af35770e0654de02a38b1d135bdef0be6bee32c3
[ffmpeg.git] / libavutil / opencl.c
1 /*
2  * Copyright (C) 2012 Peng  Gao     <peng@multicorewareinc.com>
3  * Copyright (C) 2012 Li    Cao     <li@multicorewareinc.com>
4  * Copyright (C) 2012 Wei   Gao     <weigao@multicorewareinc.com>
5  * Copyright (C) 2013 Lenny Wang    <lwanghpc@gmail.com>
6  *
7  * This file is part of FFmpeg.
8  *
9  * FFmpeg is free software; you can redistribute it and/or
10  * modify it under the terms of the GNU Lesser General Public
11  * License as published by the Free Software Foundation; either
12  * version 2.1 of the License, or (at your option) any later version.
13  *
14  * FFmpeg is distributed in the hope that it will be useful,
15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
17  * Lesser General Public License for more details.
18  *
19  * You should have received a copy of the GNU Lesser General Public
20  * License along with FFmpeg; if not, write to the Free Software
21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22  */
23
24 #include "opencl.h"
25 #include "avstring.h"
26 #include "log.h"
27 #include "avassert.h"
28 #include "opt.h"
29
30 #if HAVE_THREADS
31 #include "thread.h"
32 #include "atomic.h"
33
34 static pthread_mutex_t * volatile atomic_opencl_lock = NULL;
35 #define LOCK_OPENCL pthread_mutex_lock(atomic_opencl_lock)
36 #define UNLOCK_OPENCL pthread_mutex_unlock(atomic_opencl_lock)
37 #else
38 #define LOCK_OPENCL
39 #define UNLOCK_OPENCL
40 #endif
41
42 #define MAX_KERNEL_CODE_NUM 200
43
44 typedef struct {
45     int is_compiled;
46     const char *kernel_string;
47 } KernelCode;
48
49 typedef struct {
50     const AVClass *class;
51     int log_offset;
52     void *log_ctx;
53     int init_count;
54     int opt_init_flag;
55      /**
56      * if set to 1, the OpenCL environment was created by the user and
57      * passed as AVOpenCLExternalEnv when initing ,0:created by opencl wrapper.
58      */
59     int is_user_created;
60     int platform_idx;
61     int device_idx;
62     cl_platform_id platform_id;
63     cl_device_type device_type;
64     cl_context context;
65     cl_device_id device_id;
66     cl_command_queue command_queue;
67     int kernel_code_count;
68     KernelCode kernel_code[MAX_KERNEL_CODE_NUM];
69     AVOpenCLDeviceList device_list;
70 } OpenclContext;
71
72 #define OFFSET(x) offsetof(OpenclContext, x)
73
74 static const AVOption opencl_options[] = {
75      { "platform_idx",        "set platform index value",  OFFSET(platform_idx),  AV_OPT_TYPE_INT,    {.i64=-1}, -1, INT_MAX},
76      { "device_idx",          "set device index value",    OFFSET(device_idx),    AV_OPT_TYPE_INT,    {.i64=-1}, -1, INT_MAX},
77      { NULL }
78 };
79
80 static const AVClass openclutils_class = {
81     .class_name                = "opencl",
82     .option                    = opencl_options,
83     .item_name                 = av_default_item_name,
84     .version                   = LIBAVUTIL_VERSION_INT,
85     .log_level_offset_offset   = offsetof(OpenclContext, log_offset),
86     .parent_log_context_offset = offsetof(OpenclContext, log_ctx),
87 };
88
89 static OpenclContext opencl_ctx = {&openclutils_class};
90
91 static const cl_device_type device_type[] = {CL_DEVICE_TYPE_GPU, CL_DEVICE_TYPE_CPU};
92
93 typedef struct {
94     int err_code;
95     const char *err_str;
96 } OpenclErrorMsg;
97
98 static const OpenclErrorMsg opencl_err_msg[] = {
99     {CL_DEVICE_NOT_FOUND,                               "DEVICE NOT FOUND"},
100     {CL_DEVICE_NOT_AVAILABLE,                           "DEVICE NOT AVAILABLE"},
101     {CL_COMPILER_NOT_AVAILABLE,                         "COMPILER NOT AVAILABLE"},
102     {CL_MEM_OBJECT_ALLOCATION_FAILURE,                  "MEM OBJECT ALLOCATION FAILURE"},
103     {CL_OUT_OF_RESOURCES,                               "OUT OF RESOURCES"},
104     {CL_OUT_OF_HOST_MEMORY,                             "OUT OF HOST MEMORY"},
105     {CL_PROFILING_INFO_NOT_AVAILABLE,                   "PROFILING INFO NOT AVAILABLE"},
106     {CL_MEM_COPY_OVERLAP,                               "MEM COPY OVERLAP"},
107     {CL_IMAGE_FORMAT_MISMATCH,                          "IMAGE FORMAT MISMATCH"},
108     {CL_IMAGE_FORMAT_NOT_SUPPORTED,                     "IMAGE FORMAT NOT_SUPPORTED"},
109     {CL_BUILD_PROGRAM_FAILURE,                          "BUILD PROGRAM FAILURE"},
110     {CL_MAP_FAILURE,                                    "MAP FAILURE"},
111     {CL_MISALIGNED_SUB_BUFFER_OFFSET,                   "MISALIGNED SUB BUFFER OFFSET"},
112     {CL_EXEC_STATUS_ERROR_FOR_EVENTS_IN_WAIT_LIST,      "EXEC STATUS ERROR FOR EVENTS IN WAIT LIST"},
113     {CL_COMPILE_PROGRAM_FAILURE,                        "COMPILE PROGRAM FAILURE"},
114     {CL_LINKER_NOT_AVAILABLE,                           "LINKER NOT AVAILABLE"},
