1 /*
2 ** SPDX-License-Identifier: BSD-3-Clause
3 ** Copyright Contributors to the OpenEXR Project.
4 */
5 
6 #ifndef OPENEXR_CORE_ENCODE_H
7 #define OPENEXR_CORE_ENCODE_H
8 
9 #include "openexr_chunkio.h"
10 #include "openexr_coding.h"
11 
12 #ifdef __cplusplus
13 extern "C" {
14 #endif
15 
16 /** @file */
17 
18 /** Can be bit-wise or'ed into the decode_flags in the decode pipeline.
19  *
20  * Indicates that the sample count table should be encoded from an
21  * individual sample count list (n, m, o, ...), meaning it will have
22  * to compute the cumulative counts on the fly.
23  *
24  * Without this (i.e. a value of 0 in that bit), indicates the sample
25  * count table is already a cumulative list (n, n+m, n+m+o, ...),
26  * which is the on-disk representation.
27  */
28 #define EXR_ENCODE_DATA_SAMPLE_COUNTS_ARE_INDIVIDUAL ((uint16_t) (1 << 0))
29 
30 /** Can be bit-wise or'ed into the decode_flags in the decode pipeline.
31  *
32  * Indicates that the data in the channel pointers to encode from is not
33  * a direct pointer, but instead is a pointer-to-pointers. In this
34  * mode, the user_pixel_stride and user_line_stride are used to
35  * advance the pointer offsets for each pixel in the output, but the
36  * user_bytes_per_element and user_data_type are used to put
37  * (successive) entries into each destination.
38  *
39  * So each channel pointer must then point to an array of
40  * chunk.width * chunk.height pointers. If an entry is
41  * `NULL`, 0 samples will be placed in the output.
42  *
43  * If this is NOT set (0), the default packing routine assumes the
44  * data will be planar and contiguous (each channel is a separate
45  * memory block), ignoring user_line_stride and user_pixel_stride and
46  * advancing only by the sample counts and bytes per element.
47  */
48 #define EXR_ENCODE_NON_IMAGE_DATA_AS_POINTERS ((uint16_t) (1 << 1))
49 
50 /** Struct meant to be used on a per-thread basis for writing exr data.
51  *
52  * As should be obvious, this structure is NOT thread safe, but rather
53  * meant to be used by separate threads, which can all be accessing
54  * the same context concurrently.
55  */
56 typedef struct _exr_encode_pipeline
57 {
58     /** The output channel information for this chunk.
59      *
60      * User is expected to fill the channel pointers for the input
61      * channels. For writing, all channels must be initialized prior
62      * to using exr_encoding_choose_default_routines(). If a custom pack routine
63      * is written, that is up to the implementor.
64      *
65      * Describes the channel information. This information is
66      * allocated dynamically during exr_encoding_initialize().
67      */
68     exr_coding_channel_info_t* channels;
69     int16_t                    channel_count;
70 
71     /** Encode flags to control the behavior. */
72     uint16_t encode_flags;
73 
74     /** Copy of the parameters given to the initialize/update for convenience. */
75     int                 part_index;
76     exr_const_context_t context;
77     exr_chunk_info_t    chunk;
78 
79     /** Can be used by the user to pass custom context data through
80      * the encode pipeline.
81      */
82     void* encoding_user_data;
83 
84     /** The packed buffer where individual channels have been put into here.
85      *
86      * If `NULL`, will be allocated during the run of the pipeline.
87      *
88      * If the caller wishes to take control of the buffer, simple
89      * adopt the pointer and set it to `NULL` here. Be cognizant of any
90      * custom allocators.
91      */
92     void* packed_buffer;
93 
94     /** Differing from the allocation size, the number of actual bytes */
95     uint64_t packed_bytes;
96 
97     /** Used when re-using the same encode pipeline struct to know if
98      * chunk is changed size whether current buffer is large enough
99      *
100      * If `NULL`, will be allocated during the run of the pipeline.
101      *
102      * If the caller wishes to take control of the buffer, simple
103      * adopt the pointer and set it to `NULL` here. Be cognizant of any
104      * custom allocators.
105      */
106     size_t packed_alloc_size;
107 
108     /** For deep data. NB: the members NOT const because we need to
109      * temporarily swap it to xdr order and restore it (to avoid a
110      * duplicate buffer allocation).
