4 * This file was part of the Independent JPEG Group's software:
5 * Copyright (C) 1991-1996, Thomas G. Lane.
6 * libjpeg-turbo Modifications:
7 * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
8 * Copyright (C) 2014, MIPS Technologies, Inc., California.
9 * Copyright (C) 2015, 2019, D. R. Commander.
10 * For conditions of distribution and use, see the accompanying README.ijg
13 * This file contains downsampling routines.
15 * Downsampling input data is counted in "row groups". A row group
16 * is defined to be max_v_samp_factor pixel rows of each component,
17 * from which the downsampler produces v_samp_factor sample rows.
18 * A single row group is processed in each call to the downsampler module.
20 * The downsampler is responsible for edge-expansion of its output data
21 * to fill an integral number of DCT blocks horizontally. The source buffer
22 * may be modified if it is helpful for this purpose (the source buffer is
23 * allocated wide enough to correspond to the desired output width).
24 * The caller (the prep controller) is responsible for vertical padding.
26 * The downsampler may request "context rows" by setting need_context_rows
27 * during startup. In this case, the input arrays will contain at least
28 * one row group's worth of pixels above and below the passed-in data;
29 * the caller will create dummy rows at image top and bottom by replicating
30 * the first or last real pixel row.
32 * An excellent reference for image resampling is
33 * Digital Image Warping, George Wolberg, 1990.
34 * Pub. by IEEE Computer Society Press, Los Alamitos, CA. ISBN 0-8186-8944-7.
36 * The downsampling algorithm used here is a simple average of the source
37 * pixels covered by the output pixel. The hi-falutin sampling literature
38 * refers to this as a "box filter". In general the characteristics of a box
39 * filter are not very good, but for the specific cases we normally use (1:1
40 * and 2:1 ratios) the box is equivalent to a "triangle filter" which is not
41 * nearly so bad. If you intend to use other sampling ratios, you'd be well
42 * advised to improve this code.
44 * A simple input-smoothing capability is provided. This is mainly intended
45 * for cleaning up color-dithered GIF input files (if you find it inadequate,
46 * we suggest using an external filtering program such as pnmconvol). When
47 * enabled, each input pixel P is replaced by a weighted sum of itself and its
48 * eight neighbors. P's weight is 1-8*SF and each neighbor's weight is SF,
49 * where SF = (smoothing_factor / 1024).
50 * Currently, smoothing is only supported for 2h2v sampling factors.
53 #define JPEG_INTERNALS
59 /* Pointer to routine to downsample a single component */
60 typedef void (*downsample1_ptr
) (j_compress_ptr cinfo
,
61 jpeg_component_info
*compptr
,
62 JSAMPARRAY input_data
,
63 JSAMPARRAY output_data
);
65 /* Private subobject */
68 struct jpeg_downsampler pub
; /* public fields */
70 /* Downsampling method pointers, one per component */
71 downsample1_ptr methods
[MAX_COMPONENTS
];
74 typedef my_downsampler
*my_downsample_ptr
;
78 * Initialize for a downsampling pass.
82 start_pass_downsample(j_compress_ptr cinfo
)
89 * Expand a component horizontally from width input_cols to width output_cols,
90 * by duplicating the rightmost samples.
94 expand_right_edge(JSAMPARRAY image_data
, int num_rows
, JDIMENSION input_cols
,
95 JDIMENSION output_cols
)
97 register JSAMPROW ptr
;
98 register JSAMPLE pixval
;
101 int numcols
= (int)(output_cols
- input_cols
);
104 for (row
= 0; row
< num_rows
; row
++) {
105 ptr
= image_data
[row
] + input_cols
;
107 for (count
= numcols
; count
> 0; count
--)
115 * Do downsampling for a whole row group (all components).
117 * In this version we simply downsample each component independently.
121 sep_downsample(j_compress_ptr cinfo
, JSAMPIMAGE input_buf
,
122 JDIMENSION in_row_index
, JSAMPIMAGE output_buf
,
123 JDIMENSION out_row_group_index
)
125 my_downsample_ptr downsample
= (my_downsample_ptr
)cinfo
->downsample
;
127 jpeg_component_info
*compptr
;
128 JSAMPARRAY in_ptr
, out_ptr
;
130 for (ci
= 0, compptr
= cinfo
->comp_info
; ci
< cinfo
->num_components
;
132 in_ptr
= input_buf
[ci
] + in_row_index
;
133 out_ptr
= output_buf
[ci
] + (out_row_group_index
* compptr
->v_samp_factor
);
134 (*downsample
->methods
[ci
]) (cinfo
, compptr
, in_ptr
, out_ptr
);
140 * Downsample pixel values of a single component.
