gfx/2d/ssse3-scaler.c

   1 /*
   2  * Copyright © 2013 Soren Sandmann Pedersen
   3  * Copyright © 2013 Red Hat, Inc.
   4  * Copyright © 2016 Mozilla Foundation
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  22  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  23  * DEALINGS IN THE SOFTWARE.
  24  *
  25  * Author: Soren Sandmann (soren.sandmann@gmail.com)
  26  *         Jeff Muizelaar (jmuizelaar@mozilla.com)
  27  */
  28
  29 /* This has been adapted from the ssse3 code from pixman. It's currently
  30  * a mess as I want to try it out in practice before finalizing the details.
  31  */
  32
  33 #include <stdlib.h>
  34 #include <xmmintrin.h>
  35 #include <emmintrin.h>
  36 #include <tmmintrin.h>
  37 #include <stdint.h>
  38 #include <assert.h>
  39 #include "ssse3-scaler.h"
  40
  41 typedef int32_t pixman_fixed_16_16_t;
  42 typedef pixman_fixed_16_16_t pixman_fixed_t;
  43 #define pixman_fixed_1 (pixman_int_to_fixed(1))
  44 #define pixman_fixed_to_int(f) ((int)((f) >> 16))
  45 #define pixman_int_to_fixed(i) ((pixman_fixed_t)((i) << 16))
  46 #define pixman_double_to_fixed(d) ((pixman_fixed_t)((d) * 65536.0))
  47 #define PIXMAN_FIXED_INT_MAX 32767
  48 #define PIXMAN_FIXED_INT_MIN -32768
  49 typedef struct pixman_vector pixman_vector_t;
  50
  51 typedef int pixman_bool_t;
  52 typedef int64_t pixman_fixed_32_32_t;
  53 typedef pixman_fixed_32_32_t pixman_fixed_48_16_t;
  54 typedef struct {
  55   pixman_fixed_48_16_t v[3];
  56 } pixman_vector_48_16_t;
  57
  58 struct pixman_vector {
  59   pixman_fixed_t vector[3];
  60 };
  61 typedef struct pixman_transform pixman_transform_t;
  62
  63 struct pixman_transform {
  64   pixman_fixed_t matrix[3][3];
  65 };
  66
  67 #ifdef _MSC_VER
  68 #  define force_inline __forceinline
  69 #else
  70 #  define force_inline __inline__ __attribute__((always_inline))
  71 #endif
  72
  73 #define BILINEAR_INTERPOLATION_BITS 6
  74
  75 static force_inline int pixman_fixed_to_bilinear_weight(pixman_fixed_t x) {
  76   return (x >> (16 - BILINEAR_INTERPOLATION_BITS)) &
  77          ((1 << BILINEAR_INTERPOLATION_BITS) - 1);
  78 }
  79
  80 static void pixman_transform_point_31_16_3d(const pixman_transform_t* t,
  81                                             const pixman_vector_48_16_t* v,
  82                                             pixman_vector_48_16_t* result) {
  83   int i;
  84   int64_t tmp[3][2];
  85
  86   /* input vector values must have no more than 31 bits (including sign)
  87    * in the integer part */
  88   assert(v->v[0] < ((pixman_fixed_48_16_t)1 << (30 + 16)));
  89   assert(v->v[0] >= -((pixman_fixed_48_16_t)1 << (30 + 16)));
  90   assert(v->v[1] < ((pixman_fixed_48_16_t)1 << (30 + 16)));
  91   assert(v->v[1] >= -((pixman_fixed_48_16_t)1 << (30 + 16)));
  92   assert(v->v[2] < ((pixman_fixed_48_16_t)1 << (30 + 16)));
  93   assert(v->v[2] >= -((pixman_fixed_48_16_t)1 << (30 + 16)));
