michael@0: /*
michael@0:  *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
michael@0:  *
michael@0:  *  Use of this source code is governed by a BSD-style license
michael@0:  *  that can be found in the LICENSE file in the root of the source
michael@0:  *  tree. An additional intellectual property rights grant can be found
michael@0:  *  in the file PATENTS.  All contributing project authors may
michael@0:  *  be found in the AUTHORS file in the root of the source tree.
michael@0:  */
michael@0: 
michael@0: 
michael@0: /* MFQE: Multiframe Quality Enhancement
michael@0:  * In rate limited situations keyframes may cause significant visual artifacts
michael@0:  * commonly referred to as "popping." This file implements a postproccesing
michael@0:  * algorithm which blends data from the preceeding frame when there is no
michael@0:  * motion and the q from the previous frame is lower which indicates that it is
michael@0:  * higher quality.
michael@0:  */
michael@0: 
michael@0: #include "postproc.h"
michael@0: #include "variance.h"
michael@0: #include "vpx_mem/vpx_mem.h"
michael@0: #include "vp8_rtcd.h"
michael@0: #include "vpx_scale/yv12config.h"
michael@0: 
michael@0: #include <limits.h>
michael@0: #include <stdlib.h>
michael@0: 
michael@0: static void filter_by_weight(unsigned char *src, int src_stride,
michael@0:                              unsigned char *dst, int dst_stride,
michael@0:                              int block_size, int src_weight)
michael@0: {
michael@0:     int dst_weight = (1 << MFQE_PRECISION) - src_weight;
michael@0:     int rounding_bit = 1 << (MFQE_PRECISION - 1);
michael@0:     int r, c;
michael@0: 
michael@0:     for (r = 0; r < block_size; r++)
michael@0:     {
michael@0:         for (c = 0; c < block_size; c++)
michael@0:         {
michael@0:             dst[c] = (src[c] * src_weight +
michael@0:                       dst[c] * dst_weight +
michael@0:                       rounding_bit) >> MFQE_PRECISION;
michael@0:         }
michael@0:         src += src_stride;
michael@0:         dst += dst_stride;
michael@0:     }
michael@0: }
michael@0: 
michael@0: void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride,
michael@0:                                  unsigned char *dst, int dst_stride,
michael@0:                                  int src_weight)
michael@0: {
michael@0:     filter_by_weight(src, src_stride, dst, dst_stride, 16, src_weight);
michael@0: }
michael@0: 
michael@0: void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride,
michael@0:                                unsigned char *dst, int dst_stride,
michael@0:                                int src_weight)
michael@0: {
michael@0:     filter_by_weight(src, src_stride, dst, dst_stride, 8, src_weight);
michael@0: }
michael@0: 
michael@0: void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride,
michael@0:                                unsigned char *dst, int dst_stride,
michael@0:                                int src_weight)
michael@0: {
michael@0:     filter_by_weight(src, src_stride, dst, dst_stride, 4, src_weight);
michael@0: }
michael@0: 
michael@0: static void apply_ifactor(unsigned char *y_src,
michael@0:                           int y_src_stride,
michael@0:                           unsigned char *y_dst,
michael@0:                           int y_dst_stride,
michael@0:                           unsigned char *u_src,
michael@0:                           unsigned char *v_src,
michael@0:                           int uv_src_stride,
michael@0:                           unsigned char *u_dst,
michael@0:                           unsigned char *v_dst,
