vp9_convolve.c 10.5 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
2
3
4
5
6
7
8
9
/*
 *  Copyright (c) 2013 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */
Christian Duvivier's avatar
Christian Duvivier committed
10

John Koleszar's avatar
John Koleszar committed
11
12
13
14
15
#include <assert.h>

#include "./vpx_config.h"
#include "./vp9_rtcd.h"
#include "vp9/common/vp9_common.h"
16
#include "vp9/common/vp9_convolve.h"
17
#include "vp9/common/vp9_filter.h"
John Koleszar's avatar
John Koleszar committed
18
#include "vpx/vpx_integer.h"
Christian Duvivier's avatar
Christian Duvivier committed
19
#include "vpx_ports/mem.h"
John Koleszar's avatar
John Koleszar committed
20

Dmitry Kovalev's avatar
Dmitry Kovalev committed
21
22
static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
                           uint8_t *dst, ptrdiff_t dst_stride,
23
                           const InterpKernel *x_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
24
25
26
                           int x0_q4, int x_step_q4, int w, int h) {
  int x, y;
  src -= SUBPEL_TAPS / 2 - 1;
John Koleszar's avatar
John Koleszar committed
27
  for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
28
    int x_q4 = x0_q4;
John Koleszar's avatar
John Koleszar committed
29
    for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
30
31
32
33
34
      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_x[k] * x_filter[k];
35
      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
John Koleszar's avatar
John Koleszar committed
36
37
38
39
40
41
42
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
43
44
static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
                               uint8_t *dst, ptrdiff_t dst_stride,
45
                               const InterpKernel *x_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
46
47
48
                               int x0_q4, int x_step_q4, int w, int h) {
  int x, y;
  src -= SUBPEL_TAPS / 2 - 1;
John Koleszar's avatar
John Koleszar committed
49
  for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
50
    int x_q4 = x0_q4;
John Koleszar's avatar
John Koleszar committed
51
    for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
52
53
54
55
56
      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_x[k] * x_filter[k];
57
      dst[x] = ROUND_POWER_OF_TWO(dst[x] +
Dmitry Kovalev's avatar
Dmitry Kovalev committed
58
          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
John Koleszar's avatar
John Koleszar committed
59
60
61
62
63
64
65
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
66
67
static void convolve_vert(const uint8_t *src, ptrdiff_t src_stride,
                          uint8_t *dst, ptrdiff_t dst_stride,
68
                          const InterpKernel *y_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
69
70
71
                          int y0_q4, int y_step_q4, int w, int h) {
  int x, y;
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
John Koleszar's avatar
John Koleszar committed
72

73
  for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
74
    int y_q4 = y0_q4;
John Koleszar's avatar
John Koleszar committed
75
    for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
76
77
78
79
80
81
      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
John Koleszar's avatar
John Koleszar committed
82
83
84
85
86
87
88
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
89
90
static void convolve_avg_vert(const uint8_t *src, ptrdiff_t src_stride,
                              uint8_t *dst, ptrdiff_t dst_stride,
91
                              const InterpKernel *y_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
92
93
94
                              int y0_q4, int y_step_q4, int w, int h) {
  int x, y;
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
John Koleszar's avatar
John Koleszar committed
95

96
  for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
97
    int y_q4 = y0_q4;
John Koleszar's avatar
John Koleszar committed
98
    for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
99
100
101
102
103
      const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
      int k, sum = 0;
      for (k = 0; k < SUBPEL_TAPS; ++k)
        sum += src_y[k * src_stride] * y_filter[k];
104
      dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
Dmitry Kovalev's avatar
Dmitry Kovalev committed
105
          clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)), 1);
John Koleszar's avatar
John Koleszar committed
106
107
108
109
110
111
112
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
113
114
static void convolve(const uint8_t *src, ptrdiff_t src_stride,
                     uint8_t *dst, ptrdiff_t dst_stride,
115
                     const InterpKernel *const x_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
116
                     int x0_q4, int x_step_q4,
117
                     const InterpKernel *const y_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
118
119
                     int y0_q4, int y_step_q4,
                     int w, int h) {
120
121
122
123
124
125
126
127
128
129
130
131
132
  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
  // 2d filtering proceeds in 2 steps:
  //   (1) Interpolate horizontally into an intermediate buffer, temp.
  //   (2) Interpolate temp vertically to derive the sub-pixel result.
  // Deriving the maximum number of rows in the temp buffer (135):
  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
  // --Largest block size is 64x64 pixels.
  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
  //   original frame (in 1/16th pixel units).
  // --Must round-up because block may be located at sub-pixel position.
  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
  uint8_t temp[135 * 64];
Dmitry Kovalev's avatar
Dmitry Kovalev committed
133
  int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + SUBPEL_TAPS;
134

135
136
  assert(w <= 64);
  assert(h <= 64);
137
138
  assert(y_step_q4 <= 32);
  assert(x_step_q4 <= 32);
139
140
141

  if (intermediate_height < h)
    intermediate_height = h;
142

Dmitry Kovalev's avatar
Dmitry Kovalev committed
143
144
145
146
147
148
  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp, 64,
                 x_filters, x0_q4, x_step_q4, w, intermediate_height);
  convolve_vert(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst, dst_stride,
                y_filters, y0_q4, y_step_q4, w, h);
}

