From: Ronald S. Bultje Date: Mon, 28 Jun 2010 13:50:55 +0000 (+0000) Subject: Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but X-Git-Tag: v0.7b1~3537 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=7c4dcf81658103b9506adcbf848bd23efe3b0b4e;p=platform%2Fupstream%2Flibav.git Optimize split MC, so we don't always do 4x4 blocks of 4x4pixels each, but we apply them as 16x8/8x16/8x8 subblocks where possible. Since this allows us to use width=8/16 instead of width=4 MC functions, we can now take more advantage of SSE2/SSSE3 optimizations, leading to a total speedup for splitMV filter of about 10%. Originally committed as revision 23853 to svn://svn.ffmpeg.org/ffmpeg/trunk --- diff --git a/libavcodec/vp8.c b/libavcodec/vp8.c index 0000706..7558599 100644 --- a/libavcodec/vp8.c +++ b/libavcodec/vp8.c @@ -943,6 +943,39 @@ static inline void vp8_mc(VP8Context *s, int luma, mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my); } +static inline void vp8_mc_part(VP8Context *s, uint8_t *dst[3], + AVFrame *ref_frame, int x_off, int y_off, + int bx_off, int by_off, + int block_w, int block_h, + int width, int height, VP56mv *mv) +{ + VP56mv uvmv = *mv; + + /* Y */ + vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off, + ref_frame->data[0], mv, x_off + bx_off, y_off + by_off, + block_w, block_h, width, height, s->linesize, + s->put_pixels_tab[block_w == 8]); + + /* U/V */ + if (s->profile == 3) { + uvmv.x &= ~7; + uvmv.y &= ~7; + } + x_off >>= 1; y_off >>= 1; + bx_off >>= 1; by_off >>= 1; + width >>= 1; height >>= 1; + block_w >>= 1; block_h >>= 1; + vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off, + ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off, + block_w, block_h, width, height, s->uvlinesize, + s->put_pixels_tab[1 + (block_w == 4)]); + vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off, + ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off, + block_w, block_h, width, height, s->uvlinesize, + s->put_pixels_tab[1 + (block_w == 4)]); +} + /** * Apply motion vectors to prediction buffer, chapter 18. */ @@ -951,29 +984,14 @@ static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, { int x_off = mb_x << 4, y_off = mb_y << 4; int width = 16*s->mb_width, height = 16*s->mb_height; - VP56mv uvmv; if (mb->mode < VP8_MVMODE_SPLIT) { - /* Y */ - vp8_mc(s, 1, dst[0], s->framep[mb->ref_frame]->data[0], &mb->mv, - x_off, y_off, 16, 16, width, height, s->linesize, - s->put_pixels_tab[0]); - - /* U/V */ - uvmv = mb->mv; - if (s->profile == 3) { - uvmv.x &= ~7; - uvmv.y &= ~7; - } - x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1; - vp8_mc(s, 0, dst[1], s->framep[mb->ref_frame]->data[1], &uvmv, - x_off, y_off, 8, 8, width, height, s->uvlinesize, - s->put_pixels_tab[1]); - vp8_mc(s, 0, dst[2], s->framep[mb->ref_frame]->data[2], &uvmv, - x_off, y_off, 8, 8, width, height, s->uvlinesize, - s->put_pixels_tab[1]); - } else { + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 0, 0, 16, 16, width, height, &mb->mv); + } else switch (mb->partitioning) { + case VP8_SPLITMVMODE_4x4: { int x, y; + VP56mv uvmv; /* Y */ for (y = 0; y < 4; y++) { @@ -1016,6 +1034,30 @@ static void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, s->put_pixels_tab[2]); } } + break; + } + case VP8_SPLITMVMODE_16x8: + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 0, 0, 16, 8, width, height, &mb->bmv[0]); + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 0, 8, 16, 8, width, height, &mb->bmv[8]); + break; + case VP8_SPLITMVMODE_8x16: + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 0, 0, 8, 16, width, height, &mb->bmv[0]); + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 8, 0, 8, 16, width, height, &mb->bmv[2]); + break; + case VP8_SPLITMVMODE_8x8: + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 0, 0, 8, 8, width, height, &mb->bmv[0]); + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 8, 0, 8, 8, width, height, &mb->bmv[2]); + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 0, 8, 8, 8, width, height, &mb->bmv[8]); + vp8_mc_part(s, dst, s->framep[mb->ref_frame], x_off, y_off, + 8, 8, 8, 8, width, height, &mb->bmv[10]); + break; } } diff --git a/libavcodec/vp8data.h b/libavcodec/vp8data.h index 0cd29ed..5c9dc6e 100644 --- a/libavcodec/vp8data.h +++ b/libavcodec/vp8data.h @@ -62,6 +62,13 @@ enum inter_submvmode { VP8_SUBMVMODE_NEW4X4 }; +enum inter_splitmvmode { + VP8_SPLITMVMODE_16x8 = 0, ///< 2 16x8 blocks (vertical) + VP8_SPLITMVMODE_8x16, ///< 2 8x16 blocks (horizontal) + VP8_SPLITMVMODE_8x8, ///< 2x2 blocks of 8x8px each + VP8_SPLITMVMODE_4x4, ///< 4x4 blocks of 4x4px each +}; + static const uint8_t vp8_pred4x4_mode[] = { [DC_PRED8x8] = DC_PRED, @@ -130,10 +137,10 @@ static const uint8_t vp8_mbfirstidx[4][16] = { }; static const int8_t vp8_mbsplit_tree[3][2] = { - { -3, 1 }, // '0' - 16 individual MVs - { -2, 2 }, // '10' - quarter-based MVs - { -0, -1 } // '110' - top/bottom MVs, - // '111' - left/right MVs + { -VP8_SPLITMVMODE_4x4, 1 }, // '0' - 16 individual MVs + { -VP8_SPLITMVMODE_8x8, 2 }, // '10' - quarter-based MVs + { -VP8_SPLITMVMODE_16x8, // '110' - top/bottom MVs + -VP8_SPLITMVMODE_8x16 } // '111' - left/right MVs }; static const uint8_t vp8_mbsplit_count[4] = { 2, 2, 4, 16 }; static const uint8_t vp8_mbsplit_prob[3] = { 110, 111, 150 };