u_format_pack.py revision 8d62b2aca99ba67f794dd682ed1ec49dc8826390
1#!/usr/bin/env python
2
3'''
4/**************************************************************************
5 *
6 * Copyright 2009-2010 VMware, Inc.
7 * All Rights Reserved.
8 *
9 * Permission is hereby granted, free of charge, to any person obtaining a
10 * copy of this software and associated documentation files (the
11 * "Software"), to deal in the Software without restriction, including
12 * without limitation the rights to use, copy, modify, merge, publish,
13 * distribute, sub license, and/or sell copies of the Software, and to
14 * permit persons to whom the Software is furnished to do so, subject to
15 * the following conditions:
16 *
17 * The above copyright notice and this permission notice (including the
18 * next paragraph) shall be included in all copies or substantial portions
19 * of the Software.
20 *
21 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
22 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
23 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
24 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
25 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
26 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
27 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
28 *
29 **************************************************************************/
30
31/**
32 * @file
33 * Pixel format packing and unpacking functions.
34 *
35 * @author Jose Fonseca <jfonseca@vmware.com>
36 */
37'''
38
39
40from u_format_parse import *
41
42
43def generate_format_type(format):
44    '''Generate a structure that describes the format.'''
45
46    assert format.layout == PLAIN
47
48    print 'union util_format_%s {' % format.short_name()
49
50    if format.block_size() in (8, 16, 32, 64):
51        print '   uint%u_t value;' % (format.block_size(),)
52
53    use_bitfields = False
54    for channel in format.channels:
55        if channel.size % 8 or not is_pot(channel.size):
56            use_bitfields = True
57
58    print '   struct {'
59    for channel in format.channels:
60        if use_bitfields:
61            if channel.type == VOID:
62                if channel.size:
63                    print '      unsigned %s:%u;' % (channel.name, channel.size)
64            elif channel.type == UNSIGNED:
65                print '      unsigned %s:%u;' % (channel.name, channel.size)
66            elif channel.type in (SIGNED, FIXED):
67                print '      int %s:%u;' % (channel.name, channel.size)
68            elif channel.type == FLOAT:
69                if channel.size == 64:
70                    print '      double %s;' % (channel.name)
71                elif channel.size == 32:
72                    print '      float %s;' % (channel.name)
73                else:
74                    print '      unsigned %s:%u;' % (channel.name, channel.size)
75            else:
76                assert 0
77        else:
78            assert channel.size % 8 == 0 and is_pot(channel.size)
79            if channel.type == VOID:
80                if channel.size:
81                    print '      uint%u_t %s;' % (channel.size, channel.name)
82            elif channel.type == UNSIGNED:
83                print '      uint%u_t %s;' % (channel.size, channel.name)
84            elif channel.type in (SIGNED, FIXED):
85                print '      int%u_t %s;' % (channel.size, channel.name)
86            elif channel.type == FLOAT:
87                if channel.size == 64:
88                    print '      double %s;' % (channel.name)
89                elif channel.size == 32:
90                    print '      float %s;' % (channel.name)
91                elif channel.size == 16:
92                    print '      uint16_t %s;' % (channel.name)
93                else:
94                    assert 0
95            else:
96                assert 0
97    print '   } chan;'
98    print '};'
99    print
100
101
102def bswap_format(format):
103    '''Generate a structure that describes the format.'''
104
105    if format.is_bitmask() and not format.is_array() and format.block_size() > 8:
106        print '#ifdef PIPE_ARCH_BIG_ENDIAN'
107        print '   pixel.value = util_bswap%u(pixel.value);' % format.block_size()
108        print '#endif'
109
110
111def is_format_supported(format):
112    '''Determines whether we actually have the plumbing necessary to generate the
113    to read/write to/from this format.'''
114
115    # FIXME: Ideally we would support any format combination here.
