atl_smm4x4x128_av.c

来自「基于Blas CLapck的.用过的人知道是干啥的」· C语言 代码 · 共 3,072 行 · 第 1/5 页

C
3,072
字号
   #endif	vmaddfp	vC10, va1, vb0, vC10   #if KB > 20		addi	k2, k2, 16   #endif	vmaddfp	vC20, va2, vb0, vC20   #if KB > 20		addi	k3, k3, 16   #endif	vmaddfp	vC30, va3, vb0, vC30   #if KB > 20		lvx vb0, pB0, k0   #endif	vmaddfp	vC01, va0, vb1, vC01	vmaddfp	vC11, va1, vb1, vC11	vmaddfp	vC21, va2, vb1, vC21	vmaddfp	vC31, va3, vb1, vC31   #if KB > 20		lvx vb1, pB0, k1   #endif	vmaddfp	vC02, va0, vb2, vC02	vmaddfp	vC12, va1, vb2, vC12	vmaddfp	vC22, va2, vb2, vC22	vmaddfp	vC32, va3, vb2, vC32   #if KB > 20		lvx vb2, pB0, k2   #endif	vmaddfp	vC03, va0, vb3, vC03   #if KB > 20		lvx va0, pA0, k0   #endif	vmaddfp	vC13, va1, vb3, vC13   #if KB > 20		lvx va1, pA0, k1   #endif	vmaddfp	vC23, va2, vb3, vC23   #if KB > 20		lvx va2, pA0, k2   #endif	vmaddfp	vC33, va3, vb3, vC33   #if KB > 20		lvx va3, pA0, k3   #endif   #if KB > 20		lvx vb3, pB0, k3   #endif#endif  /* end K=12 block */#if KB > 16   #if KB > 24		addi	k0, k0, 16   #endif	vmaddfp	vC00, vA0, vB0, vC00   #if KB > 24		addi	k1, k1, 16   #endif	vmaddfp	vC10, vA1, vB0, vC10   #if KB > 24		addi	k2, k2, 16   #endif	vmaddfp	vC20, vA2, vB0, vC20   #if KB > 24		addi	k3, k3, 16   #endif	vmaddfp	vC30, vA3, vB0, vC30   #if KB > 24		lvx vB0, pB0, k0   #endif	vmaddfp	vC01, vA0, vB1, vC01	vmaddfp	vC11, vA1, vB1, vC11	vmaddfp	vC21, vA2, vB1, vC21	vmaddfp	vC31, vA3, vB1, vC31   #if KB > 24		lvx vB1, pB0, k1   #endif	vmaddfp	vC02, vA0, vB2, vC02	vmaddfp	vC12, vA1, vB2, vC12	vmaddfp	vC22, vA2, vB2, vC22	vmaddfp	vC32, vA3, vB2, vC32   #if KB > 24		lvx vB2, pB0, k2   #endif	vmaddfp	vC03, vA0, vB3, vC03   #if KB > 24		lvx vA0, pA0, k0   #endif	vmaddfp	vC13, vA1, vB3, vC13   #if KB > 24		lvx vA1, pA0, k1   #endif	vmaddfp	vC23, vA2, vB3, vC23   #if KB > 24		lvx vA2, pA0, k2   #endif	vmaddfp	vC33, vA3, vB3, vC33   #if KB > 24		lvx vA3, pA0, k3   #endif   #if KB > 24		lvx vB3, pB0, k3   #endif#endif  /* end K=16 block */#if KB > 20   #if KB > 28		addi	k0, k0, 16   #endif	vmaddfp	vC00, va0, vb0, vC00   #if KB > 28		addi	k1, k1, 16   #endif	vmaddfp	vC10, va1, vb0, vC10   #if KB > 28		addi	k2, k2, 16   #endif	vmaddfp	vC20, va2, vb0, vC20   #if KB > 28		addi	k3, k3, 16   #endif	vmaddfp	vC30, va3, vb0, vC30   #if KB > 28		lvx vb0, pB0, k0   #endif	vmaddfp	vC01, va0, vb1, vC01	vmaddfp	vC11, va1, vb1, vC11	vmaddfp	vC21, va2, vb1, vC21	vmaddfp	vC31, va3, vb1, vC31   #if KB > 28		lvx vb1, pB0, k1   #endif	vmaddfp	vC02, va0, vb2, vC02	vmaddfp	vC12, va1, vb2, vC12	vmaddfp	vC22, va2, vb2, vC22	vmaddfp	vC32, va3, vb2, vC32   #if KB > 28		lvx vb2, pB0, k2   #endif	vmaddfp	vC03, va0, vb3, vC03   #if KB > 28		lvx va0, pA0, k0   #endif	vmaddfp	vC13, va1, vb3, vC13   #if KB > 28		lvx va1, pA0, k1   #endif	vmaddfp	vC23, va2, vb3, vC23   #if KB > 28		lvx va2, pA0, k2   #endif	vmaddfp	vC33, va3, vb3, vC33   #if KB > 28		lvx va3, pA0, k3   #endif   #if KB > 28		lvx vb3, pB0, k3   #endif#endif  /* end K=20 block */#if KB > 24   #if KB > 32		addi	k0, k0, 16   #endif	vmaddfp	vC00, vA0, vB0, vC00   #if KB > 32		addi	k1, k1, 16   #endif	vmaddfp	vC10, vA1, vB0, vC10   #if KB > 32		addi	k2, k2, 16   #endif	vmaddfp	vC20, vA2, vB0, vC20   #if KB > 32		addi	k3, k3, 16   #endif	vmaddfp	vC30, vA3, vB0, vC30   #if KB > 32		lvx vB0, pB0, k0   #endif	vmaddfp	vC01, vA0, vB1, vC01	vmaddfp	vC11, vA1, vB1, vC11	vmaddfp	vC21, vA2, vB1, vC21	vmaddfp	vC31, vA3, vB1, vC31   #if KB > 32		lvx vB1, pB0, k1   #endif	vmaddfp	vC02, vA0, vB2, vC02	vmaddfp	vC12, vA1, vB2, vC12	vmaddfp	vC22, vA2, vB2, vC22	vmaddfp	vC32, vA3, vB2, vC32   #if KB > 32		lvx vB2, pB0, k2   #endif	vmaddfp	vC03, vA0, vB3, vC03   #if KB > 32		lvx vA0, pA0, k0   #endif	vmaddfp	vC13, vA1, vB3, vC13   #if KB > 32		lvx vA1, pA0, k1   #endif	vmaddfp	vC23, vA2, vB3, vC23   #if KB > 32		lvx vA2, pA0, k2   #endif	vmaddfp	vC33, vA3, vB3, vC33   #if KB > 32		lvx vA3, pA0, k3   #endif   #if KB > 32		lvx vB3, pB0, k3   #endif#endif  /* end K=24 block */#if KB > 28   #if KB > 36		addi	k0, k0, 16   #endif	vmaddfp	vC00, va0, vb0, vC00   #if KB > 36		addi	k1, k1, 16   #endif	vmaddfp	vC10, va1, vb0, vC10   #if KB > 36		addi	k2, k2, 16   #endif	vmaddfp	vC20, va2, vb0, vC20   #if KB > 36		addi	k3, k3, 16   #endif	vmaddfp	vC30, va3, vb0, vC30   #if KB > 36		lvx vb0, pB0, k0   #endif	vmaddfp	vC01, va0, vb1, vC01	vmaddfp	vC11, va1, vb1, vC11	vmaddfp	vC21, va2, vb1, vC21	vmaddfp	vC31, va3, vb1, vC31   #if KB > 36		lvx vb1, pB0, k1   #endif	vmaddfp	vC02, va0, vb2, vC02	vmaddfp	vC12, va1, vb2, vC12                        dcbt    0, pfB, 0	vmaddfp	vC22, va2, vb2, vC22                        