111      *
112      * Depending on the flag set above, will be treated either as a
113      * cumulative list (n, n+m, n+m+o, ...), or an individual table
114      * (n, m, o, ...). */
115     int32_t* sample_count_table;
116 
117     /** Allocated table size (to avoid re-allocations). Number of
118      * samples must always be width * height for the chunk.
119      */
120     size_t sample_count_alloc_size;
121 
122     /** Packed sample table (compressed, raw on disk representation)
123      * for deep or other non-image data.
124      */
125     void* packed_sample_count_table;
126 
127     /** Number of bytes to write (actual size) for the
128      * packed_sample_count_table.
129      */
130     size_t packed_sample_count_bytes;
131 
132     /** Allocated size (to avoid re-allocations) for the
133      * packed_sample_count_table.
134      */
135     size_t packed_sample_count_alloc_size;
136 
137     /** The compressed buffer, only needed for compressed files.
138      *
139      * If `NULL`, will be allocated during the run of the pipeline when
140      * needed.
141      *
142      * If the caller wishes to take control of the buffer, simple
143      * adopt the pointer and set it to `NULL` here. Be cognizant of any
144      * custom allocators.
145      */
146     void* compressed_buffer;
147 
148     /** Must be filled in as the pipeline runs to inform the writing
149      * software about the compressed size of the chunk (if it is an
150      * uncompressed file or the compression would make the file
151      * larger, it is expected to be the packed_buffer)
152      *
153      * If the caller wishes to take control of the buffer, simple
154      * adopt the pointer and set it to zero here. Be cognizant of any
155      * custom allocators.
156      */
157     size_t compressed_bytes;
158 
159     /** Used when re-using the same encode pipeline struct to know if
160      * chunk is changed size whether current buffer is large enough.
161      *
162      * If `NULL`, will be allocated during the run of the pipeline when
163      * needed.
164      *
165      * If the caller wishes to take control of the buffer, simple
166      * adopt the pointer and set it to zero here. Be cognizant of any
167      * custom allocators.
168      */
169     size_t compressed_alloc_size;
170 
171     /** A scratch buffer for intermediate results.
172      *
173      * If `NULL`, will be allocated during the run of the pipeline when
174      * needed.
175      *
176      * If the caller wishes to take control of the buffer, simple
177      * adopt the pointer and set it to `NULL` here. Be cognizant of any
178      * custom allocators.
179      */
180     void* scratch_buffer_1;
181 
182     /** Used when re-using the same encode pipeline struct to know if
183      * chunk is changed size whether current buffer is large enough.
184      *
185      * If `NULL`, will be allocated during the run of the pipeline when
186      * needed.
187      *
188      * If the caller wishes to take control of the buffer, simple
189      * adopt the pointer and set it to `NULL` here. Be cognizant of any
190      * custom allocators.
191      */
192     size_t scratch_alloc_size_1;
193 
194     /** Some compression routines may need a second scratch buffer.
195      *
196      * If `NULL`, will be allocated during the run of the pipeline when
197      * needed.
198      *
199      * If the caller wishes to take control of the buffer, simple
200      * adopt the pointer and set it to `NULL` here. Be cognizant of any
201      * custom allocators.
202      */
203     void* scratch_buffer_2;
204 
205     /** Used when re-using the same encode pipeline struct to know if
206      * chunk is changed size whether current buffer is large enough.
207      */
208     size_t scratch_alloc_size_2;
209 
210     /** Enable a custom allocator for the different buffers (if
211      * encoding on a GPU). If `NULL`, will use the allocator from the
212      * context.
213      */
214     void* (*alloc_fn) (exr_transcoding_pipeline_buffer_id_t, size_t);
215 
216     /** Enable a custom allocator for the different buffers (if
217      * encoding on a GPU). If `NULL`, will use the allocator from the
218      * context.
219      */
220     void (*free_fn) (exr_transcoding_pipeline_buffer_id_t, void*);
221 
222     /** Function chosen based on the output layout of the channels of the part to
223      * decompress data.
224      *
225      * If the user has a custom method for the
226      * compression on this part, this can be changed after
227      * initialization.
228      */
229     exr_result_t (*convert_and_pack_fn) (struct _exr_encode_pipeline* pipeline);
230 
231     /** Function chosen based on the compression type of the part to
232      * compress data.
