Mine
public text v1 · immutable;
; scale_frame( uint16_t *src, uint16_t *dst, int stride, int lshift, int rshift, int width, int height )
;
cglobal scale_frame_sse2, 7,7
; dst[k] = (src[k] << lshift) + (src[k] >> rshift);
movdqa xmm0 [r0]
movdqa xmm1 xmm0
movd xmm2 r3
movd xmm3 r4
psllw xmm0 xmm2
pslrw xmm1 xmm3
padd xmm0 xmm1
movdqa [r1] xmm0