aom_convolve.c 23.9 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
Yaowu Xu's avatar
Yaowu Xu committed
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
John Koleszar's avatar
John Koleszar committed
3
 *
Yaowu Xu's avatar
Yaowu Xu committed
4 5 6 7 8 9
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
John Koleszar's avatar
John Koleszar committed
10
 */
Christian Duvivier's avatar
Christian Duvivier committed
11

John Koleszar's avatar
John Koleszar committed
12
#include <assert.h>
Zoe Liu's avatar
Zoe Liu committed
13
#include <string.h>
John Koleszar's avatar
John Koleszar committed
14

Yaowu Xu's avatar
Yaowu Xu committed
15 16 17 18 19 20
#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
#include "aom/aom_integer.h"
#include "aom_dsp/aom_convolve.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/aom_filter.h"
21
#include "aom_ports/mem.h"
John Koleszar's avatar
John Koleszar committed
22

Dmitry Kovalev's avatar
Dmitry Kovalev committed
23 24
static void convolve_horiz(const uint8_t *src, ptrdiff_t src_stride,
                           uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
25 26
                           const InterpKernel *x_filters, int x0_q4,
                           int x_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
27 28
  int x, y;
  src -= SUBPEL_TAPS / 2 - 1;
John Koleszar's avatar
John Koleszar committed
29
  for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
30
    int x_q4 = x0_q4;
John Koleszar's avatar
John Koleszar committed
31
    for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
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;
clang-format's avatar
clang-format committed
35
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
36
      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
John Koleszar's avatar
John Koleszar committed
37 38 39 40 41 42 43
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

Dmitry Kovalev's avatar
Dmitry Kovalev committed
44 45
static void convolve_avg_horiz(const uint8_t *src, ptrdiff_t src_stride,
                               uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
46 47
                               const InterpKernel *x_filters, int x0_q4,
                               int x_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
48 49
  int x, y;
  src -= SUBPEL_TAPS / 2 - 1;
John Koleszar's avatar
John Koleszar committed
50
  for (y = 0; y < h; ++y) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
51
    int x_q4 = x0_q4;
John Koleszar's avatar
John Koleszar committed
52
    for (x = 0; x < w; ++x) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
53 54 55
      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;
clang-format's avatar
clang-format committed
56 57 58
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
      dst[x] = ROUND_POWER_OF_TWO(
          dst[x] + 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,
clang-format's avatar
clang-format committed
68 69
                          const InterpKernel *y_filters, int y0_q4,
                          int y_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
70 71
  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,
clang-format's avatar
clang-format committed
91 92
                              const InterpKernel *y_filters, int y0_q4,
                              int y_step_q4, int w, int h) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
93 94
  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];
clang-format's avatar
clang-format committed
104 105 106 107
      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
          dst[y * dst_stride] +
              clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS)),
          1);
John Koleszar's avatar
John Koleszar committed
108 109 110 111 112 113 114
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

clang-format's avatar
clang-format committed
115 116
static void convolve(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
                     ptrdiff_t dst_stride, const InterpKernel *const x_filters,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
117
                     int x0_q4, int x_step_q4,
clang-format's avatar
clang-format committed
118 119
                     const InterpKernel *const y_filters, int y0_q4,
                     int y_step_q4, int w, int h) {
120 121 122 123 124 125 126 127 128 129 130 131
  // 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.
132
  uint8_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
133
  int intermediate_height =
clang-format's avatar
clang-format committed
134
      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
135

136 137
  assert(w <= MAX_SB_SIZE);
  assert(h <= MAX_SB_SIZE);
138

139 140
  assert(y_step_q4 <= 32);
  assert(x_step_q4 <= 32);
141

clang-format's avatar
clang-format committed
142 143 144 145 146
  convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride, temp,
                 MAX_SB_SIZE, x_filters, x0_q4, x_step_q4, w,
                 intermediate_height);
  convolve_vert(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), MAX_SB_SIZE, dst,
                dst_stride, y_filters, y0_q4, y_step_q4, w, h);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
147 148
}

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
}

Yaowu Xu's avatar
Yaowu Xu committed
159
void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
160
                           uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
161
                           const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
162 163
                           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;

clang-format's avatar
clang-format 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
}

Yaowu Xu's avatar
Yaowu Xu committed
174
void aom_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
175
                               uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
176
                               const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
177 178
                               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;

clang-format's avatar
clang-format 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
}

Yaowu Xu's avatar
Yaowu Xu committed
189
void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
190
                          uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
191
                          const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
192 193
                          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;

clang-format's avatar
clang-format 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
}

Yaowu Xu's avatar
Yaowu Xu committed
204
void aom_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
205
                              uint8_t *dst, ptrdiff_t dst_stride,
John Koleszar's avatar
John Koleszar committed
206
                              const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
207 208
                              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;

clang-format's avatar
clang-format 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
}