233      *
234      * If the user has a custom compression method for the compression
235      * type on this part, this can be changed after initialization.
236      */
237     exr_result_t (*compress_fn) (struct _exr_encode_pipeline* pipeline);
238 
239     /** This routine is used when waiting for other threads to finish
240      * writing previous chunks such that this thread can write this
241      * chunk. This is used for parts which have a specified chunk
242      * ordering (increasing/decreasing y) and the chunks can not be
243      * written randomly (as could be true for uncompressed).
244      *
245      * This enables the calling application to contribute thread time
246      * to other computation as needed, or just use something like
247      * pthread_yield().
248      *
249      * By default, this routine will be assigned to a function which
250      * returns an error, failing the encode immediately. In this way,
251      * it assumes that there is only one thread being used for
252      * writing.
253      *
254      * It is up to the user to provide an appropriate routine if
255      * performing multi-threaded writing.
256      */
257     exr_result_t (*yield_until_ready_fn) (
258         struct _exr_encode_pipeline* pipeline);
259 
260     /** Function chosen to write chunk data to the context.
261      *
262      * This is allowed to be overridden, but probably is not necessary
263      * in most scenarios.
264      */
265     exr_result_t (*write_fn) (struct _exr_encode_pipeline* pipeline);
266 
267     /** Small stash of channel info values. This is faster than calling
268      * malloc when the channel count in the part is small (RGBAZ),
269      * which is super common, however if there are a large number of
270      * channels, it will allocate space for that, so do not rely on
271      * this being used.
272      */
273     exr_coding_channel_info_t _quick_chan_store[5];
274 } exr_encode_pipeline_t;
275 
276 /** @brief Simple macro to initialize an empty decode pipeline. */
277 #define EXR_ENCODE_PIPELINE_INITIALIZER                                        \
278     {                                                                          \
279         0                                                                      \
280     }
281 
282 /** Initialize the encoding pipeline structure with the channel info
283  * for the specified part based on the chunk to be written.
284  *
285  * NB: The encode_pipe->pack_and_convert_fn field will be `NULL` after this. If that
286  * stage is desired, initialize the channel output information and
287  * call exr_encoding_choose_default_routines().
288  */
289 EXR_EXPORT
290 exr_result_t exr_encoding_initialize (
291     exr_const_context_t     ctxt,
292     int                     part_index,
293     const exr_chunk_info_t* cinfo,
294     exr_encode_pipeline_t*  encode_pipe);
295 
296 /** Given an initialized encode pipeline, find an appropriate
297  * function to shuffle and convert data into the defined channel
298  * outputs.
299  *
300  * Calling this is not required if a custom routine will be used, or
301  * if just the raw decompressed data is desired.
302  */
303 EXR_EXPORT
304 exr_result_t exr_encoding_choose_default_routines (
305     exr_const_context_t    ctxt,
306     int                    part_index,
307     exr_encode_pipeline_t* encode_pipe);
308 
309 /** Given a encode pipeline previously initialized, update it for the
310  * new chunk to be written.
311  *
312  * In this manner, memory buffers can be re-used to avoid continual
313  * malloc/free calls. Further, it allows the previous choices for
314  * the various functions to be quickly re-used.
315  */
316 EXR_EXPORT
317 exr_result_t exr_encoding_update (
318     exr_const_context_t     ctxt,
319     int                     part_index,
320     const exr_chunk_info_t* cinfo,
321     exr_encode_pipeline_t*  encode_pipe);
322 
323 /** Execute the encoding pipeline. */
324 EXR_EXPORT
325 exr_result_t exr_encoding_run (
326     exr_const_context_t    ctxt,
327     int                    part_index,
328     exr_encode_pipeline_t* encode_pipe);
329 
330 /** Free any intermediate memory in the encoding pipeline.
331  *
332  * This does NOT free any pointers referred to in the channel info
333  * areas, but rather only the intermediate buffers and memory needed
334  * for the structure itself.
335  */
336 EXR_EXPORT
337 exr_result_t exr_encoding_destroy (
338     exr_const_context_t ctxt, exr_encode_pipeline_t* encode_pipe);
339 
340 #ifdef __cplusplus
341 } /* extern "C" */
342 #endif
343 
344 #endif /* OPENEXR_CORE_ENCODE_H */
345