📄 atl_ncmmjik.c
字号:
const TYPE *B, const int ldb, const SCALAR beta, TYPE *C, const int ldc); if (TA == AtlasNoTrans) { if (TB == AtlasNoTrans) { mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),NN),0x0x0_aX_bX); mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),NN),0x0x0_aX_bX); } else { mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),NT),0x0x0_aX_bX); mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),NT),0x0x0_aX_bX); } incAk = lda * KB; incAm = MB - Kb * incAk; incAn = -Mb * MB; } else { if (TB == AtlasNoTrans) { mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),TN),0x0x0_aX_bX); mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),TN),0x0x0_aX_bX); } else { mm_fixedKcu=Mjoin(Mjoin(Mjoin(NCmm00,Mjoin(0x0x,KB)),TT),0x0x0_aX_bX); mmcu = Mjoin(Mjoin(Mjoin(NCmm00,0x0x0),TT),0x0x0_aX_bX); } incAk = KB; incAm = lda*MB - Kb*KB; incAn = -lda*MB*Mb; } if (TB == AtlasNoTrans) { incBk = KB; incBm = -KB*Kb; incBn = ldb*NB; } else { incBk = KB*ldb; incBm = -Kb * incBk; incBn = NB; } if (alpha == ATL_rone) { if (TA == AtlasNoTrans) { if (TB == AtlasNoTrans) { mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,NN),0x0x0),_a1_b1); if (beta == ATL_rone) mm_bX = mm_b1; else if (beta == ATL_rzero) mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,NN),0x0x0),_a1_b0); else mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,NN),0x0x0),_a1_bX); } else { mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,NT),0x0x0),_a1_b1); if (beta == ATL_rone) mm_bX = mm_b1; else if (beta == ATL_rzero) mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,NT),0x0x0),_a1_b0); else mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,NT),0x0x0),_a1_bX); } } else { if (TB == AtlasNoTrans) { mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_a1_b1); if (beta == ATL_rone) mm_bX = mm_b1; else if (beta == ATL_rzero) mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_a1_b0); else mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_a1_bX); } else { mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_b1); if (beta == ATL_rone) mm_bX = mm_b1; else if (beta == ATL_rzero) mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_b0); else mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_a1_bX); } } } else /* non-one alpha */ { btmp = Mabs(beta); if (btmp < ATL_rone) btmp = 1.0;/* * If needed, call version that uses temp C to handle alpha & beta safely */ if (Kb >= ATL_MaxMMalpha || Mabs(alpha) < btmp) return(Mjoin(PATL,NCmmJIK_c)(TA, TB, M, N, K, alpha, A, lda, B, ldb, beta, C, ldc)); if (TA == AtlasNoTrans) { if (TB == AtlasNoTrans) { mm_bX = mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,NN),0x0x0),_aX_bX); if (beta == ATL_rzero) mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,NN),0x0x0),_aX_b0); } else { mm_bX = mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,NT),0x0x0),_aX_bX); if (beta == ATL_rzero) mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,NT),0x0x0),_aX_b0); } } else { if (TB == AtlasNoTrans) { mm_bX = mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_aX_bX); if (beta == ATL_rzero) mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TN),0x0x0),_aX_b0); } else { mm_bX = mm_b1 = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_bX); if (beta == ATL_rzero) mm_bX = Mjoin(Mjoin(Mjoin(NCmm0,TT),0x0x0),_aX_b0); } } } for (j=Nb; j; j--, a += incAn, b += incBn, c += incCn) { for (i=Mb; i; i--, a += incAm, b += incBm, c += incCm) { if (Kb) { mm_bX(MB, NB, KB, alpha, a, lda, b, ldb, beta, c, ldc); a += incAk; b += incBk; for (k=Kb-1; k; k--, a += incAk, b += incBk) mm_b1(MB, NB, KB, alpha, a, lda, b, ldb, ATL_rone, c, ldc); if (kr) mmcu(MB, NB, kr, alpha, a, lda, b, ldb, ATL_rone, c, ldc); } else if (kr) { if (BetaIsZero) Mjoin(PATL,gezero)(MB, NB, c, ldc); mmcu(MB, NB, kr, alpha, a, lda, b, ldb, beta, c, ldc); } } } if (mr && N != nr) ATL_assert(Mjoin(PATL,NCmmIJK)(TA, TB, mr, N-nr, K, alpha, A+Mb*(incAm+Kb*incAk), lda, B, ldb, beta, C+Mb*MB, ldc) ==0); if (nr) { for (i=Mb; i; i--, a += incAm, b += incBm, c += incCm) { if (BetaIsZero) Mjoin(PATL,gezero)(MB, nr, c, ldc); if (Kb) { mm_fixedKcu(MB, nr, KB, alpha, a, lda, b, ldb, beta, c, ldc); a += incAk; b += incBk; for (k=Kb-1; k; k--, a += incAk, b += incBk) mm_fixedKcu(MB, nr, KB, alpha, a, lda, b, ldb, ATL_rone, c, ldc); if (kr) mmcu(MB, nr, kr, alpha, a, lda, b, ldb, ATL_rone, c, ldc); } else if (kr) mmcu(MB, nr, kr, alpha, a, lda, b, ldb, beta, c, ldc); } if (mr) /* cleanup small mr x nr block of C */ { c = C + Mb*MB + ldc*Nb*NB; a = A + Mb*(incAm+Kb*incAk); b = B + Nb*( incBn+(Mb*(incBm+Kb*incBk)) ); if (BetaIsZero) Mjoin(PATL,gezero)(mr, nr, c, ldc); if (Kb) { mm_fixedKcu(mr, nr, KB, alpha, a, lda, b, ldb, beta, c, ldc); a += incAk; b += incBk; for (k=Kb-1; k; k--, a += incAk, b += incBk) mm_fixedKcu(mr, nr, KB, alpha, a, lda, b, ldb, ATL_rone, c, ldc); if (kr) mmcu(mr, nr, kr, alpha, a, lda, b, ldb, ATL_rone, c, ldc); } else if (kr) mmcu(mr, nr, kr, alpha, a, lda, b, ldb, beta, c, ldc); } } 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 + -