115     {CL_LINK_PROGRAM_FAILURE,                           "LINK PROGRAM FAILURE"},
116     {CL_DEVICE_PARTITION_FAILED,                        "DEVICE PARTITION FAILED"},
117     {CL_KERNEL_ARG_INFO_NOT_AVAILABLE,                  "KERNEL ARG INFO NOT AVAILABLE"},
118     {CL_INVALID_VALUE,                                  "INVALID VALUE"},
119     {CL_INVALID_DEVICE_TYPE,                            "INVALID DEVICE TYPE"},
120     {CL_INVALID_PLATFORM,                               "INVALID PLATFORM"},
121     {CL_INVALID_DEVICE,                                 "INVALID DEVICE"},
122     {CL_INVALID_CONTEXT,                                "INVALID CONTEXT"},
123     {CL_INVALID_QUEUE_PROPERTIES,                       "INVALID QUEUE PROPERTIES"},
124     {CL_INVALID_COMMAND_QUEUE,                          "INVALID COMMAND QUEUE"},
125     {CL_INVALID_HOST_PTR,                               "INVALID HOST PTR"},
126     {CL_INVALID_MEM_OBJECT,                             "INVALID MEM OBJECT"},
127     {CL_INVALID_IMAGE_FORMAT_DESCRIPTOR,                "INVALID IMAGE FORMAT DESCRIPTOR"},
128     {CL_INVALID_IMAGE_SIZE,                             "INVALID IMAGE SIZE"},
129     {CL_INVALID_SAMPLER,                                "INVALID SAMPLER"},
130     {CL_INVALID_BINARY,                                 "INVALID BINARY"},
131     {CL_INVALID_BUILD_OPTIONS,                          "INVALID BUILD OPTIONS"},
132     {CL_INVALID_PROGRAM,                                "INVALID PROGRAM"},
133     {CL_INVALID_PROGRAM_EXECUTABLE,                     "INVALID PROGRAM EXECUTABLE"},
134     {CL_INVALID_KERNEL_NAME,                            "INVALID KERNEL NAME"},
135     {CL_INVALID_KERNEL_DEFINITION,                      "INVALID KERNEL DEFINITION"},
136     {CL_INVALID_KERNEL,                                 "INVALID KERNEL"},
137     {CL_INVALID_ARG_INDEX,                              "INVALID ARG INDEX"},
138     {CL_INVALID_ARG_VALUE,                              "INVALID ARG VALUE"},
139     {CL_INVALID_ARG_SIZE,                               "INVALID ARG_SIZE"},
140     {CL_INVALID_KERNEL_ARGS,                            "INVALID KERNEL ARGS"},
141     {CL_INVALID_WORK_DIMENSION,                         "INVALID WORK DIMENSION"},
142     {CL_INVALID_WORK_GROUP_SIZE,                        "INVALID WORK GROUP SIZE"},
143     {CL_INVALID_WORK_ITEM_SIZE,                         "INVALID WORK ITEM SIZE"},
144     {CL_INVALID_GLOBAL_OFFSET,                          "INVALID GLOBAL OFFSET"},
145     {CL_INVALID_EVENT_WAIT_LIST,                        "INVALID EVENT WAIT LIST"},
146     {CL_INVALID_EVENT,                                  "INVALID EVENT"},
147     {CL_INVALID_OPERATION,                              "INVALID OPERATION"},
148     {CL_INVALID_GL_OBJECT,                              "INVALID GL OBJECT"},
149     {CL_INVALID_BUFFER_SIZE,                            "INVALID BUFFER SIZE"},
150     {CL_INVALID_MIP_LEVEL,                              "INVALID MIP LEVEL"},
151     {CL_INVALID_GLOBAL_WORK_SIZE,                       "INVALID GLOBAL WORK SIZE"},
152     {CL_INVALID_PROPERTY,                               "INVALID PROPERTY"},
153     {CL_INVALID_IMAGE_DESCRIPTOR,                       "INVALID IMAGE DESCRIPTOR"},
154     {CL_INVALID_COMPILER_OPTIONS,                       "INVALID COMPILER OPTIONS"},
155     {CL_INVALID_LINKER_OPTIONS,                         "INVALID LINKER OPTIONS"},
156     {CL_INVALID_DEVICE_PARTITION_COUNT,                 "INVALID DEVICE PARTITION COUNT"},
157 };
158
159 const char *av_opencl_errstr(cl_int status)
160 {
161     int i;
162     for (i = 0; i < FF_ARRAY_ELEMS(opencl_err_msg); i++) {
163         if (opencl_err_msg[i].err_code == status)
164             return opencl_err_msg[i].err_str;
165     }
166     return "unknown error";
167 }
168
169 static void free_device_list(AVOpenCLDeviceList *device_list)
170 {
171     int i, j;
172     if (!device_list)
173         return;
174     for (i = 0; i < device_list->platform_num; i++) {
175         if (!device_list->platform_node[i])
176             continue;
177         for (j = 0; j < device_list->platform_node[i]->device_num; j++) {
178             av_freep(&(device_list->platform_node[i]->device_node[j]->device_name));
179             av_freep(&(device_list->platform_node[i]->device_node[j]));
180         }
181         av_freep(&device_list->platform_node[i]->device_node);
182         av_freep(&(device_list->platform_node[i]->platform_name));
183         av_freep(&device_list->platform_node[i]);
184     }
185     av_freep(&device_list->platform_node);
186     device_list->platform_num = 0;
187 }
188
189 static int get_device_list(AVOpenCLDeviceList *device_list)
190 {
191     cl_int status;