116
117    if format.layout != PLAIN:
118        return False
119
120    for i in range(4):
121        channel = format.channels[i]
122        if channel.type not in (VOID, UNSIGNED, SIGNED, FLOAT, FIXED):
123            return False
124        if channel.type == FLOAT and channel.size not in (16, 32, 64):
125            return False
126
127    return True
128
129
130def native_type(format):
131    '''Get the native appropriate for a format.'''
132
133    if format.layout == PLAIN:
134        if not format.is_array():
135            # For arithmetic pixel formats return the integer type that matches the whole pixel
136            return 'uint%u_t' % format.block_size()
137        else:
138            # For array pixel formats return the integer type that matches the color channel
139            channel = format.channels[0]
140            if channel.type in (UNSIGNED, VOID):
141                return 'uint%u_t' % channel.size
142            elif channel.type in (SIGNED, FIXED):
143                return 'int%u_t' % channel.size
144            elif channel.type == FLOAT:
145                if channel.size == 16:
146                    return 'uint16_t'
147                elif channel.size == 32:
148                    return 'float'
149                elif channel.size == 64:
150                    return 'double'
151                else:
152                    assert False
153            else:
154                assert False
155    else:
156        assert False
157
158
159def intermediate_native_type(bits, sign):
160    '''Find a native type adequate to hold intermediate results of the request bit size.'''
161
162    bytes = 4 # don't use anything smaller than 32bits
163    while bytes * 8 < bits:
164        bytes *= 2
165    bits = bytes*8
166
167    if sign:
168        return 'int%u_t' % bits
169    else:
170        return 'uint%u_t' % bits
171
172
173def get_one_shift(type):
174    '''Get the number of the bit that matches unity for this type.'''
175    if type.type == 'FLOAT':
176        assert False
177    if not type.norm:
178        return 0
179    if type.type == UNSIGNED:
180        return type.size
181    if type.type == SIGNED:
182        return type.size - 1
183    if type.type == FIXED:
184        return type.size / 2
185    assert False
186
187
188def value_to_native(type, value):
189    '''Get the value of unity for this type.'''
190    if type.type == FLOAT:
191        return value
192    if type.type == FIXED:
193        return int(value * (1 << (type.size/2)))
194    if not type.norm:
195        return int(value)
196    if type.type == UNSIGNED:
197        return int(value * ((1 << type.size) - 1))
198    if type.type == SIGNED:
199        return int(value * ((1 << (type.size - 1)) - 1))
200    assert False
201
202
203def native_to_constant(type, value):
204    '''Get the value of unity for this type.'''
205    if type.type == FLOAT:
206        if type.size <= 32:
207            return "%ff" % value
208        else:
209            return "%ff" % value
210    else:
211        return str(int(value))
212
213
214def get_one(type):
215    '''Get the value of unity for this type.'''
216    return value_to_native(type, 1)
217
218
219def clamp_expr(src_channel, dst_channel, dst_native_type, value):
220    '''Generate the expression to clamp the value in the source type to the
221    destination type range.'''
222
223    if src_channel == dst_channel:
224        return value
225
226    src_min = src_channel.min()
227    src_max = src_channel.max()
228    dst_min = dst_channel.min()
229    dst_max = dst_channel.max()
230
231    # Translate the destination range to the src native value
232    dst_min_native = value_to_native(src_channel, dst_min)
233    dst_max_native = value_to_native(src_channel, dst_max)
234
235    if src_min < dst_min and src_max > dst_max:
236        return 'CLAMP(%s, %s, %s)' % (value, dst_min_native, dst_max_native)
237
238    if src_max > dst_max:
239        return 'MIN2(%s, %s)' % (value, dst_max_native)
240
241    if src_min < dst_min:
242        return 'MAX2(%s, %s)' % (value, dst_min_native)
243
244    return value
245
246
247def conversion_expr(src_channel,
248                    dst_channel, dst_native_type,
249                    value,
250                    clamp=True,
251                    src_colorspace = RGB,
252                    dst_colorspace = RGB):
253    '''Generate the expression to convert a value between two types.'''
