157             short4 in_s = convert_short4(*in);  in kernel()
158             short4 out_s = convert_short4(*out);  in kernel()
159             in_s = in_s + ((out_s * (short4)(255 - in_s.w)) >> (short4)8);  in kernel()
176             short4 in_s = convert_short4(*in);  in kernel()
177             short4 out_s = convert_short4(*out);  in kernel()
178             in_s = out_s + ((in_s * (short4)(255 - out_s.w)) >> (short4)8);  in kernel()
195             short4 in_s = convert_short4(*in);  in kernel()
196             in_s = (in_s * out->w) >> (short4)8;  in kernel()
213             short4 out_s = convert_short4(*out);  in kernel()
214             out_s = (out_s * in->w) >> (short4)8;  in kernel()
231             short4 in_s = convert_short4(*in);  in kernel()
232             in_s = (in_s * (short4)(255 - out->w)) >> (short4)8;  in kernel()
249             short4 out_s = convert_short4(*out);  in kernel()
250             out_s = (out_s * (short4)(255 - in->w)) >> (short4)8;  in kernel()
267             short4 in_s = convert_short4(*in);  in kernel()
268             short4 out_s = convert_short4(*out);  in kernel()
287             short4 in_s = convert_short4(*in);  in kernel()
288             short4 out_s = convert_short4(*out);  in kernel()
332                                 >> (short4)8);  in kernel()