Lines Matching refs:ysum

468     const Xbyak::Ymm &ysum = this->ymm9;  in generate()  local
516 this->vmulps(ysum, yc, yc); in generate()
517 this->vfmadd231ps(ysum, ya, ya); // ysum <- ysum + ya*ya in generate()
518 this->vfmadd231ps(ysum, yb, yb); in generate()
519 this->vfmadd231ps(ysum, yd, yd); in generate()
520 this->vfmadd231ps(ysum, ye, ye); in generate()
521 this->vfmadd132ps(ysum, yk_, valpha_); // ysum <- ysum*valpha_+yk_ in generate()
523 this->vmovaps(ybase, ysum); in generate()
526 this->vmulps(ysum2, ysum, ysum); in generate()
527 this->vmulps(ysum, ysum, ysum2); // ysum = ybase^3; in generate()
528 this->vsqrtps(ysum, ysum); in generate()
529 this->vsqrtps(ysum, ysum); // ysum = ybase^0.75 in generate()
530 this->vdivps(ydst, ysrc, ysum); // ydst = ysrc / ysum in generate()
706 const Xbyak::Ymm &ysum = this->ymm7; in generate() local
725 this->vxorps(ysum, ysum, ysum); in generate()
730 this->vfmadd231ps(ysum, ya, ya); // ysum <- ysum + ya^2+yb^2+yc^2+yd^2+ye^2 in generate()
735 this->vfmadd231ps(ysum, yb, yb); in generate()
744 this->vfmadd231ps(ysum, yc, yc); in generate()
745 this->vfmadd231ps(ysum, yd, yd); in generate()
746 this->vfmadd231ps(ysum, ye, ye); in generate()
748 this->vmovups(ydst, ysum); in generate()
762 this->vxorps(ysum, ysum, ysum); in generate()
769 this->vfmadd231ps(ysum, ya, ya); in generate()
771 this->vfmadd231ps(ysum, yb, yb); in generate()
778 this->vfmadd231ps(ysum, yc, yc); in generate()
783 this->vfmadd231ps(ysum, yd, yd); // ysum <- ysum + ya^2+yb^2+yc^2+yd^2+ye^2 in generate()
788 this->vfmadd231ps(ysum, ye, ye); in generate()
790 this->vmovups(ydst, ysum); in generate()
1038 Xbyak::Ymm ysum) {} in nchw_body() argument
1043 Xbyak::Ymm yc, Xbyak::Ymm yd, Xbyak::Ymm ye, Xbyak::Ymm ysum) { in nchw_body() argument
1047 this->vfmadd231ps(ysum, ye, ye); in nchw_body()
1049 this->vmovups(ydst, ysum); in nchw_body()
1070 this->vfnmadd231ps(ysum, ya, ya); in nchw_body()
1236 const Xbyak::Ymm &ysum = this->ymm8; in generate() local
1269 this->vxorps(ysum, ysum, ysum); in generate()
1270 this->vfmadd231ps(ysum, yc, yc); // ysum <- ysum + ya^2+yb^2+yc^2+yd^2+ye^2 in generate()
1271 this->vfmadd231ps(ysum, yd, yd); in generate()
1282 nchw_body(J.tail, J.HW, pk_, ymask, ya, yb, yc, yd, ye, ysum); in generate()
1293 nchw_body(J.tail, J.HW, pk_, ymask, ya, yb, yc, yd, ye, ysum); in generate()
1298 nchw_body(J.tail, J.HW, pk_, ymask, ya, yb, yc, yd, ye, ysum); in generate()
1514 const Xbyak::Ymm &ysum = this->ymm14; in generate() local
1571 this->vdivps(ysum, ydiffsrc, yws); in generate()
1572 this->vmulps(ysum, ysum, ysrc); in generate()
1589 this->vmovups(this->ptr[t + 16], ysum); in generate()
1595 this->vaddps(ysum, ysum, ya); in generate()
1597 this->vaddps(ysum, ysum, yb); in generate()
1601 this->vaddps(ysum, ysum, yd); in generate()
1602 this->vaddps(ysum, ysum, ye); in generate()
1604 this->vfmadd231ps(ydiffsrc, ysum, ysrc); in generate()