diff options
| -rw-r--r-- | vp9/common/vp9_alloccommon.c | 12 | ||||
| -rw-r--r-- | vp9/common/vp9_loopfilter.c | 400 | ||||
| -rw-r--r-- | vp9/common/vp9_loopfilter.h | 60 | ||||
| -rw-r--r-- | vp9/common/vp9_thread_common.c | 14 | ||||
| -rw-r--r-- | vp9/decoder/vp9_decodeframe.c | 8 | ||||
| -rw-r--r-- | vp9/encoder/vp9_encoder.c | 3 | ||||
| -rw-r--r-- | vp9/encoder/vp9_picklpf.c | 2 |
7 files changed, 334 insertions, 165 deletions
diff --git a/vp9/common/vp9_alloccommon.c b/vp9/common/vp9_alloccommon.c index ac417b692..24c6c54ed 100644 --- a/vp9/common/vp9_alloccommon.c +++ b/vp9/common/vp9_alloccommon.c @@ -115,6 +115,8 @@ void vp9_free_context_buffers(VP9_COMMON *cm) { cm->above_context = NULL; vpx_free(cm->above_seg_context); cm->above_seg_context = NULL; + vpx_free(cm->lf.lfm); + cm->lf.lfm = NULL; } int vp9_alloc_context_buffers(VP9_COMMON *cm, int width, int height) { @@ -149,6 +151,16 @@ int vp9_alloc_context_buffers(VP9_COMMON *cm, int width, int height) { cm->above_context_alloc_cols = cm->mi_cols; } + vpx_free(cm->lf.lfm); + + // Each lfm holds bit masks for all the 8x8 blocks in a 64x64 region. The + // stride and rows are rounded up / truncated to a multiple of 8. + cm->lf.lfm_stride = (cm->mi_cols + (MI_BLOCK_SIZE - 1)) >> 3; + cm->lf.lfm = (LOOP_FILTER_MASK *)vpx_calloc( + ((cm->mi_rows + (MI_BLOCK_SIZE - 1)) >> 3) * cm->lf.lfm_stride, + sizeof(*cm->lf.lfm)); + if (!cm->lf.lfm) goto fail; + return 0; fail: diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c index 5e090f315..92fc0f729 100644 --- a/vp9/common/vp9_loopfilter.c +++ b/vp9/common/vp9_loopfilter.c @@ -825,6 +825,120 @@ static void build_y_mask(const loop_filter_info_n *const lfi_n, *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y; } +void vp9_adjust_mask(VP9_COMMON *const cm, const int mi_row, + const int mi_col, LOOP_FILTER_MASK *lfm) { + int i; + + // The largest loopfilter we have is 16x16 so we use the 16x16 mask + // for 32x32 transforms also. + lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32]; + lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32]; + lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32]; + lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32]; + + // We do at least 8 tap filter on every 32x32 even if the transform size + // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and + // remove it from the 4x4. + lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border; + lfm->left_y[TX_4X4] &= ~left_border; + lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border; + lfm->above_y[TX_4X4] &= ~above_border; + lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv; + lfm->left_uv[TX_4X4] &= ~left_border_uv; + lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv; + lfm->above_uv[TX_4X4] &= ~above_border_uv; + + // We do some special edge handling. + if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) { + const uint64_t rows = cm->mi_rows - mi_row; + + // Each pixel inside the border gets a 1, + const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1); + const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1); + + // Remove values completely outside our border. + for (i = 0; i < TX_32X32; i++) { + lfm->left_y[i] &= mask_y; + lfm->above_y[i] &= mask_y; + lfm->left_uv[i] &= mask_uv; + lfm->above_uv[i] &= mask_uv; + } + lfm->int_4x4_y &= mask_y; + lfm->int_4x4_uv &= mask_uv; + + // We don't apply a wide loop filter on the last uv block row. If set + // apply the shorter one instead. + if (rows == 1) { + lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16]; + lfm->above_uv[TX_16X16] = 0; + } + if (rows == 5) { + lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00; + lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00); + } + } + + if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) { + const uint64_t columns = cm->mi_cols - mi_col; + + // Each pixel inside the border gets a 1, the multiply