1 #include "vaapi_jpeg_decoder.h"
3 #include "jpeg_destroyer.h"
4 #include "jpeg_frame.h"
5 #include "memcpy_interleaved.h"
21 #include <va/va_drm.h>
22 #include <va/va_x11.h>
26 static unique_ptr<VADisplayWithCleanup> va_dpy;
27 static VAConfigID config_id;
28 static VAImageFormat uyvy_format;
29 bool vaapi_jpeg_decoding_usable = false;
32 unsigned width, height;
37 static list<VAResources> va_resources_freelist;
38 static mutex va_resources_mutex;
40 #define CHECK_VASTATUS(va_status, func) \
41 if (va_status != VA_STATUS_SUCCESS) { \
42 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
46 #define CHECK_VASTATUS_RET(va_status, func) \
47 if (va_status != VA_STATUS_SUCCESS) { \
48 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
52 VAResources get_va_resources(unsigned width, unsigned height)
55 lock_guard<mutex> lock(va_resources_mutex);
56 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
57 if (it->width == width && it->height == height) {
58 VAResources ret = *it;
59 va_resources_freelist.erase(it);
70 VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
72 &ret.surface, 1, nullptr, 0);
73 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
75 va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
76 CHECK_VASTATUS(va_status, "vaCreateContext");
78 va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
79 CHECK_VASTATUS(va_status, "vaCreateImage");
84 void release_va_resources(VAResources resources)
86 lock_guard<mutex> lock(va_resources_mutex);
87 if (va_resources_freelist.size() > 10) {
88 auto it = va_resources_freelist.end();
91 VAStatus va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
92 CHECK_VASTATUS(va_status, "vaDestroyImage");
94 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
95 CHECK_VASTATUS(va_status, "vaDestroyContext");
97 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
98 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
100 va_resources_freelist.erase(it);
103 va_resources_freelist.push_front(resources);
106 // RAII wrapper to release VAResources on return (even on error).
107 class ReleaseVAResources {
109 ReleaseVAResources(const VAResources &resources)
110 : resources(resources) {}
111 ~ReleaseVAResources()
114 release_va_resources(resources);
118 void commit() { committed = true; }
121 const VAResources &resources;
122 bool committed = false;
125 VADisplayWithCleanup::~VADisplayWithCleanup()
127 if (va_dpy != nullptr) {
130 if (x11_display != nullptr) {
131 XCloseDisplay(x11_display);
138 unique_ptr<VADisplayWithCleanup> va_open_display(const string &va_display)
140 if (va_display.empty() || va_display[0] != '/') { // An X display.
141 Display *x11_display = XOpenDisplay(va_display.empty() ? nullptr : va_display.c_str());
142 if (x11_display == nullptr) {
143 fprintf(stderr, "error: can't connect to X server!\n");
147 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
148 ret->x11_display = x11_display;
149 ret->va_dpy = vaGetDisplay(x11_display);
150 if (ret->va_dpy == nullptr) {
154 } else { // A DRM node on the filesystem (e.g. /dev/dri/renderD128).
155 int drm_fd = open(va_display.c_str(), O_RDWR);
157 perror(va_display.c_str());
160 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
161 ret->drm_fd = drm_fd;
162 ret->va_dpy = vaGetDisplayDRM(drm_fd);
163 if (ret->va_dpy == nullptr) {
170 unique_ptr<VADisplayWithCleanup> try_open_va(const string &va_display, string *error)
172 unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
173 if (va_dpy == nullptr) {
175 *error = "Opening VA display failed";
178 int major_ver, minor_ver;
179 VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
180 if (va_status != VA_STATUS_SUCCESS) {
182 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
183 if (error != nullptr)
188 int num_entrypoints = vaMaxNumEntrypoints(va_dpy->va_dpy);
189 unique_ptr<VAEntrypoint[]> entrypoints(new VAEntrypoint[num_entrypoints]);
190 if (entrypoints == nullptr) {
191 if (error != nullptr)
192 *error = "Failed to allocate memory for VA entry points";
196 vaQueryConfigEntrypoints(va_dpy->va_dpy, VAProfileJPEGBaseline, entrypoints.get(), &num_entrypoints);
197 for (int slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
198 if (entrypoints[slice_entrypoint] != VAEntrypointVLD) {
202 // We found a usable decode, so return it.
206 if (error != nullptr)
207 *error = "Can't find VAEntrypointVLD for the JPEG profile";
211 string get_usable_va_display()
213 // Reduce the amount of chatter while probing,
214 // unless the user has specified otherwise.