  94
  95   for (i = 0; i < 3; i++) {
  96     tmp[i][0] = (int64_t)t->matrix[i][0] * (v->v[0] >> 16);
  97     tmp[i][1] = (int64_t)t->matrix[i][0] * (v->v[0] & 0xFFFF);
  98     tmp[i][0] += (int64_t)t->matrix[i][1] * (v->v[1] >> 16);
  99     tmp[i][1] += (int64_t)t->matrix[i][1] * (v->v[1] & 0xFFFF);
 100     tmp[i][0] += (int64_t)t->matrix[i][2] * (v->v[2] >> 16);
 101     tmp[i][1] += (int64_t)t->matrix[i][2] * (v->v[2] & 0xFFFF);
 102   }
 103
 104   result->v[0] = tmp[0][0] + ((tmp[0][1] + 0x8000) >> 16);
 105   result->v[1] = tmp[1][0] + ((tmp[1][1] + 0x8000) >> 16);
 106   result->v[2] = tmp[2][0] + ((tmp[2][1] + 0x8000) >> 16);
 107 }
 108
 109 static pixman_bool_t pixman_transform_point_3d(
 110     const struct pixman_transform* transform, struct pixman_vector* vector) {
 111   pixman_vector_48_16_t tmp;
 112   tmp.v[0] = vector->vector[0];
 113   tmp.v[1] = vector->vector[1];
 114   tmp.v[2] = vector->vector[2];
 115
 116   pixman_transform_point_31_16_3d(transform, &tmp, &tmp);
 117
 118   vector->vector[0] = tmp.v[0];
 119   vector->vector[1] = tmp.v[1];
 120   vector->vector[2] = tmp.v[2];
 121
 122   return vector->vector[0] == tmp.v[0] && vector->vector[1] == tmp.v[1] &&
 123          vector->vector[2] == tmp.v[2];
 124 }
 125
 126 struct bits_image_t {
 127   uint32_t* bits;
 128   int rowstride;
 129   pixman_transform_t* transform;
 130 };
 131
 132 typedef struct bits_image_t bits_image_t;
 133 typedef struct {
 134   int unused;
 135 } pixman_iter_info_t;
 136
 137 typedef struct pixman_iter_t pixman_iter_t;
 138 typedef void (*pixman_iter_fini_t)(pixman_iter_t* iter);
 139
 140 struct pixman_iter_t {
 141   int x, y;
 142   pixman_iter_fini_t fini;
 143   bits_image_t* image;
 144   uint32_t* buffer;
 145   int width;
 146   int height;
 147   void* data;
 148 };
 149
 150 typedef struct {
 151   int y;
 152   uint64_t* buffer;
 153 } line_t;
 154
 155 typedef struct {
 156   line_t lines[2];
 157   pixman_fixed_t y;
 158   pixman_fixed_t x;
 159   uint64_t data[1];
 160 } bilinear_info_t;
 161
 162 static void ssse3_fetch_horizontal(bits_image_t* image, line_t* line, int y,
 163                                    pixman_fixed_t x, pixman_fixed_t ux, int n) {
 164   uint32_t* bits = image->bits + y * image->rowstride;
 165   __m128i vx = _mm_set_epi16(-(x + 1), x, -(x + 1), x, -(x + ux + 1), x + ux,
 166                              -(x + ux + 1), x + ux);
 167   __m128i vux = _mm_set_epi16(-2 * ux, 2 * ux, -2 * ux, 2 * ux, -2 * ux, 2 * ux,
 168                               -2 * ux, 2 * ux);
 169   __m128i vaddc = _mm_set_epi16(1, 0, 1, 0, 1, 0, 1, 0);
 170   __m128i* b = (__m128i*)line->buffer;
 171   __m128i vrl0, vrl1;
 172
 173   while ((n -= 2) >= 0) {
 174     __m128i vw, vr, s;
 175 #ifdef HACKY_PADDING
 176     if (pixman_fixed_to_int(x + ux) >= image->rowstride) {
 177       vrl1 = _mm_setzero_si128();
