361   uint8_t* ptmp  = src - 4;  in HFilter16()  local
371   LD_UB8(ptmp, stride, row0, row1, row2, row3, row4, row5, row6, row7);  in HFilter16()
372   ptmp += (8 * stride);  in HFilter16()
373   LD_UB8(ptmp, stride, row8, row9, row10, row11, row12, row13, row14, row15);  in HFilter16()
385   ptmp = src - 3;  in HFilter16()
386   ST6x1_UB(tmp3, 0, tmp2, 0, ptmp, 4);  in HFilter16()
387   ptmp += stride;  in HFilter16()
388   ST6x1_UB(tmp3, 1, tmp2, 1, ptmp, 4);  in HFilter16()
389   ptmp += stride;  in HFilter16()
390   ST6x1_UB(tmp3, 2, tmp2, 2, ptmp, 4);  in HFilter16()
391   ptmp += stride;  in HFilter16()
392   ST6x1_UB(tmp3, 3, tmp2, 3, ptmp, 4);  in HFilter16()
393   ptmp += stride;  in HFilter16()
394   ST6x1_UB(tmp4, 0, tmp2, 4, ptmp, 4);  in HFilter16()
395   ptmp += stride;  in HFilter16()
396   ST6x1_UB(tmp4, 1, tmp2, 5, ptmp, 4);  in HFilter16()
397   ptmp += stride;  in HFilter16()
398   ST6x1_UB(tmp4, 2, tmp2, 6, ptmp, 4);  in HFilter16()
399   ptmp += stride;  in HFilter16()
400   ST6x1_UB(tmp4, 3, tmp2, 7, ptmp, 4);  in HFilter16()
401   ptmp += stride;  in HFilter16()
402   ST6x1_UB(tmp6, 0, tmp5, 0, ptmp, 4);  in HFilter16()
403   ptmp += stride;  in HFilter16()
404   ST6x1_UB(tmp6, 1, tmp5, 1, ptmp, 4);  in HFilter16()
405   ptmp += stride;  in HFilter16()
406   ST6x1_UB(tmp6, 2, tmp5, 2, ptmp, 4);  in HFilter16()
407   ptmp += stride;  in HFilter16()
408   ST6x1_UB(tmp6, 3, tmp5, 3, ptmp, 4);  in HFilter16()
409   ptmp += stride;  in HFilter16()
410   ST6x1_UB(tmp7, 0, tmp5, 4, ptmp, 4);  in HFilter16()
411   ptmp += stride;  in HFilter16()
412   ST6x1_UB(tmp7, 1, tmp5, 5, ptmp, 4);  in HFilter16()
413   ptmp += stride;  in HFilter16()
414   ST6x1_UB(tmp7, 2, tmp5, 6, ptmp, 4);  in HFilter16()
415   ptmp += stride;  in HFilter16()
416   ST6x1_UB(tmp7, 3, tmp5, 7, ptmp, 4);  in HFilter16()