git.sesse.net Git - ffmpeg/blob - 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 #if HAVE_PTHREADS
  32 #include <pthread.h>
  33 #elif HAVE_W32THREADS
  34 #include "compat/w32pthreads.h"
  35 #elif HAVE_OS2THREADS
  36 #include "compat/os2threads.h"
  37 #endif
  38 #include "atomic.h"
  39
  40 static volatile pthread_mutex_t *atomic_opencl_lock = NULL;
  41 #define LOCK_OPENCL pthread_mutex_lock(atomic_opencl_lock)
  42 #define UNLOCK_OPENCL pthread_mutex_unlock(atomic_opencl_lock)
  43 #else
  44 #define LOCK_OPENCL
  45 #define UNLOCK_OPENCL
  46 #endif
  47
  48 #define MAX_KERNEL_CODE_NUM 200
  49
  50 typedef struct {
  51     int is_compiled;
  52     const char *kernel_string;
  53 } KernelCode;
  54
  55 typedef struct {
  56     const AVClass *class;
  57     int log_offset;
  58     void *log_ctx;
  59     int init_count;
  60     int opt_init_flag;
  61      /**
  62      * if set to 1, the OpenCL environment was created by the user and
  63      * passed as AVOpenCLExternalEnv when initing ,0:created by opencl wrapper.
  64      */
  65     int is_user_created;
  66     int platform_idx;
  67     int device_idx;
  68     cl_platform_id platform_id;
  69     cl_device_type device_type;
  70     cl_context context;
  71     cl_device_id device_id;
  72     cl_command_queue command_queue;
  73     int kernel_code_count;
  74     KernelCode kernel_code[MAX_KERNEL_CODE_NUM];
  75     AVOpenCLDeviceList device_list;
  76 } OpenclContext;
  77
  78 #define OFFSET(x) offsetof(OpenclContext, x)
  79
  80 static const AVOption opencl_options[] = {
  81      { "platform_idx",        "set platform index value",  OFFSET(platform_idx),  AV_OPT_TYPE_INT,    {.i64=-1}, -1, INT_MAX},
  82      { "device_idx",          "set device index value",    OFFSET(device_idx),    AV_OPT_TYPE_INT,    {.i64=-1}, -1, INT_MAX},
  83      { NULL }
  84 };
  85
  86 static const AVClass openclutils_class = {
  87     .class_name                = "OPENCLUTILS",
  88     .option                    = opencl_options,
  89     .item_name                 = av_default_item_name,
  90     .version                   = LIBAVUTIL_VERSION_INT,
  91     .log_level_offset_offset   = offsetof(OpenclContext, log_offset),
  92     .parent_log_context_offset = offsetof(OpenclContext, log_ctx),
  93 };
  94
  95 static OpenclContext opencl_ctx = {&openclutils_class};
  96
  97 static const cl_device_type device_type[] = {CL_DEVICE_TYPE_GPU, CL_DEVICE_TYPE_CPU};
  98
  99 typedef struct {
 100     int err_code;
 101     const char *err_str;
 102 } OpenclErrorMsg;
 103
 104 static const OpenclErrorMsg opencl_err_msg[] = {
 105     {CL_DEVICE_NOT_FOUND,                               "DEVICE NOT FOUND"},
 106     {CL_DEVICE_NOT_AVAILABLE,                           "DEVICE NOT AVAILABLE"},
 107     {CL_COMPILER_NOT_AVAILABLE,                         "COMPILER NOT AVAILABLE"},
 108     {CL_MEM_OBJECT_ALLOCATION_FAILURE,                  "MEM OBJECT ALLOCATION FAILURE"},
 109     {CL_OUT_OF_RESOURCES,                               "OUT OF RESOURCES"},
 110     {CL_OUT_OF_HOST_MEMORY,                             "OUT OF HOST MEMORY"},
 111     {CL_PROFILING_INFO_NOT_AVAILABLE,                   "PROFILING INFO NOT AVAILABLE"},
 112     {CL_MEM_COPY_OVERLAP,                               "MEM COPY OVERLAP"},
 113     {CL_IMAGE_FORMAT_MISMATCH,                          "IMAGE FORMAT MISMATCH"},
 114     {CL_IMAGE_FORMAT_NOT_SUPPORTED,                     "IMAGE FORMAT NOT_SUPPORTED"},
 115     {CL_BUILD_PROGRAM_FAILURE,                          "BUILD PROGRAM FAILURE"},
 116     {CL_MAP_FAILURE,                                    "MAP FAILURE"},
 117     {CL_MISALIGNED_SUB_BUFFER_OFFSET,                   "MISALIGNED SUB BUFFER OFFSET"},
 118     {CL_EXEC_STATUS_ERROR_FOR_EVENTS_IN_WAIT_LIST,      "EXEC STATUS ERROR FOR EVENTS IN WAIT LIST"},
 119     {CL_COMPILE_PROGRAM_FAILURE,                        "COMPILE PROGRAM FAILURE"},
 120     {CL_LINKER_NOT_AVAILABLE,                           "LINKER NOT AVAILABLE"},
 121     {CL_LINK_PROGRAM_FAILURE,                           "LINK PROGRAM FAILURE"},
 122     {CL_DEVICE_PARTITION_FAILED,                        "DEVICE PARTITION FAILED"},
 123     {CL_KERNEL_ARG_INFO_NOT_AVAILABLE,                  "KERNEL ARG INFO NOT AVAILABLE"},
 124     {CL_INVALID_VALUE,                                  "INVALID VALUE"},
 125     {CL_INVALID_DEVICE_TYPE,                            "INVALID DEVICE TYPE"},
 126     {CL_INVALID_PLATFORM,                               "INVALID PLATFORM"},
 127     {CL_INVALID_DEVICE,                                 "INVALID DEVICE"},