 178       printf("overread 2loop\n");
 179     } else {
 180       if (pixman_fixed_to_int(x + ux) < 0) printf("underflow\n");
 181       vrl1 = _mm_loadl_epi64(
 182           (__m128i*)(bits + (pixman_fixed_to_int(x + ux) < 0
 183                                  ? 0
 184                                  : pixman_fixed_to_int(x + ux))));
 185     }
 186 #else
 187     vrl1 = _mm_loadl_epi64((__m128i*)(bits + pixman_fixed_to_int(x + ux)));
 188 #endif
 189     /* vrl1: R1, L1 */
 190
 191   final_pixel:
 192 #ifdef HACKY_PADDING
 193     vrl0 = _mm_loadl_epi64(
 194         (__m128i*)(bits +
 195                    (pixman_fixed_to_int(x) < 0 ? 0 : pixman_fixed_to_int(x))));
 196 #else
 197     vrl0 = _mm_loadl_epi64((__m128i*)(bits + pixman_fixed_to_int(x)));
 198 #endif
 199     /* vrl0: R0, L0 */
 200
 201     /* The weights are based on vx which is a vector of
 202      *
 203      *    - (x + 1), x, - (x + 1), x,
 204      *          - (x + ux + 1), x + ux, - (x + ux + 1), x + ux
 205      *
 206      * so the 16 bit weights end up like this:
 207      *
 208      *    iw0, w0, iw0, w0, iw1, w1, iw1, w1
 209      *
 210      * and after shifting and packing, we get these bytes:
 211      *
 212      *    iw0, w0, iw0, w0, iw1, w1, iw1, w1,
 213      *        iw0, w0, iw0, w0, iw1, w1, iw1, w1,
 214      *
 215      * which means the first and the second input pixel
 216      * have to be interleaved like this:
 217      *
 218      *    la0, ra0, lr0, rr0, la1, ra1, lr1, rr1,
 219      *        lg0, rg0, lb0, rb0, lg1, rg1, lb1, rb1
 220      *
 221      * before maddubsw can be used.
 222      */
 223
 224     vw = _mm_add_epi16(vaddc,
 225                        _mm_srli_epi16(vx, 16 - BILINEAR_INTERPOLATION_BITS));
 226     /* vw: iw0, w0, iw0, w0, iw1, w1, iw1, w1
 227      */
 228
 229     vw = _mm_packus_epi16(vw, vw);
 230     /* vw: iw0, w0, iw0, w0, iw1, w1, iw1, w1,
 231      *         iw0, w0, iw0, w0, iw1, w1, iw1, w1
 232      */
 233     vx = _mm_add_epi16(vx, vux);
 234
 235     x += 2 * ux;
 236
 237     vr = _mm_unpacklo_epi16(vrl1, vrl0);
 238     /* vr: rar0, rar1, rgb0, rgb1, lar0, lar1, lgb0, lgb1 */
 239
 240     s = _mm_shuffle_epi32(vr, _MM_SHUFFLE(1, 0, 3, 2));
 241     /* s:  lar0, lar1, lgb0, lgb1, rar0, rar1, rgb0, rgb1 */
 242
 243     vr = _mm_unpackhi_epi8(vr, s);
 244     /* vr: la0, ra0, lr0, rr0, la1, ra1, lr1, rr1,
 245      *         lg0, rg0, lb0, rb0, lg1, rg1, lb1, rb1
 246      */
 247
 248     vr = _mm_maddubs_epi16(vr, vw);
 249
 250     /* When the weight is 0, the inverse weight is
 251      * 128 which can't be represented in a signed byte.
 252      * As a result maddubsw computes the following:
 253      *
 254      *     r = l * -128 + r * 0
 255      *
 256      * rather than the desired
 257      *
 258      *     r = l * 128 + r * 0
 259      *
 260      * We fix this by taking the absolute value of the
 261      * result.