192     int i, j, k, device_num, total_devices_num, ret = 0;
193     int *devices_num;
194     cl_platform_id *platform_ids = NULL;
195     cl_device_id *device_ids = NULL;
196     AVOpenCLDeviceNode *device_node = NULL;
197     size_t platform_name_size = 0;
198     size_t device_name_size = 0;
199     status = clGetPlatformIDs(0, NULL, &device_list->platform_num);
200     if (status != CL_SUCCESS) {
201         av_log(&opencl_ctx, AV_LOG_ERROR,
202                "Could not get OpenCL platform ids: %s\n", av_opencl_errstr(status));
203         return AVERROR_EXTERNAL;
204     }
205     platform_ids = av_mallocz_array(device_list->platform_num, sizeof(cl_platform_id));
206     if (!platform_ids)
207         return AVERROR(ENOMEM);
208     status = clGetPlatformIDs(device_list->platform_num, platform_ids, NULL);
209     if (status != CL_SUCCESS) {
210         av_log(&opencl_ctx, AV_LOG_ERROR,
211                 "Could not get OpenCL platform ids: %s\n", av_opencl_errstr(status));
212         ret = AVERROR_EXTERNAL;
213         goto end;
214     }
215     device_list->platform_node = av_mallocz_array(device_list->platform_num, sizeof(AVOpenCLPlatformNode *));
216     if (!device_list->platform_node) {
217         ret = AVERROR(ENOMEM);
218         goto end;
219     }
220     devices_num = av_mallocz(sizeof(int) * FF_ARRAY_ELEMS(device_type));
221     if (!devices_num) {
222         ret = AVERROR(ENOMEM);
223         goto end;
224     }
225     for (i = 0; i < device_list->platform_num; i++) {
226         device_list->platform_node[i] = av_mallocz(sizeof(AVOpenCLPlatformNode));
227         if (!device_list->platform_node[i]) {
228             ret = AVERROR(ENOMEM);
229             goto end;
230         }
231         device_list->platform_node[i]->platform_id = platform_ids[i];
232         status = clGetPlatformInfo(platform_ids[i], CL_PLATFORM_VENDOR,
233                                    0, NULL, &platform_name_size);
234         if (status != CL_SUCCESS) {
235             av_log(&opencl_ctx, AV_LOG_WARNING,
236                     "Could not get size of platform name: %s\n", av_opencl_errstr(status));
237         } else {
238             device_list->platform_node[i]->platform_name = av_malloc(platform_name_size * sizeof(char));
239             if (!device_list->platform_node[i]->platform_name) {
240                 av_log(&opencl_ctx, AV_LOG_WARNING,
241                         "Could not allocate memory for device name: %s\n", av_opencl_errstr(status));
242             } else {
243                 status = clGetPlatformInfo(platform_ids[i], CL_PLATFORM_VENDOR,
244                                            platform_name_size * sizeof(char),
245                                            device_list->platform_node[i]->platform_name, NULL);
246                 if (status != CL_SUCCESS) {
247                     av_log(&opencl_ctx, AV_LOG_WARNING,
248                             "Could not get platform name: %s\n", av_opencl_errstr(status));
249                 }
250             }
251         }
252         total_devices_num = 0;
253         for (j = 0; j < FF_ARRAY_ELEMS(device_type); j++) {
254             status = clGetDeviceIDs(device_list->platform_node[i]->platform_id,
255                                     device_type[j], 0, NULL, &devices_num[j]);
256             total_devices_num += devices_num[j];
257         }
258         device_list->platform_node[i]->device_node = av_mallocz_array(total_devices_num, sizeof(AVOpenCLDeviceNode *));
259         if (!device_list->platform_node[i]->device_node) {
260             ret = AVERROR(ENOMEM);
261             goto end;
262         }
263         for (j = 0; j < FF_ARRAY_ELEMS(device_type); j++) {
264             if (devices_num[j]) {
265                 device_ids = av_mallocz_array(devices_num[j], sizeof(cl_device_id));
266                 if (!device_ids) {
267                     ret = AVERROR(ENOMEM);
268                     goto end;
269                 }
270                 status = clGetDeviceIDs(device_list->platform_node[i]->platform_id, device_type[j],
271                                         devices_num[j], device_ids, NULL);
272                 if (status != CL_SUCCESS) {
273                     av_log(&opencl_ctx, AV_LOG_WARNING,
274                             "Could not get device ID: %s:\n", av_opencl_errstr(status));
275                     av_freep(&device_ids);
276                     continue;
277                 }
278                 for (k = 0; k < devices_num[j]; k++) {
279                     device_num = device_list->platform_node[i]->device_num;
280                     device_list->platform_node[i]->device_node[device_num] = av_mallocz(sizeof(AVOpenCLDeviceNode));
281                     if (!device_list->platform_node[i]->device_node[device_num]) {
