src/third_party/libvpx/source/libvpx/vp8/encoder/arm/neon/denoising_neon.c

   1 /*
   2  *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
   3  *
   4  *  Use of this source code is governed by a BSD-style license
   5  *  that can be found in the LICENSE file in the root of the source
   6  *  tree. An additional intellectual property rights grant can be found
   7  *  in the file PATENTS.  All contributing project authors may
   8  *  be found in the AUTHORS file in the root of the source tree.
   9  */
  10
  11 #include <arm_neon.h>
  12
  13 #include "vp8/encoder/denoising.h"
  14 #include "vpx_mem/vpx_mem.h"
  15 #include "./vp8_rtcd.h"
  16
  17 /*
  18  * The filter function was modified to reduce the computational complexity.
  19  *
  20  * Step 1:
  21  *  Instead of applying tap coefficients for each pixel, we calculated the
  22  *  pixel adjustments vs. pixel diff value ahead of time.
  23  *     adjustment = filtered_value - current_raw
  24  *                = (filter_coefficient * diff + 128) >> 8
  25  *  where
  26  *     filter_coefficient = (255 << 8) / (256 + ((abs_diff * 330) >> 3));
  27  *     filter_coefficient += filter_coefficient /
  28  *                           (3 + motion_magnitude_adjustment);
  29  *     filter_coefficient is clamped to 0 ~ 255.
  30  *
  31  * Step 2:
  32  *  The adjustment vs. diff curve becomes flat very quick when diff increases.
  33  *  This allowed us to use only several levels to approximate the curve without
  34  *  changing the filtering algorithm too much.
  35  *  The adjustments were further corrected by checking the motion magnitude.
  36  *  The levels used are:
  37  *      diff          level       adjustment w/o       adjustment w/
  38  *                               motion correction    motion correction
  39  *      [-255, -16]     3              -6                   -7
  40  *      [-15, -8]       2              -4                   -5
  41  *      [-7, -4]        1              -3                   -4
  42  *      [-3, 3]         0              diff                 diff
  43  *      [4, 7]          1               3                    4
  44  *      [8, 15]         2               4                    5
  45  *      [16, 255]       3               6                    7
  46  */
  47
  48 int vp8_denoiser_filter_neon(YV12_BUFFER_CONFIG *mc_running_avg,
  49                              YV12_BUFFER_CONFIG *running_avg,
  50                              MACROBLOCK *signal, unsigned int motion_magnitude,
  51                              int y_offset, int uv_offset) {
  52     /* If motion_magnitude is small, making the denoiser more aggressive by
  53      * increasing the adjustment for each level, level1 adjustment is
  54      * increased, the deltas stay the same.
  55      */
  56     const uint8x16_t v_level1_adjustment = vdupq_n_u8(
  57         (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 4 : 3);
  58     const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
  59     const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
  60     const uint8x16_t v_level1_threshold = vdupq_n_u8(4);
  61     const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
  62     const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
  63
  64     /* Local variables for array pointers and strides. */
  65     unsigned char *sig = signal->thismb;
  66     int            sig_stride = 16;
  67     unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
  68     int            mc_running_avg_y_stride = mc_running_avg->y_stride;
  69     unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
  70     int            running_avg_y_stride = running_avg->y_stride;
  71     int64x2_t v_sum_diff_total = vdupq_n_s64(0);
  72
  73     /* Go over lines. */
  74     int i;
  75     for (i = 0; i < 16; ++i) {
  76         /* Load inputs. */
  77         const uint8x16_t v_sig = vld1q_u8(sig);
  78         const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
  79
  80         /* Calculate absolute difference and sign masks. */
  81         const uint8x16_t v_abs_diff      = vabdq_u8(v_sig, v_mc_running_avg_y);
  82         const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig, v_mc_running_avg_y);
  83         const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig, v_mc_running_avg_y);
  84
  85         /* Figure out which level that put us in. */
  86         const uint8x16_t v_level1_mask = vcleq_u8(v_level1_threshold,
  87                                                   v_abs_diff);
  88         const uint8x16_t v_level2_mask = vcleq_u8(v_level2_threshold,
  89                                                   v_abs_diff);
  90         const uint8x16_t v_level3_mask = vcleq_u8(v_level3_threshold,
  91                                                   v_abs_diff);
  92
  93         /* Calculate absolute adjustments for level 1, 2 and 3. */
  94         const uint8x16_t v_level2_adjustment = vandq_u8(v_level2_mask,
  95                                                         v_delta_level_1_and_2);
  96         const uint8x16_t v_level3_adjustment = vandq_u8(v_level3_mask,
  97                                                         v_delta_level_2_and_3);
  98         const uint8x16_t v_level1and2_adjustment = vaddq_u8(v_level1_adjustment,
  99             v_level2_adjustment);
 100         const uint8x16_t v_level1and2and3_adjustment = vaddq_u8(
 101             v_level1and2_adjustment, v_level3_adjustment);
 102
 103         /* Figure adjustment absolute value by selecting between the absolute
 104          * difference if in level0 or the value for level 1, 2 and 3.
 105          */
 106         const uint8x16_t v_abs_adjustment = vbslq_u8(v_level1_mask,
 107             v_level1and2and3_adjustment, v_abs_diff);
 108
 109         /* Calculate positive and negative adjustments. Apply them to the signal
 110          * and accumulate them. Adjustments are less than eight and the maximum
 111          * sum of them (7 * 16) can fit in a signed char.
 112          */
 113         const uint8x16_t v_pos_adjustment = vandq_u8(v_diff_pos_mask,
 114                                                      v_abs_adjustment);
 115         const uint8x16_t v_neg_adjustment = vandq_u8(v_diff_neg_mask,
 116                                                      v_abs_adjustment);
 117
 118         uint8x16_t v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
 119         v_running_avg_y = vqsubq_u8(v_running_avg_y, v_neg_adjustment);
 120
 121         /* Store results. */
 122         vst1q_u8(running_avg_y, v_running_avg_y);
 123
 124         /* Sum all the accumulators to have the sum of all pixel differences
 125          * for this macroblock.
 126          */
 127         {
 128             const int8x16_t v_sum_diff =
 129                 vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
 130                           vreinterpretq_s8_u8(v_neg_adjustment));
 131
 132             const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff);
 133
 134             const int32x4_t fedc_ba98_7654_3210 =
 135                 vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
 136
 137             const int64x2_t fedcba98_76543210 =
 138                 vpaddlq_s32(fedc_ba98_7654_3210);
 139
 140             v_sum_diff_total = vqaddq_s64(v_sum_diff_total, fedcba98_76543210);
 141         }
 142
 143         /* Update pointers for next iteration. */
 144         sig += sig_stride;
 145         mc_running_avg_y += mc_running_avg_y_stride;
 146         running_avg_y += running_avg_y_stride;
 147     }
 148
 149     /* Too much adjustments => copy block. */
 150     {
 151         const int64x1_t x = vqadd_s64(vget_high_s64(v_sum_diff_total),
 152                                       vget_low_s64(v_sum_diff_total));
 153         const int s0 = vget_lane_s32(vabs_s32(vreinterpret_s32_s64(x)), 0);
 154
 155         if (s0 > SUM_DIFF_THRESHOLD)
 156             return COPY_BLOCK;
 157     }
 158
 159     /* Tell above level that block was filtered. */
 160     running_avg_y -= running_avg_y_stride * 16;
 161     sig -= sig_stride * 16;
 162
 163     vp8_copy_mem16x16(running_avg_y, running_avg_y_stride, sig, sig_stride);
 164
 165     return FILTER_BLOCK;
 166 }