git.sesse.net Git - nageru/blob - futatabi/vaapi_jpeg_decoder.cpp

   1 #include "vaapi_jpeg_decoder.h"
   2
   3 #include "jpeg_destroyer.h"
   4 #include "jpeg_frame.h"
   5 #include "jpeglib_error_wrapper.h"
   6 #include "shared/memcpy_interleaved.h"
   7
   8 #include <X11/Xlib.h>
   9 #include <assert.h>
  10 #include <errno.h>
  11 #include <fcntl.h>
  12 #include <glob.h>
  13 #include <jpeglib.h>
  14 #include <list>
  15 #include <mutex>
  16 #include <stdio.h>
  17 #include <stdlib.h>
  18 #include <string.h>
  19 #include <string>
  20 #include <unistd.h>
  21 #include <va/va.h>
  22 #include <va/va_drm.h>
  23 #include <va/va_x11.h>
  24
  25 using namespace std;
  26
  27 static unique_ptr<VADisplayWithCleanup> va_dpy;
  28 static VAConfigID config_id;
  29 static VAImageFormat uyvy_format;
  30 bool vaapi_jpeg_decoding_usable = false;
  31
  32 struct VAResources {
  33         unsigned width, height;
  34         VASurfaceID surface;
  35         VAContextID context;
  36         VAImage image;
  37 };
  38 static list<VAResources> va_resources_freelist;
  39 static mutex va_resources_mutex;
  40
  41 #define CHECK_VASTATUS(va_status, func) \
  42         if (va_status != VA_STATUS_SUCCESS) { \
  43                 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  44                 abort(); \
  45         }
  46
  47 #define CHECK_VASTATUS_RET(va_status, func) \
  48         if (va_status != VA_STATUS_SUCCESS) { \
  49                 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  50                 return nullptr; \
  51         }
  52
  53 // From libjpeg (although it's of course identical between implementations).
  54 static const int jpeg_natural_order[DCTSIZE2] = {
  55          0,  1,  8, 16,  9,  2,  3, 10,
  56         17, 24, 32, 25, 18, 11,  4,  5,
  57         12, 19, 26, 33, 40, 48, 41, 34,
  58         27, 20, 13,  6,  7, 14, 21, 28,
  59         35, 42, 49, 56, 57, 50, 43, 36,
  60         29, 22, 15, 23, 30, 37, 44, 51,
  61         58, 59, 52, 45, 38, 31, 39, 46,
  62         53, 60, 61, 54, 47, 55, 62, 63,
  63 };
  64
  65 VAResources get_va_resources(unsigned width, unsigned height)
  66 {
  67         {
  68                 lock_guard<mutex> lock(va_resources_mutex);
  69                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
  70                         if (it->width == width && it->height == height) {
  71                                 VAResources ret = *it;
  72                                 va_resources_freelist.erase(it);
  73                                 return ret;
  74                         }
  75                 }
  76         }
  77
  78         VAResources ret;
  79
  80         ret.width = width;
  81         ret.height = height;
  82
  83         VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
  84                                               width, height,
  85                                               &ret.surface, 1, nullptr, 0);
  86         CHECK_VASTATUS(va_status, "vaCreateSurfaces");
  87
  88         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
  89         CHECK_VASTATUS(va_status, "vaCreateContext");
  90
  91         va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
  92         CHECK_VASTATUS(va_status, "vaCreateImage");
  93
  94         return ret;
  95 }
  96
  97 void release_va_resources(VAResources resources)
  98 {
  99         lock_guard<mutex> lock(va_resources_mutex);
 100         if (va_resources_freelist.size() > 10) {
 101                 auto it = va_resources_freelist.end();
 102                 --it;
 103
 104                 VAStatus va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
 105                 CHECK_VASTATUS(va_status, "vaDestroyImage");
 106
 107                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
 108                 CHECK_VASTATUS(va_status, "vaDestroyContext");
 109
 110                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
 111                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
 112
 113                 va_resources_freelist.erase(it);
 114         }
 115
 116         va_resources_freelist.push_front(resources);
 117 }
 118
 119 // RAII wrapper to release VAResources on return (even on error).