 128     {CL_INVALID_CONTEXT,                                "INVALID CONTEXT"},
 129     {CL_INVALID_QUEUE_PROPERTIES,                       "INVALID QUEUE PROPERTIES"},
 130     {CL_INVALID_COMMAND_QUEUE,                          "INVALID COMMAND QUEUE"},
 131     {CL_INVALID_HOST_PTR,                               "INVALID HOST PTR"},
 132     {CL_INVALID_MEM_OBJECT,                             "INVALID MEM OBJECT"},
 133     {CL_INVALID_IMAGE_FORMAT_DESCRIPTOR,                "INVALID IMAGE FORMAT DESCRIPTOR"},
 134     {CL_INVALID_IMAGE_SIZE,                             "INVALID IMAGE SIZE"},
 135     {CL_INVALID_SAMPLER,                                "INVALID SAMPLER"},
 136     {CL_INVALID_BINARY,                                 "INVALID BINARY"},
 137     {CL_INVALID_BUILD_OPTIONS,                          "INVALID BUILD OPTIONS"},
 138     {CL_INVALID_PROGRAM,                                "INVALID PROGRAM"},
 139     {CL_INVALID_PROGRAM_EXECUTABLE,                     "INVALID PROGRAM EXECUTABLE"},
 140     {CL_INVALID_KERNEL_NAME,                            "INVALID KERNEL NAME"},
 141     {CL_INVALID_KERNEL_DEFINITION,                      "INVALID KERNEL DEFINITION"},
 142     {CL_INVALID_KERNEL,                                 "INVALID KERNEL"},
 143     {CL_INVALID_ARG_INDEX,                              "INVALID ARG INDEX"},
 144     {CL_INVALID_ARG_VALUE,                              "INVALID ARG VALUE"},
 145     {CL_INVALID_ARG_SIZE,                               "INVALID ARG_SIZE"},
 146     {CL_INVALID_KERNEL_ARGS,                            "INVALID KERNEL ARGS"},
 147     {CL_INVALID_WORK_DIMENSION,                         "INVALID WORK DIMENSION"},
 148     {CL_INVALID_WORK_GROUP_SIZE,                        "INVALID WORK GROUP SIZE"},
 149     {CL_INVALID_WORK_ITEM_SIZE,                         "INVALID WORK ITEM SIZE"},
 150     {CL_INVALID_GLOBAL_OFFSET,                          "INVALID GLOBAL OFFSET"},
 151     {CL_INVALID_EVENT_WAIT_LIST,                        "INVALID EVENT WAIT LIST"},
 152     {CL_INVALID_EVENT,                                  "INVALID EVENT"},
 153     {CL_INVALID_OPERATION,                              "INVALID OPERATION"},
 154     {CL_INVALID_GL_OBJECT,                              "INVALID GL OBJECT"},
 155     {CL_INVALID_BUFFER_SIZE,                            "INVALID BUFFER SIZE"},
 156     {CL_INVALID_MIP_LEVEL,                              "INVALID MIP LEVEL"},
 157     {CL_INVALID_GLOBAL_WORK_SIZE,                       "INVALID GLOBAL WORK SIZE"},
 158     {CL_INVALID_PROPERTY,                               "INVALID PROPERTY"},
 159     {CL_INVALID_IMAGE_DESCRIPTOR,                       "INVALID IMAGE DESCRIPTOR"},
 160     {CL_INVALID_COMPILER_OPTIONS,                       "INVALID COMPILER OPTIONS"},
 161     {CL_INVALID_LINKER_OPTIONS,                         "INVALID LINKER OPTIONS"},
 162     {CL_INVALID_DEVICE_PARTITION_COUNT,                 "INVALID DEVICE PARTITION COUNT"},
 163 };
 164
 165 const char *av_opencl_errstr(cl_int status)
 166 {
 167     int i;
 168     for (i = 0; i < FF_ARRAY_ELEMS(opencl_err_msg); i++) {
 169         if (opencl_err_msg[i].err_code == status)
 170             return opencl_err_msg[i].err_str;
 171     }
 172     return "unknown error";
 173 }
 174
 175 static void free_device_list(AVOpenCLDeviceList *device_list)
 176 {
 177     int i, j;
 178     if (!device_list)
 179         return;
 180     for (i = 0; i < device_list->platform_num; i++) {
 181         if (!device_list->platform_node[i])
 182             continue;
 183         for (j = 0; j < device_list->platform_node[i]->device_num; j++) {
 184             av_freep(&(device_list->platform_node[i]->device_node[j]->device_name));
 185             av_freep(&(device_list->platform_node[i]->device_node[j]));
 186         }
 187         av_freep(&device_list->platform_node[i]->device_node);
 188         av_freep(&(device_list->platform_node[i]->platform_name));
 189         av_freep(&device_list->platform_node[i]);
 190     }
 191     av_freep(&device_list->platform_node);
 192     device_list->platform_num = 0;
 193 }
 194
 195 static int get_device_list(AVOpenCLDeviceList *device_list)
 196 {
 197     cl_int status;
 198     int i, j, k, device_num, total_devices_num, ret = 0;
 199     int *devices_num;
 200     cl_platform_id *platform_ids = NULL;
 201     cl_device_id *device_ids = NULL;
 202     AVOpenCLDeviceNode *device_node = NULL;
 203     size_t platform_name_size = 0;
 204     size_t device_name_size = 0;
 205     status = clGetPlatformIDs(0, NULL, &device_list->platform_num);