copies the border + // to where we need it. + const uint64_t mask_y = (((1 << columns) - 1)) * 0x0101010101010101ULL; + const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111; + + // Internal edges are not applied on the last column of the image so + // we mask 1 more for the internal edges + const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111; + + // Remove the bits outside the image edge. + for (i = 0; i < TX_32X32; i++) { + lfm->left_y[i] &= mask_y; + lfm->above_y[i] &= mask_y; + lfm->left_uv[i] &= mask_uv; + lfm->above_uv[i] &= mask_uv; + } + lfm->int_4x4_y &= mask_y; + lfm->int_4x4_uv &= mask_uv_int; + + // We don't apply a wide loop filter on the last uv column. If set + // apply the shorter one instead. + if (columns == 1) { + lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16]; + lfm->left_uv[TX_16X16] = 0; + } + if (columns == 5) { + lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc); + lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc); + } + } + // We don't apply a loop filter on the first column in the image, mask that + // out. + if (mi_col == 0) { + for (i = 0; i < TX_32X32; i++) { + lfm->left_y[i] &= 0xfefefefefefefefeULL; + lfm->left_uv[i] &= 0xeeee; + } + } + + // Assert if we try to apply 2 different loop filters at the same position. + assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8])); + assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4])); + assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4])); + assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16])); + assert(!(lfm->left_uv[TX_16X16]&lfm->left_uv[TX_8X8])); + assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4])); + assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4])); + assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16])); + assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8])); + assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4])); + assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4])); + assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16])); + assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8])); + assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4])); + assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4])); + assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16])); +} + // This function sets up the bit masks for the entire 64x64 region represented // by mi_row, mi_col. // TODO(JBB): This function only works for yv12. @@ -855,7 +969,6 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col, const int shift_8_y[] = {0, 1, 8, 9}; const int shift_32_uv[] = {0, 2, 8, 10}; const int shift_16_uv[] = {0, 1, 4, 5}; - int i; const int max_rows = (mi_row + MI_BLOCK_SIZE > cm->mi_rows ? cm->mi_rows - mi_row : MI_BLOCK_SIZE); const int max_cols = (mi_col + MI_BLOCK_SIZE > cm->mi_cols ? @@ -970,114 +1083,8 @@ void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col, } break; } - // The largest loopfilter we have is 16x16 so we use the 16x16 mask - // for 32x32 transforms also. - lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32]; - lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32]; - lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32]; - lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32]; - // We do at least 8 tap filter on every 32x32 even if the transform size - // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and - // remove it from the 4x4. - lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border; - lfm->left_y[TX_4X4] &= ~left_border; - lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border; - lfm->above_y[TX_4X4] &= ~above_border; - lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv; - lfm->left_uv[TX_4X4] &= ~left_border_uv; - lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv; - lfm->above_uv[TX_4X4] &= ~above_border_uv; - - // We do some special edge