From 043d348aaeb9862cb87727a68eb7ee4f84d911d1 Mon Sep 17 00:00:00 2001 From: John Koleszar Date: Thu, 30 May 2013 09:12:57 -0700 Subject: [PATCH] Reimplementation of loop filter This version of the loop filter supports non-4:2:0 subsampling and a fourth plane, as well as changing the filtering order to be more friendly to hardware implementations. The filters are applied first to all vertical edges within the 64x64 SB, followed by the top horizontal edge and any internal horizontal edges. Since filtering is applied on each 4x4 edge serially, a dependency is created from filtering one block edge to the next. It would be possible to remove this depencnecy by building all filtering decisions from the unfiltered reconstruction data. Change-Id: I08f3e9683eb7bded8a76651cbc50fc0dfdd05fa7 --- configure | 1 + vp9/common/vp9_loopfilter.c | 234 ++++++++++++++++++++++++++++++++++++ vp9/common/vp9_loopfilter_filters.c | 5 +- vp9/common/vp9_rtcd_defs.sh | 20 +++ 4 files changed, 259 insertions(+), 1 deletion(-) diff --git a/configure b/configure index 4a97889..583919c 100755 --- a/configure +++ b/configure @@ -245,6 +245,7 @@ EXPERIMENT_LIST=" non420 alpha balanced_coeftree + new_loopfilter " CONFIG_LIST=" external_build diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c index ab14994..59df0b7 100644 --- a/vp9/common/vp9_loopfilter.c +++ b/vp9/common/vp9_loopfilter.c @@ -11,6 +11,7 @@ #include "vpx_config.h" #include "vp9/common/vp9_loopfilter.h" #include "vp9/common/vp9_onyxc_int.h" +#include "vp9/common/vp9_reconinter.h" #include "vpx_mem/vpx_mem.h" #include "vp9/common/vp9_seg_common.h" @@ -159,6 +160,7 @@ void vp9_loop_filter_frame_init(VP9_COMMON *cm, } } +#if !CONFIG_NEW_LOOPFILTER // Determine if we should skip inner-MB loop filtering within a MB // The current condition is that the loop filtering is skipped only // the MB uses a prediction size of 16x16 and either 16x16 transform @@ -539,3 +541,235 @@ void vp9_loop_filter_frame(VP9_COMMON *cm, } } } + +#else +static int build_lfi(const VP9_COMMON *cm, const MB_MODE_INFO *mbmi, + struct loop_filter_info *lfi) { + const loop_filter_info_n *lfi_n = &cm->lf_info; + int mode = mbmi->mode; + int mode_index = lfi_n->mode_lf_lut[mode]; + int seg = mbmi->segment_id; + int ref_frame = mbmi->ref_frame; + int filter_level = lfi_n->lvl[seg][ref_frame][mode_index]; + + if (filter_level) { + const int hev_index = filter_level >> 4; + lfi->mblim = lfi_n->mblim[filter_level]; + lfi->blim = lfi_n->blim[filter_level]; + lfi->lim = lfi_n->lim[filter_level]; + lfi->hev_thr = lfi_n->hev_thr[hev_index]; + return 1; + } + return 0; +} + +static void filter_selectively_vert(uint8_t *s, int pitch, + unsigned int mask_16x16, + unsigned int mask_8x8, + unsigned int mask_4x4, + unsigned int mask_4x4_1, + const struct loop_filter_info *lfi) { + unsigned int mask; + + for (mask = mask_16x16 | mask_8x8 | mask_4x4; mask; mask >>= 1) { + if (mask & 1) { + if (mask_16x16 & 1) { + vp9_mb_lpf_vertical_edge_w(s, pitch, lfi->mblim, lfi->lim, + lfi->hev_thr, 1); + assert(!(mask_8x8 & 1)); + assert(!(mask_4x4 & 1)); + assert(!(mask_4x4_1 & 1)); + } else if (mask_8x8 & 1) { + vp9_mbloop_filter_vertical_edge(s, pitch, lfi->mblim, lfi->lim, + lfi->hev_thr, 1); + assert(!(mask_16x16 & 1)); + assert(!(mask_4x4 & 1)); + } else if (mask_4x4 & 1) { + vp9_loop_filter_vertical_edge(s, pitch, lfi->mblim, lfi->lim, + lfi->hev_thr, 1); + assert(!(mask_16x16 & 1)); + assert(!