 206     if (status != CL_SUCCESS) {
 207         av_log(&opencl_ctx, AV_LOG_ERROR,
 208                "Could not get OpenCL platform ids: %s\n", av_opencl_errstr(status));
 209         return AVERROR_EXTERNAL;
 210     }
 211     platform_ids = av_mallocz_array(device_list->platform_num, sizeof(cl_platform_id));
 212     if (!platform_ids)
 213         return AVERROR(ENOMEM);
 214     status = clGetPlatformIDs(device_list->platform_num, platform_ids, NULL);
 215     if (status != CL_SUCCESS) {
 216         av_log(&opencl_ctx, AV_LOG_ERROR,
 217                 "Could not get OpenCL platform ids: %s\n", av_opencl_errstr(status));
 218         ret = AVERROR_EXTERNAL;
 219         goto end;
 220     }
 221     device_list->platform_node = av_mallocz_array(device_list->platform_num, sizeof(AVOpenCLPlatformNode *));
 222     if (!device_list->platform_node) {
 223         ret = AVERROR(ENOMEM);
 224         goto end;
 225     }
 226     devices_num = av_mallocz(sizeof(int) * FF_ARRAY_ELEMS(device_type));
 227     if (!devices_num) {
 228         ret = AVERROR(ENOMEM);
 229         goto end;
 230     }
 231     for (i = 0; i < device_list->platform_num; i++) {
 232         device_list->platform_node[i] = av_mallocz(sizeof(AVOpenCLPlatformNode));
 233         if (!device_list->platform_node[i]) {
 234             ret = AVERROR(ENOMEM);
 235             goto end;
 236         }
 237         device_list->platform_node[i]->platform_id = platform_ids[i];
 238         status = clGetPlatformInfo(platform_ids[i], CL_PLATFORM_VENDOR,
 239                                    0, NULL, &platform_name_size);
 240         if (status != CL_SUCCESS) {
 241             av_log(&opencl_ctx, AV_LOG_WARNING,
 242                     "Could not get size of platform name: %s\n", av_opencl_errstr(status));
 243         } else {
 244             device_list->platform_node[i]->platform_name = av_malloc(platform_name_size * sizeof(char));
 245             if (!device_list->platform_node[i]->platform_name) {
 246                 av_log(&opencl_ctx, AV_LOG_WARNING,
 247                         "Could not allocate memory for device name: %s\n", av_opencl_errstr(status));
 248             } else {
 249                 status = clGetPlatformInfo(platform_ids[i], CL_PLATFORM_VENDOR,
 250                                            platform_name_size * sizeof(char),
 251                                            device_list->platform_node[i]->platform_name, NULL);
 252                 if (status != CL_SUCCESS) {
 253                     av_log(&opencl_ctx, AV_LOG_WARNING,
 254                             "Could not get platform name: %s\n", av_opencl_errstr(status));
 255                 }
 256             }
 257         }
 258         total_devices_num = 0;
 259         for (j = 0; j < FF_ARRAY_ELEMS(device_type); j++) {
 260             status = clGetDeviceIDs(device_list->platform_node[i]->platform_id,
 261                                     device_type[j], 0, NULL, &devices_num[j]);
 262             total_devices_num += devices_num[j];
 263         }
 264         device_list->platform_node[i]->device_node = av_mallocz_array(total_devices_num, sizeof(AVOpenCLDeviceNode *));
 265         if (!device_list->platform_node[i]->device_node) {
 266             ret = AVERROR(ENOMEM);
 267             goto end;
 268         }
 269         for (j = 0; j < FF_ARRAY_ELEMS(device_type); j++) {
 270             if (devices_num[j]) {
 271                 device_ids = av_mallocz_array(devices_num[j], sizeof(cl_device_id));
 272                 if (!device_ids) {
 273                     ret = AVERROR(ENOMEM);
 274                     goto end;
 275                 }
 276                 status = clGetDeviceIDs(device_list->platform_node[i]->platform_id, device_type[j],
 277                                         devices_num[j], device_ids, NULL);
 278                 if (status != CL_SUCCESS) {
 279                     av_log(&opencl_ctx, AV_LOG_WARNING,
 280                             "Could not get device ID: %s:\n", av_opencl_errstr(status));
 281                     av_freep(&device_ids);
 282                     continue;
 283                 }
 284                 for (k = 0; k < devices_num[j]; k++) {
 285                     device_num = device_list->platform_node[i]->device_num;
 286                     device_list->platform_node[i]->device_node[device_num] = av_mallocz(sizeof(AVOpenCLDeviceNode));
 287                     if (!device_list->platform_node[i]->device_node[device_num]) {
 288                         ret = AVERROR(ENOMEM);
 289                         goto end;