215 bool need_env_reset = false;
216 if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
217 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
218 need_env_reset = true;
221 // First try the default (ie., whatever $DISPLAY is set to).
222 unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va("", nullptr);
223 if (va_dpy != nullptr) {
224 if (need_env_reset) {
225 unsetenv("LIBVA_MESSAGING_LEVEL");
230 fprintf(stderr, "The X11 display did not expose a VA-API JPEG decoder.\n");
232 // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
234 int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
236 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
238 for (size_t i = 0; i < g.gl_pathc; ++i) {
239 string path = g.gl_pathv[i];
240 va_dpy = try_open_va(path, nullptr);
241 if (va_dpy != nullptr) {
242 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
245 if (need_env_reset) {
246 unsetenv("LIBVA_MESSAGING_LEVEL");
253 fprintf(stderr, "No suitable VA-API JPEG decoders were found in /dev/dri; giving up.\n");
254 fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
255 fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
256 fprintf(stderr, "to expose Quick Sync.\n");
260 void init_jpeg_vaapi()
262 string dpy = get_usable_va_display();
267 va_dpy = try_open_va(dpy, nullptr);
268 if (va_dpy == nullptr) {
272 VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
274 VAStatus va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointVLD,
275 &attr, 1, &config_id);
276 CHECK_VASTATUS(va_status, "vaCreateConfig");
278 int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
279 assert(num_formats > 0);
281 unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
282 va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
283 CHECK_VASTATUS(va_status, "vaQueryImageFormats");
286 for (int i = 0; i < num_formats; ++i) {
287 // Seemingly VA_FOURCC_422H is no good for vaGetImage(). :-/
288 if (formats[i].fourcc == VA_FOURCC_UYVY) {
289 memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
298 fprintf(stderr, "VA-API JPEG decoding initialized.\n");
299 vaapi_jpeg_decoding_usable = true;
302 shared_ptr<Frame> decode_jpeg_vaapi(const string &filename)
304 jpeg_decompress_struct dinfo;
306 dinfo.err = jpeg_std_error(&jerr);
307 jpeg_create_decompress(&dinfo);
308 JPEGDestroyer destroy_dinfo(&dinfo);
310 FILE *fp = fopen(filename.c_str(), "rb");
312 perror(filename.c_str());
315 jpeg_stdio_src(&dinfo, fp);
317 jpeg_read_header(&dinfo, true);
319 // Read the data that comes after the header. VA-API will destuff and all for us.
320 std::string str((const char *)dinfo.src->next_input_byte, dinfo.src->bytes_in_buffer);
323 size_t ret = fread(buf, 1, sizeof(buf), fp);
324 str.append(buf, ret);
328 if (dinfo.num_components != 3) {
329 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
330 dinfo.num_components,
331 dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
332 dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
333 dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
336 if (dinfo.comp_info[0].h_samp_factor != 2 ||
337 dinfo.comp_info[1].h_samp_factor != 1 ||
338 dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[0].v_samp_factor ||
339 dinfo.comp_info[2].h_samp_factor != 1 ||
340 dinfo.comp_info[2].v_samp_factor != dinfo.comp_info[0].v_samp_factor) {
341 fprintf(stderr, "Not 4:2:2. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
342 dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
343 dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
344 dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
348 // Picture parameters.
349 VAPictureParameterBufferJPEGBaseline pic_param;
350 memset(&pic_param, 0, sizeof(pic_param));
351 pic_param.picture_width = dinfo.image_width;
352 pic_param.picture_height = dinfo.image_height;
353 for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
354 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
355 pic_param.components[component_idx].component_id = comp->component_id;
356 pic_param.components[component_idx].h_sampling_factor = comp->h_samp_factor;
357 pic_param.components[component_idx].v_sampling_factor = comp->v_samp_factor;
358 pic_param.components[component_idx].quantiser_table_selector = comp->quant_tbl_no;
360 pic_param.num_components = dinfo.num_components;
361 pic_param.color_space = 0; // YUV.
362 pic_param.rotation = VA_ROTATION_NONE;
364 VABufferID pic_param_buffer;
365 VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAPictureParameterBufferType, sizeof(pic_param), 1, &pic_param, &pic_param_buffer);
366 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
368 // Quantization matrices.