254
255    if src_colorspace != dst_colorspace:
256        if src_colorspace == SRGB:
257            assert src_channel.type == UNSIGNED
258            assert src_channel.norm
259            assert src_channel.size == 8
260            assert dst_colorspace == RGB
261            if dst_channel.type == FLOAT:
262                return 'util_format_srgb_8unorm_to_linear_float(%s)' % value
263            else:
264                assert dst_channel.type == UNSIGNED
265                assert dst_channel.norm
266                assert dst_channel.size == 8
267                return 'util_format_srgb_to_linear_8unorm(%s)' % value
268        elif dst_colorspace == SRGB:
269            assert dst_channel.type == UNSIGNED
270            assert dst_channel.norm
271            assert dst_channel.size == 8
272            assert src_colorspace == RGB
273            if src_channel.type == FLOAT:
274                return 'util_format_linear_float_to_srgb_8unorm(%s)' % value
275            else:
276                assert src_channel.type == UNSIGNED
277                assert src_channel.norm
278                assert src_channel.size == 8
279                return 'util_format_linear_to_srgb_8unorm(%s)' % value
280        elif src_colorspace == ZS:
281            pass
282        elif dst_colorspace == ZS:
283            pass
284        else:
285            assert 0
286
287    if src_channel == dst_channel:
288        return value
289
290    src_type = src_channel.type
291    src_size = src_channel.size
292    src_norm = src_channel.norm
293
294    # Promote half to float
295    if src_type == FLOAT and src_size == 16:
296        value = 'util_half_to_float(%s)' % value
297        src_size = 32
298
299    # Special case for float <-> ubytes for more accurate results
300    # Done before clamping since these functions already take care of that
301    if src_type == UNSIGNED and src_norm and src_size == 8 and dst_channel.type == FLOAT and dst_channel.size == 32:
302        return 'ubyte_to_float(%s)' % value
303    if src_type == FLOAT and src_size == 32 and dst_channel.type == UNSIGNED and dst_channel.norm and dst_channel.size == 8:
304        return 'float_to_ubyte(%s)' % value
305
306    if clamp:
307        if dst_channel.type != FLOAT or src_type != FLOAT:
308            value = clamp_expr(src_channel, dst_channel, dst_native_type, value)
309
310    if src_type in (SIGNED, UNSIGNED) and dst_channel.type in (SIGNED, UNSIGNED):
311        if not src_norm and not dst_channel.norm:
312            # neither is normalized -- just cast
313            return '(%s)%s' % (dst_native_type, value)
314
315        src_one = get_one(src_channel)
316        dst_one = get_one(dst_channel)
317
318        if src_one > dst_one and src_norm and dst_channel.norm:
319            # We can just bitshift
320            src_shift = get_one_shift(src_channel)
321            dst_shift = get_one_shift(dst_channel)
322            value = '(%s >> %s)' % (value, src_shift - dst_shift)
323        else:
324            # We need to rescale using an intermediate type big enough to hold the multiplication of both
325            tmp_native_type = intermediate_native_type(src_size + dst_channel.size, src_channel.sign and dst_channel.sign)
326            value = '((%s)%s)' % (tmp_native_type, value)
327            value = '(%s * 0x%x / 0x%x)' % (value, dst_one, src_one)
328        value = '(%s)%s' % (dst_native_type, value)
329        return value
330
331    # Promote to either float or double
332    if src_type != FLOAT:
333        if src_norm or src_type == FIXED:
334            one = get_one(src_channel)
335            if src_size <= 23:
336                value = '(%s * (1.0f/0x%x))' % (value, one)
337                if dst_channel.size <= 32:
338                    value = '(float)%s' % value
339                src_size = 32
340            else:
341                # bigger than single precision mantissa, use double
342                value = '(%s * (1.0/0x%x))' % (value, one)
343                src_size = 64
344            src_norm = False
345        else:
346            if src_size <= 23 or dst_channel.size <= 32:
347                value = '(float)%s' % value
348                src_size = 32
349            else:
350                # bigger than single precision mantissa, use double