 290                     }
 291                     device_node = device_list->platform_node[i]->device_node[device_num];
 292                     device_node->device_id = device_ids[k];
 293                     device_node->device_type = device_type[j];
 294                     status = clGetDeviceInfo(device_node->device_id, CL_DEVICE_NAME,
 295                                              0, NULL, &device_name_size);
 296                     if (status != CL_SUCCESS) {
 297                         av_log(&opencl_ctx, AV_LOG_WARNING,
 298                                 "Could not get size of device name: %s\n", av_opencl_errstr(status));
 299                         continue;
 300                     }
 301                     device_node->device_name = av_malloc(device_name_size * sizeof(char));
 302                     if (!device_node->device_name) {
 303                         av_log(&opencl_ctx, AV_LOG_WARNING,
 304                                 "Could not allocate memory for device name: %s\n", av_opencl_errstr(status));
 305                         continue;
 306                     }
 307                     status = clGetDeviceInfo(device_node->device_id, CL_DEVICE_NAME,
 308                                              device_name_size * sizeof(char),
 309                                              device_node->device_name, NULL);
 310                     if (status != CL_SUCCESS) {
 311                         av_log(&opencl_ctx, AV_LOG_WARNING,
 312                                 "Could not get device name: %s\n", av_opencl_errstr(status));
 313                         continue;
 314                     }
 315                     device_list->platform_node[i]->device_num++;
 316                 }
 317                 av_freep(&device_ids);
 318             }
 319         }
 320     }
 321 end:
 322     av_freep(&platform_ids);
 323     av_freep(&devices_num);
 324     av_freep(&device_ids);
 325     if (ret < 0)
 326         free_device_list(device_list);
 327     return ret;
 328 }
 329
 330 int av_opencl_get_device_list(AVOpenCLDeviceList **device_list)
 331 {
 332     int ret = 0;
 333     *device_list = av_mallocz(sizeof(AVOpenCLDeviceList));
 334     if (!(*device_list)) {
 335         av_log(&opencl_ctx, AV_LOG_ERROR, "Could not allocate opencl device list\n");
 336         return AVERROR(ENOMEM);
 337     }
 338     ret = get_device_list(*device_list);
 339     if (ret < 0) {
 340         av_log(&opencl_ctx, AV_LOG_ERROR, "Could not get device list from environment\n");
 341         free_device_list(*device_list);
 342         av_freep(device_list);
 343         return ret;
 344     }
 345     return ret;
 346 }
 347
 348 void av_opencl_free_device_list(AVOpenCLDeviceList **device_list)
 349 {
 350     free_device_list(*device_list);
 351     av_freep(device_list);
 352 }
 353
 354 static inline int init_opencl_mtx(void)
 355 {
 356 #if HAVE_THREADS
 357     if (!atomic_opencl_lock) {
 358         int err;
 359         pthread_mutex_t *tmp = av_malloc(sizeof(pthread_mutex_t));
 360         if (!tmp)
 361             return AVERROR(ENOMEM);
 362         if ((err = pthread_mutex_init(tmp, NULL))) {
 363             av_free(tmp);
 364             return AVERROR(err);
 365         }
 366         if (avpriv_atomic_ptr_cas(&atomic_opencl_lock, NULL, tmp)) {
 367             pthread_mutex_destroy(tmp);
 368             av_free(tmp);
 369         }
 370     }
 371 #endif
 372     return 0;
 373 }
 374
 375 int av_opencl_set_option(const char *key, const char *val)
 376 {
 377     int ret = init_opencl_mtx( );
 378     if (ret < 0)
 379         return ret;
 380     LOCK_OPENCL;
 381     if (!opencl_ctx.opt_init_flag) {
 382         av_opt_set_defaults(&opencl_ctx);
 383         opencl_ctx.opt_init_flag = 1;
 384     }
 385     ret = av_opt_set(&opencl_ctx, key, val, 0);
 386     UNLOCK_OPENCL;
 387     return ret;
 388 }
 389
 390 int av_opencl_get_option(const char *key, uint8_t **out_val)
 391 {
 392     int ret = 0;
 393     LOCK_OPENCL;
 394     ret = av_opt_get(&opencl_ctx, key, 0, out_val);
 395     UNLOCK_OPENCL;
 396     return ret;
 397 }
 398
 399 void av_opencl_free_option(void)
 400 {
 401     /*FIXME: free openclutils context*/
 402     LOCK_OPENCL;
 403     av_opt_free(&opencl_ctx);
 404     UNLOCK_OPENCL;
 405 }
 406
 407 AVOpenCLExternalEnv *av_opencl_alloc_external_env(void)
 408 {
 409     AVOpenCLExternalEnv *ext = av_mallocz(sizeof(AVOpenCLExternalEnv));
 410     if (!ext) {
 411         av_log(&opencl_ctx, AV_LOG_ERROR,
 412                "Could not malloc external opencl environment data space\n");
 413     }