141 * One row group is processed per call.
142 * This version handles arbitrary integral sampling ratios, without smoothing.
143 * Note that this version is not actually used for customary sampling ratios.
147 int_downsample(j_compress_ptr cinfo
, jpeg_component_info
*compptr
,
148 JSAMPARRAY input_data
, JSAMPARRAY output_data
)
150 int inrow
, outrow
, h_expand
, v_expand
, numpix
, numpix2
, h
, v
;
151 JDIMENSION outcol
, outcol_h
; /* outcol_h == outcol*h_expand */
152 JDIMENSION output_cols
= compptr
->width_in_blocks
* DCTSIZE
;
153 JSAMPROW inptr
, outptr
;
156 h_expand
= cinfo
->max_h_samp_factor
/ compptr
->h_samp_factor
;
157 v_expand
= cinfo
->max_v_samp_factor
/ compptr
->v_samp_factor
;
158 numpix
= h_expand
* v_expand
;
159 numpix2
= numpix
/ 2;
161 /* Expand input data enough to let all the output samples be generated
162 * by the standard loop. Special-casing padded output would be more
165 expand_right_edge(input_data
, cinfo
->max_v_samp_factor
, cinfo
->image_width
,
166 output_cols
* h_expand
);
169 for (outrow
= 0; outrow
< compptr
->v_samp_factor
; outrow
++) {
170 outptr
= output_data
[outrow
];
171 for (outcol
= 0, outcol_h
= 0; outcol
< output_cols
;
172 outcol
++, outcol_h
+= h_expand
) {
174 for (v
= 0; v
< v_expand
; v
++) {
175 inptr
= input_data
[inrow
+ v
] + outcol_h
;
176 for (h
= 0; h
< h_expand
; h
++) {
177 outvalue
+= (JLONG
)(*inptr
++);
180 *outptr
++ = (JSAMPLE
)((outvalue
+ numpix2
) / numpix
);
188 * Downsample pixel values of a single component.
189 * This version handles the special case of a full-size component,
194 fullsize_downsample(j_compress_ptr cinfo
, jpeg_component_info
*compptr
,
195 JSAMPARRAY input_data
, JSAMPARRAY output_data
)
198 jcopy_sample_rows(input_data
, 0, output_data
, 0, cinfo
->max_v_samp_factor
,
201 expand_right_edge(output_data
, cinfo
->max_v_samp_factor
, cinfo
->image_width
,
202 compptr
->width_in_blocks
* DCTSIZE
);
207 * Downsample pixel values of a single component.
208 * This version handles the common case of 2:1 horizontal and 1:1 vertical,
211 * A note about the "bias" calculations: when rounding fractional values to
212 * integer, we do not want to always round 0.5 up to the next integer.
213 * If we did that, we'd introduce a noticeable bias towards larger values.
214 * Instead, this code is arranged so that 0.5 will be rounded up or down at
215 * alternate pixel locations (a simple ordered dither pattern).
219 h2v1_downsample(j_compress_ptr cinfo
, jpeg_component_info
*compptr
,
220 JSAMPARRAY input_data
, JSAMPARRAY output_data
)
224 JDIMENSION output_cols
= compptr
->width_in_blocks
* DCTSIZE
;
225 register JSAMPROW inptr
, outptr
;
228 /* Expand input data enough to let all the output samples be generated
229 * by the standard loop. Special-casing padded output would be more
232 expand_right_edge(input_data
, cinfo
->max_v_samp_factor
, cinfo
->image_width
,
235 for (outrow
= 0; outrow
< compptr
->v_samp_factor
; outrow
++) {
236 outptr
= output_data
[outrow
];
237 inptr
= input_data
[outrow
];
238 bias
= 0; /* bias = 0,1,0,1,... for successive samples */
239 for (outcol
= 0; outcol
< output_cols
; outcol
++) {
240 *outptr
++ = (JSAMPLE
)((inptr
[0] + inptr
[1] + bias
) >> 1);
241 bias
^= 1; /* 0=>1, 1=>0 */
249 * Downsample pixel values of a single component.
