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 "pbo_pool.h"
   7 #include "shared/memcpy_interleaved.h"
   8
   9 #include <X11/Xlib.h>
  10 #include <assert.h>
  11 #include <errno.h>
  12 #include <fcntl.h>
  13 #include <glob.h>
  14 #include <jpeglib.h>
  15 #include <list>
  16 #include <mutex>
  17 #include <stdio.h>
  18 #include <stdlib.h>
  19 #include <string.h>
  20 #include <string>
  21 #include <unistd.h>
  22 #include <va/va.h>
  23 #include <va/va_drm.h>
  24 #include <va/va_x11.h>
  25
  26 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  27
  28 using namespace std;
  29
  30 static unique_ptr<VADisplayWithCleanup> va_dpy;
  31 static VAConfigID config_id;
  32 static VAImageFormat uyvy_format;
  33 bool vaapi_jpeg_decoding_usable = false;
  34
  35 struct VAResources {
  36         unsigned width, height;
  37         VASurfaceID surface;
  38         VAContextID context;
  39         VAImage image;
  40 };
  41 static list<VAResources> va_resources_freelist;
  42 static mutex va_resources_mutex;
  43
  44 #define CHECK_VASTATUS(va_status, func) \
  45         if (va_status != VA_STATUS_SUCCESS) { \
  46                 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  47                 abort(); \
  48         }
  49
  50 #define CHECK_VASTATUS_RET(va_status, func) \
  51         if (va_status != VA_STATUS_SUCCESS) { \
  52                 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
  53                 return nullptr; \
  54         }
  55
  56 // From libjpeg (although it's of course identical between implementations).
  57 static const int jpeg_natural_order[DCTSIZE2] = {
  58          0,  1,  8, 16,  9,  2,  3, 10,
  59         17, 24, 32, 25, 18, 11,  4,  5,
  60         12, 19, 26, 33, 40, 48, 41, 34,
  61         27, 20, 13,  6,  7, 14, 21, 28,
  62         35, 42, 49, 56, 57, 50, 43, 36,
  63         29, 22, 15, 23, 30, 37, 44, 51,
  64         58, 59, 52, 45, 38, 31, 39, 46,
  65         53, 60, 61, 54, 47, 55, 62, 63,
  66 };
  67
  68 VAResources get_va_resources(unsigned width, unsigned height)
  69 {
  70         {
  71                 lock_guard<mutex> lock(va_resources_mutex);
  72                 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
  73                         if (it->width == width && it->height == height) {
  74                                 VAResources ret = *it;
  75                                 va_resources_freelist.erase(it);
  76                                 return ret;
  77                         }
  78                 }
  79         }
  80
  81         VAResources ret;
  82
  83         ret.width = width;
  84         ret.height = height;
  85
  86         VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
  87                                               width, height,
  88                                               &ret.surface, 1, nullptr, 0);
  89         CHECK_VASTATUS(va_status, "vaCreateSurfaces");
  90
  91         va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
  92         CHECK_VASTATUS(va_status, "vaCreateContext");
  93
  94         va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
  95         CHECK_VASTATUS(va_status, "vaCreateImage");
  96
  97         return ret;
  98 }
  99
 100 void release_va_resources(VAResources resources)
 101 {
 102         lock_guard<mutex> lock(va_resources_mutex);
 103         if (va_resources_freelist.size() > 10) {
 104                 auto it = va_resources_freelist.end();
 105                 --it;
 106
 107                 VAStatus va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
 108                 CHECK_VASTATUS(va_status, "vaDestroyImage");
 109
 110                 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
 111                 CHECK_VASTATUS(va_status, "vaDestroyContext");
 112
 113                 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
 114                 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
 115
 116                 va_resources_freelist.erase(it);
 117         }
 118
 119         va_resources_freelist.push_front(resources);
 120 }
 121
 122 // RAII wrapper to release VAResources on return (even on error).
 123 class ReleaseVAResources {
 124 public:
 125         ReleaseVAResources(const VAResources &resources)
 126                 : resources(resources) {}
 127         ~ReleaseVAResources()
 128         {
 129                 if (!committed) {
 130                         release_va_resources(resources);
 131                 }
 132         }
 133
 134         void commit() { committed = true; }
 135
 136 private:
 137         const VAResources &resources;
 138         bool committed = false;
 139 };
 140
 141 VADisplayWithCleanup::~VADisplayWithCleanup()
 142 {
 143         if (va_dpy != nullptr) {
 144                 vaTerminate(va_dpy);
 145         }
 146         if (x11_display != nullptr) {
 147                 XCloseDisplay(x11_display);
 148         }
 149         if (drm_fd != -1) {
 150                 close(drm_fd);
 151         }
 152 }
 153
 154 unique_ptr<VADisplayWithCleanup> va_open_display(const string &va_display)
 155 {
 156         if (va_display.empty() || va_display[0] != '/') {  // An X display.