handling. - if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) { - const uint64_t rows = cm->mi_rows - mi_row; - - // Each pixel inside the border gets a 1, - const uint64_t mask_y = (((uint64_t) 1 << (rows << 3)) - 1); - const uint16_t mask_uv = (((uint16_t) 1 << (((rows + 1) >> 1) << 2)) - 1); - - // Remove values completely outside our border. - for (i = 0; i < TX_32X32; i++) { - lfm->left_y[i] &= mask_y; - lfm->above_y[i] &= mask_y; - lfm->left_uv[i] &= mask_uv; - lfm->above_uv[i] &= mask_uv; - } - lfm->int_4x4_y &= mask_y; - lfm->int_4x4_uv &= mask_uv; - - // We don't apply a wide loop filter on the last uv block row. If set - // apply the shorter one instead. - if (rows == 1) { - lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16]; - lfm->above_uv[TX_16X16] = 0; - } - if (rows == 5) { - lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00; - lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00); - } - } - - if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) { - const uint64_t columns = cm->mi_cols - mi_col; - - // Each pixel inside the border gets a 1, the multiply copies the border - // to where we need it. - const uint64_t mask_y = (((1 << columns) - 1)) * 0x0101010101010101ULL; - const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111; - - // Internal edges are not applied on the last column of the image so - // we mask 1 more for the internal edges - const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111; - - // Remove the bits outside the image edge. - for (i = 0; i < TX_32X32; i++) { - lfm->left_y[i] &= mask_y; - lfm->above_y[i] &= mask_y; - lfm->left_uv[i] &= mask_uv; - lfm->above_uv[i] &= mask_uv; - } - lfm->int_4x4_y &= mask_y; - lfm->int_4x4_uv &= mask_uv_int; - - // We don't apply a wide loop filter on the last uv column. If set - // apply the shorter one instead. - if (columns == 1) { - lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16]; - lfm->left_uv[TX_16X16] = 0; - } - if (columns == 5) { - lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc); - lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc); - } - } - // We don't apply a loop filter on the first column in the image, mask that - // out. - if (mi_col == 0) { - for (i = 0; i < TX_32X32; i++) { - lfm->left_y[i] &= 0xfefefefefefefefeULL; - lfm->left_uv[i] &= 0xeeee; - } - } - - // Assert if we try to apply 2 different loop filters at the same position. - assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8])); - assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4])); - assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4])); - assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16])); - assert(!(lfm->left_uv[TX_16X16]&lfm->left_uv[TX_8X8])); - assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4])); - assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4])); - assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16])); - assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8])); - assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4])); - assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4])); - assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16])); - assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8])); - assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4])); - assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4])); - assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16])); + vp9_adjust_mask(cm, mi_row, mi_col, lfm); } static void filter_selectively_vert(uint8_t *s, int pitch, @@ -1428,6 +1435,7 @@ void vp9_filter_block_plane_ss11(VP9_COMMON *const cm, struct buf_2d *const dst = &plane->dst; uint8_t *const dst0 = dst->buf; int r, c; + uint8_t lfl_uv[16]; uint16_t mask_16x16 = lfm->left_uv[TX_16X16]; uint16_t mask_8x8 = lfm->left_uv[TX_8X8]; @@ -1440,8 +1448,8 @@ void vp9_filter_block_plane_ss11(VP9_COMMON *const cm, for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 