 120 class ReleaseVAResources {
 121 public:
 122         ReleaseVAResources(const VAResources &resources)
 123                 : resources(resources) {}
 124         ~ReleaseVAResources()
 125         {
 126                 if (!committed) {
 127                         release_va_resources(resources);
 128                 }
 129         }
 130
 131         void commit() { committed = true; }
 132
 133 private:
 134         const VAResources &resources;
 135         bool committed = false;
 136 };
 137
 138 VADisplayWithCleanup::~VADisplayWithCleanup()
 139 {
 140         if (va_dpy != nullptr) {
 141                 vaTerminate(va_dpy);
 142         }
 143         if (x11_display != nullptr) {
 144                 XCloseDisplay(x11_display);
 145         }
 146         if (drm_fd != -1) {
 147                 close(drm_fd);
 148         }
 149 }
 150
 151 unique_ptr<VADisplayWithCleanup> va_open_display(const string &va_display)
 152 {
 153         if (va_display.empty() || va_display[0] != '/') {  // An X display.
 154                 Display *x11_display = XOpenDisplay(va_display.empty() ? nullptr : va_display.c_str());
 155                 if (x11_display == nullptr) {
 156                         fprintf(stderr, "error: can't connect to X server!\n");
 157                         return nullptr;
 158                 }
 159
 160                 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
 161                 ret->x11_display = x11_display;
 162                 ret->va_dpy = vaGetDisplay(x11_display);
 163                 if (ret->va_dpy == nullptr) {
 164                         return nullptr;
 165                 }
 166                 return ret;
 167         } else {  // A DRM node on the filesystem (e.g. /dev/dri/renderD128).
 168                 int drm_fd = open(va_display.c_str(), O_RDWR);
 169                 if (drm_fd == -1) {
 170                         perror(va_display.c_str());
 171                         return nullptr;
 172                 }
 173                 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
 174                 ret->drm_fd = drm_fd;
 175                 ret->va_dpy = vaGetDisplayDRM(drm_fd);
 176                 if (ret->va_dpy == nullptr) {
 177                         return nullptr;
 178                 }
 179                 return ret;
 180         }
 181 }
 182
 183 unique_ptr<VADisplayWithCleanup> try_open_va(const string &va_display, string *error)
 184 {
 185         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
 186         if (va_dpy == nullptr) {
 187                 if (error)
 188                         *error = "Opening VA display failed";
 189                 return nullptr;
 190         }
 191         int major_ver, minor_ver;
 192         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
 193         if (va_status != VA_STATUS_SUCCESS) {
 194                 char buf[256];
 195                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
 196                 if (error != nullptr)
 197                         *error = buf;
 198                 return nullptr;
 199         }
 200
 201         int num_entrypoints = vaMaxNumEntrypoints(va_dpy->va_dpy);
 202         unique_ptr<VAEntrypoint[]> entrypoints(new VAEntrypoint[num_entrypoints]);
 203         if (entrypoints == nullptr) {
 204                 if (error != nullptr)
 205                         *error = "Failed to allocate memory for VA entry points";
 206                 return nullptr;
 207         }
 208
 209         vaQueryConfigEntrypoints(va_dpy->va_dpy, VAProfileJPEGBaseline, entrypoints.get(), &num_entrypoints);
 210         for (int slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 211                 if (entrypoints[slice_entrypoint] != VAEntrypointVLD) {
 212                         continue;
 213                 }
 214
 215                 // We found a usable decode, so return it.
 216                 return va_dpy;
 217         }
 218
 219         if (error != nullptr)
 220                 *error = "Can't find VAEntrypointVLD for the JPEG profile";
 221         return nullptr;
 222 }
 223
 224 string get_usable_va_display()
 225 {
 226         // Reduce the amount of chatter while probing,
 227         // unless the user has specified otherwise.
 228         bool need_env_reset = false;
 229         if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
 230                 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
 231                 need_env_reset = true;
 232         }
 233
 234         // First try the default (ie., whatever $DISPLAY is set to).