 414     return ext;
 415 }
 416
 417 void av_opencl_free_external_env(AVOpenCLExternalEnv **ext_opencl_env)
 418 {
 419     av_freep(ext_opencl_env);
 420 }
 421
 422 int av_opencl_register_kernel_code(const char *kernel_code)
 423 {
 424     int i, ret = init_opencl_mtx( );
 425     if (ret < 0)
 426         return ret;
 427     LOCK_OPENCL;
 428     if (opencl_ctx.kernel_code_count >= MAX_KERNEL_CODE_NUM) {
 429         av_log(&opencl_ctx, AV_LOG_ERROR,
 430                "Could not register kernel code, maximum number of registered kernel code %d already reached\n",
 431                MAX_KERNEL_CODE_NUM);
 432         ret = AVERROR(EINVAL);
 433         goto end;
 434     }
 435     for (i = 0; i < opencl_ctx.kernel_code_count; i++) {
 436         if (opencl_ctx.kernel_code[i].kernel_string == kernel_code) {
 437             av_log(&opencl_ctx, AV_LOG_WARNING, "Same kernel code has been registered\n");
 438             goto end;
 439         }
 440     }
 441     opencl_ctx.kernel_code[opencl_ctx.kernel_code_count].kernel_string = kernel_code;
 442     opencl_ctx.kernel_code[opencl_ctx.kernel_code_count].is_compiled = 0;
 443     opencl_ctx.kernel_code_count++;
 444 end:
 445     UNLOCK_OPENCL;
 446     return ret;
 447 }
 448
 449 cl_program av_opencl_compile(const char *program_name, const char *build_opts)
 450 {
 451     int i;
 452     cl_int status;
 453     int kernel_code_idx = 0;
 454     const char *kernel_source;
 455     size_t kernel_code_len;
 456     char* ptr = NULL;
 457     cl_program program = NULL;
 458
 459     LOCK_OPENCL;
 460     for (i = 0; i < opencl_ctx.kernel_code_count; i++) {
 461         // identify a program using a unique name within the kernel source
 462         ptr = av_stristr(opencl_ctx.kernel_code[i].kernel_string, program_name);
 463         if (ptr && !opencl_ctx.kernel_code[i].is_compiled) {
 464             kernel_source = opencl_ctx.kernel_code[i].kernel_string;
 465             kernel_code_len = strlen(opencl_ctx.kernel_code[i].kernel_string);
 466             kernel_code_idx = i;
 467             break;
 468         }
 469     }
 470     if (!kernel_source) {
 471         av_log(&opencl_ctx, AV_LOG_ERROR,
 472                "Unable to find OpenCL kernel source '%s'\n", program_name);
 473         goto end;
 474     }
 475
 476     /* create a CL program from kernel source */
 477     program = clCreateProgramWithSource(opencl_ctx.context, 1, &kernel_source, &kernel_code_len, &status);
 478     if(status != CL_SUCCESS) {
 479         av_log(&opencl_ctx, AV_LOG_ERROR,
 480                "Unable to create OpenCL program '%s': %s\n", program_name, av_opencl_errstr(status));
 481         program = NULL;
 482         goto end;
 483     }
 484     status = clBuildProgram(program, 1, &(opencl_ctx.device_id), build_opts, NULL, NULL);
 485     if (status != CL_SUCCESS) {
 486         av_log(&opencl_ctx, AV_LOG_ERROR,
 487                "Compilation failed with OpenCL program: %s\n", program_name);
 488         program = NULL;
 489         goto end;
 490     }
 491
 492     opencl_ctx.kernel_code[kernel_code_idx].is_compiled = 1;
 493 end:
 494     UNLOCK_OPENCL;
 495     return program;
 496 }
 497
 498 cl_command_queue av_opencl_get_command_queue(void)
 499 {
 500     return opencl_ctx.command_queue;
 501 }
 502
 503 static int init_opencl_env(OpenclContext *opencl_ctx, AVOpenCLExternalEnv *ext_opencl_env)
 504 {
 505     cl_int status;
 506     cl_context_properties cps[3];
 507     int i, ret = 0;
 508     AVOpenCLDeviceNode *device_node = NULL;
 509
 510     if (ext_opencl_env) {
 511         if (opencl_ctx->is_user_created)
 512             return 0;
 513         opencl_ctx->platform_id     = ext_opencl_env->platform_id;
 514         opencl_ctx->is_user_created = 1;
 515         opencl_ctx->command_queue   = ext_opencl_env->command_queue;
 516         opencl_ctx->context         = ext_opencl_env->context;
 517         opencl_ctx->device_id       = ext_opencl_env->device_id;
 518         opencl_ctx->device_type     = ext_opencl_env->device_type;
 519     } else {
 520         if (!opencl_ctx->is_user_created) {
 521             if (!opencl_ctx->device_list.platform_num) {
 522                 ret = get_device_list(&opencl_ctx->device_list);
 523                 if (ret < 0) {
 524                     return ret;
 525                 }
 526             }
 527             if (opencl_ctx->platform_idx >= 0) {
 528                 if (opencl_ctx->device_list.platform_num < opencl_ctx->platform_idx + 1) {
 529                     av_log(opencl_ctx, AV_LOG_ERROR, "User set platform index not exist\n");