4) { if (plane->plane_type == 1) { for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) { - lfm->lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)]; - lfm->lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) + (c << 1)]; + lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)]; + lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) + (c << 1)]; } } @@ -1457,18 +1465,18 @@ void vp9_filter_block_plane_ss11(VP9_COMMON *const cm, highbd_filter_selectively_vert_row2( plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info, - &lfm->lfl_uv[r << 1], (int)cm->bit_depth); + &lfl_uv[r << 1], (int)cm->bit_depth); } else { filter_selectively_vert_row2( plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info, - &lfm->lfl_uv[r << 1]); + &lfl_uv[r << 1]); } #else filter_selectively_vert_row2( plane->subsampling_x, dst->buf, dst->stride, mask_16x16_l, mask_8x8_l, mask_4x4_l, mask_4x4_int_l, &cm->lf_info, - &lfm->lfl_uv[r << 1]); + &lfl_uv[r << 1]); #endif // CONFIG_VP9_HIGHBITDEPTH dst->buf += 16 * dst->stride; @@ -1509,16 +1517,16 @@ void vp9_filter_block_plane_ss11(VP9_COMMON *const cm, highbd_filter_selectively_horiz(CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r, mask_4x4_r, mask_4x4_int_r, &cm->lf_info, - &lfm->lfl_uv[r << 1], (int)cm->bit_depth); + &lfl_uv[r << 1], (int)cm->bit_depth); } else { filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r, mask_4x4_r, mask_4x4_int_r, &cm->lf_info, - &lfm->lfl_uv[r << 1]); + &lfl_uv[r << 1]); } #else filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r, mask_4x4_r, mask_4x4_int_r, &cm->lf_info, - &lfm->lfl_uv[r << 1]); + &lfl_uv[r << 1]); #endif // CONFIG_VP9_HIGHBITDEPTH dst->buf += 8 * dst->stride; @@ -1529,13 +1537,11 @@ void vp9_filter_block_plane_ss11(VP9_COMMON *const cm, } } -void vp9_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer, - VP9_COMMON *cm, - struct macroblockd_plane planes[MAX_MB_PLANE], - int start, int stop, int y_only) { +static void loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer, VP9_COMMON *cm, + struct macroblockd_plane planes[MAX_MB_PLANE], + int start, int stop, int y_only) { const int num_planes = y_only ? 1 : MAX_MB_PLANE; enum lf_path path; - LOOP_FILTER_MASK lfm; int mi_row, mi_col; if (y_only) @@ -1549,24 +1555,24 @@ void vp9_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer, for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) { MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride; + LOOP_FILTER_MASK *lfm = get_lfm(&cm->lf, mi_row, 0); - for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) { + for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE, ++lfm) { int plane; vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col); // TODO(JBB): Make setup_mask work for non 420. - vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, - &lfm); + vp9_adjust_mask(cm, mi_row, mi_col, lfm); - vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm); + vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, lfm); for (plane = 1; plane < num_planes; ++plane) { switch (path) { case LF_PATH_420: - vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm); + vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, lfm); break; case LF_PATH_444: - vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm); + vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, lfm); break; case LF_PATH_SLOW: vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col, @@ -1592,10 +1598,132 @@ void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame, mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8); } end_mi_row = start_mi_row + mi_rows_to_filter; + loop_filter_rows(frame, cm, xd->plane, start_mi_row, end_mi_row, y_only); +} + +// Used by the encoder to build the loopfilter masks. +void vp9_build_mask_frame(VP9_COMMON *cm, int frame_filter_level, + int partial_frame) { + int start_mi_row, end_mi_row, mi_rows_to_filter; + int