335     uint32_t x1 = xstart;  in kernelU4()  local
339     if (gArchUseSIMD && x2 > x1 && cp->scaleX < 4.0f) {  in kernelU4()
340         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelU4()
346         int len = x2 - x1;  in kernelU4()
361         x1 += len;  in kernelU4()
365     while(x1 < x2) {  in kernelU4()
366         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelU4()
369         x1++;  in kernelU4()
402     uint32_t x1 = xstart;  in kernelU2()  local
406     if (gArchUseSIMD && x2 > x1 && cp->scaleX < 4.0f) {  in kernelU2()
407         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelU2()
413         int len = x2 - x1;  in kernelU2()
428         x1 += len;  in kernelU2()
432     while(x1 < x2) {  in kernelU2()
433         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelU2()
436         x1++;  in kernelU2()
469     uint32_t x1 = xstart;  in kernelU1()  local
473     if (gArchUseSIMD && x2 > x1 && cp->scaleX < 4.0f) {  in kernelU1()
474         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelU1()
480         int len = x2 - x1;  in kernelU1()
495         x1 += len;  in kernelU1()
499     while(x1 < x2) {  in kernelU1()
500         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelU1()
503         x1++;  in kernelU1()
536     uint32_t x1 = xstart;  in kernelF4()  local
539     while(x1 < x2) {  in kernelF4()
540         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelF4()
543         x1++;  in kernelF4()
576     uint32_t x1 = xstart;  in kernelF2()  local
579     while(x1 < x2) {  in kernelF2()
580         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelF2()
583         x1++;  in kernelF2()
616     uint32_t x1 = xstart;  in kernelF1()  local
619     while(x1 < x2) {  in kernelF1()
620         float xf = (x1 + 0.5f) * cp->scaleX - 0.5f;  in kernelF1()
623         x1++;  in kernelF1()