4
* Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5
* Copyright 2009-2011 D. R. Commander
7
* Based on the x86 SIMD extension for IJG JPEG library,
8
* Copyright (C) 1999-2006, MIYASAKA Masaru.
9
* For conditions of distribution and use, see copyright notice in jsimdext.inc
11
* This file contains the interface between the "normal" portions
12
* of the library and the SIMD implementations when running on
15
* Based on the stubs from 'jsimd_none.c'
18
#define JPEG_INTERNALS
19
#include "../jinclude.h"
20
#include "../jpeglib.h"
23
#include "../jsimddct.h"
30
static unsigned int simd_support = ~0;
32
#if defined(__linux__) || defined(ANDROID) || defined(__ANDROID__)
34
#define SOMEWHAT_SANE_PROC_CPUINFO_SIZE_LIMIT (1024 * 1024)
37
check_feature (char *buffer, char *feature)
42
if (strncmp(buffer, "Features", 8) != 0)
45
while (isspace(*buffer))
48
/* Check if 'feature' is present in the buffer as a separate word */
49
while ((p = strstr(buffer, feature))) {
50
if (p > buffer && !isspace(*(p - 1))) {
55
if (*p != 0 && !isspace(*p)) {
65
parse_proc_cpuinfo (int bufsize)
67
char *buffer = (char *)malloc(bufsize);
74
fd = fopen("/proc/cpuinfo", "r");
76
while (fgets(buffer, bufsize, fd)) {
77
if (!strchr(buffer, '\n') && !feof(fd)) {
78
/* "impossible" happened - insufficient size of the buffer! */
83
if (check_feature(buffer, "neon"))
84
simd_support |= JSIMD_ARM_NEON;
95
* Check what SIMD accelerations are supported.
97
* FIXME: This code is racy under a multi-threaded environment.
103
#if !defined(__ARM_NEON__) && defined(__linux__) || defined(ANDROID) || defined(__ANDROID__)
104
int bufsize = 1024; /* an initial guess for the line buffer size limit */
107
if (simd_support != ~0)
112
#if defined(__ARM_NEON__)
113
simd_support |= JSIMD_ARM_NEON;
114
#elif defined(__linux__) || defined(ANDROID) || defined(__ANDROID__)
115
/* We still have a chance to use NEON regardless of globally used
116
* -mcpu/-mfpu options passed to gcc by performing runtime detection via
117
* /proc/cpuinfo parsing on linux/android */
118
while (!parse_proc_cpuinfo(bufsize)) {
120
if (bufsize > SOMEWHAT_SANE_PROC_CPUINFO_SIZE_LIMIT)
125
/* Force different settings through environment variables */
126
env = getenv("JSIMD_FORCE_ARM_NEON");
127
if ((env != NULL) && (strcmp(env, "1") == 0))
128
simd_support &= JSIMD_ARM_NEON;
129
env = getenv("JSIMD_FORCE_NO_SIMD");
130
if ((env != NULL) && (strcmp(env, "1") == 0))
135
jsimd_can_rgb_ycc (void)
139
/* The code is optimised for these values only */
140
if (BITS_IN_JSAMPLE != 8)
142
if (sizeof(JDIMENSION) != 4)
144
if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
147
if (simd_support & JSIMD_ARM_NEON)
154
jsimd_can_rgb_gray (void)
162
jsimd_can_ycc_rgb (void)
166
/* The code is optimised for these values only */
167
if (BITS_IN_JSAMPLE != 8)
169
if (sizeof(JDIMENSION) != 4)
171
if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
173
if (simd_support & JSIMD_ARM_NEON)
180
jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
181
JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
182
JDIMENSION output_row, int num_rows)
184
void (*neonfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