(mask_8x8 & 1)); + } else { + assert(0); + } + + if (mask_4x4_1 & 1) + vp9_loop_filter_vertical_edge(s + 4, pitch, lfi->mblim, lfi->lim, + lfi->hev_thr, 1); + } + s += 8; + lfi++; + mask_16x16 >>= 1; + mask_8x8 >>= 1; + mask_4x4 >>= 1; + mask_4x4_1 >>= 1; + } +} + +static void filter_selectively_horiz(uint8_t *s, int pitch, + unsigned int mask_16x16, + unsigned int mask_8x8, + unsigned int mask_4x4, + unsigned int mask_4x4_1, + int only_4x4_1, + const struct loop_filter_info *lfi) { + unsigned int mask; + + for (mask = mask_16x16 | mask_8x8 | mask_4x4; mask; mask >>= 1) { + if (mask & 1) { + if (!only_4x4_1) { + if (mask_16x16 & 1) { + vp9_mb_lpf_horizontal_edge_w(s, pitch, lfi->mblim, lfi->lim, + lfi->hev_thr, 1); + assert(!(mask_8x8 & 1)); + assert(!(mask_4x4 & 1)); + assert(!(mask_4x4_1 & 1)); + } else if (mask_8x8 & 1) { + vp9_mbloop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim, + lfi->hev_thr, 1); + assert(!(mask_16x16 & 1)); + assert(!(mask_4x4 & 1)); + } else if (mask_4x4 & 1) { + vp9_loop_filter_horizontal_edge(s, pitch, lfi->mblim, lfi->lim, + lfi->hev_thr, 1); + assert(!(mask_16x16 & 1)); + assert(!(mask_8x8 & 1)); + } else { + assert(0); + } + } + + if (mask_4x4_1 & 1) + vp9_loop_filter_horizontal_edge(s + 4 * pitch, pitch, lfi->mblim, + lfi->lim, lfi->hev_thr, 1); + } + s += 8; + lfi++; + mask_16x16 >>= 1; + mask_8x8 >>= 1; + mask_4x4 >>= 1; + mask_4x4_1 >>= 1; + } +} + +static void filter_block_plane(VP9_COMMON *cm, MACROBLOCKD *xd, + int plane, int mi_row, int mi_col) { + const int ss_x = xd->plane[plane].subsampling_x; + const int row_step = 1 << xd->plane[plane].subsampling_y; + const int col_step = 1 << xd->plane[plane].subsampling_x; + struct buf_2d * const dst = &xd->plane[plane].dst; + uint8_t* const dst0 = dst->buf; + MODE_INFO* const mi0 = xd->mode_info_context; + unsigned int mask_16x16[64 / MI_SIZE] = {0}; + unsigned int mask_8x8[64 / MI_SIZE] = {0}; + unsigned int mask_4x4[64 / MI_SIZE] = {0}; + unsigned int mask_4x4_1[64 / MI_SIZE] = {0}; + struct loop_filter_info lfi[64 / MI_SIZE][64 / MI_SIZE]; + int r, c; + + for (r = 0; r < 64 / MI_SIZE && mi_row + r < cm->mi_rows; r += row_step) { + unsigned int mask_16x16_c = 0; + unsigned int mask_8x8_c = 0; + unsigned int mask_4x4_c = 0; + unsigned int border_mask; + + // Determine the vertical edges that need filtering + for (c = 0; c < 64 / MI_SIZE && mi_col + c < cm->mi_cols; c += col_step) { + const MODE_INFO const *mi = xd->mode_info_context; + const MODE_INFO const *mi_above = xd->mode_info_context - + cm->mode_info_stride; + const int skip_above = + (r + mi_row > 0) ? mi_above[c].mbmi.mb_skip_coeff : 0; + const int skip_left = + (c + mi_col > 0) ? mi[c - 1].mbmi.mb_skip_coeff : 0; + const int skip_this = mi[c].mbmi.mb_skip_coeff; + const int skip_this_c = skip_this && skip_left; + const int skip_this_r = skip_this && skip_above; + const TX_SIZE tx_size = plane ? get_uv_tx_size(xd) : mi[c].mbmi.txfm_size; + + // Filter level can vary per MI + if (!build_lfi(cm, &mi[c].mbmi, + lfi[r] + (c >> xd->plane[plane].subsampling_x))) + continue; + + // Build masks based on the transform size of each block + if (tx_size == TX_32X32) { + if (!skip_this_c && (c & 3) == 0) + mask_16x16_c |= 1 << (c >> ss_x); + if (!skip_this_r && (r & 3) == 0) + mask_16x16[r] |= 1 << (c >> ss_x); + } else if (tx_size == TX_16X16) { + if (!skip_this_c && (c & 1) == 0) + mask_16x16_c |= 1 << (c >> ss_x); + if (!skip_this_r && (r & 1) == 0) + mask_16x16[r] |= 1 << (c >> ss_x); + } else { + // force 8x8 filtering on 32x32 boundaries + if (!skip_this_c) { + if (tx_size == TX_8X8 || (c & 3) == 0) + mask_8x8_c |= 1 << (c >> ss_x); + else + mask_4x4_c |= 1 << (c >> ss_x); + } + + if (!skip_this_r) { + if (tx_size == TX_8X8 || (r & 3) == 0) + mask_8x8[r] |= 1 << (c >> ss_x); + else + mask_4x4[r] |= 1 << (c >> ss_x); + } + + if (!skip_this && tx_size < TX_8X8) + mask_4x4_1[r] |= 1 << (c >> ss_x); + } + } + + // Disable filtering on the leftmost column + border_mask = ~(mi_col == 0); + filter_selectively_vert(dst->buf, dst->stride, + mask_16x16_c & border_mask, + mask_8x8_c & border_mask, + mask_4x4_c & border_mask, + mask_4x4_1[r], lfi[r]); + dst->buf += 8 * dst->stride; + xd->mode_info_context += cm->mode_info_stride * row_step; + } + + // Now do horizontal pass + dst->buf = dst0; + xd->mode_info_context = mi0; + for (r = 0; r < 64 / MI_SIZE && mi_row + r < cm->mi_rows; r += row_step) { + filter_selectively_horiz(dst->buf, dst->stride, + mask_16x16[r], + mask_8x8[r], + mask_4x4[r], + mask_4x4_1[r], mi_row + r == 0, lfi[r]); + dst->buf += 8 * dst->stride; + xd->mode_info_context += cm->mode_info_stride * row_step; + } +} + +void vp9_loop_filter_frame(VP9_COMMON *cm, + MACROBLOCKD *xd, + int frame_filter_level, + int y_only) { + int mi_row, mi_col; + + // Initialize the loop filter for this frame. + vp9_loop_filter_frame_init(cm, xd, frame_filter_level); + + for (mi_row = 0; mi_row < cm->mi_rows; mi_row += 64 / MI_SIZE) { + MODE_INFO* const mi = cm->mi + mi_row * cm->mode_info_stride; + + for (mi_col = 0; mi_col < cm->mi_cols; mi_col += 64 / MI_SIZE) { + int plane; + + setup_dst_planes(xd, cm->frame_to_show, mi_row, mi_col); + for (plane = 0; plane < (y_only ? 1 : MAX_MB_PLANE); plane++) { + xd->mode_info_context = mi + mi_col; + filter_block_plane(cm, xd, plane, mi_row, mi_col); + } + } + } +} +#endif diff --git a/vp9/common/vp9_loopfilter_filters.c b/vp9/common/vp9_loopfilter_filters.c index fc7fbc4..3f90304 100644 --- a/vp9/common/vp9_loopfilter_filters.c +++ b/vp9/common/vp9_loopfilter_filters.c @@ -206,6 +206,7 @@ void vp9_mbloop_filter_vertical_edge_c(uint8_t *s, int pitch, } } +#if !CONFIG_NEW_LOOPFILTER /* Vertical MB Filtering */ void vp9_loop_filter_mbv_c(uint8_t *y_ptr, uint8_t *u_ptr, uint8_t *v_ptr, int y_stride, int uv_stride, @@ -308,6 +309,7 @@ void vp9_loop_filter_bv8x8_c(uint8_t *y, uint8_t *u, uint8_t *v, vp9_loop_filter_vertical_edge_c(v + 4, uv_stride, lfi->blim, lfi->lim, lfi->hev_thr, 1); } +#endif static INLINE void wide_mbfilter(int8_t mask, uint8_t hev, uint8_t flat, uint8_t flat2, @@ -412,6 +414,7 @@ void vp9_mb_lpf_vertical_edge_w(uint8_t *s, int p, } } +#if !CONFIG_NEW_LOOPFILTER void vp9_lpf_mbv_w_c(uint8_t *y, uint8_t *u, uint8_t *v, int y_stride, int uv_stride, struct loop_filter_info *lfi) { @@ -441,4 +444,4 @@ void vp9_lpf_mbh_w_c(uint8_t *y, uint8_t *u, uint8_t *v, vp9_mbloop_filter_horizontal_edge_c(v, uv_stride, lfi->mblim, lfi->lim, lfi->hev_thr, 1); } - +#endif diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh index bb49169..d48d66a 100644 --- a/vp9/common/vp9_rtcd_defs.sh +++ b/vp9/common/vp9_rtcd_defs.sh @@ -86,6 +86,25 @@ fi # # Loopfilter # +if [ "$CONFIG_NEW_LOOPFILTER" = "yes" ]; then +prototype void vp9_mb_lpf_vertical_edge_w "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count" +specialize vp9_mb_lpf_vertical_edge_w + +prototype void vp9_mbloop_filter_vertical_edge "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count" +specialize vp9_mbloop_filter_vertical_edge + +prototype void vp9_loop_filter_vertical_edge "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count" +specialize vp9_loop_filter_vertical_edge + +prototype void vp9_mb_lpf_horizontal_edge_w "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count" +specialize vp9_mb_lpf_horizontal_edge_w + +prototype void vp9_mbloop_filter_horizontal_edge "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count" +specialize vp9_mbloop_filter_horizontal_edge + +prototype void vp9_loop_filter_horizontal_edge "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count" +specialize vp9_loop_filter_horizontal_edge +else prototype void vp9_loop_filter_mbv "uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi" specialize vp9_loop_filter_mbv sse2 @@ -109,6 +128,7 @@ specialize vp9_lpf_mbh_w sse2 prototype void vp9_lpf_mbv_w "unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi" specialize vp9_lpf_mbv_w sse2 +fi # # post proc -- 2.7.4