michael@0: /*
michael@0:  *  Copyright (c) 2010 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: #include "./vpx_scale_rtcd.h"
michael@0: #include "vp8/common/onyxc_int.h"
michael@0: #include "onyx_int.h"
michael@0: #include "quantize.h"
michael@0: #include "vpx_mem/vpx_mem.h"
michael@0: #include "vpx_scale/vpx_scale.h"
michael@0: #include "vp8/common/alloccommon.h"
michael@0: #include "vp8/common/loopfilter.h"
michael@0: #if ARCH_ARM
michael@0: #include "vpx_ports/arm.h"
michael@0: #endif
michael@0: 
michael@0: extern int vp8_calc_ss_err(YV12_BUFFER_CONFIG *source, YV12_BUFFER_CONFIG *dest);
michael@0: 
michael@0: void vp8_yv12_copy_partial_frame_c(YV12_BUFFER_CONFIG *src_ybc,
michael@0:                                    YV12_BUFFER_CONFIG *dst_ybc)
michael@0: {
michael@0:     unsigned char *src_y, *dst_y;
michael@0:     int yheight;
michael@0:     int ystride;
michael@0:     int yoffset;
michael@0:     int linestocopy;
michael@0: 
michael@0:     yheight  = src_ybc->y_height;
michael@0:     ystride  = src_ybc->y_stride;
michael@0: 
michael@0:     /* number of MB rows to use in partial filtering */
michael@0:     linestocopy = (yheight >> 4) / PARTIAL_FRAME_FRACTION;
michael@0:     linestocopy = linestocopy ? linestocopy << 4 : 16;     /* 16 lines per MB */
michael@0: 
michael@0:     /* Copy extra 4 so that full filter context is available if filtering done
michael@0:      * on the copied partial frame and not original. Partial filter does mb
michael@0:      * filtering for top row also, which can modify3 pixels above.
michael@0:      */
michael@0:     linestocopy += 4;
michael@0:     /* partial image starts at ~middle of frame (macroblock border)*/
michael@0:     yoffset  = ystride * (((yheight >> 5) * 16) - 4);
michael@0:     src_y = src_ybc->y_buffer + yoffset;
michael@0:     dst_y = dst_ybc->y_buffer + yoffset;
michael@0: 
michael@0:     vpx_memcpy(dst_y, src_y, ystride * linestocopy);
michael@0: }
michael@0: 
michael@0: static int calc_partial_ssl_err(YV12_BUFFER_CONFIG *source,
michael@0:                                 YV12_BUFFER_CONFIG *dest)
michael@0: {
michael@0:     int i, j;
michael@0:     int Total = 0;
michael@0:     int srcoffset, dstoffset;
michael@0:     unsigned char *src = source->y_buffer;
michael@0:     unsigned char *dst = dest->y_buffer;
michael@0: 
michael@0:     int linestocopy;
michael@0: 
michael@0:     /* number of MB rows to use in partial filtering */
michael@0:     linestocopy = (source->y_height >> 4) / PARTIAL_FRAME_FRACTION;
michael@0:     linestocopy = linestocopy ? linestocopy << 4 : 16;     /* 16 lines per MB */
michael@0: 
michael@0: 
michael@0:     /* partial image starts at ~middle of frame (macroblock border)*/
michael@0:     srcoffset = source->y_stride * ((dest->y_height >> 5) * 16);
michael@0:     dstoffset = dest->y_stride   * ((dest->y_height >> 5) * 16);
michael@0: 
michael@0:     src += srcoffset;
michael@0:     dst += dstoffset;
michael@0: 
michael@0:     /* Loop through the Y plane raw and reconstruction data summing
michael@0:      * (square differences)
michael@0:      */
michael@0:     for (i = 0; i < linestocopy; i += 16)
michael@0:     {
michael@0:         for (j = 0; j < source->y_width; j += 16)
michael@0:         {
michael@0:             unsigned int sse;
michael@0:             Total += vp8_mse16x16(src + j, source->y_stride,
michael@0:                                                      dst + j, dest->y_stride,
michael@0:                                                      &sse);
michael@0:         }
michael@0: 
michael@0:         src += 16 * source->y_stride;
michael@0:         dst += 16 * dest->y_stride;
michael@0:     }
michael@0: 
michael@0:     return Total;
michael@0: }
michael@0: 
michael@0: /* Enforce a minimum filter level based upon baseline Q */
michael@0: static int get_min_filter_level(VP8_COMP *cpi, int base_qindex)
michael@0: {
michael@0:     int min_filter_level;
michael@0: 
michael@0:     if (cpi->source_alt_ref_active && cpi->common.refresh_golden_frame &&
michael@0:         !cpi->common.refresh_alt_ref_frame)
michael@0:         min_filter_level = 0;
michael@0:     else
michael@0:     {
michael@0:         if (base_qindex <= 6)
michael@0:             min_filter_level = 0;
michael@0:         else if (base_qindex <= 16)
michael@0:             min_filter_level = 1;
michael@0:         else
michael@0:             min_filter_level = (base_qindex / 8);
michael@0:     }
michael@0: 
michael@0:     return min_filter_level;
michael@0: }
michael@0: 
michael@0: /* Enforce a maximum filter level based upon baseline Q */
michael@0: static int get_max_filter_level(VP8_COMP *cpi, int base_qindex)
michael@0: {
michael@0:     /* PGW August 2006: Highest filter values almost always a bad idea */
michael@0: 
michael@0:     /* jbb chg: 20100118 - not so any more with this overquant stuff allow
michael@0:      * high values with lots of intra coming in.
