shithub: libvpx

Info  •  Files  •  Log  •  Branches

intrapred/d135: flatten border results before storing

the results along the top and left border are then stored with a moving
window into the vector.
~40-67% faster on ARM, ~40-77+% on x86 depending on the block size.

Change-Id: Iab369aa2946a3ae4eb7290d512868fe5db92dbc8

--- a/vpx_dsp/intrapred.c

+++ b/vpx_dsp/intrapred.c

@@ -158,20 +158,29 @@

 static INLINE void d135_predictor(uint8_t *dst, ptrdiff_t stride, int bs,

                                   const uint8_t *above, const uint8_t *left) {

-  int r, c;

-  dst[0] = AVG3(left[0], above[-1], above[0]);

-  for (c = 1; c < bs; c++)

-    dst[c] = AVG3(above[c - 2], above[c - 1], above[c]);

+  int i;

+#if defined(__GNUC__) && __GNUC__ == 4 && __GNUC_MINOR__ > 7

+  // silence a spurious -Warray-bounds warning, possibly related to:

+  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56273

+  uint8_t border[69];

+#else

+  uint8_t border[32 + 32 - 1];  // outer border from bottom-left to top-right

+#endif

-  dst[stride] = AVG3(above[-1], left[0], left[1]);

-  for (r = 2; r < bs; ++r)

-    dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]);

+  // dst(bs, bs - 2)[0], i.e., border starting at bottom-left

+  for (i = 0; i < bs - 2; ++i) {

+    border[i] = AVG3(left[bs - 3 - i], left[bs - 2 - i], left[bs - 1 - i]);

+  }

+  border[bs - 2] = AVG3(above[-1], left[0], left[1]);

+  border[bs - 1] = AVG3(left[0], above[-1], above[0]);

+  border[bs - 0] = AVG3(above[-1], above[0], above[1]);

+  // dst[0][2, size), i.e., remaining top border ascending

+  for (i = 0; i < bs - 2; ++i) {

+    border[bs + 1 + i] = AVG3(above[i], above[i + 1], above[i + 2]);

+  }

-  dst += stride;

-  for (r = 1; r < bs; ++r) {

-    for (c = 1; c < bs; c++)

-      dst[c] = dst[-stride + c - 1];

-    dst += stride;

+  for (i = 0; i < bs; ++i) {

+    memcpy(dst + i * stride, border + bs - 1 - i, bs);

   }

 }