351                value = '(double)%s' % value
352                src_size = 64
353        src_type = FLOAT
354
355    # Convert double or float to non-float
356    if dst_channel.type != FLOAT:
357        if dst_channel.norm or dst_channel.type == FIXED:
358            dst_one = get_one(dst_channel)
359            if dst_channel.size <= 23:
360                value = '(%s * 0x%x)' % (value, dst_one)
361            else:
362                # bigger than single precision mantissa, use double
363                value = '(%s * (double)0x%x)' % (value, dst_one)
364        value = '(%s)%s' % (dst_native_type, value)
365    else:
366        # Cast double to float when converting to either half or float
367        if dst_channel.size <= 32 and src_size > 32:
368            value = '(float)%s' % value
369            src_size = 32
370
371        if dst_channel.size == 16:
372            value = 'util_float_to_half(%s)' % value
373        elif dst_channel.size == 64 and src_size < 64:
374            value = '(double)%s' % value
375
376    return value
377
378
379def generate_unpack_kernel(format, dst_channel, dst_native_type):
380
381    if not is_format_supported(format):
382        return
383
384    assert format.layout == PLAIN
385
386    src_native_type = native_type(format)
387
388    if format.is_bitmask():
389        depth = format.block_size()
390        print '         uint%u_t value = *(const uint%u_t *)src;' % (depth, depth)
391
392        # Declare the intermediate variables
393        for i in range(format.nr_channels()):
394            src_channel = format.channels[i]
395            if src_channel.type == UNSIGNED:
396                print '         uint%u_t %s;' % (depth, src_channel.name)
397            elif src_channel.type == SIGNED:
398                print '         int%u_t %s;' % (depth, src_channel.name)
399
400        if depth > 8:
401            print '#ifdef PIPE_ARCH_BIG_ENDIAN'
402            print '         value = util_bswap%u(value);' % depth
403            print '#endif'
404
405        # Compute the intermediate unshifted values
406        shift = 0
407        for i in range(format.nr_channels()):
408            src_channel = format.channels[i]
409            value = 'value'
410            if src_channel.type == UNSIGNED:
411                if shift:
412                    value = '%s >> %u' % (value, shift)
413                if shift + src_channel.size < depth:
414                    value = '(%s) & 0x%x' % (value, (1 << src_channel.size) - 1)
415            elif src_channel.type == SIGNED:
416                if shift + src_channel.size < depth:
417                    # Align the sign bit
418                    lshift = depth - (shift + src_channel.size)
419                    value = '%s << %u' % (value, lshift)
420                # Cast to signed
421                value = '(int%u_t)(%s) ' % (depth, value)
422                if src_channel.size < depth:
423                    # Align the LSB bit
424                    rshift = depth - src_channel.size
425                    value = '(%s) >> %u' % (value, rshift)
426            else:
427                value = None
428
429            if value is not None:
430                print '         %s = %s;' % (src_channel.name, value)
431
432            shift += src_channel.size
433
434        # Convert, swizzle, and store final values
435        for i in range(4):
436            swizzle = format.swizzles[i]
437            if swizzle < 4:
438                src_channel = format.channels[swizzle]
439                src_colorspace = format.colorspace
440                if src_colorspace == SRGB and i == 3:
441                    # Alpha channel is linear
442                    src_colorspace = RGB
443                value = src_channel.name
444                value = conversion_expr(src_channel,
445                                        dst_channel, dst_native_type,
446                                        value,
447                                        src_colorspace = src_colorspace)
448            elif swizzle == SWIZZLE_0:
449                value = '0'
450            elif swizzle == SWIZZLE_1:
451                value = get_one(dst_channel)
452            elif swizzle == SWIZZLE_NONE:
453                value = '0'
