shithub: dav1d

Info  •  Files  •  Log  •  Branches

Extract y related operations out of warp_affine inner loop

--- a/src/recon_tmpl.c

+++ b/src/recon_tmpl.c

@@ -1071,15 +1071,15 @@

     const int height = (refp->p.p.h + ss_ver) >> ss_ver;

     for (int y = 0; y < b_dim[1] * v_mul; y += 8) {

+        const int src_y = t->by * 4 + ((y + 4) << ss_ver);

+        const int64_t mat3_y = (int64_t) mat[3] * src_y + mat[0];

+        const int64_t mat5_y = (int64_t) mat[5] * src_y + mat[1];

         for (int x = 0; x < b_dim[0] * h_mul; x += 8) {

             // calculate transformation relative to center of 8x8 block in

             // luma pixel units

             const int src_x = t->bx * 4 + ((x + 4) << ss_hor);

-            const int src_y = t->by * 4 + ((y + 4) << ss_ver);

-            const int64_t mvx = ((int64_t) mat[2] * src_x +

-                                 (int64_t) mat[3] * src_y + mat[0]) >> ss_hor;

-            const int64_t mvy = ((int64_t) mat[4] * src_x +

-                                 (int64_t) mat[5] * src_y + mat[1]) >> ss_ver;

+            const int64_t mvx = ((int64_t) mat[2] * src_x + mat3_y) >> ss_hor;

+            const int64_t mvy = ((int64_t) mat[4] * src_x + mat5_y) >> ss_ver;

             const int dx = (int) (mvx >> 16) - 4;

             const int mx = (((int) mvx & 0xffff) - wmp->alpha * 4 -