282                         ret = AVERROR(ENOMEM);
283                         goto end;
284                     }
285                     device_node = device_list->platform_node[i]->device_node[device_num];
286                     device_node->device_id = device_ids[k];
287                     device_node->device_type = device_type[j];
288                     status = clGetDeviceInfo(device_node->device_id, CL_DEVICE_NAME,
289                                              0, NULL, &device_name_size);
290                     if (status != CL_SUCCESS) {
291                         av_log(&opencl_ctx, AV_LOG_WARNING,
292                                 "Could not get size of device name: %s\n", av_opencl_errstr(status));
293                         continue;
294                     }
295                     device_node->device_name = av_malloc(device_name_size * sizeof(char));
296                     if (!device_node->device_name) {
297                         av_log(&opencl_ctx, AV_LOG_WARNING,
298                                 "Could not allocate memory for device name: %s\n", av_opencl_errstr(status));
299                         continue;
300                     }
301                     status = clGetDeviceInfo(device_node->device_id, CL_DEVICE_NAME,
302                                              device_name_size * sizeof(char),
303                                              device_node->device_name, NULL);
304                     if (status != CL_SUCCESS) {
305                         av_log(&opencl_ctx, AV_LOG_WARNING,
306                                 "Could not get device name: %s\n", av_opencl_errstr(status));
307                         continue;
308                     }
309                     device_list->platform_node[i]->device_num++;
310                 }
311                 av_freep(&device_ids);
312             }
313         }
314     }
315 end:
316     av_freep(&platform_ids);
317     av_freep(&devices_num);
318     av_freep(&device_ids);
319     if (ret < 0)
320         free_device_list(device_list);
321     return ret;
322 }
323
324 int av_opencl_get_device_list(AVOpenCLDeviceList **device_list)
325 {
326     int ret = 0;
327     *device_list = av_mallocz(sizeof(AVOpenCLDeviceList));
328     if (!(*device_list)) {
329         av_log(&opencl_ctx, AV_LOG_ERROR, "Could not allocate opencl device list\n");
330         return AVERROR(ENOMEM);
331     }
332     ret = get_device_list(*device_list);
333     if (ret < 0) {
334         av_log(&opencl_ctx, AV_LOG_ERROR, "Could not get device list from environment\n");
335         free_device_list(*device_list);
336         av_freep(device_list);
337         return ret;
338     }
339     return ret;
340 }
341
342 void av_opencl_free_device_list(AVOpenCLDeviceList **device_list)
343 {
344     free_device_list(*device_list);
345     av_freep(device_list);
346 }
347
348 static inline int init_opencl_mtx(void)
349 {
350 #if HAVE_THREADS
351     if (!atomic_opencl_lock) {
352         int err;
353         pthread_mutex_t *tmp = av_malloc(sizeof(pthread_mutex_t));
354         if (!tmp)
355             return AVERROR(ENOMEM);
356         if ((err = pthread_mutex_init(tmp, NULL))) {
357             av_free(tmp);
358             return AVERROR(err);
359         }
360         if (avpriv_atomic_ptr_cas((void * volatile *)&atomic_opencl_lock, NULL, tmp)) {
361             pthread_mutex_destroy(tmp);
362             av_free(tmp);
363         }
364     }
365 #endif
366     return 0;
367 }
368
369 int av_opencl_set_option(const char *key, const char *val)
370 {
371     int ret = init_opencl_mtx( );
372     if (ret < 0)
373         return ret;
374     LOCK_OPENCL;
375     if (!opencl_ctx.opt_init_flag) {
376         av_opt_set_defaults(&opencl_ctx);
377         opencl_ctx.opt_init_flag = 1;
378     }
379     ret = av_opt_set(&opencl_ctx, key, val, 0);
380     UNLOCK_OPENCL;
381     return ret;
382 }
383
384 int av_opencl_get_option(const char *key, uint8_t **out_val)
385 {
386     int ret = 0;
387     LOCK_OPENCL;
388     ret = av_opt_get(&opencl_ctx, key, 0, out_val);
389     UNLOCK_OPENCL;
390     return ret;
391 }
392
393 void av_opencl_free_option(void)
394 {
395     /*FIXME: free openclutils context*/
396     LOCK_OPENCL;
397     av_opt_free(&opencl_ctx);
398     UNLOCK_OPENCL;
399 }
400
401 AVOpenCLExternalEnv *av_opencl_alloc_external_env(void)
402 {
403     AVOpenCLExternalEnv *ext = av_mallocz(sizeof(AVOpenCLExternalEnv));
404     if (!ext) {
405         av_log(&opencl_ctx, AV_LOG_ERROR,
406                "Could not malloc external opencl environment data space\n");
407     }
408     return ext;
409 }
410
411 void av_opencl_free_external_env(AVOpenCLExternalEnv **ext_opencl_env)
412 {
413     av_freep(ext_opencl_env);
414 }
415
416 int av_opencl_register_kernel_code(const char *kernel_code)