454            else:
455                assert False
456            print '         dst[%u] = %s; /* %s */' % (i, value, 'rgba'[i])
457
458    else:
459        print '         union util_format_%s pixel;' % format.short_name()
460        print '         memcpy(&pixel, src, sizeof pixel);'
461        bswap_format(format)
462
463        for i in range(4):
464            swizzle = format.swizzles[i]
465            if swizzle < 4:
466                src_channel = format.channels[swizzle]
467                src_colorspace = format.colorspace
468                if src_colorspace == SRGB and i == 3:
469                    # Alpha channel is linear
470                    src_colorspace = RGB
471                value = 'pixel.chan.%s' % src_channel.name
472                value = conversion_expr(src_channel,
473                                        dst_channel, dst_native_type,
474                                        value,
475                                        src_colorspace = src_colorspace)
476            elif swizzle == SWIZZLE_0:
477                value = '0'
478            elif swizzle == SWIZZLE_1:
479                value = get_one(dst_channel)
480            elif swizzle == SWIZZLE_NONE:
481                value = '0'
482            else:
483                assert False
484            print '         dst[%u] = %s; /* %s */' % (i, value, 'rgba'[i])
485
486
487def generate_pack_kernel(format, src_channel, src_native_type):
488
489    if not is_format_supported(format):
490        return
491
492    dst_native_type = native_type(format)
493
494    assert format.layout == PLAIN
495
496    inv_swizzle = format.inv_swizzles()
497
498    if format.is_bitmask():
499        depth = format.block_size()
500        print '         uint%u_t value = 0;' % depth
501
502        shift = 0
503        for i in range(4):
504            dst_channel = format.channels[i]
505            if inv_swizzle[i] is not None:
506                value ='src[%u]' % inv_swizzle[i]
507                dst_colorspace = format.colorspace
508                if dst_colorspace == SRGB and inv_swizzle[i] == 3:
509                    # Alpha channel is linear
510                    dst_colorspace = RGB
511                value = conversion_expr(src_channel,
512                                        dst_channel, dst_native_type,
513                                        value,
514                                        dst_colorspace = dst_colorspace)
515                if dst_channel.type in (UNSIGNED, SIGNED):
516                    if shift + dst_channel.size < depth:
517                        value = '(%s) & 0x%x' % (value, (1 << dst_channel.size) - 1)
518                    if shift:
519                        value = '(%s) << %u' % (value, shift)
520                    if dst_channel.type == SIGNED:
521                        # Cast to unsigned
522                        value = '(uint%u_t)(%s) ' % (depth, value)
523                else:
524                    value = None
525                if value is not None:
526                    print '         value |= %s;' % (value)
527
528            shift += dst_channel.size
529
530        if depth > 8:
531            print '#ifdef PIPE_ARCH_BIG_ENDIAN'
532            print '         value = util_bswap%u(value);' % depth
533            print '#endif'
534
535        print '         *(uint%u_t *)dst = value;' % depth
536
537    else:
538        print '         union util_format_%s pixel;' % format.short_name()
539
540        for i in range(4):
541            dst_channel = format.channels[i]
542            width = dst_channel.size
543            if inv_swizzle[i] is None:
544                continue
545            dst_colorspace = format.colorspace
546            if dst_colorspace == SRGB and inv_swizzle[i] == 3:
547                # Alpha channel is linear
548                dst_colorspace = RGB
549            value ='src[%u]' % inv_swizzle[i]
550            value = conversion_expr(src_channel,
551                                    dst_channel, dst_native_type,
552                                    value,
553                                    dst_colorspace = dst_colorspace)
554            print '         pixel.chan.%s = %s;' % (dst_channel.name, value)
555
556        bswap_format(format)