149
static const InterpKernel *get_filter_base(const int16_t *filter) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
150
151
  // NOTE: This assumes that the filter table is 256-byte aligned.
  // TODO(agrange) Modify to make independent of table alignment.
152
  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
Dmitry Kovalev's avatar
Dmitry Kovalev committed
153
154
}

155
static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
156
  return (int)((const InterpKernel *)(intptr_t)f - base);
John Koleszar's avatar
John Koleszar committed
157
158
}

159
160
void vp9_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
                           uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
161
162
163
                           const int16_t *filter_x, int x_step_q4,
                           const int16_t *filter_y, int y_step_q4,
                           int w, int h) {
164
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
165
166
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

167
168
169
  (void)filter_y;
  (void)y_step_q4;

Dmitry Kovalev's avatar
Dmitry Kovalev committed
170
171
  convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
                 x0_q4, x_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
172
173
}

174
175
void vp9_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
                               uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
176
177
178
                               const int16_t *filter_x, int x_step_q4,
                               const int16_t *filter_y, int y_step_q4,
                               int w, int h) {
179
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
180
181
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

182
183
184
  (void)filter_y;
  (void)y_step_q4;

Dmitry Kovalev's avatar
Dmitry Kovalev committed
185
186
  convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
                     x0_q4, x_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
187
188
}

189
190
void vp9_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
                          uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
191
192
193
                          const int16_t *filter_x, int x_step_q4,
                          const int16_t *filter_y, int y_step_q4,
                          int w, int h) {
194
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
195
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
196
197
198
199

  (void)filter_x;
  (void)x_step_q4;

Dmitry Kovalev's avatar
Dmitry Kovalev committed
200
201
  convolve_vert(src, src_stride, dst, dst_stride, filters_y,
                y0_q4, y_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
202
203
}

204
205
void vp9_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
                              uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
206
207
208
                              const int16_t *filter_x, int x_step_q4,
                              const int16_t *filter_y, int y_step_q4,
                              int w, int h) {
209
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
210
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
211
212
213
214

  (void)filter_x;
  (void)x_step_q4;

Dmitry Kovalev's avatar
Dmitry Kovalev committed
215
216
  convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
                    y0_q4, y_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
217
218
}

219
220
void vp9_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
                     uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
221
222
223
                     const int16_t *filter_x, int x_step_q4,
                     const int16_t *filter_y, int y_step_q4,
                     int w, int h) {
224
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
225
226
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

227
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
228
229
230
231
232
  const int y0_q4 = get_filter_offset(filter_y, filters_y);

  convolve(src, src_stride, dst, dst_stride,
           filters_x, x0_q4, x_step_q4,
           filters_y, y0_q4, y_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
233
234
}

235
236
void vp9_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
                         uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
237
238
239
                         const int16_t *filter_x, int x_step_q4,
                         const int16_t *filter_y, int y_step_q4,
                         int w, int h) {
Christian Duvivier's avatar
Christian Duvivier committed
240
  /* Fixed size intermediate buffer places limits on parameters. */
241
242
243
  DECLARE_ALIGNED_ARRAY(16, uint8_t, temp, 64 * 64);
  assert(w <= 64);
  assert(h <= 64);
Christian Duvivier's avatar
Christian Duvivier committed
244

Dmitry Kovalev's avatar
Dmitry Kovalev committed
245
246
247
  vp9_convolve8_c(src, src_stride, temp, 64,
                  filter_x, x_step_q4, filter_y, y_step_q4, w, h);
  vp9_convolve_avg_c(temp, 64, dst, dst_stride, NULL, 0, NULL, 0, w, h);
John Koleszar's avatar
John Koleszar committed
248
}
249

250
251
252
253
254
255
256
void vp9_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride,
                         uint8_t *dst, ptrdiff_t dst_stride,
                         const int16_t *filter_x, int filter_x_stride,
                         const int16_t *filter_y, int filter_y_stride,
                         int w, int h) {
  int r;

257
258
259
  (void)filter_x;  (void)filter_x_stride;
  (void)filter_y;  (void)filter_y_stride;

260
  for (r = h; r > 0; --r) {
Yaowu Xu's avatar
Yaowu Xu committed
261
    vpx_memcpy(dst, src, w);
262
263
    src += src_stride;
    dst += dst_stride;
264
265
266
  }
}

267
268
269
270
271
void vp9_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride,
                        uint8_t *dst, ptrdiff_t dst_stride,
                        const int16_t *filter_x, int filter_x_stride,
                        const int16_t *filter_y, int filter_y_stride,
                        int w, int h) {
272
273
  int x, y;

274
275
276
  (void)filter_x;  (void)filter_x_stride;
  (void)filter_y;  (void)filter_y_stride;

277
  for (y = 0; y < h; ++y) {
278
279
280
    for (x = 0; x < w; ++x)
      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);

281
282
283
284
    src += src_stride;
    dst += dst_stride;
  }
}