michael@0:                           int uv_dst_stride,
michael@0:                           int block_size,
michael@0:                           int src_weight)
michael@0: {
michael@0:     if (block_size == 16)
michael@0:     {
michael@0:         vp8_filter_by_weight16x16(y_src, y_src_stride, y_dst, y_dst_stride, src_weight);
michael@0:         vp8_filter_by_weight8x8(u_src, uv_src_stride, u_dst, uv_dst_stride, src_weight);
michael@0:         vp8_filter_by_weight8x8(v_src, uv_src_stride, v_dst, uv_dst_stride, src_weight);
michael@0:     }
michael@0:     else /* if (block_size == 8) */
michael@0:     {
michael@0:         vp8_filter_by_weight8x8(y_src, y_src_stride, y_dst, y_dst_stride, src_weight);
michael@0:         vp8_filter_by_weight4x4(u_src, uv_src_stride, u_dst, uv_dst_stride, src_weight);
michael@0:         vp8_filter_by_weight4x4(v_src, uv_src_stride, v_dst, uv_dst_stride, src_weight);
michael@0:     }
michael@0: }
michael@0: 
michael@0: static unsigned int int_sqrt(unsigned int x)
michael@0: {
michael@0:     unsigned int y = x;
michael@0:     unsigned int guess;
michael@0:     int p = 1;
michael@0:     while (y>>=1) p++;
michael@0:     p>>=1;
michael@0: 
michael@0:     guess=0;
michael@0:     while (p>=0)
michael@0:     {
michael@0:         guess |= (1<<p);
michael@0:         if (x<guess*guess)
michael@0:             guess -= (1<<p);
michael@0:         p--;
michael@0:     }
michael@0:     /* choose between guess or guess+1 */
michael@0:     return guess+(guess*guess+guess+1<=x);
michael@0: }
michael@0: 
michael@0: #define USE_SSD
michael@0: static void multiframe_quality_enhance_block
michael@0: (
michael@0:     int blksize, /* Currently only values supported are 16, 8 */
michael@0:     int qcurr,
michael@0:     int qprev,
michael@0:     unsigned char *y,
michael@0:     unsigned char *u,
michael@0:     unsigned char *v,
michael@0:     int y_stride,
michael@0:     int uv_stride,
michael@0:     unsigned char *yd,
michael@0:     unsigned char *ud,
michael@0:     unsigned char *vd,
michael@0:     int yd_stride,
michael@0:     int uvd_stride
michael@0: )
michael@0: {
michael@0:     static const unsigned char VP8_ZEROS[16]=
michael@0:     {
michael@0:          0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
michael@0:     };
michael@0:     int uvblksize = blksize >> 1;
michael@0:     int qdiff = qcurr - qprev;
michael@0: 
michael@0:     int i;
michael@0:     unsigned char *up;
michael@0:     unsigned char *udp;
michael@0:     unsigned char *vp;
michael@0:     unsigned char *vdp;
michael@0: 
michael@0:     unsigned int act, actd, sad, usad, vsad, sse, thr, thrsq, actrisk;
michael@0: 
michael@0:     if (blksize == 16)
michael@0:     {
michael@0:         actd = (vp8_variance16x16(yd, yd_stride, VP8_ZEROS, 0, &sse)+128)>>8;
michael@0:         act = (vp8_variance16x16(y, y_stride, VP8_ZEROS, 0, &sse)+128)>>8;
michael@0: #ifdef USE_SSD
michael@0:         sad = (vp8_variance16x16(y, y_stride, yd, yd_stride, &sse));
michael@0:         sad = (sse + 128)>>8;
michael@0:         usad = (vp8_variance8x8(u, uv_stride, ud, uvd_stride, &sse));
michael@0:         usad = (sse + 32)>>6;
michael@0:         vsad = (vp8_variance8x8(v, uv_stride, vd, uvd_stride, &sse));
michael@0:         vsad = (sse + 32)>>6;
michael@0: #else
michael@0:         sad = (vp8_sad16x16(y, y_stride, yd, yd_stride, UINT_MAX) + 128) >> 8;
michael@0:         usad = (vp8_sad8x8(u, uv_stride, ud, uvd_stride, UINT_MAX) + 32) >> 6;
michael@0:         vsad = (vp8_sad8x8(v, uv_stride, vd, uvd_stride, UINT_MAX)+ 32) >> 6;
michael@0: #endif
michael@0:     }
michael@0:     else /* if (blksize == 8) */
michael@0:     {