250 * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
255 h2v2_downsample(j_compress_ptr cinfo
, jpeg_component_info
*compptr
,
256 JSAMPARRAY input_data
, JSAMPARRAY output_data
)
260 JDIMENSION output_cols
= compptr
->width_in_blocks
* DCTSIZE
;
261 register JSAMPROW inptr0
, inptr1
, outptr
;
264 /* Expand input data enough to let all the output samples be generated
265 * by the standard loop. Special-casing padded output would be more
268 expand_right_edge(input_data
, cinfo
->max_v_samp_factor
, cinfo
->image_width
,
272 for (outrow
= 0; outrow
< compptr
->v_samp_factor
; outrow
++) {
273 outptr
= output_data
[outrow
];
274 inptr0
= input_data
[inrow
];
275 inptr1
= input_data
[inrow
+ 1];
276 bias
= 1; /* bias = 1,2,1,2,... for successive samples */
277 for (outcol
= 0; outcol
< output_cols
; outcol
++) {
279 (JSAMPLE
)((inptr0
[0] + inptr0
[1] + inptr1
[0] + inptr1
[1] + bias
) >> 2);
280 bias
^= 3; /* 1=>2, 2=>1 */
281 inptr0
+= 2; inptr1
+= 2;
288 #ifdef INPUT_SMOOTHING_SUPPORTED
291 * Downsample pixel values of a single component.
292 * This version handles the standard case of 2:1 horizontal and 2:1 vertical,
293 * with smoothing. One row of context is required.
297 h2v2_smooth_downsample(j_compress_ptr cinfo
, jpeg_component_info
*compptr
,
298 JSAMPARRAY input_data
, JSAMPARRAY output_data
)
302 JDIMENSION output_cols
= compptr
->width_in_blocks
* DCTSIZE
;
303 register JSAMPROW inptr0
, inptr1
, above_ptr
, below_ptr
, outptr
;
304 JLONG membersum
, neighsum
, memberscale
, neighscale
;
306 /* Expand input data enough to let all the output samples be generated
307 * by the standard loop. Special-casing padded output would be more
310 expand_right_edge(input_data
- 1, cinfo
->max_v_samp_factor
+ 2,
311 cinfo
->image_width
, output_cols
* 2);
313 /* We don't bother to form the individual "smoothed" input pixel values;
314 * we can directly compute the output which is the average of the four
315 * smoothed values. Each of the four member pixels contributes a fraction
316 * (1-8*SF) to its own smoothed image and a fraction SF to each of the three
317 * other smoothed pixels, therefore a total fraction (1-5*SF)/4 to the final
318 * output. The four corner-adjacent neighbor pixels contribute a fraction
319 * SF to just one smoothed pixel, or SF/4 to the final output; while the
320 * eight edge-adjacent neighbors contribute SF to each of two smoothed
321 * pixels, or SF/2 overall. In order to use integer arithmetic, these
322 * factors are scaled by 2^16 = 65536.
323 * Also recall that SF = smoothing_factor / 1024.
326 memberscale
= 16384 - cinfo
->smoothing_factor
* 80; /* scaled (1-5*SF)/4 */
327 neighscale
= cinfo
->smoothing_factor
* 16; /* scaled SF/4 */
330 for (outrow
= 0; outrow
< compptr
->v_samp_factor
; outrow
++) {
331 outptr
= output_data
[outrow
];
332 inptr0
= input_data
[inrow
];
333 inptr1
= input_data
[inrow
+ 1];
334 above_ptr
= input_data
[inrow
- 1];
335 below_ptr
= input_data
[inrow
+ 2];
337 /* Special case for first column: pretend column -1 is same as column 0 */
338 membersum
= inptr0
[0] + inptr0
[1] + inptr1
[0] + inptr1
[1];
339 neighsum
= above_ptr
[0] + above_ptr
[1] + below_ptr
[0] + below_ptr
[1] +
340 inptr0
[0] + inptr0
[2] + inptr1
[0] + inptr1
[2];
341 neighsum
+= neighsum
;
342 neighsum
+= above_ptr
[0] + above_ptr
[2] + below_ptr
[0] + below_ptr
[2];
343 membersum
= membersum
* memberscale
+ neighsum
* neighscale
;
344 *outptr
++ = (JSAMPLE
)((membersum
+ 32768) >> 16);
345 inptr0
+= 2; inptr1
+= 2; above_ptr
+= 2; below_ptr
+= 2;
347 for (colctr
= output_cols
- 2; colctr
> 0; colctr
--) {
348 /* sum of pixels directly mapped to this output element */
349 membersum
= inptr0
[0] + inptr0
[1] + inptr1
[0] + inptr1