 530                     return AVERROR(EINVAL);
 531                 }
 532                 if (!opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->device_num) {
 533                     av_log(opencl_ctx, AV_LOG_ERROR, "No devices in user specific platform with index %d\n",
 534                            opencl_ctx->platform_idx);
 535                     return AVERROR(EINVAL);
 536                 }
 537                 opencl_ctx->platform_id = opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->platform_id;
 538             } else {
 539                 /* get a usable platform by default*/
 540                 for (i = 0; i < opencl_ctx->device_list.platform_num; i++) {
 541                     if (opencl_ctx->device_list.platform_node[i]->device_num) {
 542                         opencl_ctx->platform_id = opencl_ctx->device_list.platform_node[i]->platform_id;
 543                         opencl_ctx->platform_idx = i;
 544                         break;
 545                     }
 546                 }
 547             }
 548             if (!opencl_ctx->platform_id) {
 549                 av_log(opencl_ctx, AV_LOG_ERROR, "Could not get OpenCL platforms\n");
 550                 return AVERROR_EXTERNAL;
 551             }
 552             /* get a usable device*/
 553             if (opencl_ctx->device_idx >= 0) {
 554                 if (opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->device_num < opencl_ctx->device_idx + 1) {
 555                     av_log(opencl_ctx, AV_LOG_ERROR,
 556                            "Could not get OpenCL device idx %d in the user set platform\n", opencl_ctx->platform_idx);
 557                     return AVERROR(EINVAL);
 558                 }
 559             } else {
 560                 opencl_ctx->device_idx = 0;
 561             }
 562
 563             device_node = opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->device_node[opencl_ctx->device_idx];
 564             opencl_ctx->device_id = device_node->device_id;
 565             opencl_ctx->device_type = device_node->device_type;
 566
 567             /*
 568              * Use available platform.
 569              */
 570             av_log(opencl_ctx, AV_LOG_VERBOSE, "Platform Name: %s, Device Name: %s\n",
 571                    opencl_ctx->device_list.platform_node[opencl_ctx->platform_idx]->platform_name,
 572                    device_node->device_name);
 573             cps[0] = CL_CONTEXT_PLATFORM;
 574             cps[1] = (cl_context_properties)opencl_ctx->platform_id;
 575             cps[2] = 0;
 576
 577             opencl_ctx->context = clCreateContextFromType(cps, opencl_ctx->device_type,
 578                                                        NULL, NULL, &status);
 579             if (status != CL_SUCCESS) {
 580                 av_log(opencl_ctx, AV_LOG_ERROR,
 581                        "Could not get OpenCL context from device type: %s\n", av_opencl_errstr(status));
 582                 return AVERROR_EXTERNAL;
 583             }
 584             opencl_ctx->command_queue = clCreateCommandQueue(opencl_ctx->context, opencl_ctx->device_id,
 585                                                           0, &status);
 586             if (status != CL_SUCCESS) {
 587                 av_log(opencl_ctx, AV_LOG_ERROR,
 588                        "Could not create OpenCL command queue: %s\n", av_opencl_errstr(status));
 589                 return AVERROR_EXTERNAL;
 590             }
 591         }
 592     }
 593     return ret;
 594 }
 595
 596 int av_opencl_init(AVOpenCLExternalEnv *ext_opencl_env)
 597 {
 598     int ret = init_opencl_mtx( );
 599     if (ret < 0)
 600         return ret;
 601     LOCK_OPENCL;
 602     if (!opencl_ctx.init_count) {
 603         if (!opencl_ctx.opt_init_flag) {
 604             av_opt_set_defaults(&opencl_ctx);
 605             opencl_ctx.opt_init_flag = 1;
 606         }
 607         ret = init_opencl_env(&opencl_ctx, ext_opencl_env);
 608         if (ret < 0)
 609             goto end;
 610         if (opencl_ctx.kernel_code_count <= 0) {
 611             av_log(&opencl_ctx, AV_LOG_ERROR,
 612                    "No kernel code is registered, compile kernel file failed\n");
 613             ret = AVERROR(EINVAL);
 614             goto end;
 615         }
 616     }
 617     opencl_ctx.init_count++;
 618 end:
 619     UNLOCK_OPENCL;
 620     return ret;
 621 }
 622
 623 void av_opencl_uninit(void)
 624 {
 625     int i;
 626     cl_int status;
 627     LOCK_OPENCL;
 628     opencl_ctx.init_count--;
 629     if (opencl_ctx.is_user_created)
 630         goto end;
