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 #include "shared/va_display.h"
   9 #include "shared/va_resource_pool.h"
  10
  11 #include <X11/Xlib.h>
  12 #include <assert.h>
  13 #include <errno.h>
  14 #include <fcntl.h>
  15 #include <glob.h>
  16 #include <jpeglib.h>
  17 #include <list>
  18 #include <mutex>
  19 #include <stdio.h>
  20 #include <stdlib.h>
  21 #include <string.h>
  22 #include <string>
  23 #include <unistd.h>
  24 #include <va/va.h>
  25 #include <va/va_drm.h>
  26 #include <va/va_x11.h>
  27
  28 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
  29
  30 using namespace std;
  31
  32 static unique_ptr<VADisplayWithCleanup> va_dpy;
  33 static unique_ptr<VAResourcePool> va_pool;
  34
  35 bool vaapi_jpeg_decoding_usable = false;
  36
  37 // From libjpeg (although it's of course identical between implementations).
  38 static const int jpeg_natural_order[DCTSIZE2] = {
  39          0,  1,  8, 16,  9,  2,  3, 10,
  40         17, 24, 32, 25, 18, 11,  4,  5,
  41         12, 19, 26, 33, 40, 48, 41, 34,
  42         27, 20, 13,  6,  7, 14, 21, 28,
  43         35, 42, 49, 56, 57, 50, 43, 36,
  44         29, 22, 15, 23, 30, 37, 44, 51,
  45         58, 59, 52, 45, 38, 31, 39, 46,
  46         53, 60, 61, 54, 47, 55, 62, 63,
  47 };
  48
  49 static unique_ptr<VADisplayWithCleanup> try_open_va_mjpeg(const string &va_display)
  50 {
  51         VAConfigID config_id_422, config_id_420;
  52         VAImageFormat uyvy_format, nv12_format;
  53
  54         // Seemingly VA_FOURCC_422H is no good for vaGetImage(). :-/
  55         unique_ptr<VADisplayWithCleanup> va_dpy =
  56                 try_open_va(va_display, { VAProfileJPEGBaseline }, VAEntrypointVLD,
  57                         { { "4:2:2", VA_RT_FORMAT_YUV422, VA_FOURCC_UYVY, &config_id_422, &uyvy_format },
  58                           { "4:2:0", VA_RT_FORMAT_YUV420, VA_FOURCC_NV12, &config_id_420, &nv12_format } },
  59                         /*chosen_profile=*/nullptr, /*error=*/nullptr);
  60         if (va_dpy == nullptr) {
  61                 return va_dpy;
  62         }
  63
  64         va_pool.reset(new VAResourcePool(va_dpy->va_dpy, uyvy_format, nv12_format, config_id_422, config_id_420, /*with_data_buffer=*/false));
  65
  66         return va_dpy;
  67 }
  68
  69 string get_usable_va_display()
  70 {
  71         // Reduce the amount of chatter while probing,
  72         // unless the user has specified otherwise.
  73         bool need_env_reset = false;
  74         if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
  75                 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
  76                 need_env_reset = true;
  77         }
  78
  79         // First try the default (ie., whatever $DISPLAY is set to).
  80         unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va_mjpeg("");
  81         if (va_dpy != nullptr) {
  82                 if (need_env_reset) {
  83                         unsetenv("LIBVA_MESSAGING_LEVEL");
  84                 }
  85                 return "";
  86         }
  87
  88         fprintf(stderr, "The X11 display did not expose a VA-API JPEG decoder.\n");
  89
  90         // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
  91         glob_t g;
  92         int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
  93         if (err != 0) {
  94                 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
  95         } else {
  96                 for (size_t i = 0; i < g.gl_pathc; ++i) {
  97                         string path = g.gl_pathv[i];
  98                         va_dpy = try_open_va_mjpeg(path);
  99                         if (va_dpy != nullptr) {
 100                                 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
 101                                         path.c_str());
 102                                 globfree(&g);
 103                                 if (need_env_reset) {
 104                                         unsetenv("LIBVA_MESSAGING_LEVEL");
 105                                 }
 106                                 return path;
 107                         }
 108                 }
 109         }
 110
 111         fprintf(stderr, "No suitable VA-API JPEG decoders were found in /dev/dri; giving up.\n");
 112         fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
 113         fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
 114         fprintf(stderr, "to expose Quick Sync.\n");
 115         return "none";
 116 }
 117
 118 void init_jpeg_vaapi()
 119 {
 120         string dpy = get_usable_va_display();
 121         if (dpy == "none") {
 122                 return;
 123         }
 124
 125         va_dpy = try_open_va_mjpeg(dpy);
 126         if (va_dpy == nullptr) {
 127                 return;
 128         }
 129
 130         fprintf(stderr, "VA-API JPEG decoding initialized.\n");
 131         vaapi_jpeg_decoding_usable = true;
 132 }
 133
 134 class VABufferDestroyer {
 135 public:
 136         VABufferDestroyer(VADisplay dpy, VABufferID buf)