michael@0:      */
michael@0:     int max_filter_level = MAX_LOOP_FILTER;
michael@0:     (void)base_qindex;
michael@0: 
michael@0:     if (cpi->twopass.section_intra_rating > 8)
michael@0:         max_filter_level = MAX_LOOP_FILTER * 3 / 4;
michael@0: 
michael@0:     return max_filter_level;
michael@0: }
michael@0: 
michael@0: void vp8cx_pick_filter_level_fast(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi)
michael@0: {
michael@0:     VP8_COMMON *cm = &cpi->common;
michael@0: 
michael@0:     int best_err = 0;
michael@0:     int filt_err = 0;
michael@0:     int min_filter_level = get_min_filter_level(cpi, cm->base_qindex);
michael@0:     int max_filter_level = get_max_filter_level(cpi, cm->base_qindex);
michael@0:     int filt_val;
michael@0:     int best_filt_val = cm->filter_level;
michael@0:     YV12_BUFFER_CONFIG * saved_frame = cm->frame_to_show;
michael@0: 
michael@0:     /* Replace unfiltered frame buffer with a new one */
michael@0:     cm->frame_to_show = &cpi->pick_lf_lvl_frame;
michael@0: 
michael@0:     if (cm->frame_type == KEY_FRAME)
michael@0:         cm->sharpness_level = 0;
michael@0:     else
michael@0:         cm->sharpness_level = cpi->oxcf.Sharpness;
michael@0: 
michael@0:     if (cm->sharpness_level != cm->last_sharpness_level)
michael@0:     {
michael@0:         vp8_loop_filter_update_sharpness(&cm->lf_info, cm->sharpness_level);
michael@0:         cm->last_sharpness_level = cm->sharpness_level;
michael@0:     }
michael@0: 
michael@0:     /* Start the search at the previous frame filter level unless it is
michael@0:      * now out of range.
michael@0:      */
michael@0:     if (cm->filter_level < min_filter_level)
michael@0:         cm->filter_level = min_filter_level;
michael@0:     else if (cm->filter_level > max_filter_level)
michael@0:         cm->filter_level = max_filter_level;
michael@0: 
michael@0:     filt_val = cm->filter_level;
michael@0:     best_filt_val = filt_val;
michael@0: 
michael@0:     /* Get the err using the previous frame's filter value. */
michael@0: 
michael@0:     /* Copy the unfiltered / processed recon buffer to the new buffer */
michael@0:     vp8_yv12_copy_partial_frame(saved_frame, cm->frame_to_show);
michael@0:     vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val);
michael@0: 
michael@0:     best_err = calc_partial_ssl_err(sd, cm->frame_to_show);
michael@0: 
michael@0:     filt_val -= 1 + (filt_val > 10);
michael@0: 
michael@0:     /* Search lower filter levels */
michael@0:     while (filt_val >= min_filter_level)
michael@0:     {
michael@0:         /* Apply the loop filter */
michael@0:         vp8_yv12_copy_partial_frame(saved_frame, cm->frame_to_show);
michael@0:         vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val);
michael@0: 
michael@0:         /* Get the err for filtered frame */
michael@0:         filt_err = calc_partial_ssl_err(sd, cm->frame_to_show);
michael@0: 
michael@0:         /* Update the best case record or exit loop. */
michael@0:         if (filt_err < best_err)
michael@0:         {
michael@0:             best_err = filt_err;
michael@0:             best_filt_val = filt_val;
michael@0:         }
michael@0:         else
michael@0:             break;
michael@0: 
michael@0:         /* Adjust filter level */
michael@0:         filt_val -= 1 + (filt_val > 10);
michael@0:     }
michael@0: 
michael@0:     /* Search up (note that we have already done filt_val = cm->filter_level) */
michael@0:     filt_val = cm->filter_level + 1 + (filt_val > 10);
michael@0: 
michael@0:     if (best_filt_val == cm->filter_level)
michael@0:     {
michael@0:         /* Resist raising filter level for very small gains */
michael@0:         best_err -= (best_err >> 10);
michael@0: 