 631     if (opencl_ctx.init_count > 0)
 632         goto end;
 633     if (opencl_ctx.command_queue) {
 634         status = clReleaseCommandQueue(opencl_ctx.command_queue);
 635         if (status != CL_SUCCESS) {
 636             av_log(&opencl_ctx, AV_LOG_ERROR,
 637                    "Could not release OpenCL command queue: %s\n", av_opencl_errstr(status));
 638         }
 639         opencl_ctx.command_queue = NULL;
 640     }
 641     if (opencl_ctx.context) {
 642         status = clReleaseContext(opencl_ctx.context);
 643         if (status != CL_SUCCESS) {
 644             av_log(&opencl_ctx, AV_LOG_ERROR,
 645                    "Could not release OpenCL context: %s\n", av_opencl_errstr(status));
 646         }
 647         opencl_ctx.context = NULL;
 648     }
 649     for (i = 0; i < opencl_ctx.kernel_code_count; i++) {
 650         opencl_ctx.kernel_code[i].is_compiled = 0;
 651     }
 652     free_device_list(&opencl_ctx.device_list);
 653 end:
 654     if (opencl_ctx.init_count <= 0)
 655         av_opt_free(&opencl_ctx); //FIXME: free openclutils context
 656     UNLOCK_OPENCL;
 657 }
 658
 659 int av_opencl_buffer_create(cl_mem *cl_buf, size_t cl_buf_size, int flags, void *host_ptr)
 660 {
 661     cl_int status;
 662     *cl_buf = clCreateBuffer(opencl_ctx.context, flags, cl_buf_size, host_ptr, &status);
 663     if (status != CL_SUCCESS) {
 664         av_log(&opencl_ctx, AV_LOG_ERROR, "Could not create OpenCL buffer: %s\n", av_opencl_errstr(status));
 665         return AVERROR_EXTERNAL;
 666     }
 667     return 0;
 668 }
 669
 670 void av_opencl_buffer_release(cl_mem *cl_buf)
 671 {
 672     cl_int status = 0;
 673     if (!cl_buf)
 674         return;
 675     status = clReleaseMemObject(*cl_buf);
 676     if (status != CL_SUCCESS) {
 677         av_log(&opencl_ctx, AV_LOG_ERROR,
 678                "Could not release OpenCL buffer: %s\n", av_opencl_errstr(status));
 679     }
 680     memset(cl_buf, 0, sizeof(*cl_buf));
 681 }
 682
 683 int av_opencl_buffer_write(cl_mem dst_cl_buf, uint8_t *src_buf, size_t buf_size)
 684 {
 685     cl_int status;
 686     void *mapped = clEnqueueMapBuffer(opencl_ctx.command_queue, dst_cl_buf,
 687                                       CL_TRUE, CL_MAP_WRITE, 0, sizeof(uint8_t) * buf_size,
 688                                       0, NULL, NULL, &status);
 689
 690     if (status != CL_SUCCESS) {
 691         av_log(&opencl_ctx, AV_LOG_ERROR,
 692                "Could not map OpenCL buffer: %s\n", av_opencl_errstr(status));
 693         return AVERROR_EXTERNAL;
 694     }
 695     memcpy(mapped, src_buf, buf_size);
 696
 697     status = clEnqueueUnmapMemObject(opencl_ctx.command_queue, dst_cl_buf, mapped, 0, NULL, NULL);
 698     if (status != CL_SUCCESS) {
 699         av_log(&opencl_ctx, AV_LOG_ERROR,
 700                "Could not unmap OpenCL buffer: %s\n", av_opencl_errstr(status));
 701         return AVERROR_EXTERNAL;
 702     }
 703     return 0;
 704 }
 705
 706 int av_opencl_buffer_read(uint8_t *dst_buf, cl_mem src_cl_buf, size_t buf_size)
 707 {
 708     cl_int status;
 709     void *mapped = clEnqueueMapBuffer(opencl_ctx.command_queue, src_cl_buf,
 710                                       CL_TRUE, CL_MAP_READ, 0, buf_size,
 711                                       0, NULL, NULL, &status);
 712
 713     if (status != CL_SUCCESS) {
 714         av_log(&opencl_ctx, AV_LOG_ERROR,
 715                "Could not map OpenCL buffer: %s\n", av_opencl_errstr(status));
 716         return AVERROR_EXTERNAL;
 717     }
 718     memcpy(dst_buf, mapped, buf_size);
 719
 720     status = clEnqueueUnmapMemObject(opencl_ctx.command_queue, src_cl_buf, mapped, 0, NULL, NULL);
 721     if (status != CL_SUCCESS) {
 722         av_log(&opencl_ctx, AV_LOG_ERROR,
 723                "Could not unmap OpenCL buffer: %s\n", av_opencl_errstr(status));
 724         return AVERROR_EXTERNAL;
 725     }
 726     return 0;
 727 }
 728
 729 int av_opencl_buffer_write_image(cl_mem dst_cl_buf, size_t cl_buffer_size, int dst_cl_offset,
 730                                  uint8_t **src_data, int *plane_size, int plane_num)
 731 {
 732     int i, buffer_size = 0;
 733     uint8_t *temp;
 734     cl_int status;
 735     void *mapped;
 736     if ((unsigned int)plane_num > 8) {
 737         return AVERROR(EINVAL);
 738     }
 739     for (i = 0;i < plane_num;i++) {
 740         buffer_size += plane_size[i];
 741     }
 742     if (buffer_size > cl_buffer_size) {
 743         av_log(&opencl_ctx, AV_LOG_ERROR,
 744                "Cannot write image to OpenCL buffer: buffer too small\n");
 745         return AVERROR(EINVAL);