 137                 : dpy(dpy), buf(buf) {}
 138
 139         ~VABufferDestroyer()
 140         {
 141                 VAStatus va_status = vaDestroyBuffer(dpy, buf);
 142                 CHECK_VASTATUS(va_status, "vaDestroyBuffer");
 143         }
 144
 145 private:
 146         VADisplay dpy;
 147         VABufferID buf;
 148 };
 149
 150 shared_ptr<Frame> decode_jpeg_vaapi(const string &jpeg)
 151 {
 152         jpeg_decompress_struct dinfo;
 153         JPEGWrapErrorManager error_mgr(&dinfo);
 154         if (!error_mgr.run([&dinfo] { jpeg_create_decompress(&dinfo); })) {
 155                 return nullptr;
 156         }
 157         JPEGDestroyer destroy_dinfo(&dinfo);
 158
 159         jpeg_save_markers(&dinfo, JPEG_APP0 + 1, 0xFFFF);
 160
 161         jpeg_mem_src(&dinfo, reinterpret_cast<const unsigned char *>(jpeg.data()), jpeg.size());
 162         if (!error_mgr.run([&dinfo] { jpeg_read_header(&dinfo, true); })) {
 163                 return nullptr;
 164         }
 165
 166         if (dinfo.num_components != 3) {
 167                 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 168                         dinfo.num_components,
 169                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 170                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 171                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 172                 return nullptr;
 173         }
 174         if (dinfo.comp_info[0].h_samp_factor != 2 ||
 175             dinfo.comp_info[1].h_samp_factor != 1 ||
 176             dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[0].v_samp_factor ||
 177             dinfo.comp_info[2].h_samp_factor != 1 ||
 178             dinfo.comp_info[2].v_samp_factor != dinfo.comp_info[0].v_samp_factor) {
 179                 fprintf(stderr, "Not 4:2:2. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
 180                         dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
 181                         dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
 182                         dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
 183                 return nullptr;
 184         }
 185
 186         // Picture parameters.
 187         VAPictureParameterBufferJPEGBaseline pic_param;
 188         memset(&pic_param, 0, sizeof(pic_param));
 189         pic_param.picture_width = dinfo.image_width;
 190         pic_param.picture_height = dinfo.image_height;
 191         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 192                 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
 193                 pic_param.components[component_idx].component_id = comp->component_id;
 194                 pic_param.components[component_idx].h_sampling_factor = comp->h_samp_factor;
 195                 pic_param.components[component_idx].v_sampling_factor = comp->v_samp_factor;
 196                 pic_param.components[component_idx].quantiser_table_selector = comp->quant_tbl_no;
 197         }
 198         pic_param.num_components = dinfo.num_components;
 199         pic_param.color_space = 0;  // YUV.
 200         pic_param.rotation = VA_ROTATION_NONE;
 201
 202         VAResourcePool::VAResources resources = va_pool->get_va_resources(dinfo.image_width, dinfo.image_height, VA_FOURCC_UYVY);
 203         ReleaseVAResources release(va_pool.get(), resources);
 204
 205         VABufferID pic_param_buffer;
 206         VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAPictureParameterBufferType, sizeof(pic_param), 1, &pic_param, &pic_param_buffer);
 207         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 208         VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
 209
 210         // Quantization matrices.
 211         VAIQMatrixBufferJPEGBaseline iq;
 212         memset(&iq, 0, sizeof(iq));
 213
 214         for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
 215                 const JQUANT_TBL *qtbl = dinfo.quant_tbl_ptrs[quant_tbl_idx];
 216                 if (qtbl == nullptr) {
 217                         iq.load_quantiser_table[quant_tbl_idx] = 0;
 218                 } else {
 219                         iq.load_quantiser_table[quant_tbl_idx] = 1;
 220                         for (int i = 0; i < 64; ++i) {
 221                                 if (qtbl->quantval[i] > 255) {
 222                                         fprintf(stderr, "Baseline JPEG only!\n");
 223                                         return nullptr;
 224                                 }
 225                                 iq.quantiser_table[quant_tbl_idx][i] = qtbl->quantval[jpeg_natural_order[i]];
 226                         }
 227                 }
 228         }
 229
 230         VABufferID iq_buffer;
 231         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAIQMatrixBufferType, sizeof(iq), 1, &iq, &iq_buffer);
 232         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 233         VABufferDestroyer destroy_iq(va_dpy->va_dpy, iq_buffer);
 234
 235         // Huffman tables (arithmetic is not supported).