417 {
418     int i, ret = init_opencl_mtx( );
419     if (ret < 0)
420         return ret;
421     LOCK_OPENCL;
422     if (opencl_ctx.kernel_code_count >= MAX_KERNEL_CODE_NUM) {
423         av_log(&opencl_ctx, AV_LOG_ERROR,
424                "Could not register kernel code, maximum number of registered kernel code %d already reached\n",
425                MAX_KERNEL_CODE_NUM);
426         ret = AVERROR(EINVAL);
427         goto end;
428     }
429     for (i = 0; i < opencl_ctx.kernel_code_count; i++) {
430         if (opencl_ctx.kernel_code[i].kernel_string == kernel_code) {
431             av_log(&opencl_ctx, AV_LOG_WARNING, "Same kernel code has been registered\n");
432             goto end;
433         }
434     }
435     opencl_ctx.kernel_code[opencl_ctx.kernel_code_count].kernel_string = kernel_code;
436     opencl_ctx.kernel_code[opencl_ctx.kernel_code_count].is_compiled = 0;
437     opencl_ctx.kernel_code_count++;
438 end:
439     UNLOCK_OPENCL;
440     return ret;
441 }
442
443 cl_program av_opencl_compile(const char *program_name, const char *build_opts)
444 {
445     int i;
446     cl_int status, build_status;
447     int kernel_code_idx = 0;
448     const char *kernel_source = NULL;
449     size_t kernel_code_len;
450     char* ptr = NULL;
451     cl_program program = NULL;
452     size_t log_size;
453     char *log = NULL;
454
455     LOCK_OPENCL;
456     for (i = 0; i < opencl_ctx.kernel_code_count; i++) {
457         // identify a program using a unique name within the kernel source
458         ptr = av_stristr(opencl_ctx.kernel_code[i].kernel_string, program_name);
459         if (ptr && !opencl_ctx.kernel_code[i].is_compiled) {
460             kernel_source = opencl_ctx.kernel_code[i].kernel_string;
461             kernel_code_len = strlen(opencl_ctx.kernel_code[i].kernel_string);
462             kernel_code_idx = i;
463             break;
464         }
465     }
466     if (!kernel_source) {
467         av_log(&opencl_ctx, AV_LOG_ERROR,
468                "Unable to find OpenCL kernel source '%s'\n", program_name);
469         goto end;
470     }
471
472     /* create a CL program from kernel source */
473     program = clCreateProgramWithSource(opencl_ctx.context, 1, &kernel_source, &kernel_code_len, &status);
474     if(status != CL_SUCCESS) {
475         av_log(&opencl_ctx, AV_LOG_ERROR,
476                "Unable to create OpenCL program '%s': %s\n", program_name, av_opencl_errstr(status));
477         program = NULL;
478         goto end;
479     }
480
481     build_status = clBuildProgram(program, 1, &(opencl_ctx.device_id), build_opts, NULL, NULL);
482     status = clGetProgramBuildInfo(program, opencl_ctx.device_id,
483                                    CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
484     if (status != CL_SUCCESS) {
485         av_log(&opencl_ctx, AV_LOG_WARNING,
486                "Failed to get compilation log: %s\n",
487                av_opencl_errstr(status));
488     } else {
489         log = av_malloc(log_size);
490         if (log) {
491             status = clGetProgramBuildInfo(program, opencl_ctx.device_id,
492                                            CL_PROGRAM_BUILD_LOG, log_size,
493                                            log, NULL);
494             if (status != CL_SUCCESS) {
495                 av_log(&opencl_ctx, AV_LOG_WARNING,
496                        "Failed to get compilation log: %s\n",
497                        av_opencl_errstr(status));
498             } else {
499                 int level = build_status == CL_SUCCESS ? AV_LOG_DEBUG :
500                                                          AV_LOG_ERROR;
501                 av_log(&opencl_ctx, level, "Compilation log:\n%s\n", log);
502             }
503         }
504         av_freep(&log);
505     }
506     if (build_status != CL_SUCCESS) {
507         av_log(&opencl_ctx, AV_LOG_ERROR,
508                "Compilation failed with OpenCL program '%s': %s\n",
509                program_name, av_opencl_errstr(build_status));
510         program = NULL;
511         goto end;
512     }
513
514     opencl_ctx.kernel_code[kernel_code_idx].is_compiled = 1;
515 end:
516     UNLOCK_OPENCL;
517     return program;
518 }
519
520 cl_command_queue av_opencl_get_command_queue(void)
521 {
522     return opencl_ctx.command_queue;
523 }
524
525 static int init_opencl_env(OpenclContext *opencl_ctx, AVOpenCLExternalEnv *ext_opencl_env)
526 {
527     cl_int status;
528     cl_context_properties cps[3];
529     int i, ret = 0;
530     AVOpenCLDeviceNode *device_node = NULL;
531
532     if (ext_opencl_env) {
533         if (opencl_ctx->is_user_created)
534             return 0;
535         opencl_ctx->platform_id     = ext_opencl_env->platform_id;