186
switch(cinfo->in_color_space)
189
neonfct=jsimd_extrgb_ycc_convert_neon;
193
neonfct=jsimd_extrgbx_ycc_convert_neon;
196
neonfct=jsimd_extbgr_ycc_convert_neon;
200
neonfct=jsimd_extbgrx_ycc_convert_neon;
204
neonfct=jsimd_extxbgr_ycc_convert_neon;
208
neonfct=jsimd_extxrgb_ycc_convert_neon;
211
neonfct=jsimd_extrgb_ycc_convert_neon;
215
if (simd_support & JSIMD_ARM_NEON)
216
neonfct(cinfo->image_width, input_buf,
217
output_buf, output_row, num_rows);
221
jsimd_rgb_gray_convert (j_compress_ptr cinfo,
222
JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
223
JDIMENSION output_row, int num_rows)
228
jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
229
JSAMPIMAGE input_buf, JDIMENSION input_row,
230
JSAMPARRAY output_buf, int num_rows)
232
void (*neonfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
234
switch(cinfo->out_color_space)
237
neonfct=jsimd_ycc_extrgb_convert_neon;
241
neonfct=jsimd_ycc_extrgbx_convert_neon;
244
neonfct=jsimd_ycc_extbgr_convert_neon;
248
neonfct=jsimd_ycc_extbgrx_convert_neon;
252
neonfct=jsimd_ycc_extxbgr_convert_neon;
256
neonfct=jsimd_ycc_extxrgb_convert_neon;
259
neonfct=jsimd_ycc_extrgb_convert_neon;
263
if (simd_support & JSIMD_ARM_NEON)
264
neonfct(cinfo->output_width, input_buf,
265
input_row, output_buf, num_rows);
269
jsimd_can_h2v2_downsample (void)
277
jsimd_can_h2v1_downsample (void)
285
jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
286
JSAMPARRAY input_data, JSAMPARRAY output_data)
291
jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
292
JSAMPARRAY input_data, JSAMPARRAY output_data)
297
jsimd_can_h2v2_upsample (void)
305
jsimd_can_h2v1_upsample (void)
313
jsimd_h2v2_upsample (j_decompress_ptr cinfo,
314
jpeg_component_info * compptr,
315
JSAMPARRAY input_data,
316
JSAMPARRAY * output_data_ptr)
321
jsimd_h2v1_upsample (j_decompress_ptr cinfo,
322
jpeg_component_info * compptr,
323
JSAMPARRAY input_data,
324
JSAMPARRAY * output_data_ptr)
329
jsimd_can_h2v2_fancy_upsample (void)
337
jsimd_can_h2v1_fancy_upsample (void)
345
jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
346
jpeg_component_info * compptr,
347
JSAMPARRAY input_data,
348
JSAMPARRAY * output_data_ptr)
353
jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
354
jpeg_component_info * compptr,
355
JSAMPARRAY input_data,
356
JSAMPARRAY * output_data_ptr)
361
jsimd_can_h2v2_merged_upsample (void)
369
jsimd_can_h2v1_merged_upsample (void)
377
jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
378
JSAMPIMAGE input_buf,
379
JDIMENSION in_row_group_ctr,
380
JSAMPARRAY output_buf)
385
jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
386
JSAMPIMAGE input_buf,
387
JDIMENSION in_row_group_ctr,
388
JSAMPARRAY output_buf)
393
jsimd_can_convsamp (void)
397
/* The code is optimised for these values only */
400
if (BITS_IN_JSAMPLE != 8)
402
if (sizeof(JDIMENSION) != 4)
404
if (sizeof(DCTELEM) != 2)
407
if (simd_support & JSIMD_ARM_NEON)
414
jsimd_can_convsamp_float (void)
422
jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
425
if (simd_support & JSIMD_ARM_NEON)
426
jsimd_convsamp_neon(sample_data, start_col, workspace);
430
jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