addi    pfB, pfB, 64	vmaddfp	vC32, va3, vb2, vC32   #if KB > 36		lvx vb2, pB0, k2   #endif	vmaddfp	vC03, va0, vb3, vC03   #if KB > 36		lvx va0, pA0, k0   #endif	vmaddfp	vC13, va1, vb3, vC13   #if KB > 36		lvx va1, pA0, k1   #endif	vmaddfp	vC23, va2, vb3, vC23   #if KB > 36		lvx va2, pA0, k2   #endif	vmaddfp	vC33, va3, vb3, vC33   #if KB > 36		lvx va3, pA0, k3   #endif   #if KB > 36		lvx vb3, pB0, k3   #endif#endif  /* end K=28 block */#if KB > 32   #if KB > 40		addi	k0, k0, 16   #endif	vmaddfp	vC00, vA0, vB0, vC00   #if KB > 40		addi	k1, k1, 16   #endif	vmaddfp	vC10, vA1, vB0, vC10   #if KB > 40		addi	k2, k2, 16   #endif	vmaddfp	vC20, vA2, vB0, vC20   #if KB > 40		addi	k3, k3, 16   #endif	vmaddfp	vC30, vA3, vB0, vC30   #if KB > 40		lvx vB0, pB0, k0   #endif	vmaddfp	vC01, vA0, vB1, vC01	vmaddfp	vC11, vA1, vB1, vC11	vmaddfp	vC21, vA2, vB1, vC21	vmaddfp	vC31, vA3, vB1, vC31   #if KB > 40		lvx vB1, pB0, k1   #endif	vmaddfp	vC02, vA0, vB2, vC02	vmaddfp	vC12, vA1, vB2, vC12	vmaddfp	vC22, vA2, vB2, vC22	vmaddfp	vC32, vA3, vB2, vC32   #if KB > 40		lvx vB2, pB0, k2   #endif	vmaddfp	vC03, vA0, vB3, vC03   #if KB > 40		lvx vA0, pA0, k0   #endif	vmaddfp	vC13, vA1, vB3, vC13   #if KB > 40		lvx vA1, pA0, k1   #endif	vmaddfp	vC23, vA2, vB3, vC23   #if KB > 40		lvx vA2, pA0, k2   #endif	vmaddfp	vC33, vA3, vB3, vC33   #if KB > 40		lvx vA3, pA0, k3   #endif   #if KB > 40		lvx vB3, pB0, k3   #endif#endif  /* end K=32 block */#if KB > 36   #if KB > 44		addi	k0, k0, 16   #endif	vmaddfp	vC00, va0, vb0, vC00   #if KB > 44		addi	k1, k1, 16   #endif	vmaddfp	vC10, va1, vb0, vC10   #if KB > 44		addi	k2, k2, 16   #endif	vmaddfp	vC20, va2, vb0, vC20   #if KB > 44		addi	k3, k3, 16   #endif	vmaddfp	vC30, va3, vb0, vC30   #if KB > 44		lvx vb0, pB0, k0   #endif	vmaddfp	vC01, va0, vb1, vC01	vmaddfp	vC11, va1, vb1, vC11	vmaddfp	vC21, va2, vb1, vC21	vmaddfp	vC31, va3, vb1, vC31   #if KB > 44		lvx vb1, pB0, k1   #endif	vmaddfp	vC02, va0, vb2, vC02	vmaddfp	vC12, va1, vb2, vC12	vmaddfp	vC22, va2, vb2, vC22	vmaddfp	vC32, va3, vb2, vC32   #if KB > 44		lvx vb2, pB0, k2   #endif	vmaddfp	vC03, va0, vb3, vC03   #if KB > 44		lvx va0, pA0, k0   #endif	vmaddfp	vC13, va1, vb3, vC13   #if KB > 44		lvx va1, pA0, k1   #endif	vmaddfp	vC23, va2, vb3, vC23   #if KB > 44		lvx va2, pA0, k2   #endif	