 746     }
 747     mapped = clEnqueueMapBuffer(opencl_ctx.command_queue, dst_cl_buf,
 748                                 CL_TRUE, CL_MAP_WRITE, 0, buffer_size + dst_cl_offset,
 749                                 0, NULL, NULL, &status);
 750     if (status != CL_SUCCESS) {
 751         av_log(&opencl_ctx, AV_LOG_ERROR,
 752                "Could not map OpenCL buffer: %s\n", av_opencl_errstr(status));
 753         return AVERROR_EXTERNAL;
 754     }
 755     temp = mapped;
 756     temp += dst_cl_offset;
 757     for (i = 0; i < plane_num; i++) {
 758         memcpy(temp, src_data[i], plane_size[i]);
 759         temp += plane_size[i];
 760     }
 761     status = clEnqueueUnmapMemObject(opencl_ctx.command_queue, dst_cl_buf, mapped, 0, NULL, NULL);
 762     if (status != CL_SUCCESS) {
 763         av_log(&opencl_ctx, AV_LOG_ERROR,
 764                "Could not unmap OpenCL buffer: %s\n", av_opencl_errstr(status));
 765         return AVERROR_EXTERNAL;
 766     }
 767     return 0;
 768 }
 769
 770 int av_opencl_buffer_read_image(uint8_t **dst_data, int *plane_size, int plane_num,
 771                                 cl_mem src_cl_buf, size_t cl_buffer_size)
 772 {
 773     int i,buffer_size = 0,ret = 0;
 774     uint8_t *temp;
 775     void *mapped;
 776     cl_int status;
 777     if ((unsigned int)plane_num > 8) {
 778         return AVERROR(EINVAL);
 779     }
 780     for (i = 0; i < plane_num; i++) {
 781         buffer_size += plane_size[i];
 782     }
 783     if (buffer_size > cl_buffer_size) {
 784         av_log(&opencl_ctx, AV_LOG_ERROR,
 785                "Cannot write image to CPU buffer: OpenCL buffer too small\n");
 786         return AVERROR(EINVAL);
 787     }
 788     mapped = clEnqueueMapBuffer(opencl_ctx.command_queue, src_cl_buf,
 789                                 CL_TRUE, CL_MAP_READ, 0, buffer_size,
 790                                 0, NULL, NULL, &status);
 791
 792     if (status != CL_SUCCESS) {
 793         av_log(&opencl_ctx, AV_LOG_ERROR,
 794                "Could not map OpenCL buffer: %s\n", av_opencl_errstr(status));
 795         return AVERROR_EXTERNAL;
 796     }
 797     temp = mapped;
 798     if (ret >= 0) {
 799         for (i = 0; i < plane_num; i++) {
 800             memcpy(dst_data[i], temp, plane_size[i]);
 801             temp += plane_size[i];
 802         }
 803     }
 804     status = clEnqueueUnmapMemObject(opencl_ctx.command_queue, src_cl_buf, mapped, 0, NULL, NULL);
 805     if (status != CL_SUCCESS) {
 806         av_log(&opencl_ctx, AV_LOG_ERROR,
 807                "Could not unmap OpenCL buffer: %s\n", av_opencl_errstr(status));
 808         return AVERROR_EXTERNAL;
 809     }
 810     return 0;
 811 }
 812
 813 int64_t av_opencl_benchmark(AVOpenCLDeviceNode *device_node, cl_platform_id platform,
 814                             int64_t (*benchmark)(AVOpenCLExternalEnv *ext_opencl_env))
 815 {
 816     int64_t ret = 0;
 817     cl_int status;
 818     cl_context_properties cps[3];
 819     AVOpenCLExternalEnv *ext_opencl_env = NULL;
 820
 821     ext_opencl_env = av_opencl_alloc_external_env();
 822     ext_opencl_env->device_id = device_node->device_id;
 823     ext_opencl_env->device_type = device_node->device_type;
 824     av_log(&opencl_ctx, AV_LOG_VERBOSE, "Performing test on OpenCL device %s\n",
 825            device_node->device_name);
 826
 827     cps[0] = CL_CONTEXT_PLATFORM;
 828     cps[1] = (cl_context_properties)platform;
 829     cps[2] = 0;
 830     ext_opencl_env->context = clCreateContextFromType(cps, ext_opencl_env->device_type,
 831                                                       NULL, NULL, &status);
 832     if (status != CL_SUCCESS || !ext_opencl_env->context) {
 833         ret = AVERROR_EXTERNAL;
 834         goto end;
 835     }
 836     ext_opencl_env->command_queue = clCreateCommandQueue(ext_opencl_env->context,
 837                                                          ext_opencl_env->device_id, 0, &status);
 838     if (status != CL_SUCCESS || !ext_opencl_env->command_queue) {
 839         ret = AVERROR_EXTERNAL;
 840         goto end;
 841     }
 842     ret = benchmark(ext_opencl_env);
 843     if (ret < 0)
 844         av_log(&opencl_ctx, AV_LOG_ERROR, "Benchmark failed with OpenCL device %s\n",
 845                device_node->device_name);
 846 end:
 847     if (ext_opencl_env->command_queue)
 848         clReleaseCommandQueue(ext_opencl_env->command_queue);
 849     if (ext_opencl_env->context)
 850         clReleaseContext(ext_opencl_env->context);
 851     av_opencl_free_external_env(&ext_opencl_env);
 852     return ret;
 853 }