431
FAST_FLOAT * workspace)
436
jsimd_can_fdct_islow (void)
444
jsimd_can_fdct_ifast (void)
448
/* The code is optimised for these values only */
451
if (sizeof(DCTELEM) != 2)
454
if (simd_support & JSIMD_ARM_NEON)
461
jsimd_can_fdct_float (void)
469
jsimd_fdct_islow (DCTELEM * data)
474
jsimd_fdct_ifast (DCTELEM * data)
476
if (simd_support & JSIMD_ARM_NEON)
477
jsimd_fdct_ifast_neon(data);
481
jsimd_fdct_float (FAST_FLOAT * data)
486
jsimd_can_quantize (void)
490
/* The code is optimised for these values only */
493
if (sizeof(JCOEF) != 2)
495
if (sizeof(DCTELEM) != 2)
498
if (simd_support & JSIMD_ARM_NEON)
505
jsimd_can_quantize_float (void)
513
jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors,
516
if (simd_support & JSIMD_ARM_NEON)
517
jsimd_quantize_neon(coef_block, divisors, workspace);
521
jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors,
522
FAST_FLOAT * workspace)
527
jsimd_can_idct_2x2 (void)
531
/* The code is optimised for these values only */
534
if (sizeof(JCOEF) != 2)
536
if (BITS_IN_JSAMPLE != 8)
538
if (sizeof(JDIMENSION) != 4)
540
if (sizeof(ISLOW_MULT_TYPE) != 2)
543
if ((simd_support & JSIMD_ARM_NEON))
550
jsimd_can_idct_4x4 (void)
554
/* The code is optimised for these values only */
557
if (sizeof(JCOEF) != 2)
559
if (BITS_IN_JSAMPLE != 8)
561
if (sizeof(JDIMENSION) != 4)
563
if (sizeof(ISLOW_MULT_TYPE) != 2)
566
if ((simd_support & JSIMD_ARM_NEON))
573
jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
574
JCOEFPTR coef_block, JSAMPARRAY output_buf,
575
JDIMENSION output_col)
577
if ((simd_support & JSIMD_ARM_NEON))
578
jsimd_idct_2x2_neon(compptr->dct_table, coef_block, output_buf, output_col);
582
jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
583
JCOEFPTR coef_block, JSAMPARRAY output_buf,
584
JDIMENSION output_col)
586
if ((simd_support & JSIMD_ARM_NEON))
587
jsimd_idct_4x4_neon(compptr->dct_table, coef_block, output_buf, output_col);
591
jsimd_can_idct_islow (void)
595
/* The code is optimised for these values only */
598
if (sizeof(JCOEF) != 2)
600
if (BITS_IN_JSAMPLE != 8)
602
if (sizeof(JDIMENSION) != 4)
604
if (sizeof(ISLOW_MULT_TYPE) != 2)
607
if (simd_support & JSIMD_ARM_NEON)
614
jsimd_can_idct_ifast (void)
618
/* The code is optimised for these values only */
621
if (sizeof(JCOEF) != 2)
623
if (BITS_IN_JSAMPLE != 8)
625
if (sizeof(JDIMENSION) != 4)
627
if (sizeof(IFAST_MULT_TYPE) != 2)
629
if (IFAST_SCALE_BITS != 2)
632
if ((simd_support & JSIMD_ARM_NEON))
639
jsimd_can_idct_float (void)
647
jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
648
JCOEFPTR coef_block, JSAMPARRAY output_buf,
649
JDIMENSION output_col)
651
if ((simd_support & JSIMD_ARM_NEON))
652
jsimd_idct_islow_neon(compptr->dct_table, coef_block, output_buf, output_col);
656
jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
657
JCOEFPTR coef_block, JSAMPARRAY output_buf,
658
JDIMENSION output_col)
660
if ((simd_support & JSIMD_ARM_NEON))
661
jsimd_idct_ifast_neon(compptr->dct_table, coef_block, output_buf, output_col);
665
jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
666
JCOEFPTR coef_block, JSAMPARRAY output_buf,
667
JDIMENSION output_col)