 236         VAHuffmanTableBufferJPEGBaseline huff;
 237         memset(&huff, 0, sizeof(huff));
 238
 239         for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
 240                 const JHUFF_TBL *ac_hufftbl = dinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
 241                 const JHUFF_TBL *dc_hufftbl = dinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
 242                 if (ac_hufftbl == nullptr) {
 243                         assert(dc_hufftbl == nullptr);
 244                         huff.load_huffman_table[huff_tbl_idx] = 0;
 245                 } else {
 246                         assert(dc_hufftbl != nullptr);
 247                         huff.load_huffman_table[huff_tbl_idx] = 1;
 248
 249                         for (int i = 0; i < 16; ++i) {
 250                                 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
 251                         }
 252                         for (int i = 0; i < 12; ++i) {
 253                                 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
 254                         }
 255                         for (int i = 0; i < 16; ++i) {
 256                                 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
 257                         }
 258                         for (int i = 0; i < 162; ++i) {
 259                                 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
 260                         }
 261                 }
 262         }
 263
 264         VABufferID huff_buffer;
 265         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VAHuffmanTableBufferType, sizeof(huff), 1, &huff, &huff_buffer);
 266         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 267         VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
 268
 269         // Slice parameters (metadata about the slice).
 270         VASliceParameterBufferJPEGBaseline parms;
 271         memset(&parms, 0, sizeof(parms));
 272         parms.slice_data_size = dinfo.src->bytes_in_buffer;
 273         parms.slice_data_offset = 0;
 274         parms.slice_data_flag = VA_SLICE_DATA_FLAG_ALL;
 275         parms.slice_horizontal_position = 0;
 276         parms.slice_vertical_position = 0;
 277         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 278                 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
 279                 parms.components[component_idx].component_selector = comp->component_id;
 280                 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
 281                 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
 282                 if (parms.components[component_idx].dc_table_selector > 1 ||
 283                     parms.components[component_idx].ac_table_selector > 1) {
 284                         fprintf(stderr, "Uses too many Huffman tables\n");
 285                         return nullptr;
 286                 }
 287         }
 288         parms.num_components = dinfo.num_components;
 289         parms.restart_interval = dinfo.restart_interval;
 290         int horiz_mcus = (dinfo.image_width + (DCTSIZE * 2) - 1) / (DCTSIZE * 2);
 291         int vert_mcus = (dinfo.image_height + DCTSIZE - 1) / DCTSIZE;
 292         parms.num_mcus = horiz_mcus * vert_mcus;
 293
 294         VABufferID slice_param_buffer;
 295         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VASliceParameterBufferType, sizeof(parms), 1, &parms, &slice_param_buffer);
 296         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 297         VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
 298
 299         // The actual data. VA-API will destuff and all for us.
 300         VABufferID data_buffer;
 301         va_status = vaCreateBuffer(va_dpy->va_dpy, resources.context, VASliceDataBufferType, dinfo.src->bytes_in_buffer, 1, const_cast<unsigned char *>(dinfo.src->next_input_byte), &data_buffer);
 302         CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
 303         VABufferDestroyer destroy_data(va_dpy->va_dpy, data_buffer);
 304
 305         va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
 306         CHECK_VASTATUS_RET(va_status, "vaBeginPicture");
 307         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
 308         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(pic_param)");
 309         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &iq_buffer, 1);
 310         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(iq)");
 311         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
 312         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(huff)");
 313         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
 314         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(slice_param)");
 315         va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &data_buffer, 1);
 316         CHECK_VASTATUS_RET(va_status, "vaRenderPicture(data)");
 317         va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
 318         CHECK_VASTATUS_RET(va_status, "vaEndPicture");
 319
 320         // vaDeriveImage() works, but the resulting image seems to live in
 321         // uncached memory, which makes copying data out from it very, very slow.
 322         // Thanks to FFmpeg for the observation that you can vaGetImage() the
 323         // surface onto your own image (although then, it can't be planar, which
 324         // is unfortunate for us).