mi_col, mi_row; + if (!frame_filter_level) return; + start_mi_row = 0; + mi_rows_to_filter = cm->mi_rows; + if (partial_frame && cm->mi_rows > 8) { + start_mi_row = cm->mi_rows >> 1; + start_mi_row &= 0xfffffff8; + mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8); + } + end_mi_row = start_mi_row + mi_rows_to_filter; + vp9_loop_filter_frame_init(cm, frame_filter_level); - vp9_loop_filter_rows(frame, cm, xd->plane, - start_mi_row, end_mi_row, - y_only); + + for (mi_row = start_mi_row; mi_row < end_mi_row; mi_row += MI_BLOCK_SIZE) { + MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride; + for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) { + // vp9_setup_mask() zeros lfm + vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, + get_lfm(&cm->lf, mi_row, mi_col)); + } + } +} + +// 8x8 blocks in a superblock. A "1" represents the first block in a 16x16 +// or greater area. +static const uint8_t first_block_in_16x16[8][8] = { + {1, 0, 1, 0, 1, 0, 1, 0}, + {0, 0, 0, 0, 0, 0, 0, 0}, + {1, 0, 1, 0, 1, 0, 1, 0}, + {0, 0, 0, 0, 0, 0, 0, 0}, + {1, 0, 1, 0, 1, 0, 1, 0}, + {0, 0, 0, 0, 0, 0, 0, 0}, + {1, 0, 1, 0, 1, 0, 1, 0}, + {0, 0, 0, 0, 0, 0, 0, 0} +}; + +// This function sets up the bit masks for a block represented +// by mi_row, mi_col in a 64x64 region. +// TODO(SJL): This function only works for yv12. +void vp9_build_mask(VP9_COMMON *cm, const MB_MODE_INFO *mbmi, int mi_row, + int mi_col, int bw, int bh) { + const BLOCK_SIZE block_size = mbmi->sb_type; + const TX_SIZE tx_size_y = mbmi->tx_size; + const loop_filter_info_n *const lfi_n = &cm->lf_info; + const int filter_level = get_filter_level(lfi_n, mbmi); + const TX_SIZE tx_size_uv = get_uv_tx_size_impl(tx_size_y, block_size, 1, 1); + LOOP_FILTER_MASK *const lfm = get_lfm(&cm->lf, mi_row, mi_col); + uint64_t *const left_y = &lfm->left_y[tx_size_y]; + uint64_t *const above_y = &lfm->above_y[tx_size_y]; + uint64_t *const int_4x4_y = &lfm->int_4x4_y; + uint16_t *const left_uv = &lfm->left_uv[tx_size_uv]; + uint16_t *const above_uv = &lfm->above_uv[tx_size_uv]; + uint16_t *const int_4x4_uv = &lfm->int_4x4_uv; + const int row_in_sb = (mi_row & 7); + const int col_in_sb = (mi_col & 7); + const int shift_y = col_in_sb + (row_in_sb << 3); + const int shift_uv = (col_in_sb >> 1) + ((row_in_sb >> 1) << 2); + const int build_uv = first_block_in_16x16[row_in_sb][col_in_sb]; + + if (!filter_level) { + return; + } else { + int index = shift_y; + int i; + for (i = 0; i < bh; i++) { + memset(&lfm->lfl_y[index], filter_level, bw); + index += 8; + } + } + + // These set 1 in the current block size for the block size edges. + // For instance if the block size is 32x16, we'll set: + // above = 1111 + // 0000 + // and + // left = 1000 + // = 1000 + // NOTE : In this example the low bit is left most ( 1000 ) is stored as + // 1, not 8... + // + // U and V set things on a 16 bit scale. + // + *above_y |= above_prediction_mask[block_size] << shift_y; + *left_y |= left_prediction_mask[block_size] << shift_y; + + if (build_uv) { + *above_uv |= above_prediction_mask_uv[block_size] << shift_uv; + *left_uv |= left_prediction_mask_uv[block_size] << shift_uv; + } + + // If the block has no coefficients and is not intra we skip applying + // the loop filter on block edges. + if (mbmi->skip && is_inter_block(mbmi)) + return; + + // Add a mask for the transform size. The transform size mask is set to + // be correct for a 64x64 prediction block size. Mask to match the size of + // the block we are working on and then shift it into place. + *above_y |= (size_mask[block_size] & + above_64x64_txform_mask[tx_size_y]) << shift_y; + *left_y |= (size_mask[block_size] & + left_64x64_txform_mask[tx_size_y]) << shift_y; + + if (build_uv) { + *above_uv |= (size_mask_uv[block_size] & + above_64x64_txform_mask_uv[tx_size_uv]) << shift_uv; + + *left_uv |= (size_mask_uv[block_size] & + left_64x64_txform_mask_uv[tx_size_uv]) << shift_uv; + } + + // Try