michael@0:         actd = (vp8_variance8x8(yd, yd_stride, VP8_ZEROS, 0, &sse)+32)>>6;
michael@0:         act = (vp8_variance8x8(y, y_stride, VP8_ZEROS, 0, &sse)+32)>>6;
michael@0: #ifdef USE_SSD
michael@0:         sad = (vp8_variance8x8(y, y_stride, yd, yd_stride, &sse));
michael@0:         sad = (sse + 32)>>6;
michael@0:         usad = (vp8_variance4x4(u, uv_stride, ud, uvd_stride, &sse));
michael@0:         usad = (sse + 8)>>4;
michael@0:         vsad = (vp8_variance4x4(v, uv_stride, vd, uvd_stride, &sse));
michael@0:         vsad = (sse + 8)>>4;
michael@0: #else
michael@0:         sad = (vp8_sad8x8(y, y_stride, yd, yd_stride, UINT_MAX) + 32) >> 6;
michael@0:         usad = (vp8_sad4x4(u, uv_stride, ud, uvd_stride, UINT_MAX) + 8) >> 4;
michael@0:         vsad = (vp8_sad4x4(v, uv_stride, vd, uvd_stride, UINT_MAX) + 8) >> 4;
michael@0: #endif
michael@0:     }
michael@0: 
michael@0:     actrisk = (actd > act * 5);
michael@0: 
michael@0:     /* thr = qdiff/16 + log2(act) + log4(qprev) */
michael@0:     thr = (qdiff >> 4);
michael@0:     while (actd >>= 1) thr++;
michael@0:     while (qprev >>= 2) thr++;
michael@0: 
michael@0: #ifdef USE_SSD
michael@0:     thrsq = thr * thr;
michael@0:     if (sad < thrsq &&
michael@0:         /* additional checks for color mismatch and excessive addition of
michael@0:          * high-frequencies */
michael@0:         4 * usad < thrsq && 4 * vsad < thrsq && !actrisk)
michael@0: #else
michael@0:     if (sad < thr &&
michael@0:         /* additional checks for color mismatch and excessive addition of
michael@0:          * high-frequencies */
michael@0:         2 * usad < thr && 2 * vsad < thr && !actrisk)
michael@0: #endif
michael@0:     {
michael@0:         int ifactor;
michael@0: #ifdef USE_SSD
michael@0:         /* TODO: optimize this later to not need sqr root */
michael@0:         sad = int_sqrt(sad);
michael@0: #endif
michael@0:         ifactor = (sad << MFQE_PRECISION) / thr;
michael@0:         ifactor >>= (qdiff >> 5);
michael@0: 
michael@0:         if (ifactor)
michael@0:         {
michael@0:             apply_ifactor(y, y_stride, yd, yd_stride,
michael@0:                           u, v, uv_stride,
michael@0:                           ud, vd, uvd_stride,
michael@0:                           blksize, ifactor);
michael@0:         }
michael@0:     }
michael@0:     else  /* else implicitly copy from previous frame */
michael@0:     {
michael@0:         if (blksize == 16)
michael@0:         {
michael@0:             vp8_copy_mem16x16(y, y_stride, yd, yd_stride);
michael@0:             vp8_copy_mem8x8(u, uv_stride, ud, uvd_stride);
michael@0:             vp8_copy_mem8x8(v, uv_stride, vd, uvd_stride);
michael@0:         }
michael@0:         else  /* if (blksize == 8) */
michael@0:         {
michael@0:             vp8_copy_mem8x8(y, y_stride, yd, yd_stride);
michael@0:             for (up = u, udp = ud, i = 0; i < uvblksize; ++i, up += uv_stride, udp += uvd_stride)
michael@0:                 vpx_memcpy(udp, up, uvblksize);
michael@0:             for (vp = v, vdp = vd, i = 0; i < uvblksize; ++i, vp += uv_stride, vdp += uvd_stride)
michael@0:                 vpx_memcpy(vdp, vp, uvblksize);
michael@0:         }
michael@0:     }
michael@0: }
michael@0: 
michael@0: static int qualify_inter_mb(const MODE_INFO *mode_info_context, int *map)
michael@0: {
michael@0:     if (mode_info_context->mbmi.mb_skip_coeff)
michael@0:         map[0] = map[1] = map[2] = map[3] = 1;
michael@0:     else if (mode_info_context->mbmi.mode==SPLITMV)
michael@0:     {
michael@0:         static int ndx[4][4] =
michael@0:         {
michael@0:             {0, 1, 4, 5},
michael@0:             {2, 3, 6, 7},