 325 #if 0
 326         VAImage image;
 327         va_status = vaDeriveImage(va_dpy->va_dpy, surf, &image);
 328         CHECK_VASTATUS_RET(va_status, "vaDeriveImage");
 329 #else
 330         va_status = vaSyncSurface(va_dpy->va_dpy, resources.surface);
 331         CHECK_VASTATUS_RET(va_status, "vaSyncSurface");
 332
 333         va_status = vaGetImage(va_dpy->va_dpy, resources.surface, 0, 0, dinfo.image_width, dinfo.image_height, resources.image.image_id);
 334         CHECK_VASTATUS_RET(va_status, "vaGetImage");
 335 #endif
 336
 337         void *mapped;
 338         va_status = vaMapBuffer(va_dpy->va_dpy, resources.image.buf, &mapped);
 339         CHECK_VASTATUS_RET(va_status, "vaMapBuffer");
 340
 341         shared_ptr<Frame> frame(new Frame);
 342 #if 0
 343         // 4:2:2 planar (for vaDeriveImage).
 344         frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
 345         frame->cb.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
 346         frame->cr.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
 347         for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
 348                 uint8_t *dptr;
 349                 size_t width;
 350                 if (component_idx == 0) {
 351                         dptr = frame->y.get();
 352                         width = dinfo.image_width;
 353                 } else if (component_idx == 1) {
 354                         dptr = frame->cb.get();
 355                         width = dinfo.image_width / 2;
 356                 } else if (component_idx == 2) {
 357                         dptr = frame->cr.get();
 358                         width = dinfo.image_width / 2;
 359                 } else {
 360                         assert(false);
 361                 }
 362                 const uint8_t *sptr = (const uint8_t *)mapped + image.offsets[component_idx];
 363                 size_t spitch = image.pitches[component_idx];
 364                 for (size_t y = 0; y < dinfo.image_height; ++y) {
 365                         memcpy(dptr + y * width, sptr + y * spitch, width);
 366                 }
 367         }
 368 #else
 369         // Convert Y'CbCr to separate Y' and CbCr.
 370         frame->is_semiplanar = true;
 371
 372         PBO pbo = global_pbo_pool->alloc_pbo();
 373         size_t cbcr_offset = dinfo.image_width * dinfo.image_height;
 374         uint8_t *y_pix = pbo.ptr;
 375         uint8_t *cbcr_pix = pbo.ptr + cbcr_offset;
 376
 377         const uint8_t *src = (const uint8_t *)mapped + resources.image.offsets[0];
 378         if (resources.image.pitches[0] == dinfo.image_width * 2) {
 379                 memcpy_interleaved(cbcr_pix, y_pix, src, dinfo.image_width * dinfo.image_height * 2);
 380         } else {
 381                 for (unsigned y = 0; y < dinfo.image_height; ++y) {
 382                         memcpy_interleaved(cbcr_pix + y * dinfo.image_width, y_pix + y * dinfo.image_width,
 383                                            src + y * resources.image.pitches[0], dinfo.image_width * 2);
 384                 }
 385         }
 386
 387         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo.pbo);
 388         frame->y = create_texture_2d(dinfo.image_width, dinfo.image_height, GL_R8, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
 389         frame->cbcr = create_texture_2d(dinfo.image_width / 2, dinfo.image_height, GL_RG8, GL_RG, GL_UNSIGNED_BYTE, BUFFER_OFFSET(cbcr_offset));
 390         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 391
 392         glFlushMappedNamedBufferRange(pbo.pbo, 0, dinfo.image_width * dinfo.image_height * 2);
 393         glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT);
 394         pbo.upload_done = RefCountedGLsync(GL_SYNC_GPU_COMMANDS_COMPLETE, /*flags=*/0);
 395         frame->uploaded_ui_thread = pbo.upload_done;
 396         frame->uploaded_interpolation = pbo.upload_done;
 397         global_pbo_pool->release_pbo(move(pbo));
 398 #endif
 399         frame->width = dinfo.image_width;
 400         frame->height = dinfo.image_height;
 401         frame->chroma_subsampling_x = 2;
 402         frame->chroma_subsampling_y = 1;
 403
 404         if (dinfo.marker_list != nullptr &&
 405             dinfo.marker_list->marker == JPEG_APP0 + 1 &&
 406             dinfo.marker_list->data_length >= 4 &&
 407             memcmp(dinfo.marker_list->data, "Exif", 4) == 0) {
 408                 frame->exif_data.assign(reinterpret_cast<char *>(dinfo.marker_list->data),
 409                         dinfo.marker_list->data_length);
 410         }
 411
 412         va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
 413         CHECK_VASTATUS_RET(va_status, "vaUnmapBuffer");
 414
 415         return frame;
 416 }