vmaddfp	vC33, va3, vb3, vC33   #if KB > 44		lvx va3, pA0, k3   #endif   #if KB > 44		lvx vb3, pB0, k3   #endif#endif  /* end K=36 block */#if KB > 40   #if KB > 48		addi	k0, k0, 16   #endif	vmaddfp	vC00, vA0, vB0, vC00   #if KB > 48		addi	k1, k1, 16   #endif	vmaddfp	vC10, vA1, vB0, vC10   #if KB > 48		addi	k2, k2, 16   #endif	vmaddfp	vC20, vA2, vB0, vC20   #if KB > 48		addi	k3, k3, 16   #endif	vmaddfp	vC30, vA3, vB0, vC30   #if KB > 48		lvx vB0, pB0, k0   #endif	vmaddfp	vC01, vA0, vB1, vC01	vmaddfp	vC11, vA1, vB1, vC11	vmaddfp	vC21, vA2, vB1, vC21	vmaddfp	vC31, vA3, vB1, vC31   #if KB > 48		lvx vB1, pB0, k1   #endif	vmaddfp	vC02, vA0, vB2, vC02	vmaddfp	vC12, vA1, vB2, vC12	vmaddfp	vC22, vA2, vB2, vC22	vmaddfp	vC32, vA3, vB2, vC32   #if KB > 48		lvx vB2, pB0, k2   #endif	vmaddfp	vC03, vA0, vB3, vC03   #if KB > 48		lvx vA0, pA0, k0   #endif	vmaddfp	vC13, vA1, vB3, vC13   #if KB > 48		lvx vA1, pA0, k1   #endif	vmaddfp	vC23, vA2, vB3, vC23   #if KB > 48		lvx vA2, pA0, k2   #endif	vmaddfp	vC33, vA3, vB3, vC33   #if KB > 48		lvx vA3, pA0, k3   #endif   #if KB > 48		lvx vB3, pB0, k3   #endif#endif  /* end K=40 block */#if KB > 44   #if KB > 52		addi	k0, k0, 16   #endif	vmaddfp	vC00, va0, vb0, vC00   #if KB > 52		addi	k1, k1, 16   #endif	vmaddfp	vC10, va1, vb0, vC10   #if KB > 52		addi	k2, k2, 16   #endif	vmaddfp	vC20, va2, vb0, vC20   #if KB > 52		addi	k3, k3, 16   #endif	vmaddfp	vC30, va3, vb0, vC30   #if KB > 52		lvx vb0, pB0, k0   #endif	vmaddfp	vC01, va0, vb1, vC01	vmaddfp	vC11, va1, vb1, vC11	vmaddfp	vC21, va2, vb1, vC21	vmaddfp	vC31, va3, vb1, vC31   #if KB > 52		lvx vb1, pB0, k1   #endif	vmaddfp	vC02, va0, vb2, vC02	vmaddfp	vC12, va1, vb2, vC12	vmaddfp	vC22, va2, vb2, vC22	vmaddfp	vC32, va3, vb2, vC32   #if KB > 52		lvx vb2, pB0, k2   #endif	vmaddfp	vC03, va0, vb3, vC03   #if KB > 52		lvx va0, pA0, k0   #endif	vmaddfp	vC13, va1, vb3, vC13   #if KB > 52		lvx va1, pA0, k1   #endif	vmaddfp	vC23, va2, vb3, vC23   #if KB > 52		lvx va2, pA0, k2   #endif	vmaddfp	vC33, va3, vb3, vC33   #if KB > 52		lvx va3, pA0, k3   #endif   #if KB > 52		lvx vb3, pB0, k3   #endif#endif  /* end K=44 block */#if KB > 48   #if KB > 56		addi	k0, k0, 16   #endif	vmaddfp	vC00, vA0, vB0, vC00   #if KB > 56		addi	k1, k1, 16   #endif	vmaddfp	vC10, vA1, vB0, vC10   #if KB > 56		addi	