Yaowu Xu's avatar
Yaowu Xu committed
219
void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
220 221
                     ptrdiff_t dst_stride, const int16_t *filter_x,
                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
John Koleszar's avatar
John Koleszar committed
222
                     int w, int h) {
223
  const InterpKernel *const filters_x = get_filter_base(filter_x);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
224 225
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

226
  const InterpKernel *const filters_y = get_filter_base(filter_y);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
227 228
  const int y0_q4 = get_filter_offset(filter_y, filters_y);

clang-format's avatar
clang-format committed
229
  convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
230
           filters_y, y0_q4, y_step_q4, w, h);
John Koleszar's avatar
John Koleszar committed
231 232
}

Yaowu Xu's avatar
Yaowu Xu committed
233
void aom_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
234 235
                         ptrdiff_t dst_stride, const int16_t *filter_x,
                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
John Koleszar's avatar
John Koleszar committed
236
                         int w, int h) {
Christian Duvivier's avatar
Christian Duvivier committed
237
  /* Fixed size intermediate buffer places limits on parameters. */
238 239 240
  DECLARE_ALIGNED(16, uint8_t, temp[MAX_SB_SIZE * MAX_SB_SIZE]);
  assert(w <= MAX_SB_SIZE);
  assert(h <= MAX_SB_SIZE);
Christian Duvivier's avatar
Christian Duvivier committed
241

Yaowu Xu's avatar
Yaowu Xu committed
242
  aom_convolve8_c(src, src_stride, temp, MAX_SB_SIZE, filter_x, x_step_q4,
clang-format's avatar
clang-format committed
243
                  filter_y, y_step_q4, w, h);
Yaowu Xu's avatar
Yaowu Xu committed
244
  aom_convolve_avg_c(temp, MAX_SB_SIZE, dst, dst_stride, NULL, 0, NULL, 0, w,
clang-format's avatar
clang-format committed
245
                     h);
John Koleszar's avatar
John Koleszar committed
246
}
247

Yaowu Xu's avatar
Yaowu Xu committed
248
void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
249 250 251
                         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) {
252 253
  int r;

clang-format's avatar
clang-format committed
254 255 256 257
  (void)filter_x;
  (void)filter_x_stride;
  (void)filter_y;
  (void)filter_y_stride;
258

259
  for (r = h; r > 0; --r) {
James Zern's avatar
James Zern committed
260
    memcpy(dst, src, w);
261 262
    src += src_stride;
    dst += dst_stride;
263 264 265
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
266
void aom_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
267 268 269
                        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) {
270 271
  int x, y;

clang-format's avatar
clang-format committed
272 273 274 275
  (void)filter_x;
  (void)filter_x_stride;
  (void)filter_y;
  (void)filter_y_stride;
276

277
  for (y = 0; y < h; ++y) {
clang-format's avatar
clang-format committed
278
    for (x = 0; x < w; ++x) dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
279

280 281 282 283
    src += src_stride;
    dst += dst_stride;
  }
}
284

Yaowu Xu's avatar
Yaowu Xu committed
285
void aom_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
286 287
                        ptrdiff_t dst_stride, const int16_t *filter_x,
                        int x_step_q4, const int16_t *filter_y, int y_step_q4,
288
                        int w, int h) {
Yaowu Xu's avatar
Yaowu Xu committed
289
  aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
290 291 292
                        filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
293
void aom_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
294 295
                       ptrdiff_t dst_stride, const int16_t *filter_x,
                       int x_step_q4, const int16_t *filter_y, int y_step_q4,
296
                       int w, int h) {
Yaowu Xu's avatar
Yaowu Xu committed
297
  aom_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
298 299 300
                       filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
301
void aom_scaled_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
302 303
                     ptrdiff_t dst_stride, const int16_t *filter_x,
                     int x_step_q4, const int16_t *filter_y, int y_step_q4,
304
                     int w, int h) {
Yaowu Xu's avatar
Yaowu Xu committed
305
  aom_convolve8_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
306 307 308
                  filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
309
void aom_scaled_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
310 311
                            uint8_t *dst, ptrdiff_t dst_stride,
                            const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
312 313
                            const int16_t *filter_y, int y_step_q4, int w,
                            int h) {
Yaowu Xu's avatar
Yaowu Xu committed
314
  aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
315 316 317
                            x_step_q4, filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
318
void aom_scaled_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
319 320
                           uint8_t *dst, ptrdiff_t dst_stride,
                           const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
321 322
                           const int16_t *filter_y, int y_step_q4, int w,
                           int h) {
Yaowu Xu's avatar
Yaowu Xu committed
323
  aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
324 325 326
                           x_step_q4, filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
327
void aom_scaled_avg_2d_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
clang-format's avatar
clang-format committed
328 329 330
                         ptrdiff_t dst_stride, const int16_t *filter_x,
                         int x_step_q4, const int16_t *filter_y, int y_step_q4,
                         int w, int h) {
Yaowu Xu's avatar
Yaowu Xu committed
331
  aom_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
332 333 334
                      filter_y, y_step_q4, w, h);
}