 157                 Display *x11_display = XOpenDisplay(va_display.empty() ? nullptr : va_display.c_str());
 158                 if (x11_display == nullptr) {
 159                         fprintf(stderr, "error: can't connect to X server!\n");
 160                         return nullptr;
 161                 }
 162
 163                 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
 164                 ret->x11_display = x11_display;
 165                 ret->va_dpy = vaGetDisplay(x11_display);
 166                 if (ret->va_dpy == nullptr) {
 167                         return nullptr;
 168                 }
 169                 return ret;
 170         } else {  // A DRM node on the filesystem (e.g. /dev/dri/renderD128).
 171                 int drm_fd = open(va_display.c_str(), O_RDWR);
 172                 if (drm_fd == -1) {
 173                         perror(va_display.c_str());
 174                         return nullptr;
 175                 }
 176                 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
 177                 ret->drm_fd = drm_fd;
 178                 ret->va_dpy = vaGetDisplayDRM(drm_fd);
 179                 if (ret->va_dpy == nullptr) {
 180                         return nullptr;
 181                 }
 182                 return ret;
 183         }
 184 }
 185
 186 unique_ptr<VADisplayWithCleanup> try_open_va(const string &va_display, string *error)
 187 {
 188         unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
 189         if (va_dpy == nullptr) {
 190                 if (error)
 191                         *error = "Opening VA display failed";
 192                 return nullptr;
 193         }
 194         int major_ver, minor_ver;
 195         VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
 196         if (va_status != VA_STATUS_SUCCESS) {
 197                 char buf[256];
 198                 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
 199                 if (error != nullptr)
 200                         *error = buf;
 201                 return nullptr;
 202         }
 203
 204         int num_entrypoints = vaMaxNumEntrypoints(va_dpy->va_dpy);
 205         unique_ptr<VAEntrypoint[]> entrypoints(new VAEntrypoint[num_entrypoints]);
 206         if (entrypoints == nullptr) {
 207                 if (error != nullptr)
 208                         *error = "Failed to allocate memory for VA entry points";
 209                 return nullptr;
 210         }
 211
 212         vaQueryConfigEntrypoints(va_dpy->va_dpy, VAProfileJPEGBaseline, entrypoints.get(), &num_entrypoints);
 213         for (int slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
 214                 if (entrypoints[slice_entrypoint] != VAEntrypointVLD) {
 215                         continue;
 216                 }
 217
 218                 // We found a usable decode, so return it.
 219                 return va_dpy;
 220         }
 221
 222         if (error != nullptr)
 223                 *error = "Can't find VAEntrypointVLD for the JPEG profile";
 224         return nullptr;
 225 }
 226
 227 string get_usable_va_display()
 228 {
 229         // Reduce the amount of chatter while probing,
 230         // unless the user has specified otherwise.
 231         bool need_env_reset = false;
 232         if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
 233                 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
 234                 need_env_reset = true;
 235         }
 236
 237         // First try the default (ie., whatever $DISPLAY is set to).