 235         unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va("", nullptr);
 236         if (va_dpy != nullptr) {
 237                 if (need_env_reset) {
 238                         unsetenv("LIBVA_MESSAGING_LEVEL");
 239                 }
 240                 return "";
 241         }
 242
 243         fprintf(stderr, "The X11 display did not expose a VA-API JPEG decoder.\n");
 244
 245         // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
 246         glob_t g;
 247         int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
 248         if (err != 0) {
 249                 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
 250         } else {
 251                 for (size_t i = 0; i < g.gl_pathc; ++i) {
 252                         string path = g.gl_pathv[i];
 253                         va_dpy = try_open_va(path, nullptr);
 254                         if (va_dpy != nullptr) {
 255                                 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
 256                                         path.c_str());
 257                                 globfree(&g);
 258                                 if (need_env_reset) {
 259                                         unsetenv("LIBVA_MESSAGING_LEVEL");
 260                                 }
 261                                 return path;
 262                         }
 263                 }
 264         }
 265
 266         fprintf(stderr, "No suitable VA-API JPEG decoders were found in /dev/dri; giving up.\n");
 267         fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
 268         fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
 269         fprintf(stderr, "to expose Quick Sync.\n");
 270         return "none";
 271 }
 272
 273 void init_jpeg_vaapi()
 274 {
 275         string dpy = get_usable_va_display();
 276         if (dpy == "none") {
 277                 return;
 278         }
 279
 280         va_dpy = try_open_va(dpy, nullptr);
 281         if (va_dpy == nullptr) {
 282                 return;
 283         }
 284
 285         VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
 286
 287         VAStatus va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointVLD,
 288                                             &attr, 1, &config_id);
 289         CHECK_VASTATUS(va_status, "vaCreateConfig");
 290
 291         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
 292         assert(num_formats > 0);
 293
 294         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
 295         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
 296         CHECK_VASTATUS(va_status, "vaQueryImageFormats");
 297
 298         bool found = false;
 299         for (int i = 0; i < num_formats; ++i) {
 300                 // Seemingly VA_FOURCC_422H is no good for vaGetImage(). :-/
 301                 if (formats[i].fourcc == VA_FOURCC_UYVY) {
 302                         memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
 303                         found = true;
 304                         break;
 305                 }
 306         }
 307         if (!found) {
 308                 return;
 309         }
 310
 311         fprintf(stderr, "VA-API JPEG decoding initialized.\n");
 312         vaapi_jpeg_decoding_usable = true;
 313 }
 314
 315 class VABufferDestroyer {
 316 public:
 317         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 318                 : dpy(dpy), buf(buf) {}
 319
 320         ~VABufferDestroyer()
 321         {
 322                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 323                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 324         }
 325
 326 private:
 327         VADisplay dpy;
 328         VABufferID buf;
 329 };
 330
 331 shared_ptr<Frame> decode_jpeg_vaapi(const string &jpeg)
 332 {
 333         jpeg_decompress_struct dinfo;
 334         JPEGWrapErrorManager error_mgr(&dinfo);
 335         if (!error_mgr.run([&dinfo] { jpeg_create_decompress(&dinfo); })) {
 336                 return nullptr;
 337         }
 338         JPEGDestroyer destroy_dinfo(&dinfo);
 339
 340         jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
 341         if (!error_mgr.run([&dinfo] { jpeg_read_header(&dinfo, true); })) {
 342                 return nullptr;
 343         }
 344
 345         if (dinfo.num_components != 3) {
 346                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 347                         dinfo.num_components,
 348                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 349                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 350                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 351                 return nullptr;
 352         }
 353         if (dinfo.comp_info[0].h_samp_factor != 2 ||
 354             dinfo.comp_info[1].h_samp_factor != 1 ||
 355             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[0].v_samp_factor ||
 356             dinfo.comp_info[2].h_samp_factor != 1 ||
 357             dinfo.comp_info[2].v_samp_factor != dinfo.comp_info[0].v_samp_factor) {
 358                 fprintf(stderr, "Not 4:2:2. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 359                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 360                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 361                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 362                 return nullptr;
 363         }
 364
 365         // Picture parameters.
 366         VAPictureParameterBufferJPEGBaseline pic_param;
 367         memset(&pic_param, 0, sizeof(pic_param));
 368         pic_param.picture_width = dinfo.image_width;
 369         pic_param.picture_height = dinfo.image_height;
 370         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 371                 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
 372                 pic_param.components[component_idx].component_id = comp->component_id;
 373                 pic_param.components[component_idx].h_sampling_factor = comp->h_samp_factor;
 374                 pic_param.components[component_idx].v_sampling_factor = comp->v_samp_factor;
 375                 pic_param.components[component_idx].quantiser_table_selector = comp->quant_tbl_no;
 376         }
 377         pic_param.num_components = dinfo.num_components;
 378         pic_param.color_space = 0;  // YUV.