to determine what to do with the internal 4x4 block boundaries. These + // differ from the 4x4 boundaries on the outside edge of an 8x8 in that the + // internal ones can be skipped and don't depend on the prediction block size. + if (tx_size_y == TX_4X4) + *int_4x4_y |= (size_mask[block_size] & -1ULL) << shift_y; + + if (build_uv && tx_size_uv == TX_4X4) + *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv; } void vp9_loop_filter_data_reset( @@ -1609,9 +1737,17 @@ void vp9_loop_filter_data_reset( memcpy(lf_data->planes, planes, sizeof(lf_data->planes)); } +void vp9_reset_lfm(VP9_COMMON *cm) { + if (cm->lf.filter_level) { + memset(cm->lf.lfm, 0, + ((cm->mi_rows + (MI_BLOCK_SIZE - 1)) >> 3) * cm->lf.lfm_stride * + sizeof(*cm->lf.lfm)); + } +} + int vp9_loop_filter_worker(LFWorkerData *const lf_data, void *unused) { (void)unused; - vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes, - lf_data->start, lf_data->stop, lf_data->y_only); + loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes, + lf_data->start, lf_data->stop, lf_data->y_only); return 1; } diff --git a/vp9/common/vp9_loopfilter.h b/vp9/common/vp9_loopfilter.h index f7cbde678..7f943ea09 100644 --- a/vp9/common/vp9_loopfilter.h +++ b/vp9/common/vp9_loopfilter.h @@ -35,24 +35,6 @@ enum lf_path { LF_PATH_SLOW, }; -struct loopfilter { - int filter_level; - - int sharpness_level; - int last_sharpness_level; - - uint8_t mode_ref_delta_enabled; - uint8_t mode_ref_delta_update; - - // 0 = Intra, Last, GF, ARF - signed char ref_deltas[MAX_REF_LF_DELTAS]; - signed char last_ref_deltas[MAX_REF_LF_DELTAS]; - - // 0 = ZERO_MV, MV - signed char mode_deltas[MAX_MODE_LF_DELTAS]; - signed char last_mode_deltas[MAX_MODE_LF_DELTAS]; -}; - // Need to align this structure so when it is declared and // passed it can be loaded into vector registers. typedef struct { @@ -83,9 +65,29 @@ typedef struct { uint16_t above_uv[TX_SIZES]; uint16_t int_4x4_uv; uint8_t lfl_y[64]; - uint8_t lfl_uv[16]; } LOOP_FILTER_MASK; +struct loopfilter { + int filter_level; + + int sharpness_level; + int last_sharpness_level; + + uint8_t mode_ref_delta_enabled; + uint8_t mode_ref_delta_update; + + // 0 = Intra, Last, GF, ARF + signed char ref_deltas[MAX_REF_LF_DELTAS]; + signed char last_ref_deltas[MAX_REF_LF_DELTAS]; + + // 0 = ZERO_MV, MV + signed char mode_deltas[MAX_MODE_LF_DELTAS]; + signed char last_mode_deltas[MAX_MODE_LF_DELTAS]; + + LOOP_FILTER_MASK *lfm; + int lfm_stride; +}; + /* assorted loopfilter functions which get used elsewhere */ struct VP9Common; struct macroblockd; @@ -116,7 +118,7 @@ void vp9_filter_block_plane_non420(struct VP9Common *cm, void vp9_loop_filter_init(struct VP9Common *cm); // Update the loop filter for the current frame. -// This should be called before vp9_loop_filter_rows(), vp9_loop_filter_frame() +// This should be called before vp9_loop_filter_frame(), vp9_build_mask_frame() // calls this function directly. void vp9_loop_filter_frame_init(struct VP9Common *cm, int default_filt_lvl); @@ -126,11 +128,19 @@ void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame, int filter_level, int y_only, int partial_frame); -// Apply the loop filter to [start, stop) macro block rows in frame_buffer. -void vp9_loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer, - struct VP9Common *cm, - struct macroblockd_plane planes[MAX_MB_PLANE], - int start, int stop, int y_only); +// Get the superblock lfm for a given mi_row, mi_col. +static INLINE LOOP_FILTER_MASK *get_lfm(const struct loopfilter *lf, + const int mi_row, const int mi_col) { + return &lf->lfm[(mi_col >> 3) + ((mi_row >> 3) * lf->lfm_stride)]; +} + +void vp9_build_mask(struct VP9Common *cm, const MB_MODE_INFO *mbmi, int mi_row, + int mi_col, int bw, int bh); +void vp9_adjust_mask(struct VP9Common *const cm, const int mi_row, + const int mi_col, LOOP_FILTER_MASK *lfm); +void vp9_build_mask_frame(struct VP9Common *cm, int