 238         unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va("", nullptr);
 239         if (va_dpy != nullptr) {
 240                 if (need_env_reset) {
 241                         unsetenv("LIBVA_MESSAGING_LEVEL");
 242                 }
 243                 return "";
 244         }
 245
 246         fprintf(stderr, "The X11 display did not expose a VA-API JPEG decoder.\n");
 247
 248         // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
 249         glob_t g;
 250         int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
 251         if (err != 0) {
 252                 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
 253         } else {
 254                 for (size_t i = 0; i < g.gl_pathc; ++i) {
 255                         string path = g.gl_pathv[i];
 256                         va_dpy = try_open_va(path, nullptr);
 257                         if (va_dpy != nullptr) {
 258                                 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
 259                                         path.c_str());
 260                                 globfree(&g);
 261                                 if (need_env_reset) {
 262                                         unsetenv("LIBVA_MESSAGING_LEVEL");
 263                                 }
 264                                 return path;
 265                         }
 266                 }
 267         }
 268
 269         fprintf(stderr, "No suitable VA-API JPEG decoders were found in /dev/dri; giving up.\n");
 270         fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
 271         fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
 272         fprintf(stderr, "to expose Quick Sync.\n");
 273         return "none";
 274 }
 275
 276 void init_jpeg_vaapi()
 277 {
 278         string dpy = get_usable_va_display();
 279         if (dpy == "none") {
 280                 return;
 281         }
 282
 283         va_dpy = try_open_va(dpy, nullptr);
 284         if (va_dpy == nullptr) {
 285                 return;
 286         }
 287
 288         VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
 289
 290         VAStatus va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointVLD,
 291                                             &attr, 1, &config_id);
 292         CHECK_VASTATUS(va_status, "vaCreateConfig");
 293
 294         int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
 295         assert(num_formats > 0);
 296
 297         unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
 298         va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
 299         CHECK_VASTATUS(va_status, "vaQueryImageFormats");
 300
 301         bool found = false;
 302         for (int i = 0; i < num_formats; ++i) {
 303                 // Seemingly VA_FOURCC_422H is no good for vaGetImage(). :-/
 304                 if (formats[i].fourcc == VA_FOURCC_UYVY) {
 305                         memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
 306                         found = true;
 307                         break;
 308                 }
 309         }
 310         if (!found) {
 311                 return;
 312         }
 313
 314         fprintf(stderr, "VA-API JPEG decoding initialized.\n");
 315         vaapi_jpeg_decoding_usable = true;
 316 }
 317
 318 class VABufferDestroyer {
 319 public:
 320         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 321                 : dpy(dpy), buf(buf) {}
 322
 323         ~VABufferDestroyer()
 324         {
 325                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 326                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 327         }
 328
 329 private:
 330         VADisplay dpy;
 331         VABufferID buf;
 332 };
 333
 334 shared_ptr<Frame> decode_jpeg_vaapi(const string &jpeg)
 335 {
 336         jpeg_decompress_struct dinfo;
 337         JPEGWrapErrorManager error_mgr(&dinfo);
 338         if (!error_mgr.run([&dinfo] { jpeg_create_decompress(&dinfo); })) {
 339                 return nullptr;
 340         }
 341         JPEGDestroyer destroy_dinfo(&dinfo);
 342
 343         jpeg_save_markers(&dinfo, JPEG_APP0 + 1, 0xFFFF);
 344
 345         jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
 346         if (!error_mgr.run([&dinfo] { jpeg_read_header(&dinfo, true); })) {
 347                 return nullptr;
 348         }
 349
 350         if (dinfo.num_components != 3) {
 351                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 352                         dinfo.num_components,
 353                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 354                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 355                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 356                 return nullptr;
 357         }
 358         if (dinfo.comp_info[0].h_samp_factor != 2 ||
 359             dinfo.comp_info[1].h_samp_factor != 1 ||
 360             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[0].v_samp_factor ||
 361             dinfo.comp_info[2].h_samp_factor != 1 ||
 362             dinfo.comp_info[2].v_samp_factor != dinfo.comp_info[0].v_samp_factor) {
 363                 fprintf(stderr, "Not 4:2:2. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 364                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 365                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 366                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 367                 return nullptr;
 368         }
 369
 370         // Picture parameters.
 371         VAPictureParameterBufferJPEGBaseline pic_param;
 372         memset(&pic_param, 0, sizeof(pic_param));
 373         pic_param.picture_width = dinfo.image_width;
 374         pic_param.picture_height = dinfo.image_height;
 375         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 376                 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
 377                 pic_param.components[component_idx].component_id = comp->component_id;
 378                 pic_param.components[component_idx].h_sampling_factor = comp->h_samp_factor;
 379                 pic_param.components[component_idx].v_sampling_factor = comp->v_samp_factor;
 380                 pic_param.components[component_idx].quantiser_table_selector = comp->quant_tbl_no;
 381         }
 382         pic_param.num_components = dinfo.num_components;
 383         pic_param.color_space = 0;  // YUV.
 384         pic_param.rotation = VA_ROTATION_NONE;
 385
 386         VABufferID pic_param_buffer;
 387         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAPictureParameterBufferType, sizeof(pic_param), 1, &pic_param, &pic_param_buffer);
 388         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 389         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 390
 391         // Quantization matrices.
