vpx_dsp/avg.c

   1 /*
   2  *  Copyright (c) 2014 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 #include <stdlib.h>
  11
  12 #include "./vpx_dsp_rtcd.h"
  13 #include "vpx_ports/mem.h"
  14
  15 unsigned int vpx_avg_8x8_c(const uint8_t *s, int p) {
  16   int i, j;
  17   int sum = 0;
  18   for (i = 0; i < 8; ++i, s+=p)
  19     for (j = 0; j < 8; sum += s[j], ++j) {}
  20
  21   return (sum + 32) >> 6;
  22 }
  23
  24 unsigned int vpx_avg_4x4_c(const uint8_t *s, int p) {
  25   int i, j;
  26   int sum = 0;
  27   for (i = 0; i < 4; ++i, s+=p)
  28     for (j = 0; j < 4; sum += s[j], ++j) {}
  29
  30   return (sum + 8) >> 4;
  31 }
  32
  33 // src_diff: first pass, 9 bit, dynamic range [-255, 255]
  34 //           second pass, 12 bit, dynamic range [-2040, 2040]
  35 static void hadamard_col8(const int16_t *src_diff, int src_stride,
  36                           int16_t *coeff) {
  37   int16_t b0 = src_diff[0 * src_stride] + src_diff[1 * src_stride];
  38   int16_t b1 = src_diff[0 * src_stride] - src_diff[1 * src_stride];
  39   int16_t b2 = src_diff[2 * src_stride] + src_diff[3 * src_stride];
  40   int16_t b3 = src_diff[2 * src_stride] - src_diff[3 * src_stride];
  41   int16_t b4 = src_diff[4 * src_stride] + src_diff[5 * src_stride];
  42   int16_t b5 = src_diff[4 * src_stride] - src_diff[5 * src_stride];
  43   int16_t b6 = src_diff[6 * src_stride] + src_diff[7 * src_stride];
  44   int16_t b7 = src_diff[6 * src_stride] - src_diff[7 * src_stride];
  45
  46   int16_t c0 = b0 + b2;
  47   int16_t c1 = b1 + b3;
  48   int16_t c2 = b0 - b2;
  49   int16_t c3 = b1 - b3;
  50   int16_t c4 = b4 + b6;
  51   int16_t c5 = b5 + b7;
  52   int16_t c6 = b4 - b6;
  53   int16_t c7 = b5 - b7;
  54
  55   coeff[0] = c0 + c4;
  56   coeff[7] = c1 + c5;
  57   coeff[3] = c2 + c6;
  58   coeff[4] = c3 + c7;
  59   coeff[2] = c0 - c4;
  60   coeff[6] = c1 - c5;
  61   coeff[1] = c2 - c6;
  62   coeff[5] = c3 - c7;
  63 }
  64
  65 // The order of the output coeff of the hadamard is not important. For
  66 // optimization purposes the final transpose may be skipped.
  67 void vpx_hadamard_8x8_c(const int16_t *src_diff, int src_stride,
  68                         int16_t *coeff) {
  69   int idx;
  70   int16_t buffer[64];
  71   int16_t *tmp_buf = &buffer[0];
  72   for (idx = 0; idx < 8; ++idx) {
  73     hadamard_col8(src_diff, src_stride, tmp_buf);  // src_diff: 9 bit
  74                                                    // dynamic range [-255, 255]
  75     tmp_buf += 8;
  76     ++src_diff;
  77   }
  78
  79   tmp_buf = &buffer[0];
  80   for (idx = 0; idx < 8; ++idx) {
  81     hadamard_col8(tmp_buf, 8, coeff);  // tmp_buf: 12 bit
  82                                        // dynamic range [-2040, 2040]
  83     coeff += 8;  // coeff: 15 bit
  84                  // dynamic range [-16320, 16320]
  85     ++tmp_buf;
  86   }
  87 }
  88
  89 // In place 16x16 2D Hadamard transform
  90 void vpx_hadamard_16x16_c(const int16_t *src_diff, int src_stride,
  91                           int16_t *coeff) {
  92   int idx;
  93   for (idx = 0; idx < 4; ++idx) {
  94     // src_diff: 9 bit, dynamic range [-255, 255]
  95     const int16_t *src_ptr = src_diff + (idx >> 1) * 8 * src_stride
  96                                 + (idx & 0x01) * 8;
  97     vpx_hadamard_8x8_c(src_ptr, src_stride, coeff + idx * 64);
  98   }
  99
 100   // coeff: 15 bit, dynamic range [-16320, 16320]
 101   for (idx = 0; idx < 64; ++idx) {
 102     int16_t a0 = coeff[0];
 103     int16_t a1 = coeff[64];
 104     int16_t a2 = coeff[128];
 105     int16_t a3 = coeff[192];
 106
 107     int16_t b0 = (a0 + a1) >> 1;  // (a0 + a1): 16 bit, [-32640, 32640]
 108     int16_t b1 = (a0 - a1) >> 1;  // b0-b3: 15 bit, dynamic range
 109     int16_t b2 = (a2 + a3) >> 1;  // [-16320, 16320]
 110     int16_t b3 = (a2 - a3) >> 1;
 111
 112     coeff[0]   = b0 + b2;  // 16 bit, [-32640, 32640]
 113     coeff[64]  = b1 + b3;
 114     coeff[128] = b0 - b2;
 115     coeff[192] = b1 - b3;
 116
 117     ++coeff;
 118   }
 119 }
 120
 121 // coeff: 16 bits, dynamic range [-32640, 32640].