frame_filter_level, + int partial_frame); +void vp9_reset_lfm(struct VP9Common *const cm); typedef struct LoopFilterWorkerData { YV12_BUFFER_CONFIG *frame_buffer; diff --git a/vp9/common/vp9_thread_common.c b/vp9/common/vp9_thread_common.c index 978e0bfe9..db78d6be8 100644 --- a/vp9/common/vp9_thread_common.c +++ b/vp9/common/vp9_thread_common.c @@ -109,29 +109,27 @@ void thread_loop_filter_rows(const YV12_BUFFER_CONFIG *const frame_buffer, for (mi_row = start; mi_row < stop; mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) { MODE_INFO **const mi = cm->mi_grid_visible + mi_row * cm->mi_stride; + LOOP_FILTER_MASK *lfm = get_lfm(&cm->lf, mi_row, 0); - for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) { + for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE, ++lfm) { const int r = mi_row >> MI_BLOCK_SIZE_LOG2; const int c = mi_col >> MI_BLOCK_SIZE_LOG2; - LOOP_FILTER_MASK lfm; int plane; sync_read(lf_sync, r, c); vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col); - // TODO(JBB): Make setup_mask work for non 420. - vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, - &lfm); + vp9_adjust_mask(cm, mi_row, mi_col, lfm); - vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, &lfm); + vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, lfm); for (plane = 1; plane < num_planes; ++plane) { switch (path) { case LF_PATH_420: - vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, &lfm); + vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, lfm); break; case LF_PATH_444: - vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, &lfm); + vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, lfm); break; case LF_PATH_SLOW: vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col, diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c index ba8c8ddc4..4e216121e 100644 --- a/vp9/decoder/vp9_decodeframe.c +++ b/vp9/decoder/vp9_decodeframe.c @@ -896,6 +896,10 @@ static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd, } xd->corrupted |= vpx_reader_has_error(r); + + if (cm->lf.filter_level) { + vp9_build_mask(cm, mbmi, mi_row, mi_col, bw, bh); + } } static INLINE int dec_partition_plane_context(const MACROBLOCKD *xd, @@ -1464,6 +1468,8 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi, memset(cm->above_seg_context, 0, sizeof(*cm->above_seg_context) * aligned_cols); + vp9_reset_lfm(cm); + get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers); if (pbi->tile_data == NULL || @@ -1662,6 +1668,8 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi, memset(cm->above_seg_context, 0, sizeof(*cm->above_seg_context) * aligned_mi_cols); + vp9_reset_lfm(cm); + // Load tile data into tile_buffers get_tile_buffers(pbi, data, data_end, tile_cols, tile_rows, tile_buffers); diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c index 91e92ff24..aa110f625 100644 --- a/vp9/encoder/vp9_encoder.c +++ b/vp9/encoder/vp9_encoder.c @@ -2755,6 +2755,7 @@ void vp9_update_reference_frames(VP9_COMP *cpi) { static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) { MACROBLOCKD *xd = &cpi->td.mb.e_mbd; struct loopfilter *lf = &cm->lf; + if (xd->lossless) { lf->filter_level = 0; } else { @@ -2771,6 +2772,8 @@ static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) { } if (lf->filter_level > 0) { + vp9_build_mask_frame(cm, lf->filter_level, 0); + if (cpi->num_workers > 1) vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane, lf->filter_level, 0, 0, diff --git a/vp9/encoder/vp9_picklpf.c b/vp9/encoder/vp9_picklpf.c index 1b1068e37..5444bc89f 100644 --- a/vp9/encoder/vp9_picklpf.c +++ b/vp9/encoder/vp9_picklpf.c @@ -40,6 +40,8 @@ static int64_t try_filter_frame(const YV12_BUFFER_CONFIG *sd, VP9_COMMON *const cm = &cpi->common; int64_t filt_err; + vp9_build_mask_frame(cm, filt_level, partial_frame); + if (cpi->num_workers > 1) vp9_loop_filter_frame_mt(cm->frame_to_show, cm, cpi->td.mb.e_mbd.plane, filt_level, 1, partial_frame, |