 392         VAIQMatrixBufferJPEGBaseline iq;
 393         memset(&iq, 0, sizeof(iq));
 394
 395         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 396                 const JQUANT_TBL *qtbl = dinfo.quant_tbl_ptrs[quant_tbl_idx];
 397                 if (qtbl == nullptr) {
 398                         iq.load_quantiser_table[quant_tbl_idx] = 0;
 399                 } else {
 400                         iq.load_quantiser_table[quant_tbl_idx] = 1;
 401                         for (int i = 0; i < 64; ++i) {
 402                                 if (qtbl->quantval[i] > 255) {
 403                                         fprintf(stderr, "Baseline JPEG only!\n");
 404                                         return nullptr;
 405                                 }
 406                                 iq.quantiser_table[quant_tbl_idx][i] = qtbl->quantval[jpeg_natural_order[i]];
 407                         }
 408                 }
 409         }
 410
 411         VABufferID iq_buffer;
 412         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAIQMatrixBufferType, sizeof(iq), 1, &iq, &iq_buffer);
 413         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 414         VABufferDestroyer destroy_iq(va_dpy->va_dpy, iq_buffer);
 415
 416         // Huffman tables (arithmetic is not supported).
 417         VAHuffmanTableBufferJPEGBaseline huff;
 418         memset(&huff, 0, sizeof(huff));
 419
 420         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 421                 const JHUFF_TBL *ac_hufftbl = dinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 422                 const JHUFF_TBL *dc_hufftbl = dinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 423                 if (ac_hufftbl == nullptr) {
 424                         assert(dc_hufftbl == nullptr);
 425                         huff.load_huffman_table[huff_tbl_idx] = 0;
 426                 } else {
 427                         assert(dc_hufftbl != nullptr);
 428                         huff.load_huffman_table[huff_tbl_idx] = 1;
 429
 430                         for (int i = 0; i < 16; ++i) {
 431                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 432                         }
 433                         for (int i = 0; i < 12; ++i) {
 434                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 435                         }
 436                         for (int i = 0; i < 16; ++i) {
 437                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 438                         }
 439                         for (int i = 0; i < 162; ++i) {
 440                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 441                         }
 442                 }
 443         }
 444
 445         VABufferID huff_buffer;
 446         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(huff), 1, &huff, &huff_buffer);
 447         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 448         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 449
 450         // Slice parameters (metadata about the slice).
 451         VASliceParameterBufferJPEGBaseline parms;
 452         memset(&parms, 0, sizeof(parms));
 453         parms.slice_data_size = dinfo.src->bytes_in_buffer;
 454         parms.slice_data_offset = 0;
 455         parms.slice_data_flag = VA_SLICE_DATA_FLAG_ALL;
 456         parms.slice_horizontal_position = 0;
 457         parms.slice_vertical_position = 0;
 458         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 459                 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
 460                 parms.components[component_idx].component_selector = comp->component_id;
 461                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 462                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 463                 if (parms.components[component_idx].dc_table_selector > 1 ||
 464                     parms.components[component_idx].ac_table_selector > 1) {
 465                         fprintf(stderr, "Uses too many Huffman tables\n");
 466                         return nullptr;
 467                 }
 468         }
 469         parms.num_components = dinfo.num_components;
 470         parms.restart_interval = dinfo.restart_interval;
 471         int horiz_mcus = (dinfo.image_width + (DCTSIZE * 2) - 1) / (DCTSIZE * 2);
 472         int vert_mcus = (dinfo.image_height + DCTSIZE - 1) / DCTSIZE;
 473         parms.num_mcus = horiz_mcus * vert_mcus;
 474
 475         VABufferID slice_param_buffer;
 476         va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceParameterBufferType, sizeof(parms), 1, &parms, &slice_param_buffer);
 477         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 478         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 479
 480         // The actual data. VA-API will destuff and all for us.
 481         VABufferID data_buffer;
 482         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);
 483         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 484         VABufferDestroyer destroy_data(va_dpy->va_dpy, data_buffer);
 485
 486         VAResources resources = get_va_resources(dinfo.image_width, dinfo.image_height);
 487         ReleaseVAResources release(resources);
 488
 489         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 490         CHECK_VASTATUS_RET(va_status, "vaBeginPicture");
 491         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 492         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(pic_param)");
 493         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &iq_buffer, 1);
 494         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(iq)");
 495         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 496         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(huff)");
 497         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 498         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(slice_param)");
 499         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &data_buffer, 1);
 500         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(data)");
 501         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 502         CHECK_VASTATUS_RET(va_status, "vaEndPicture");
 503
 504         // vaDeriveImage() works, but the resulting image seems to live in
 505         // uncached memory, which makes copying data out from it very, very slow.
 506         // Thanks to FFmpeg for the observation that you can vaGetImage() the
 507         // surface onto your own image (although then, it can't be planar, which
 508         // is unfortunate for us).
