Merge "Make CLPF handle frame widths and heights not divisible by 8." into nextgenv2

av1/common/clpf.c

@@ -59,12 +59,12 @@ int av1_clpf_frame(const YV12_BUFFER_CONFIG *dst, const YV12_BUFFER_CONFIG *rec,
   /* Constrained low-pass filter (CLPF) */
   int c, k, l, m, n;
   const int bs = MI_SIZE;
-  int width = cm->mi_cols * bs;
-  int height = cm->mi_rows * bs;
+  int width = rec->y_crop_width;
+  int height = rec->y_crop_height;
   int xpos, ypos;
   int stride_y = rec->y_stride;
-  int num_fb_hor = (width + (1 << fb_size_log2) - bs) >> fb_size_log2;
-  int num_fb_ver = (height + (1 << fb_size_log2) - bs) >> fb_size_log2;
+  int num_fb_hor = (width + (1 << fb_size_log2) - 1) >> fb_size_log2;
+  int num_fb_ver = (height + (1 << fb_size_log2) - 1) >> fb_size_log2;
   int block_index = 0;
 
   // Iterate over all filter blocks

av1/common/clpf_simd.h

@@ -187,7 +187,8 @@ void SIMD_FUNC(aom_clpf_block)(const uint8_t *src, uint8_t *dst, int stride,
   // This will only be used if 4:2:0 and width not a multiple of 16 and along
   // the right edge only, so we can fall back to the plain C implementation in
   // this case.  If not extended to chroma, this test will be redundant.
-  if (sizex != 8 || width < 16) {  // Fallback to C if frame width < 16
+  if (sizex != 8 || width < 16 || y0 + 8 > height || x0 + 8 > width) {
+    // Fallback to C for odd sizes
     aom_clpf_block_c(src, dst, stride, x0, y0, sizex, sizey, width, height,
                      strength);
   } else {