michael@0:         while (filt_val < max_filter_level)
michael@0:         {
michael@0:             /* Apply the loop filter */
michael@0:             vp8_yv12_copy_partial_frame(saved_frame, cm->frame_to_show);
michael@0: 
michael@0:             vp8_loop_filter_partial_frame(cm, &cpi->mb.e_mbd, filt_val);
michael@0: 
michael@0:             /* Get the err for filtered frame */
michael@0:             filt_err = calc_partial_ssl_err(sd, cm->frame_to_show);
michael@0: 
michael@0:             /* Update the best case record or exit loop. */
michael@0:             if (filt_err < best_err)
michael@0:             {
michael@0:                 /* Do not raise filter level if improvement is < 1 part
michael@0:                  * in 4096
michael@0:                  */
michael@0:                 best_err = filt_err - (filt_err >> 10);
michael@0: 
michael@0:                 best_filt_val = filt_val;
michael@0:             }
michael@0:             else
michael@0:                 break;
michael@0: 
michael@0:             /* Adjust filter level */
michael@0:             filt_val += 1 + (filt_val > 10);
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     cm->filter_level = best_filt_val;
michael@0: 
michael@0:     if (cm->filter_level < min_filter_level)
michael@0:         cm->filter_level = min_filter_level;
michael@0: 
michael@0:     if (cm->filter_level > max_filter_level)
michael@0:         cm->filter_level = max_filter_level;
michael@0: 
michael@0:     /* restore unfiltered frame pointer */
michael@0:     cm->frame_to_show = saved_frame;
michael@0: }
michael@0: 
michael@0: /* Stub function for now Alt LF not used */
michael@0: void vp8cx_set_alt_lf_level(VP8_COMP *cpi, int filt_val)
michael@0: {
michael@0:     MACROBLOCKD *mbd = &cpi->mb.e_mbd;
michael@0:     (void) filt_val;
michael@0: 
michael@0:     mbd->segment_feature_data[MB_LVL_ALT_LF][0] = cpi->segment_feature_data[MB_LVL_ALT_LF][0];
michael@0:     mbd->segment_feature_data[MB_LVL_ALT_LF][1] = cpi->segment_feature_data[MB_LVL_ALT_LF][1];
michael@0:     mbd->segment_feature_data[MB_LVL_ALT_LF][2] = cpi->segment_feature_data[MB_LVL_ALT_LF][2];
michael@0:     mbd->segment_feature_data[MB_LVL_ALT_LF][3] = cpi->segment_feature_data[MB_LVL_ALT_LF][3];
michael@0: }
michael@0: 
michael@0: void vp8cx_pick_filter_level(YV12_BUFFER_CONFIG *sd, VP8_COMP *cpi)
michael@0: {
michael@0:     VP8_COMMON *cm = &cpi->common;
michael@0: 
michael@0:     int best_err = 0;
michael@0:     int filt_err = 0;
michael@0:     int min_filter_level = get_min_filter_level(cpi, cm->base_qindex);
michael@0:     int max_filter_level = get_max_filter_level(cpi, cm->base_qindex);
michael@0: 
michael@0:     int filter_step;
michael@0:     int filt_high = 0;
michael@0:     /* Start search at previous frame filter level */
michael@0:     int filt_mid = cm->filter_level;
michael@0:     int filt_low = 0;
michael@0:     int filt_best;
michael@0:     int filt_direction = 0;
michael@0: 
michael@0:     /* Bias against raising loop filter and in favor of lowering it */
michael@0:     int Bias = 0;
michael@0: 
michael@0:     int ss_err[MAX_LOOP_FILTER + 1];
michael@0: 
michael@0:     YV12_BUFFER_CONFIG * saved_frame = cm->frame_to_show;
michael@0: 
michael@0:     vpx_memset(ss_err, 0, sizeof(ss_err));
michael@0: 
michael@0:     /* Replace unfiltered frame buffer with a new one */
michael@0:     cm->frame_to_show = &cpi->pick_lf_lvl_frame;
michael@0: 
michael@0:     if (cm->frame_type == KEY_FRAME)
michael@0:         cm->sharpness_level = 0;
michael@0:     else
michael@0:         cm->sharpness_level = cpi->oxcf.Sharpness;
michael@0: 
michael@0:     /* Start the search at the previous frame filter level unless it is
michael@0:      * now out of range.