557        print '         memcpy(dst, &pixel, sizeof pixel);'
558
559
560def generate_format_unpack(format, dst_channel, dst_native_type, dst_suffix):
561    '''Generate the function to unpack pixels from a particular format'''
562
563    name = format.short_name()
564
565    print 'static INLINE void'
566    print 'util_format_%s_unpack_%s(%s *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)' % (name, dst_suffix, dst_native_type)
567    print '{'
568
569    if is_format_supported(format):
570        print '   unsigned x, y;'
571        print '   for(y = 0; y < height; y += %u) {' % (format.block_height,)
572        print '      %s *dst = dst_row;' % (dst_native_type)
573        print '      const uint8_t *src = src_row;'
574        print '      for(x = 0; x < width; x += %u) {' % (format.block_width,)
575
576        generate_unpack_kernel(format, dst_channel, dst_native_type)
577
578        print '         src += %u;' % (format.block_size() / 8,)
579        print '         dst += 4;'
580        print '      }'
581        print '      src_row += src_stride;'
582        print '      dst_row += dst_stride/sizeof(*dst_row);'
583        print '   }'
584
585    print '}'
586    print
587
588
589def generate_format_pack(format, src_channel, src_native_type, src_suffix):
590    '''Generate the function to pack pixels to a particular format'''
591
592    name = format.short_name()
593
594    print 'static INLINE void'
595    print 'util_format_%s_pack_%s(uint8_t *dst_row, unsigned dst_stride, const %s *src_row, unsigned src_stride, unsigned width, unsigned height)' % (name, src_suffix, src_native_type)
596    print '{'
597
598    if is_format_supported(format):
599        print '   unsigned x, y;'
600        print '   for(y = 0; y < height; y += %u) {' % (format.block_height,)
601        print '      const %s *src = src_row;' % (src_native_type)
602        print '      uint8_t *dst = dst_row;'
603        print '      for(x = 0; x < width; x += %u) {' % (format.block_width,)
604
605        generate_pack_kernel(format, src_channel, src_native_type)
606
607        print '         src += 4;'
608        print '         dst += %u;' % (format.block_size() / 8,)
609        print '      }'
610        print '      dst_row += dst_stride;'
611        print '      src_row += src_stride/sizeof(*src_row);'
612        print '   }'
613
614    print '}'
615    print
616
617
618def generate_format_fetch(format, dst_channel, dst_native_type, dst_suffix):
619    '''Generate the function to unpack pixels from a particular format'''
620
621    name = format.short_name()
622
623    print 'static INLINE void'
624    print 'util_format_%s_fetch_%s(%s *dst, const uint8_t *src, unsigned i, unsigned j)' % (name, dst_suffix, dst_native_type)
625    print '{'
626
627    if is_format_supported(format):
628        generate_unpack_kernel(format, dst_channel, dst_native_type)
629
630    print '}'
631    print
632
633
634def is_format_hand_written(format):
635    return format.layout in ('s3tc', 'rgtc', 'subsampled', 'other') or format.colorspace == ZS
636
637
638def generate(formats):
639    print
640    print '#include "pipe/p_compiler.h"'
641    print '#include "u_math.h"'
642    print '#include "u_half.h"'
643    print '#include "u_format.h"'
644    print '#include "u_format_other.h"'
645    print '#include "u_format_srgb.h"'
646    print '#include "u_format_yuv.h"'
647    print '#include "u_format_zs.h"'
648    print
649
650    for format in formats:
651        if not is_format_hand_written(format):
652
653            if is_format_supported(format):
654                generate_format_type(format)
655
656            channel = Channel(FLOAT, False, 32)
657            native_type = 'float'
658            suffix = 'rgba_float'
659
660            generate_format_unpack(format, channel, native_type, suffix)
661            generate_format_pack(format, channel, native_type, suffix)
662            generate_format_fetch(format, channel, native_type, suffix)
663
664            channel = Channel(UNSIGNED, True, 8)
665            native_type = 'uint8_t'
666            suffix = 'rgba_8unorm'
667
668            generate_format_unpack(format, channel, native_type, suffix)
669            generate_format_pack(format, channel, native_type, suffix)
670
671