 262      */
 263     // we can drop this if we use lower precision
 264
 265     vr = _mm_shuffle_epi32(vr, _MM_SHUFFLE(2, 0, 3, 1));
 266     /* vr: A0, R0, A1, R1, G0, B0, G1, B1 */
 267     _mm_store_si128(b++, vr);
 268   }
 269
 270   if (n == -1) {
 271     vrl1 = _mm_setzero_si128();
 272     goto final_pixel;
 273   }
 274
 275   line->y = y;
 276 }
 277
 278 // scale a line of destination pixels
 279 static uint32_t* ssse3_fetch_bilinear_cover(pixman_iter_t* iter,
 280                                             const uint32_t* mask) {
 281   pixman_fixed_t fx, ux;
 282   bilinear_info_t* info = iter->data;
 283   line_t *line0, *line1;
 284   int y0, y1;
 285   int32_t dist_y;
 286   __m128i vw, uvw;
 287   int i;
 288
 289   fx = info->x;
 290   ux = iter->image->transform->matrix[0][0];
 291
 292   y0 = pixman_fixed_to_int(info->y);
 293   if (y0 < 0) *(volatile char*)0 = 9;
 294   y1 = y0 + 1;
 295
 296   // clamping in y direction
 297   if (y1 >= iter->height) {
 298     y1 = iter->height - 1;
 299   }
 300
 301   line0 = &info->lines[y0 & 0x01];
 302   line1 = &info->lines[y1 & 0x01];
 303
 304   if (line0->y != y0) {
 305     ssse3_fetch_horizontal(iter->image, line0, y0, fx, ux, iter->width);
 306   }
 307
 308   if (line1->y != y1) {
 309     ssse3_fetch_horizontal(iter->image, line1, y1, fx, ux, iter->width);
 310   }
 311
 312 #ifdef PIXMAN_STYLE_INTERPOLATION
 313   dist_y = pixman_fixed_to_bilinear_weight(info->y);
 314   dist_y <<= (16 - BILINEAR_INTERPOLATION_BITS);
 315
 316   vw = _mm_set_epi16(dist_y, dist_y, dist_y, dist_y, dist_y, dist_y, dist_y,
 317                      dist_y);
 318
 319 #else
 320   // setup the weights for the top (vw) and bottom (uvw) lines
 321   dist_y = pixman_fixed_to_bilinear_weight(info->y);
 322   // we use 15 instead of 16 because we need an extra bit to handle when the
 323   // weights are 0 and 1
 324   dist_y <<= (15 - BILINEAR_INTERPOLATION_BITS);
 325
 326   vw = _mm_set_epi16(dist_y, dist_y, dist_y, dist_y, dist_y, dist_y, dist_y,
 327                      dist_y);
 328
 329   dist_y = (1 << BILINEAR_INTERPOLATION_BITS) -
 330            pixman_fixed_to_bilinear_weight(info->y);
 331   dist_y <<= (15 - BILINEAR_INTERPOLATION_BITS);
 332   uvw = _mm_set_epi16(dist_y, dist_y, dist_y, dist_y, dist_y, dist_y, dist_y,
 333                       dist_y);
 334 #endif
 335
 336   for (i = 0; i + 3 < iter->width; i += 4) {
 337     __m128i top0 = _mm_load_si128((__m128i*)(line0->buffer + i));
 338     __m128i bot0 = _mm_load_si128((__m128i*)(line1->buffer + i));
 339     __m128i top1 = _mm_load_si128((__m128i*)(line0->buffer + i + 2));
 340     __m128i bot1 = _mm_load_si128((__m128i*)(line1->buffer + i + 2));
 341 #ifdef PIXMAN_STYLE_INTERPOLATION
 342     __m128i r0, r1, tmp, p;
 343
 344     r0 = _mm_mulhi_epu16(_mm_sub_epi16(bot0, top0), vw);
 345     tmp = _mm_cmplt_epi16(bot0, top0);
 346     tmp = _mm_and_si128(tmp, vw);
 347     r0 = _mm_sub_epi16(r0, tmp);
 348     r0 = _mm_add_epi16(r0, top0);
 349     r0 = _mm_srli_epi16(r0, BILINEAR_INTERPOLATION_BITS);
 350     /* r0:  A0 R0 A1 R1 G0 B0 G1 B1 */
 351     // r0 = _mm_shuffle_epi32 (r0, _MM_SHUFFLE (2, 0, 3, 1));
 352     /* r0:  A1 R1 G1 B1 A0 R0 G0 B0 */
