generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

intrapred/d135: flatten border results before storing

Browse Source 
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
This commit is contained in:
James Zern
2016-02-05 12:31:48 -08:00
parent f7b4364cbb
commit 05437805f7
1 changed files with 21 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
vpx_dsp/intrapred.c
View File

@@ -158,20 +158,29 @@ static INLINE void d117_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
static INLINE void d135_predictor(uint8_t *dst, ptrdiff_t stride, int bs, static INLINE void d135_predictor(uint8_t *dst, ptrdiff_t stride, int bs,
const uint8_t *above, const uint8_t *left) { const uint8_t *above, const uint8_t *left) {
int r, c; int i;
dst[0] = AVG3(left[0], above[-1], above[0]); #if defined(__GNUC__) && __GNUC__ == 4 && __GNUC_MINOR__ > 7
for (c = 1; c < bs; c++) // silence a spurious -Warray-bounds warning, possibly related to:
dst[c] = AVG3(above[c - 2], above[c - 1], above[c]); // 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]); // dst(bs, bs - 2)[0], i.e., border starting at bottom-left
for (r = 2; r < bs; ++r) for (i = 0; i < bs - 2; ++i) {
dst[r * stride] = AVG3(left[r - 2], left[r - 1], left[r]); 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 (i = 0; i < bs; ++i) {
for (r = 1; r < bs; ++r) { memcpy(dst + i * stride, border + bs - 1 - i, bs);
for (c = 1; c < bs; c++)
dst[c] = dst[-stride + c - 1];
dst += stride;
} }
} }