 509 #if 0
 510         VAImage image;
 511         va_status = vaDeriveImage(va_dpy->va_dpy, surf, &image);
 512         CHECK_VASTATUS_RET(va_status, "vaDeriveImage");
 513 #else
 514         va_status = vaSyncSurface(va_dpy->va_dpy, resources.surface);
 515         CHECK_VASTATUS_RET(va_status, "vaSyncSurface");
 516
 517         va_status = vaGetImage(va_dpy->va_dpy, resources.surface, 0, 0, dinfo.image_width, dinfo.image_height, resources.image.image_id);
 518         CHECK_VASTATUS_RET(va_status, "vaGetImage");
 519 #endif
 520
 521         void *mapped;
 522         va_status = vaMapBuffer(va_dpy->va_dpy, resources.image.buf, &mapped);
 523         CHECK_VASTATUS_RET(va_status, "vaMapBuffer");
 524
 525         shared_ptr<Frame> frame(new Frame);
 526 #if 0
 527         // 4:2:2 planar (for vaDeriveImage).
 528         frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
 529         frame->cb.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
 530         frame->cr.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
 531         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 532                 uint8_t *dptr;
 533                 size_t width;
 534                 if (component_idx == 0) {
 535                         dptr = frame->y.get();
 536                         width = dinfo.image_width;
 537                 } else if (component_idx == 1) {
 538                         dptr = frame->cb.get();
 539                         width = dinfo.image_width / 2;
 540                 } else if (component_idx == 2) {
 541                         dptr = frame->cr.get();
 542                         width = dinfo.image_width / 2;
 543                 } else {
 544                         assert(false);
 545                 }
 546                 const uint8_t *sptr = (const uint8_t *)mapped + image.offsets[component_idx];
 547                 size_t spitch = image.pitches[component_idx];
 548                 for (size_t y = 0; y < dinfo.image_height; ++y) {
 549                         memcpy(dptr + y * width, sptr + y * spitch, width);
 550                 }
 551         }
 552 #else
 553         // Convert Y'CbCr to separate Y' and CbCr.
 554         frame->is_semiplanar = true;
 555
 556         PBO pbo = global_pbo_pool->alloc_pbo();
 557         size_t cbcr_offset = dinfo.image_width * dinfo.image_height;
 558         uint8_t *y_pix = pbo.ptr;
 559         uint8_t *cbcr_pix = pbo.ptr + cbcr_offset;
 560
 561         const uint8_t *src = (const uint8_t *)mapped + resources.image.offsets[0];
 562         if (resources.image.pitches[0] == dinfo.image_width * 2) {
 563                 memcpy_interleaved(cbcr_pix, y_pix, src, dinfo.image_width * dinfo.image_height * 2);
 564         } else {
 565                 for (unsigned y = 0; y < dinfo.image_height; ++y) {
 566                         memcpy_interleaved(cbcr_pix + y * dinfo.image_width, y_pix + y * dinfo.image_width,
 567                                            src + y * resources.image.pitches[0], dinfo.image_width * 2);
 568                 }
 569         }
 570
 571         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo.pbo);
 572         frame->y = create_texture_2d(dinfo.image_width, dinfo.image_height, GL_R8, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
 573         frame->cbcr = create_texture_2d(dinfo.image_width / 2, dinfo.image_height, GL_RG8, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset));
 574         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 575
 576         glFlushMappedNamedBufferRange(pbo.pbo, 0, dinfo.image_width * dinfo.image_height * 2);
 577         glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT);
 578         pbo.upload_done = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 579         frame->uploaded_ui_thread = pbo.upload_done;
 580         frame->uploaded_interpolation = pbo.upload_done;
 581         global_pbo_pool->release_pbo(move(pbo));
 582 #endif
 583         frame->width = dinfo.image_width;
 584         frame->height = dinfo.image_height;
 585         frame->chroma_subsampling_x = 2;
 586         frame->chroma_subsampling_y = 1;
 587
 588         if (dinfo.marker_list != nullptr &&
 589             dinfo.marker_list->marker == JPEG_APP0 + 1 &&
 590             dinfo.marker_list->data_length >= 4 &&
 591             memcmp(dinfo.marker_list->data, "Exif", 4) == 0) {
 592                 frame->exif_data.assign(reinterpret_cast<char *>(dinfo.marker_list->data),
 593                         dinfo.marker_list->data_length);
 594         }
 595
 596         va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 597         CHECK_VASTATUS_RET(va_status, "vaUnmapBuffer");
 598
 599         return frame;
 600 }