Yaowu Xu's avatar
Yaowu Xu committed
335
#if CONFIG_AOM_HIGHBITDEPTH
336 337
static void highbd_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
                                  uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
338 339
                                  const InterpKernel *x_filters, int x0_q4,
                                  int x_step_q4, int w, int h, int bd) {
340 341 342 343 344 345 346 347 348 349
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= SUBPEL_TAPS / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_q4 = x0_q4;
    for (x = 0; x < w; ++x) {
      const uint16_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;
clang-format's avatar
clang-format committed
350
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
351
      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
352 353 354 355 356 357 358
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

359 360
static void highbd_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
                                      uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
361 362
                                      const InterpKernel *x_filters, int x0_q4,
                                      int x_step_q4, int w, int h, int bd) {
363 364 365 366 367 368 369 370 371 372
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= SUBPEL_TAPS / 2 - 1;
  for (y = 0; y < h; ++y) {
    int x_q4 = x0_q4;
    for (x = 0; x < w; ++x) {
      const uint16_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;
clang-format's avatar
clang-format committed
373 374 375 376
      for (k = 0; k < SUBPEL_TAPS; ++k) sum += src_x[k] * x_filter[k];
      dst[x] = ROUND_POWER_OF_TWO(
          dst[x] + clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
          1);
377 378 379 380 381 382 383
      x_q4 += x_step_q4;
    }
    src += src_stride;
    dst += dst_stride;
  }
}

384 385
static void highbd_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
                                 uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
386 387
                                 const InterpKernel *y_filters, int y0_q4,
                                 int y_step_q4, int w, int h, int bd) {
388 389 390 391 392 393 394 395 396 397 398 399
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
  for (x = 0; x < w; ++x) {
    int y_q4 = y0_q4;
    for (y = 0; y < h; ++y) {
      const uint16_t *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];
clang-format's avatar
clang-format committed
400 401
      dst[y * dst_stride] =
          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
402 403 404 405 406 407 408
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

409 410
static void highbd_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
                                     uint8_t *dst8, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
411 412
                                     const InterpKernel *y_filters, int y0_q4,
                                     int y_step_q4, int w, int h, int bd) {
413 414 415 416 417 418 419 420 421 422 423 424
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
  for (x = 0; x < w; ++x) {
    int y_q4 = y0_q4;
    for (y = 0; y < h; ++y) {
      const uint16_t *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];
clang-format's avatar
clang-format committed
425 426 427 428
      dst[y * dst_stride] = ROUND_POWER_OF_TWO(
          dst[y * dst_stride] +
              clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
          1);
429 430 431 432 433 434 435
      y_q4 += y_step_q4;
    }
    ++src;
    ++dst;
  }
}

436 437
static void highbd_convolve(const uint8_t *src, ptrdiff_t src_stride,
                            uint8_t *dst, ptrdiff_t dst_stride,
clang-format's avatar
clang-format committed
438 439 440
                            const InterpKernel *const x_filters, int x0_q4,
                            int x_step_q4, const InterpKernel *const y_filters,
                            int y0_q4, int y_step_q4, int w, int h, int bd) {
441 442 443 444 445 446 447 448 449 450 451 452
  // 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.
453
  uint16_t temp[MAX_EXT_SIZE * MAX_SB_SIZE];
454
  int intermediate_height =
clang-format's avatar
clang-format committed
455
      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
456

457 458
  assert(w <= MAX_SB_SIZE);
  assert(h <= MAX_SB_SIZE);
459 460 461
  assert(y_step_q4 <= 32);
  assert(x_step_q4 <= 32);

462
  highbd_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1), src_stride,
clang-format's avatar
clang-format committed
463 464
                        CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE, x_filters, x0_q4,
                        x_step_q4, w, intermediate_height, bd);
465
  highbd_convolve_vert(
clang-format's avatar
clang-format committed
466 467
      CONVERT_TO_BYTEPTR(temp) + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
      MAX_SB_SIZE, dst, dst_stride, y_filters, y0_q4, y_step_q4, w, h, bd);
468 469
}