536         opencl_ctx->is_user_created = 1;
537         opencl_ctx->command_queue   = ext_opencl_env->command_queue;
538         opencl_ctx->context         = ext_opencl_env->context;
539         opencl_ctx->device_id       = ext_opencl_env->device_id;
540         opencl_ctx->device_type     = ext_opencl_env->device_type;
541     } else {
542         if (!opencl_ctx->is_user_created) {
543             if (!opencl_ctx->device_list.platform_num) {
544                 ret = get_device_list(&opencl_ctx->device_list);
545                 if (ret < 0) {
546                     return ret;
547                 }
548             }
549             if (opencl_ctx->platform_idx >= 0) {
550                 if (opencl_ctx->device_list.platform_num < opencl_ctx->platform_idx + 1) {
551                     av_log(opencl_ctx, AV_LOG_ERROR, "User set platform index not exist\n");
552                     return AVERROR(EINVAL);
553                 }
554                 if (!opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->device_num) {
555                     av_log(opencl_ctx, AV_LOG_ERROR, "No devices in user specific platform with index %d\n",
556                            opencl_ctx->platform_idx);
557                     return AVERROR(EINVAL);
558                 }
559                 opencl_ctx->platform_id = opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->platform_id;
560             } else {
561                 /* get a usable platform by default*/
562                 for (i = 0; i < opencl_ctx->device_list.platform_num; i++) {
563                     if (opencl_ctx->device_list.platform_node[i]->device_num) {
564                         opencl_ctx->platform_id = opencl_ctx->device_list.platform_node[i]->platform_id;
565                         opencl_ctx->platform_idx = i;
566                         break;
567                     }
568                 }
569             }
570             if (!opencl_ctx->platform_id) {
571                 av_log(opencl_ctx, AV_LOG_ERROR, "Could not get OpenCL platforms\n");
572                 return AVERROR_EXTERNAL;
573             }
574             /* get a usable device*/
575             if (opencl_ctx->device_idx >= 0) {
576                 if (opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->device_num < opencl_ctx->device_idx + 1) {
577                     av_log(opencl_ctx, AV_LOG_ERROR,
578                            "Could not get OpenCL device idx %d in the user set platform\n", opencl_ctx->platform_idx);
579                     return AVERROR(EINVAL);
580                 }
581             } else {
582                 opencl_ctx->device_idx = 0;
583             }
584
585             device_node = opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->device_node[opencl_ctx->device_idx];
586             opencl_ctx->device_id = device_node->device_id;
587             opencl_ctx->device_type = device_node->device_type;
588
589             /*
590              * Use available platform.
591              */
592             av_log(opencl_ctx, AV_LOG_VERBOSE, "Platform Name: %s, Device Name: %s\n",
593                    opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->platform_name,
594                    device_node->device_name);
595             cps[0] = CL_CONTEXT_PLATFORM;
596             cps[1] = (cl_context_properties)opencl_ctx->platform_id;
597             cps[2] = 0;
598
599             opencl_ctx->context = clCreateContextFromType(cps, opencl_ctx->device_type,
600                                                        NULL, NULL, &status);
601             if (status != CL_SUCCESS) {
602                 av_log(opencl_ctx, AV_LOG_ERROR,
603                        "Could not get OpenCL context from device type: %s\n", av_opencl_errstr(status));
604                 return AVERROR_EXTERNAL;
605             }
606             opencl_ctx->command_queue = clCreateCommandQueue(opencl_ctx->context, opencl_ctx->device_id,
607                                                           0, &status);
608             if (status != CL_SUCCESS) {
609                 av_log(opencl_ctx, AV_LOG_ERROR,
610                        "Could not create OpenCL command queue: %s\n", av_opencl_errstr(status));
611                 return AVERROR_EXTERNAL;
612             }
613         }
614     }
615     return ret;
616 }
617
618 int av_opencl_init(AVOpenCLExternalEnv *ext_opencl_env)
619 {
620     int ret = init_opencl_mtx( );
621     if (ret < 0)
622         return ret;
623     LOCK_OPENCL;
624     if (!opencl_ctx.init_count) {
625         if (!opencl_ctx.opt_init_flag) {
626             av_opt_set_defaults(&opencl_ctx);
627             opencl_ctx.opt_init_flag = 1;
628         }
629         ret = init_opencl_env(&opencl_ctx, ext_opencl_env);
630         if (ret < 0)
631             goto end;