michael@0:      */
michael@0:     filt_mid = cm->filter_level;
michael@0: 
michael@0:     if (filt_mid < min_filter_level)
michael@0:         filt_mid = min_filter_level;
michael@0:     else if (filt_mid > max_filter_level)
michael@0:         filt_mid = max_filter_level;
michael@0: 
michael@0:     /* Define the initial step size */
michael@0:     filter_step = (filt_mid < 16) ? 4 : filt_mid / 4;
michael@0: 
michael@0:     /* Get baseline error score */
michael@0: 
michael@0:     /* Copy the unfiltered / processed recon buffer to the new buffer */
michael@0:     vpx_yv12_copy_y(saved_frame, cm->frame_to_show);
michael@0: 
michael@0:     vp8cx_set_alt_lf_level(cpi, filt_mid);
michael@0:     vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_mid);
michael@0: 
michael@0:     best_err = vp8_calc_ss_err(sd, cm->frame_to_show);
michael@0: 
michael@0:     ss_err[filt_mid] = best_err;
michael@0: 
michael@0:     filt_best = filt_mid;
michael@0: 
michael@0:     while (filter_step > 0)
michael@0:     {
michael@0:         Bias = (best_err >> (15 - (filt_mid / 8))) * filter_step;
michael@0: 
michael@0:         if (cpi->twopass.section_intra_rating < 20)
michael@0:             Bias = Bias * cpi->twopass.section_intra_rating / 20;
michael@0: 
michael@0:         filt_high = ((filt_mid + filter_step) > max_filter_level) ? max_filter_level : (filt_mid + filter_step);
michael@0:         filt_low = ((filt_mid - filter_step) < min_filter_level) ? min_filter_level : (filt_mid - filter_step);
michael@0: 
michael@0:         if ((filt_direction <= 0) && (filt_low != filt_mid))
michael@0:         {
michael@0:             if(ss_err[filt_low] == 0)
michael@0:             {
michael@0:                 /* Get Low filter error score */
michael@0:                 vpx_yv12_copy_y(saved_frame, cm->frame_to_show);
michael@0:                 vp8cx_set_alt_lf_level(cpi, filt_low);
michael@0:                 vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_low);
michael@0: 
michael@0:                 filt_err = vp8_calc_ss_err(sd, cm->frame_to_show);
michael@0:                 ss_err[filt_low] = filt_err;
michael@0:             }
michael@0:             else
michael@0:                 filt_err = ss_err[filt_low];
michael@0: 
michael@0:             /* If value is close to the best so far then bias towards a
michael@0:              * lower loop filter value.
michael@0:              */
michael@0:             if ((filt_err - Bias) < best_err)
michael@0:             {
michael@0:                 /* Was it actually better than the previous best? */
michael@0:                 if (filt_err < best_err)
michael@0:                     best_err = filt_err;
michael@0: 
michael@0:                 filt_best = filt_low;
michael@0:             }
michael@0:         }
michael@0: 
michael@0:         /* Now look at filt_high */
michael@0:         if ((filt_direction >= 0) && (filt_high != filt_mid))
michael@0:         {
michael@0:             if(ss_err[filt_high] == 0)
michael@0:             {
michael@0:                 vpx_yv12_copy_y(saved_frame, cm->frame_to_show);
michael@0:                 vp8cx_set_alt_lf_level(cpi, filt_high);
michael@0:                 vp8_loop_filter_frame_yonly(cm, &cpi->mb.e_mbd, filt_high);
michael@0: 
michael@0:                 filt_err = vp8_calc_ss_err(sd, cm->frame_to_show);
michael@0:                 ss_err[filt_high] = filt_err;
michael@0:             }
michael@0:             else
michael@0:                 filt_err = ss_err[filt_high];
michael@0: 
michael@0:             /* Was it better than the previous best? */
michael@0:             if (filt_err < (best_err - Bias))
michael@0:             {
michael@0:                 best_err = filt_err;
michael@0:                 filt_best = filt_high;
michael@0:             }
michael@0:         }
michael@0: 
michael@0:         /* Half the step distance if the best filter value was the same
michael@0:          * as last time
michael@0:          */
michael@0:         if (filt_best == filt_mid)
michael@0:         {
michael@0:             filter_step = filter_step / 2;
michael@0:             filt_direction = 0;
michael@0:         }
michael@0:         else
michael@0:         {
michael@0:             filt_direction = (filt_best < filt_mid) ? -1 : 1;
michael@0:             filt_mid = filt_best;
michael@0:         }
michael@0:     }
michael@0: 
michael@0:     cm->filter_level = filt_best;
michael@0: 
michael@0:     /* restore unfiltered frame pointer */
michael@0:     cm->frame_to_show = saved_frame;
michael@0: }