 122 // length: value range {16, 64, 256, 1024}.
 123 int vpx_satd_c(const int16_t *coeff, int length) {
 124   int i;
 125   int satd = 0;
 126   for (i = 0; i < length; ++i)
 127     satd += abs(coeff[i]);
 128
 129   // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
 130   return satd;
 131 }
 132
 133 // Integer projection onto row vectors.
 134 // height: value range {16, 32, 64}.
 135 void vpx_int_pro_row_c(int16_t hbuf[16], const uint8_t *ref,
 136                        const int ref_stride, const int height) {
 137   int idx;
 138   const int norm_factor = height >> 1;
 139   for (idx = 0; idx < 16; ++idx) {
 140     int i;
 141     hbuf[idx] = 0;
 142     // hbuf[idx]: 14 bit, dynamic range [0, 16320].
 143     for (i = 0; i < height; ++i)
 144       hbuf[idx] += ref[i * ref_stride];
 145     // hbuf[idx]: 9 bit, dynamic range [0, 510].
 146     hbuf[idx] /= norm_factor;
 147     ++ref;
 148   }
 149 }
 150
 151 // width: value range {16, 32, 64}.
 152 int16_t vpx_int_pro_col_c(const uint8_t *ref, const int width) {
 153   int idx;
 154   int16_t sum = 0;
 155   // sum: 14 bit, dynamic range [0, 16320]
 156   for (idx = 0; idx < width; ++idx)
 157     sum += ref[idx];
 158   return sum;
 159 }
 160
 161 // ref: [0 - 510]
 162 // src: [0 - 510]
 163 // bwl: {2, 3, 4}
 164 int vpx_vector_var_c(const int16_t *ref, const int16_t *src,
 165                      const int bwl) {
 166   int i;
 167   int width = 4 << bwl;
 168   int sse = 0, mean = 0, var;
 169
 170   for (i = 0; i < width; ++i) {
 171     int diff = ref[i] - src[i];  // diff: dynamic range [-510, 510], 10 bits.
 172     mean += diff;                // mean: dynamic range 16 bits.
 173     sse += diff * diff;          // sse:  dynamic range 26 bits.
 174   }
 175
 176   // (mean * mean): dynamic range 31 bits.
 177   var = sse - ((mean * mean) >> (bwl + 2));
 178   return var;
 179 }
 180
 181 void vpx_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp,
 182                       int *min, int *max) {
 183   int i, j;
 184   *min = 255;
 185   *max = 0;
 186   for (i = 0; i < 8; ++i, s += p, d += dp) {
 187     for (j = 0; j < 8; ++j) {
 188       int diff = abs(s[j]-d[j]);
 189       *min = diff < *min ? diff : *min;
 190       *max = diff > *max ? diff : *max;
 191     }
 192   }
 193 }
 194
 195 #if CONFIG_VP9_HIGHBITDEPTH
 196 unsigned int vpx_highbd_avg_8x8_c(const uint8_t *s8, int p) {
 197   int i, j;
 198   int sum = 0;
 199   const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
 200   for (i = 0; i < 8; ++i, s+=p)
 201     for (j = 0; j < 8; sum += s[j], ++j) {}
 202
 203   return (sum + 32) >> 6;
 204 }
 205
 206 unsigned int vpx_highbd_avg_4x4_c(const uint8_t *s8, int p) {
 207   int i, j;
 208   int sum = 0;
 209   const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
 210   for (i = 0; i < 4; ++i, s+=p)
 211     for (j = 0; j < 4; sum += s[j], ++j) {}
 212
 213   return (sum + 8) >> 4;
 214 }
 215
 216 void vpx_highbd_minmax_8x8_c(const uint8_t *s8, int p, const uint8_t *d8,
 217                              int dp, int *min, int *max) {
 218   int i, j;
 219   const uint16_t* s = CONVERT_TO_SHORTPTR(s8);
 220   const uint16_t* d = CONVERT_TO_SHORTPTR(d8);
 221   *min = 255;
 222   *max = 0;
 223   for (i = 0; i < 8; ++i, s += p, d += dp) {
 224     for (j = 0; j < 8; ++j) {
 225       int diff = abs(s[j]-d[j]);
 226       *min = diff < *min ? diff : *min;
 227       *max = diff > *max ? diff : *max;
 228     }
 229   }
 230 }
 231 #endif  // CONFIG_VP9_HIGHBITDEPTH
 232
 233