632         if (opencl_ctx.kernel_code_count <= 0) {
633             av_log(&opencl_ctx, AV_LOG_ERROR,
634                    "No kernel code is registered, compile kernel file failed\n");
635             ret = AVERROR(EINVAL);
636             goto end;
637         }
638     }
639     opencl_ctx.init_count++;
640 end:
641     UNLOCK_OPENCL;
642     return ret;
643 }
644
645 void av_opencl_uninit(void)
646 {
647     int i;
648     cl_int status;
649     LOCK_OPENCL;
650     opencl_ctx.init_count--;
651     if (opencl_ctx.is_user_created)
652         goto end;
653     if (opencl_ctx.init_count > 0)
654         goto end;
655     if (opencl_ctx.command_queue) {
656         status = clReleaseCommandQueue(opencl_ctx.command_queue);
657         if (status != CL_SUCCESS) {
658             av_log(&opencl_ctx, AV_LOG_ERROR,
659                    "Could not release OpenCL command queue: %s\n", av_opencl_errstr(status));
660         }
661         opencl_ctx.command_queue = NULL;
662     }
663     if (opencl_ctx.context) {
664         status = clReleaseContext(opencl_ctx.context);
665         if (status != CL_SUCCESS) {
666             av_log(&opencl_ctx, AV_LOG_ERROR,
667                    "Could not release OpenCL context: %s\n", av_opencl_errstr(status));
668         }
669         opencl_ctx.context = NULL;
670     }
671     for (i = 0; i < opencl_ctx.kernel_code_count; i++) {
672         opencl_ctx.kernel_code[i].is_compiled = 0;
673     }
674     free_device_list(&opencl_ctx.device_list);
675 end:
676     if (opencl_ctx.init_count <= 0)
677         av_opt_free(&opencl_ctx); //FIXME: free openclutils context
678     UNLOCK_OPENCL;
679 }
680
681 int av_opencl_buffer_create(cl_mem *cl_buf, size_t cl_buf_size, int flags, void *host_ptr)
682 {
683     cl_int status;
684     *cl_buf = clCreateBuffer(opencl_ctx.context, flags, cl_buf_size, host_ptr, &status);
685     if (status != CL_SUCCESS) {
686         av_log(&opencl_ctx, AV_LOG_ERROR, "Could not create OpenCL buffer: %s\n", av_opencl_errstr(status));
687         return AVERROR_EXTERNAL;
688     }
689     return 0;
690 }
691
692 void av_opencl_buffer_release(cl_mem *cl_buf)
693 {
694     cl_int status = 0;
695     if (!cl_buf)
696         return;
697     status = clReleaseMemObject(*cl_buf);
698     if (status != CL_SUCCESS) {
699         av_log(&opencl_ctx, AV_LOG_ERROR,
700                "Could not release OpenCL buffer: %s\n", av_opencl_errstr(status));
701     }
702     memset(cl_buf, 0, sizeof(*cl_buf));
703 }
704
705 int av_opencl_buffer_write(cl_mem dst_cl_buf, uint8_t *src_buf, size_t buf_size)
706 {
707     cl_int status;
708     void *mapped = clEnqueueMapBuffer(opencl_ctx.command_queue, dst_cl_buf,
709                                       CL_TRUE, CL_MAP_WRITE, 0, sizeof(uint8_t) * buf_size,
710                                       0, NULL, NULL, &status);
711
712     if (status != CL_SUCCESS) {
713         av_log(&opencl_ctx, AV_LOG_ERROR,
714                "Could not map OpenCL buffer: %s\n", av_opencl_errstr(status));
715         return AVERROR_EXTERNAL;
716     }
717     memcpy(mapped, src_buf, buf_size);
718
719     status = clEnqueueUnmapMemObject(opencl_ctx.command_queue, dst_cl_buf, mapped, 0, NULL, NULL);
720     if (status != CL_SUCCESS) {
721         av_log(&opencl_ctx, AV_LOG_ERROR,
722                "Could not unmap OpenCL buffer: %s\n", av_opencl_errstr(status));
723         return AVERROR_EXTERNAL;
724     }
725     return 0;
726 }
727
728 int av_opencl_buffer_read(uint8_t *dst_buf, cl_mem src_cl_buf, size_t buf_size)
729 {
730     cl_int status;
731     void *mapped = clEnqueueMapBuffer(opencl_ctx.command_queue, src_cl_buf,
732                                       CL_TRUE, CL_MAP_READ, 0, buf_size,
733                                       0, NULL, NULL, &status);
734
735     if (status != CL_SUCCESS) {
736         av_log(&opencl_ctx, AV_LOG_ERROR,
737                "Could not map OpenCL buffer: %s\n", av_opencl_errstr(status));
738         return AVERROR_EXTERNAL;
739     }
740     memcpy(dst_buf, mapped, buf_size);
741
742     status = clEnqueueUnmapMemObject(opencl_ctx.command_queue, src_cl_buf, mapped, 0, NULL, NULL);
743     if (status != CL_SUCCESS) {
744         av_log(&opencl_ctx, AV_LOG_ERROR,
745                "Could not unmap OpenCL buffer: %s\n", av_opencl_errstr(status));
746         return AVERROR_EXTERNAL;
747     }
748     return 0;
749 }
750
751 int av_opencl_buffer_write_image(cl_mem dst_cl_buf, size_t cl_buffer_size, int dst_cl_offset,
752                                  uint8_t **src_data, int *plane_size, int plane_num)
753 {
754     int i, buffer_size = 0;
755     uint8_t *temp;
756     cl_int status;
757     void *mapped;
758     if ((unsigned int)plane_num > 8) {
759         return AVERROR(EINVAL);
760     }
