1, MB);            }            if (kr)            {               mmcu(MB, nr, kr, ar0, a+1, lda, b+1, ldb, ATL_rnone, cp, MB);               mmcu(MB, nr, kr, ai0, a+1, lda, b, ldb, ATL_rone, cp+1, MB);               mmcu(MB, nr, kr, ATL_rone, a, lda, b, ldb, ATL_rnone, cp, MB);               mmcu(MB, nr, kr, ai1, a, lda, b+1, ldb, ATL_rone, cp+1, MB);            }         }         else if (kr)         {            mmcu(MB, nr, kr, ar0, a+1, lda, b+1, ldb, ATL_rzero, cp, MB);            mmcu(MB, nr, kr, ai0, a+1, lda, b, ldb, ATL_rzero, cp+1, MB);            mmcu(MB, nr, kr, ATL_rone, a, lda, b, ldb, ATL_rnone, cp, MB);            mmcu(MB, nr, kr, ai1, a, lda, b+1, ldb, ATL_rone, cp+1, MB);         }         geadd(MB, nr, alpha, cp, MB, beta, c, ldc);      }      if (mr)  /* cleanup small mr x nr block of C */      {         c = C + ((Mb*MB + ldc*Nb*NB)<<1);         a = A + Mb*(incAm+Kb*incAk);         b = B + Nb*( incBn+(Mb*(incBm+Kb*incBk)) );         Mjoin(PATL,zero)(MB*nr, cp, 1); /* kill NaN/INF from before */         if (Kb)         {            mm_fixedKcu(mr, nr, KB, ar0, a+1, lda, b+1, ldb, ATL_rzero,                        cp, MB);            mm_fixedKcu(mr, nr, KB, ai0, a+1, lda, b, ldb, ATL_rzero,                        cp+1, MB);            mm_fixedKcu(mr, nr, KB, ATL_rone, a, lda, b, ldb, ATL_rnone,                        cp, MB);            mm_fixedKcu(mr, nr, KB, ai1, a, lda, b+1, ldb, ATL_rone,                        cp+1, MB);            a += incAk;  b += incBk;            for (k=Kb-1; k; k--, a += incAk, b += incBk)            {               mm_fixedKcu(mr, nr, KB, ar0, a+1, lda, b+1, ldb, ATL_rnone,                           cp, MB);               mm_fixedKcu(mr, nr, KB, ai0, a+1, lda, b, ldb, ATL_rone,                           cp+1, MB);               mm_fixedKcu(mr, nr, KB, ATL_rone, a, lda, b, ldb, ATL_rnone,                           cp, MB);               mm_fixedKcu(mr, nr, KB, ai1, a, lda, b+1, ldb, ATL_rone,                           cp+1, MB);            }            if (kr)            {               mmcu(mr, nr, kr, ar0, a+1, lda, b+1, ldb, ATL_rnone, cp, MB);               mmcu(mr, nr, kr, ai0, a+1, lda, b, ldb, ATL_rone, cp+1, MB);               mmcu(mr, nr, kr, ATL_rone, a, lda, b, ldb, ATL_rnone, cp, MB);               mmcu(mr, nr, kr, ai1, a, lda, b+1, ldb, ATL_rone, cp+1, MB);            }         }         else if (kr)         {            mmcu(mr, nr, kr, ar0, a+1, lda, b+1, ldb, ATL_rzero, cp, MB);            mmcu(mr, nr, kr, ai0, a+1, lda, b, ldb, ATL_rzero, cp+1, MB);            mmcu(mr, nr, kr, ATL_rone, a, lda, b, ldb, ATL_rnone, cp, MB);            mmcu(mr, nr, kr, ai1, a, lda, b+1, ldb, ATL_rone, cp+1, MB);         }         geadd(mr, nr, alpha, cp, MB, beta, c, ldc);      }   }   free(vp);   return(0);}

⌨️ 快捷键说明

复制代码 Ctrl + C
搜索代码 Ctrl + F
全屏模式 F11
切换主题 Ctrl + Shift + D
显示快捷键 ?
增大字号 Ctrl + =
减小字号 Ctrl + -