Yaowu Xu's avatar
Yaowu Xu committed
470
void aom_highbd_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
471 472
                                  uint8_t *dst, ptrdiff_t dst_stride,
                                  const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
473 474
                                  const int16_t *filter_y, int y_step_q4, int w,
                                  int h, int bd) {
475 476 477 478 479
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
480 481
  highbd_convolve_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                        x_step_q4, w, h, bd);
482 483
}

Yaowu Xu's avatar
Yaowu Xu committed
484
void aom_highbd_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
485 486 487 488
                                      uint8_t *dst, ptrdiff_t dst_stride,
                                      const int16_t *filter_x, int x_step_q4,
                                      const int16_t *filter_y, int y_step_q4,
                                      int w, int h, int bd) {
489 490 491 492 493
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);
  (void)filter_y;
  (void)y_step_q4;

clang-format's avatar
clang-format committed
494 495
  highbd_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x, x0_q4,
                            x_step_q4, w, h, bd);
496 497
}

Yaowu Xu's avatar
Yaowu Xu committed
498
void aom_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
499 500
                                 uint8_t *dst, ptrdiff_t dst_stride,
                                 const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
501 502
                                 const int16_t *filter_y, int y_step_q4, int w,
                                 int h, int bd) {
503 504 505 506 507
  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
508 509
  highbd_convolve_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
                       y_step_q4, w, h, bd);
510 511
}

Yaowu Xu's avatar
Yaowu Xu committed
512
void aom_highbd_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
513 514 515 516
                                     uint8_t *dst, ptrdiff_t dst_stride,
                                     const int16_t *filter_x, int x_step_q4,
                                     const int16_t *filter_y, int y_step_q4,
                                     int w, int h, int bd) {
517 518 519 520 521
  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);
  (void)filter_x;
  (void)x_step_q4;

clang-format's avatar
clang-format committed
522 523
  highbd_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y, y0_q4,
                           y_step_q4, w, h, bd);
524 525
}

Yaowu Xu's avatar
Yaowu Xu committed
526
void aom_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
527 528
                            uint8_t *dst, ptrdiff_t dst_stride,
                            const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
529 530
                            const int16_t *filter_y, int y_step_q4, int w,
                            int h, int bd) {
531 532 533 534 535 536
  const InterpKernel *const filters_x = get_filter_base(filter_x);
  const int x0_q4 = get_filter_offset(filter_x, filters_x);

  const InterpKernel *const filters_y = get_filter_base(filter_y);
  const int y0_q4 = get_filter_offset(filter_y, filters_y);

clang-format's avatar
clang-format committed
537
  highbd_convolve(src, src_stride, dst, dst_stride, filters_x, x0_q4, x_step_q4,
538
                  filters_y, y0_q4, y_step_q4, w, h, bd);
539 540
}

Yaowu Xu's avatar
Yaowu Xu committed
541
void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
542 543
                                uint8_t *dst, ptrdiff_t dst_stride,
                                const int16_t *filter_x, int x_step_q4,
clang-format's avatar
clang-format committed
544 545
                                const int16_t *filter_y, int y_step_q4, int w,
                                int h, int bd) {
546
  // Fixed size intermediate buffer places limits on parameters.
547 548 549
  DECLARE_ALIGNED(16, uint16_t, temp[MAX_SB_SIZE * MAX_SB_SIZE]);
  assert(w <= MAX_SB_SIZE);
  assert(h <= MAX_SB_SIZE);
550

Yaowu Xu's avatar
Yaowu Xu committed
551
  aom_highbd_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE,
552
                         filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
Yaowu Xu's avatar
Yaowu Xu committed
553
  aom_highbd_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), MAX_SB_SIZE, dst,
clang-format's avatar
clang-format committed
554
                            dst_stride, NULL, 0, NULL, 0, w, h, bd);
555 556
}

Yaowu Xu's avatar
Yaowu Xu committed
557
void aom_highbd_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
558 559 560 561
                                uint8_t *dst8, 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 bd) {
562 563 564 565 566 567 568 569 570 571
  int r;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  (void)filter_x;
  (void)filter_y;
  (void)filter_x_stride;
  (void)filter_y_stride;
  (void)bd;

  for (r = h; r > 0; --r) {
James Zern's avatar
James Zern committed
572
    memcpy(dst, src, w * sizeof(uint16_t));
573 574 575 576 577
    src += src_stride;
    dst += dst_stride;
  }
}

Yaowu Xu's avatar
Yaowu Xu committed
578
void aom_highbd_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
579 580 581 582
                               uint8_t *dst8, 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 bd) {
583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600
  int x, y;
  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  (void)filter_x;
  (void)filter_y;
  (void)filter_x_stride;
  (void)filter_y_stride;
  (void)bd;

  for (y = 0; y < h; ++y) {
    for (x = 0; x < w; ++x) {
      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
    }
    src += src_stride;
    dst += dst_stride;
  }
}
#endif