761     for (i = 0;i < plane_num;i++) {
762         buffer_size += plane_size[i];
763     }
764     if (buffer_size > cl_buffer_size) {
765         av_log(&opencl_ctx, AV_LOG_ERROR,
766                "Cannot write image to OpenCL buffer: buffer too small\n");
767         return AVERROR(EINVAL);
768     }
769     mapped = clEnqueueMapBuffer(opencl_ctx.command_queue, dst_cl_buf,
770                                 CL_TRUE, CL_MAP_WRITE, 0, buffer_size + dst_cl_offset,
771                                 0, NULL, NULL, &status);
772     if (status != CL_SUCCESS) {
773         av_log(&opencl_ctx, AV_LOG_ERROR,
774                "Could not map OpenCL buffer: %s\n", av_opencl_errstr(status));
775         return AVERROR_EXTERNAL;
776     }
777     temp = mapped;
778     temp += dst_cl_offset;
779     for (i = 0; i < plane_num; i++) {
780         memcpy(temp, src_data[i], plane_size[i]);
781         temp += plane_size[i];
782     }
783     status = clEnqueueUnmapMemObject(opencl_ctx.command_queue, dst_cl_buf, mapped, 0, NULL, NULL);
784     if (status != CL_SUCCESS) {
785         av_log(&opencl_ctx, AV_LOG_ERROR,
786                "Could not unmap OpenCL buffer: %s\n", av_opencl_errstr(status));
787         return AVERROR_EXTERNAL;
788     }
789     return 0;
790 }
791
792 int av_opencl_buffer_read_image(uint8_t **dst_data, int *plane_size, int plane_num,
793                                 cl_mem src_cl_buf, size_t cl_buffer_size)
794 {
795     int i,buffer_size = 0,ret = 0;
796     uint8_t *temp;
797     void *mapped;
798     cl_int status;
799     if ((unsigned int)plane_num > 8) {
800         return AVERROR(EINVAL);
801     }
802     for (i = 0; i < plane_num; i++) {
803         buffer_size += plane_size[i];
804     }
805     if (buffer_size > cl_buffer_size) {
806         av_log(&opencl_ctx, AV_LOG_ERROR,
807                "Cannot write image to CPU buffer: OpenCL buffer too small\n");
808         return AVERROR(EINVAL);
809     }
810     mapped = clEnqueueMapBuffer(opencl_ctx.command_queue, src_cl_buf,
811                                 CL_TRUE, CL_MAP_READ, 0, buffer_size,
812                                 0, NULL, NULL, &status);
813
814     if (status != CL_SUCCESS) {
815         av_log(&opencl_ctx, AV_LOG_ERROR,
816                "Could not map OpenCL buffer: %s\n", av_opencl_errstr(status));
817         return AVERROR_EXTERNAL;
818     }
819     temp = mapped;
820     if (ret >= 0) {
821         for (i = 0; i < plane_num; i++) {
822             memcpy(dst_data[i], temp, plane_size[i]);
823             temp += plane_size[i];
824         }
825     }
826     status = clEnqueueUnmapMemObject(opencl_ctx.command_queue, src_cl_buf, mapped, 0, NULL, NULL);
827     if (status != CL_SUCCESS) {
828         av_log(&opencl_ctx, AV_LOG_ERROR,
829                "Could not unmap OpenCL buffer: %s\n", av_opencl_errstr(status));
830         return AVERROR_EXTERNAL;
831     }
832     return 0;
833 }
834
835 int64_t av_opencl_benchmark(AVOpenCLDeviceNode *device_node, cl_platform_id platform,
836                             int64_t (*benchmark)(AVOpenCLExternalEnv *ext_opencl_env))
837 {
838     int64_t ret = 0;
839     cl_int status;
840     cl_context_properties cps[3];
841     AVOpenCLExternalEnv *ext_opencl_env = NULL;
842
843     ext_opencl_env = av_opencl_alloc_external_env();
844     ext_opencl_env->device_id = device_node->device_id;
845     ext_opencl_env->device_type = device_node->device_type;
846     av_log(&opencl_ctx, AV_LOG_VERBOSE, "Performing test on OpenCL device %s\n",
847            device_node->device_name);
848
849     cps[0] = CL_CONTEXT_PLATFORM;
850     cps[1] = (cl_context_properties)platform;
851     cps[2] = 0;
852     ext_opencl_env->context = clCreateContextFromType(cps, ext_opencl_env->device_type,
853                                                       NULL, NULL, &status);
854     if (status != CL_SUCCESS || !ext_opencl_env->context) {
855         ret = AVERROR_EXTERNAL;
856         goto end;
857     }
858     ext_opencl_env->command_queue = clCreateCommandQueue(ext_opencl_env->context,
859                                                          ext_opencl_env->device_id, 0, &status);
860     if (status != CL_SUCCESS || !ext_opencl_env->command_queue) {
861         ret = AVERROR_EXTERNAL;
862         goto end;
863     }
864     ret = benchmark(ext_opencl_env);
865     if (ret < 0)
866         av_log(&opencl_ctx, AV_LOG_ERROR, "Benchmark failed with OpenCL device %s\n",
867                device_node->device_name);
868 end:
869     if (ext_opencl_env->command_queue)
870         clReleaseCommandQueue(ext_opencl_env->command_queue);
871     if (ext_opencl_env->context)
872         clReleaseContext(ext_opencl_env->context);
873     av_opencl_free_external_env(&ext_opencl_env);
874     return ret;
875 }