 379         pic_param.rotation = VA_ROTATION_NONE;
 380
 381         VABufferID pic_param_buffer;
 382         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAPictureParameterBufferType, sizeof(pic_param), 1, &pic_param, &pic_param_buffer);
 383         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 384         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 385
 386         // Quantization matrices.
 387         VAIQMatrixBufferJPEGBaseline iq;
 388         memset(&iq, 0, sizeof(iq));
 389
 390         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 391                 const JQUANT_TBL *qtbl = dinfo.quant_tbl_ptrs[quant_tbl_idx];
 392                 if (qtbl == nullptr) {
 393                         iq.load_quantiser_table[quant_tbl_idx] = 0;
 394                 } else {
 395                         iq.load_quantiser_table[quant_tbl_idx] = 1;
 396                         for (int i = 0; i < 64; ++i) {
 397                                 if (qtbl->quantval[i] > 255) {
 398                                         fprintf(stderr, "Baseline JPEG only!\n");
 399                                         return nullptr;
 400                                 }
 401                                 iq.quantiser_table[quant_tbl_idx][i] = qtbl->quantval[jpeg_natural_order[i]];
 402                         }
 403                 }
 404         }
 405
 406         VABufferID iq_buffer;
 407         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAIQMatrixBufferType, sizeof(iq), 1, &iq, &iq_buffer);
 408         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 409         VABufferDestroyer destroy_iq(va_dpy->va_dpy, iq_buffer);
 410
 411         // Huffman tables (arithmetic is not supported).
 412         VAHuffmanTableBufferJPEGBaseline huff;
 413         memset(&huff, 0, sizeof(huff));
 414
 415         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 416                 const JHUFF_TBL *ac_hufftbl = dinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 417                 const JHUFF_TBL *dc_hufftbl = dinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 418                 if (ac_hufftbl == nullptr) {
 419                         assert(dc_hufftbl == nullptr);
 420                         huff.load_huffman_table[huff_tbl_idx] = 0;
 421                 } else {
 422                         assert(dc_hufftbl != nullptr);
 423                         huff.load_huffman_table[huff_tbl_idx] = 1;
 424
 425                         for (int i = 0; i < 16; ++i) {
 426                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 427                         }
 428                         for (int i = 0; i < 12; ++i) {
 429                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 430                         }
 431                         for (int i = 0; i < 16; ++i) {
 432                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 433                         }
 434                         for (int i = 0; i < 162; ++i) {
 435                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 436                         }
 437                 }
 438         }
 439
 440         VABufferID huff_buffer;
 441         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(huff), 1, &huff, &huff_buffer);
 442         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 443         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 444
 445         // Slice parameters (metadata about the slice).
 446         VASliceParameterBufferJPEGBaseline parms;
 447         memset(&parms, 0, sizeof(parms));
 448         parms.slice_data_size = dinfo.src->bytes_in_buffer;
 449         parms.slice_data_offset = 0;
 450         parms.slice_data_flag = VA_SLICE_DATA_FLAG_ALL;
 451         parms.slice_horizontal_position = 0;
 452         parms.slice_vertical_position = 0;
 453         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 454                 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
 455                 parms.components[component_idx].component_selector = comp->component_id;
 456                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 457                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 458                 if (parms.components[component_idx].dc_table_selector > 1 ||
 459                     parms.components[component_idx].ac_table_selector > 1) {
 460                         fprintf(stderr, "Uses too many Huffman tables\n");
 461                         return nullptr;
 462                 }
 463         }
 464         parms.num_components = dinfo.num_components;
 465         parms.restart_interval = dinfo.restart_interval;
 466         int horiz_mcus = (dinfo.image_width + (DCTSIZE * 2) - 1) / (DCTSIZE * 2);
 467         int vert_mcus = (dinfo.image_height + DCTSIZE - 1) / DCTSIZE;
 468         parms.num_mcus = horiz_mcus * vert_mcus;
 469
 470         VABufferID slice_param_buffer;
 471         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceParameterBufferType, sizeof(parms), 1, &parms, &slice_param_buffer);
 472         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 473         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 474
 475         // The actual data. VA-API will destuff and all for us.