michael@0:             {8, 9, 12, 13},
michael@0:             {10, 11, 14, 15}
michael@0:         };
michael@0:         int i, j;
michael@0:         for (i=0; i<4; ++i)
michael@0:         {
michael@0:             map[i] = 1;
michael@0:             for (j=0; j<4 && map[j]; ++j)
michael@0:                 map[i] &= (mode_info_context->bmi[ndx[i][j]].mv.as_mv.row <= 2 &&
michael@0:                            mode_info_context->bmi[ndx[i][j]].mv.as_mv.col <= 2);
michael@0:         }
michael@0:     }
michael@0:     else
michael@0:     {
michael@0:         map[0] = map[1] = map[2] = map[3] =
michael@0:             (mode_info_context->mbmi.mode > B_PRED &&
michael@0:              abs(mode_info_context->mbmi.mv.as_mv.row) <= 2 &&
michael@0:              abs(mode_info_context->mbmi.mv.as_mv.col) <= 2);
michael@0:     }
michael@0:     return (map[0]+map[1]+map[2]+map[3]);
michael@0: }
michael@0: 
michael@0: void vp8_multiframe_quality_enhance
michael@0: (
michael@0:     VP8_COMMON *cm
michael@0: )
michael@0: {
michael@0:     YV12_BUFFER_CONFIG *show = cm->frame_to_show;
michael@0:     YV12_BUFFER_CONFIG *dest = &cm->post_proc_buffer;
michael@0: 
michael@0:     FRAME_TYPE frame_type = cm->frame_type;
michael@0:     /* Point at base of Mb MODE_INFO list has motion vectors etc */
michael@0:     const MODE_INFO *mode_info_context = cm->show_frame_mi;
michael@0:     int mb_row;
michael@0:     int mb_col;
michael@0:     int totmap, map[4];
michael@0:     int qcurr = cm->base_qindex;
michael@0:     int qprev = cm->postproc_state.last_base_qindex;
michael@0: 
michael@0:     unsigned char *y_ptr, *u_ptr, *v_ptr;
michael@0:     unsigned char *yd_ptr, *ud_ptr, *vd_ptr;
michael@0: 
michael@0:     /* Set up the buffer pointers */
michael@0:     y_ptr = show->y_buffer;
michael@0:     u_ptr = show->u_buffer;
michael@0:     v_ptr = show->v_buffer;
michael@0:     yd_ptr = dest->y_buffer;
michael@0:     ud_ptr = dest->u_buffer;
michael@0:     vd_ptr = dest->v_buffer;
michael@0: 
michael@0:     /* postprocess each macro block */
michael@0:     for (mb_row = 0; mb_row < cm->mb_rows; mb_row++)
michael@0:     {
michael@0:         for (mb_col = 0; mb_col < cm->mb_cols; mb_col++)
michael@0:         {
michael@0:             /* if motion is high there will likely be no benefit */
michael@0:             if (frame_type == INTER_FRAME) totmap = qualify_inter_mb(mode_info_context, map);
michael@0:             else totmap = (frame_type == KEY_FRAME ? 4 : 0);
michael@0:             if (totmap)
michael@0:             {
michael@0:                 if (totmap < 4)
michael@0:                 {
michael@0:                     int i, j;
michael@0:                     for (i=0; i<2; ++i)
michael@0:                         for (j=0; j<2; ++j)
michael@0:                         {
michael@0:                             if (map[i*2+j])
michael@0:                             {
michael@0:                                 multiframe_quality_enhance_block(8, qcurr, qprev,
michael@0:                                                                  y_ptr + 8*(i*show->y_stride+j),
michael@0:                                                                  u_ptr + 4*(i*show->uv_stride+j),
michael@0:                                                                  v_ptr + 4*(i*show->uv_stride+j),
michael@0:                                                                  show->y_stride,
michael@0:                                                                  show->uv_stride,
michael@0:                                                                  yd_ptr + 8*(i*dest->y_stride+j),
michael@0:                                                                  ud_ptr + 4*(i*dest->uv_stride+j),