 353
 354     // tmp = bot1 < top1 ? vw : 0;
 355     // r1 = (bot1 - top1)*vw + top1 - tmp
 356     // r1 = bot1*vw - vw*top1 + top1 - tmp
 357     // r1 = bot1*vw + top1 - vw*top1 - tmp
 358     // r1 = bot1*vw + top1*(1 - vw) - tmp
 359     r1 = _mm_mulhi_epu16(_mm_sub_epi16(bot1, top1), vw);
 360     tmp = _mm_cmplt_epi16(bot1, top1);
 361     tmp = _mm_and_si128(tmp, vw);
 362     r1 = _mm_sub_epi16(r1, tmp);
 363     r1 = _mm_add_epi16(r1, top1);
 364     r1 = _mm_srli_epi16(r1, BILINEAR_INTERPOLATION_BITS);
 365     // r1 = _mm_shuffle_epi32 (r1, _MM_SHUFFLE (2, 0, 3, 1));
 366     /* r1: A3 R3 G3 B3 A2 R2 G2 B2 */
 367 #else
 368     __m128i r0, r1, p;
 369     top0 = _mm_mulhi_epu16(top0, uvw);
 370     bot0 = _mm_mulhi_epu16(bot0, vw);
 371     r0 = _mm_add_epi16(top0, bot0);
 372     r0 = _mm_srli_epi16(r0, BILINEAR_INTERPOLATION_BITS - 1);
 373
 374     top1 = _mm_mulhi_epu16(top1, uvw);
 375     bot1 = _mm_mulhi_epu16(bot1, vw);
 376     r1 = _mm_add_epi16(top1, bot1);
 377     r1 = _mm_srli_epi16(r1, BILINEAR_INTERPOLATION_BITS - 1);
 378 #endif
 379
 380     p = _mm_packus_epi16(r0, r1);
 381     _mm_storeu_si128((__m128i*)(iter->buffer + i), p);
 382   }
 383
 384   while (i < iter->width) {
 385     __m128i top0 = _mm_load_si128((__m128i*)(line0->buffer + i));
 386     __m128i bot0 = _mm_load_si128((__m128i*)(line1->buffer + i));
 387
 388 #ifdef PIXMAN_STYLE_INTERPOLATION
 389     __m128i r0, tmp, p;
 390     r0 = _mm_mulhi_epu16(_mm_sub_epi16(bot0, top0), vw);
 391     tmp = _mm_cmplt_epi16(bot0, top0);
 392     tmp = _mm_and_si128(tmp, vw);
 393     r0 = _mm_sub_epi16(r0, tmp);
 394     r0 = _mm_add_epi16(r0, top0);
 395     r0 = _mm_srli_epi16(r0, BILINEAR_INTERPOLATION_BITS);
 396     /* r0:  A0 R0 A1 R1 G0 B0 G1 B1 */
 397     r0 = _mm_shuffle_epi32(r0, _MM_SHUFFLE(2, 0, 3, 1));
 398     /* r0:  A1 R1 G1 B1 A0 R0 G0 B0 */
 399 #else
 400     __m128i r0, p;
 401     top0 = _mm_mulhi_epu16(top0, uvw);
 402     bot0 = _mm_mulhi_epu16(bot0, vw);
 403     r0 = _mm_add_epi16(top0, bot0);
 404     r0 = _mm_srli_epi16(r0, BILINEAR_INTERPOLATION_BITS - 1);
 405 #endif
 406
 407     p = _mm_packus_epi16(r0, r0);
 408
 409     if (iter->width - i == 1) {
 410       *(uint32_t*)(iter->buffer + i) = _mm_cvtsi128_si32(p);
 411       i++;
 412     } else {
 413       _mm_storel_epi64((__m128i*)(iter->buffer + i), p);
 414       i += 2;
 415     }
 416   }
 417
 418   info->y += iter->image->transform->matrix[1][1];
 419
 420   return iter->buffer;
 421 }
 422
 423 static void ssse3_bilinear_cover_iter_fini(pixman_iter_t* iter) {
 424   free(iter->data);
 425 }
 426
 427 static void ssse3_bilinear_cover_iter_init(pixman_iter_t* iter) {
 428   int width = iter->width;
 429   bilinear_info_t* info;
 430   pixman_vector_t v;
 431
 432   if (iter->x > PIXMAN_FIXED_INT_MAX || iter->x < PIXMAN_FIXED_INT_MIN ||
 433       iter->y > PIXMAN_FIXED_INT_MAX || iter->y < PIXMAN_FIXED_INT_MIN)
 434     goto fail;
 435
 436   /* Reference point is the center of the pixel */
 437   v.vector[0] = pixman_int_to_fixed(iter->x) + pixman_fixed_1 / 2;
 438   v.vector[1] = pixman_int_to_fixed(iter->y) + pixman_fixed_1 / 2;
 439   v.vector[2] = pixman_fixed_1;