[1];
350 /* sum of edge-neighbor pixels */
351 neighsum
= above_ptr
[0] + above_ptr
[1] + below_ptr
[0] + below_ptr
[1] +
352 inptr0
[-1] + inptr0
[2] + inptr1
[-1] + inptr1
[2];
353 /* The edge-neighbors count twice as much as corner-neighbors */
354 neighsum
+= neighsum
;
355 /* Add in the corner-neighbors */
356 neighsum
+= above_ptr
[-1] + above_ptr
[2] + below_ptr
[-1] + below_ptr
[2];
357 /* form final output scaled up by 2^16 */
358 membersum
= membersum
* memberscale
+ neighsum
* neighscale
;
359 /* round, descale and output it */
360 *outptr
++ = (JSAMPLE
)((membersum
+ 32768) >> 16);
361 inptr0
+= 2; inptr1
+= 2; above_ptr
+= 2; below_ptr
+= 2;
364 /* Special case for last column */
365 membersum
= inptr0
[0] + inptr0
[1] + inptr1
[0] + inptr1
[1];
366 neighsum
= above_ptr
[0] + above_ptr
[1] + below_ptr
[0] + below_ptr
[1] +
367 inptr0
[-1] + inptr0
[1] + inptr1
[-1] + inptr1
[1];
368 neighsum
+= neighsum
;
369 neighsum
+= above_ptr
[-1] + above_ptr
[1] + below_ptr
[-1] + below_ptr
[1];
370 membersum
= membersum
* memberscale
+ neighsum
* neighscale
;
371 *outptr
= (JSAMPLE
)((membersum
+ 32768) >> 16);
379 * Downsample pixel values of a single component.
380 * This version handles the special case of a full-size component,
381 * with smoothing. One row of context is required.
385 fullsize_smooth_downsample(j_compress_ptr cinfo
, jpeg_component_info
*compptr
,
386 JSAMPARRAY input_data
, JSAMPARRAY output_data
)
390 JDIMENSION output_cols
= compptr
->width_in_blocks
* DCTSIZE
;
391 register JSAMPROW inptr
, above_ptr
, below_ptr
, outptr
;
392 JLONG membersum
, neighsum
, memberscale
, neighscale
;
393 int colsum
, lastcolsum
, nextcolsum
;
395 /* Expand input data enough to let all the output samples be generated
396 * by the standard loop. Special-casing padded output would be more
399 expand_right_edge(input_data
- 1, cinfo
->max_v_samp_factor
+ 2,
400 cinfo
->image_width
, output_cols
);
402 /* Each of the eight neighbor pixels contributes a fraction SF to the
403 * smoothed pixel, while the main pixel contributes (1-8*SF). In order
404 * to use integer arithmetic, these factors are multiplied by 2^16 = 65536.
405 * Also recall that SF = smoothing_factor / 1024.
408 memberscale
= 65536L - cinfo
->smoothing_factor
* 512L; /* scaled 1-8*SF */
409 neighscale
= cinfo
->smoothing_factor
* 64; /* scaled SF */
411 for (outrow
= 0; outrow
< compptr
->v_samp_factor
; outrow
++) {
412 outptr
= output_data
[outrow
];
413 inptr
= input_data
[outrow
];
414 above_ptr
= input_data
[outrow
- 1];
415 below_ptr
= input_data
[outrow
+ 1];
417 /* Special case for first column */
418 colsum
= (*above_ptr
++) + (*below_ptr
++) + inptr
[0];
419 membersum
= *inptr
++;
420 nextcolsum
= above_ptr
[0] + below_ptr
[0] + inptr
[0];
421 neighsum
= colsum
+ (colsum
- membersum
) + nextcolsum
;
422 membersum
= membersum
* memberscale
+ neighsum
* neighscale
;
423 *outptr
++ = (JSAMPLE
)((membersum
+ 32768) >> 16);
424 lastcolsum
= colsum
; colsum
= nextcolsum
;
426 for (colctr
= output_cols
- 2; colctr
> 0; colctr
--) {
427 membersum
= *inptr
++;
428 above_ptr
++; below_ptr
++;
429 nextcolsum
= above_ptr
[0] + below_ptr
[0] + inptr
[0];
430 neighsum
= lastcolsum
+ (colsum
- membersum
) + nextcolsum
;
431 membersum
= membersum
* memberscale
+ neighsum
* neighscale
;
432 *outptr
++ = (JSAMPLE
)((membersum
+ 32768) >> 16);
433 lastcolsum
= colsum
; colsum
= nextcolsum
;
436 /* Special case for last column */
438 neighsum
= lastcolsum
+ (colsum
- membersum
) + colsum
;
439 membersum
= membersum
* memberscale
+ neighsum
* neighscale
;
440 *outptr
= (JSAMPLE
)((membersum
+ 32768) >> 16);
445 #endif /* INPUT_SMOOTHING_SUPPORTED */
449 * Module initialization routine for downsampling.