 476         VABufferID data_buffer;
 477         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceDataBufferType, dinfo.src->bytes_in_buffer, 1, const_cast<unsigned char *>(dinfo.src->next_input_byte), &data_buffer);
 478         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 479         VABufferDestroyer destroy_data(va_dpy->va_dpy, data_buffer);
 480
 481         VAResources resources = get_va_resources(dinfo.image_width, dinfo.image_height);
 482         ReleaseVAResources release(resources);
 483
 484         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 485         CHECK_VASTATUS_RET(va_status, "vaBeginPicture");
 486         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 487         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(pic_param)");
 488         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &iq_buffer, 1);
 489         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(iq)");
 490         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 491         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(huff)");
 492         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 493         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(slice_param)");
 494         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &data_buffer, 1);
 495         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(data)");
 496         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 497         CHECK_VASTATUS_RET(va_status, "vaEndPicture");
 498
 499         // vaDeriveImage() works, but the resulting image seems to live in
 500         // uncached memory, which makes copying data out from it very, very slow.
 501         // Thanks to FFmpeg for the observation that you can vaGetImage() the
 502         // surface onto your own image (although then, it can't be planar, which
 503         // is unfortunate for us).
 504 #if 0
 505         VAImage image;
 506         va_status = vaDeriveImage(va_dpy->va_dpy, surf, &image);
 507         CHECK_VASTATUS_RET(va_status, "vaDeriveImage");
 508 #else
 509         va_status = vaSyncSurface(va_dpy->va_dpy, resources.surface);
 510         CHECK_VASTATUS_RET(va_status, "vaSyncSurface");
 511
 512         va_status = vaGetImage(va_dpy->va_dpy, resources.surface, 0, 0, dinfo.image_width, dinfo.image_height, resources.image.image_id);
 513         CHECK_VASTATUS_RET(va_status, "vaGetImage");
 514 #endif
 515
 516         void *mapped;
 517         va_status = vaMapBuffer(va_dpy->va_dpy, resources.image.buf, &mapped);
 518         CHECK_VASTATUS_RET(va_status, "vaMapBuffer");
 519
 520         shared_ptr<Frame> frame(new Frame);
 521 #if 0
 522         // 4:2:2 planar (for vaDeriveImage).
 523         frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
 524         frame->cb.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
 525         frame->cr.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
 526         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 527                 uint8_t *dptr;
 528                 size_t width;
 529                 if (component_idx == 0) {
 530                         dptr = frame->y.get();
 531                         width = dinfo.image_width;
 532                 } else if (component_idx == 1) {
 533                         dptr = frame->cb.get();
 534                         width = dinfo.image_width / 2;
 535                 } else if (component_idx == 2) {
 536                         dptr = frame->cr.get();
 537                         width = dinfo.image_width / 2;
 538                 } else {
 539                         assert(false);
 540                 }
 541                 const uint8_t *sptr = (const uint8_t *)mapped + image.offsets[component_idx];
 542                 size_t spitch = image.pitches[component_idx];
 543                 for (size_t y = 0; y < dinfo.image_height; ++y) {
 544                         memcpy(dptr + y * width, sptr + y * spitch, width);
 545                 }
 546         }
 547 #else
 548         // Convert Y'CbCr to separate Y' and CbCr.
 549         frame->is_semiplanar = true;
 550         frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
 551         frame->cbcr.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
 552         const uint8_t *src = (const uint8_t *)mapped + resources.image.offsets[0];
 553         if (resources.image.pitches[0] == dinfo.image_width * 2) {
 554                 memcpy_interleaved(frame->cbcr.get(), frame->y.get(), src, dinfo.image_width * dinfo.image_height * 2);
 555         } else {
 556                 for (unsigned y = 0; y < dinfo.image_height; ++y) {
 557                         memcpy_interleaved(frame->cbcr.get() + y * dinfo.image_width, frame->y.get() + y * dinfo.image_width,
 558                                            src + y * resources.image.pitches[0], dinfo.image_width * 2);
 559                 }
 560         }
 561 #endif
 562         frame->width = dinfo.image_width;
 563         frame->height = dinfo.image_height;
 564         frame->chroma_subsampling_x = 2;
 565         frame->chroma_subsampling_y = 1;
 566         frame->pitch_y = dinfo.image_width;
 567         frame->pitch_chroma = dinfo.image_width / 2;
 568
 569         va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 570         CHECK_VASTATUS_RET(va_status, "vaUnmapBuffer");
 571
 572         return frame;
 573 }