 440
 441   if (!pixman_transform_point_3d(iter->image->transform, &v)) goto fail;
 442
 443   info = malloc(sizeof(*info) + (2 * width - 1) * sizeof(uint64_t) + 64);
 444   if (!info) goto fail;
 445
 446   info->x = v.vector[0] - pixman_fixed_1 / 2;
 447   info->y = v.vector[1] - pixman_fixed_1 / 2;
 448
 449 #define ALIGN(addr) ((void*)((((uintptr_t)(addr)) + 15) & (~15)))
 450
 451   /* It is safe to set the y coordinates to -1 initially
 452    * because COVER_CLIP_BILINEAR ensures that we will only
 453    * be asked to fetch lines in the [0, height) interval
 454    */
 455   info->lines[0].y = -1;
 456   info->lines[0].buffer = ALIGN(&(info->data[0]));
 457   info->lines[1].y = -1;
 458   info->lines[1].buffer = ALIGN(info->lines[0].buffer + width);
 459
 460   iter->fini = ssse3_bilinear_cover_iter_fini;
 461
 462   iter->data = info;
 463   return;
 464
 465 fail:
 466   /* Something went wrong, either a bad matrix or OOM; in such cases,
 467    * we don't guarantee any particular rendering.
 468    */
 469   iter->fini = NULL;
 470 }
 471
 472 /* scale the src from src_width/height to dest_width/height drawn
 473  * into the rectangle x,y width,height
 474  * src_stride and dst_stride are 4 byte units */
 475 bool ssse3_scale_data(uint32_t* src, int src_width, int src_height,
 476                       int src_stride, uint32_t* dest, int dest_width,
 477                       int dest_height, int dest_stride, int x, int y, int width,
 478                       int height) {
 479   // XXX: assert(src_width > 1)
 480   pixman_transform_t transform = {
 481       {{pixman_fixed_1, 0, 0}, {0, pixman_fixed_1, 0}, {0, 0, pixman_fixed_1}}};
 482   double width_scale = ((double)src_width) / dest_width;
 483   double height_scale = ((double)src_height) / dest_height;
 484 #define AVOID_PADDING
 485 #ifdef AVOID_PADDING
 486   // scale up by enough that we don't read outside of the bounds of the source
 487   // surface currently this is required to avoid reading out of bounds.
 488   if (width_scale < 1) {
 489     width_scale = (double)(src_width - 1) / dest_width;
 490     transform.matrix[0][2] = pixman_fixed_1 / 2;
 491   }
 492   if (height_scale < 1) {
 493     height_scale = (double)(src_height - 1) / dest_height;
 494     transform.matrix[1][2] = pixman_fixed_1 / 2;
 495   }
 496 #endif
 497   transform.matrix[0][0] = pixman_double_to_fixed(width_scale);
 498   transform.matrix[1][1] = pixman_double_to_fixed(height_scale);
 499   transform.matrix[2][2] = pixman_fixed_1;
 500
 501   bits_image_t image;
 502   image.bits = src;
 503   image.transform = &transform;
 504   image.rowstride = src_stride;
 505
 506   pixman_iter_t iter;
 507   iter.image = &image;
 508   iter.x = x;
 509   iter.y = y;
 510   iter.width = width;
 511   iter.height = src_height;
 512   iter.buffer = dest;
 513   iter.data = NULL;
 514
 515   ssse3_bilinear_cover_iter_init(&iter);
 516
 517   if (!iter.fini) return false;
 518
 519   if (iter.data) {
 520     for (int iy = 0; iy < height; iy++) {
 521       ssse3_fetch_bilinear_cover(&iter, NULL);
 522       iter.buffer += dest_stride;
 523     }
 524     ssse3_bilinear_cover_iter_fini(&iter);
 525   }
 526   return true;
 527 }