450 * Note that we must select a routine for each component.
454 jinit_downsampler(j_compress_ptr cinfo
)
456 my_downsample_ptr downsample
;
458 jpeg_component_info
*compptr
;
459 boolean smoothok
= TRUE
;
461 downsample
= (my_downsample_ptr
)
462 (*cinfo
->mem
->alloc_small
) ((j_common_ptr
)cinfo
, JPOOL_IMAGE
,
463 sizeof(my_downsampler
));
464 cinfo
->downsample
= (struct jpeg_downsampler
*)downsample
;
465 downsample
->pub
.start_pass
= start_pass_downsample
;
466 downsample
->pub
.downsample
= sep_downsample
;
467 downsample
->pub
.need_context_rows
= FALSE
;
469 if (cinfo
->CCIR601_sampling
)
470 ERREXIT(cinfo
, JERR_CCIR601_NOTIMPL
);
472 /* Verify we can handle the sampling factors, and set up method pointers */
473 for (ci
= 0, compptr
= cinfo
->comp_info
; ci
< cinfo
->num_components
;
475 if (compptr
->h_samp_factor
== cinfo
->max_h_samp_factor
&&
476 compptr
->v_samp_factor
== cinfo
->max_v_samp_factor
) {
477 #ifdef INPUT_SMOOTHING_SUPPORTED
478 if (cinfo
->smoothing_factor
) {
479 downsample
->methods
[ci
] = fullsize_smooth_downsample
;
480 downsample
->pub
.need_context_rows
= TRUE
;
483 downsample
->methods
[ci
] = fullsize_downsample
;
484 } else if (compptr
->h_samp_factor
* 2 == cinfo
->max_h_samp_factor
&&
485 compptr
->v_samp_factor
== cinfo
->max_v_samp_factor
) {
487 if (jsimd_can_h2v1_downsample())
488 downsample
->methods
[ci
] = jsimd_h2v1_downsample
;
490 downsample
->methods
[ci
] = h2v1_downsample
;
491 } else if (compptr
->h_samp_factor
* 2 == cinfo
->max_h_samp_factor
&&
492 compptr
->v_samp_factor
* 2 == cinfo
->max_v_samp_factor
) {
493 #ifdef INPUT_SMOOTHING_SUPPORTED
494 if (cinfo
->smoothing_factor
) {
495 #if defined(__mips__)
496 if (jsimd_can_h2v2_smooth_downsample())
497 downsample
->methods
[ci
] = jsimd_h2v2_smooth_downsample
;
500 downsample
->methods
[ci
] = h2v2_smooth_downsample
;
501 downsample
->pub
.need_context_rows
= TRUE
;
505 if (jsimd_can_h2v2_downsample())
506 downsample
->methods
[ci
] = jsimd_h2v2_downsample
;
508 downsample
->methods
[ci
] = h2v2_downsample
;
510 } else if ((cinfo
->max_h_samp_factor
% compptr
->h_samp_factor
) == 0 &&
511 (cinfo
->max_v_samp_factor
% compptr
->v_samp_factor
) == 0) {
513 downsample
->methods
[ci
] = int_downsample
;
515 ERREXIT(cinfo
, JERR_FRACT_SAMPLE_NOTIMPL
);
518 #ifdef INPUT_SMOOTHING_SUPPORTED
519 if (cinfo
->smoothing_factor
&& !smoothok
)
520 TRACEMS(cinfo
, 0, JTRC_SMOOTH_NOTIMPL
);