michael@0:                                                                  vd_ptr + 4*(i*dest->uv_stride+j),
michael@0:                                                                  dest->y_stride,
michael@0:                                                                  dest->uv_stride);
michael@0:                             }
michael@0:                             else
michael@0:                             {
michael@0:                                 /* copy a 8x8 block */
michael@0:                                 int k;
michael@0:                                 unsigned char *up = u_ptr + 4*(i*show->uv_stride+j);
michael@0:                                 unsigned char *udp = ud_ptr + 4*(i*dest->uv_stride+j);
michael@0:                                 unsigned char *vp = v_ptr + 4*(i*show->uv_stride+j);
michael@0:                                 unsigned char *vdp = vd_ptr + 4*(i*dest->uv_stride+j);
michael@0:                                 vp8_copy_mem8x8(y_ptr + 8*(i*show->y_stride+j), show->y_stride,
michael@0:                                                 yd_ptr + 8*(i*dest->y_stride+j), dest->y_stride);
michael@0:                                 for (k = 0; k < 4; ++k, up += show->uv_stride, udp += dest->uv_stride,
michael@0:                                                         vp += show->uv_stride, vdp += dest->uv_stride)
michael@0:                                 {
michael@0:                                     vpx_memcpy(udp, up, 4);
michael@0:                                     vpx_memcpy(vdp, vp, 4);
michael@0:                                 }
michael@0:                             }
michael@0:                         }
michael@0:                 }
michael@0:                 else /* totmap = 4 */
michael@0:                 {
michael@0:                     multiframe_quality_enhance_block(16, qcurr, qprev, y_ptr,
michael@0:                                                      u_ptr, v_ptr,
michael@0:                                                      show->y_stride,
michael@0:                                                      show->uv_stride,
michael@0:                                                      yd_ptr, ud_ptr, vd_ptr,
michael@0:                                                      dest->y_stride,
michael@0:                                                      dest->uv_stride);
michael@0:                 }
michael@0:             }
michael@0:             else
michael@0:             {
michael@0:                 vp8_copy_mem16x16(y_ptr, show->y_stride, yd_ptr, dest->y_stride);
michael@0:                 vp8_copy_mem8x8(u_ptr, show->uv_stride, ud_ptr, dest->uv_stride);
michael@0:                 vp8_copy_mem8x8(v_ptr, show->uv_stride, vd_ptr, dest->uv_stride);
michael@0:             }
michael@0:             y_ptr += 16;
michael@0:             u_ptr += 8;
michael@0:             v_ptr += 8;
michael@0:             yd_ptr += 16;
michael@0:             ud_ptr += 8;
michael@0:             vd_ptr += 8;
michael@0:             mode_info_context++;     /* step to next MB */
michael@0:         }
michael@0: 
michael@0:         y_ptr += show->y_stride  * 16 - 16 * cm->mb_cols;
michael@0:         u_ptr += show->uv_stride *  8 - 8 * cm->mb_cols;
michael@0:         v_ptr += show->uv_stride *  8 - 8 * cm->mb_cols;
michael@0:         yd_ptr += dest->y_stride  * 16 - 16 * cm->mb_cols;
michael@0:         ud_ptr += dest->uv_stride *  8 - 8 * cm->mb_cols;
michael@0:         vd_ptr += dest->uv_stride *  8 - 8 * cm->mb_cols;
michael@0: 
michael@0:         mode_info_context++;         /* Skip border mb */
michael@0:     }
michael@0: }