k2, k2, 16   #endif	vmaddfp	vC20, vA2, vB0, vC20   #if KB > 56		addi	k3, k3, 16   #endif	vmaddfp	vC30, vA3, vB0, vC30   #if KB > 56		lvx vB0, pB0, k0   #endif	vmaddfp	vC01, vA0, vB1, vC01	vmaddfp	vC11, vA1, vB1, vC11	vmaddfp	vC21, vA2, vB1, vC21	vmaddfp	vC31, vA3, vB1, vC31   #if KB > 56		lvx vB1, pB0, k1   #endif	vmaddfp	vC02, vA0, vB2, vC02	vmaddfp	vC12, vA1, vB2, vC12	vmaddfp	vC22, vA2, vB2, vC22	vmaddfp	vC32, vA3, vB2, vC32   #if KB > 56		lvx vB2, pB0, k2   #endif	vmaddfp	vC03, vA0, vB3, vC03   #if KB > 56		lvx vA0, pA0, k0   #endif	vmaddfp	vC13, vA1, vB3, vC13   #if KB > 56		lvx vA1, pA0, k1   #endif	vmaddfp	vC23, vA2, vB3, vC23   #if KB > 56		lvx vA2, pA0, k2   #endif	vmaddfp	vC33, vA3, vB3, vC33   #if KB > 56		lvx vA3, pA0, k3   #endif   #if KB > 56		lvx vB3, pB0, k3   #endif#endif  /* end K=48 block */#if KB > 52   #if KB > 60		addi	k0, k0, 16   #endif	vmaddfp	vC00, va0, vb0, vC00   #if KB > 60		addi	k1, k1, 16   #endif	vmaddfp	vC10, va1, vb0, vC10   #if KB > 60		addi	k2, k2, 16   #endif	vmaddfp	vC20, va2, vb0, vC20   #if KB > 60		addi	k3, k3, 16   #endif	vmaddfp	vC30, va3, vb0, vC30   #if KB > 60		lvx vb0, pB0, k0   #endif	vmaddfp	vC01, va0, vb1, vC01	vmaddfp	vC11, va1, vb1, vC11	vmaddfp	vC21, va2, vb1, vC21	vmaddfp	vC31, va3, vb1, vC31   #if KB > 60		lvx vb1, pB0, k1   #endif	vmaddfp	vC02, va0, vb2, vC02	vmaddfp	vC12, va1, vb2, vC12	vmaddfp	vC22, va2, vb2, vC22	vmaddfp	vC32, va3, vb2, vC32   #if KB > 60		lvx vb2, pB0, k2   #endif	vmaddfp	vC03, va0, vb3, vC03   #if KB > 60		lvx va0, pA0, k0   #endif	vmaddfp	vC13, va1, vb3, vC13   #if KB > 60		lvx va1, pA0, k1   #endif	vmaddfp	vC23, va2, vb3, vC23   #if KB > 60		lvx va2, pA0, k2   #endif	vmaddfp	vC33, va3, vb3, vC33   #if KB > 60		lvx va3, pA0, k3   #endif   #if KB > 60		lvx vb3, pB0, k3   #endif#endif  /* end K=52 block */#if KB > 56   #if KB > 64		addi	k0, k0, 16   #endif	vmaddfp	vC00, vA0, vB0, vC00   #if KB > 64		addi	k1, k1, 16   #endif	vmaddfp	vC10, vA1, vB0, vC10   #if KB > 64		addi	k2, k2, 16   #endif	vmaddfp	vC20, vA2, vB0, vC20   #if KB > 64		addi	k3, k3, 16   #endif	vmaddfp	vC30, vA3, vB0, vC30   #if KB > 64		lvx vB0, pB0, k0   #endif	vmaddfp	vC01, vA0, vB1, vC01	vmaddfp	vC11, vA1, vB1, vC11	vmaddfp	vC21, vA2, vB1, v

⌨️ 快捷键说明

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