369 VAIQMatrixBufferJPEGBaseline iq;
370 memset(&iq, 0, sizeof(iq));
372 for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
373 const JQUANT_TBL *qtbl = dinfo.quant_tbl_ptrs[quant_tbl_idx];
374 if (qtbl == nullptr) {
375 iq.load_quantiser_table[quant_tbl_idx] = 0;
377 iq.load_quantiser_table[quant_tbl_idx] = 1;
378 for (int i = 0; i < 64; ++i) {
379 if (qtbl->quantval[i] > 255) {
380 fprintf(stderr, "Baseline JPEG only!\n");
383 iq.quantiser_table[quant_tbl_idx][i] = qtbl->quantval[i];
388 VABufferID iq_buffer;
389 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAIQMatrixBufferType, sizeof(iq), 1, &iq, &iq_buffer);
390 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
392 // Huffman tables (arithmetic is not supported).
393 VAHuffmanTableBufferJPEGBaseline huff;
394 memset(&huff, 0, sizeof(huff));
396 for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
397 const JHUFF_TBL *ac_hufftbl = dinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
398 const JHUFF_TBL *dc_hufftbl = dinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
399 if (ac_hufftbl == nullptr) {
400 assert(dc_hufftbl == nullptr);
401 huff.load_huffman_table[huff_tbl_idx] = 0;
403 assert(dc_hufftbl != nullptr);
404 huff.load_huffman_table[huff_tbl_idx] = 1;
406 for (int i = 0; i < 16; ++i) {
407 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
409 for (int i = 0; i < 12; ++i) {
410 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
412 for (int i = 0; i < 16; ++i) {
413 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
415 for (int i = 0; i < 162; ++i) {
416 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
421 VABufferID huff_buffer;
422 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(huff), 1, &huff, &huff_buffer);
423 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
425 // Slice parameters (metadata about the slice).
426 VASliceParameterBufferJPEGBaseline parms;
427 memset(&parms, 0, sizeof(parms));
428 parms.slice_data_size = str.size();
429 parms.slice_data_offset = 0;
430 parms.slice_data_flag = VA_SLICE_DATA_FLAG_ALL;
431 parms.slice_horizontal_position = 0;
432 parms.slice_vertical_position = 0;
433 for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
434 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
435 parms.components[component_idx].component_selector = comp->component_id;
436 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
437 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
438 if (parms.components[component_idx].dc_table_selector > 1 ||
439 parms.components[component_idx].ac_table_selector > 1) {
440 fprintf(stderr, "Uses too many Huffman tables\n");
444 parms.num_components = dinfo.num_components;
445 parms.restart_interval = dinfo.restart_interval;
446 int horiz_mcus = (dinfo.image_width + (DCTSIZE * 2) - 1) / (DCTSIZE * 2);
447 int vert_mcus = (dinfo.image_height + DCTSIZE - 1) / DCTSIZE;
448 parms.num_mcus = horiz_mcus * vert_mcus;
450 VABufferID slice_param_buffer;
451 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceParameterBufferType, sizeof(parms), 1, &parms, &slice_param_buffer);
452 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
455 VABufferID data_buffer;
456 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceDataBufferType, str.size(), 1, &str[0], &data_buffer);
457 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
459 VAResources resources = get_va_resources(dinfo.image_width, dinfo.image_height);
460 ReleaseVAResources release(resources);
462 va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
463 CHECK_VASTATUS_RET(va_status, "vaBeginPicture");
464 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
465 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(pic_param)");
466 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &iq_buffer, 1);
467 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(iq)");
468 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
469 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(huff)");
470 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
471 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(slice_param)");
472 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &data_buffer, 1);
473 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(data)");
474 va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
475 CHECK_VASTATUS_RET(va_status, "vaEndPicture");
477 // vaDeriveImage() works, but the resulting image seems to live in
478 // uncached memory, which makes copying data out from it very, very slow.
479 // Thanks to FFmpeg for the observation that you can vaGetImage() the
480 // surface onto your own image (although then, it can't be planar, which
481 // is unfortunate for us).
484 va_status = vaDeriveImage(va_dpy->va_dpy, surf, &image);
485 CHECK_VASTATUS_RET(va_status, "vaDeriveImage");
487 va_status = vaSyncSurface(va_dpy->va_dpy, resources.surface);
488 CHECK_VASTATUS_RET(va_status, "vaSyncSurface");
490 va_status = vaGetImage(va_dpy->va_dpy, resources.surface, 0, 0, dinfo.image_width, dinfo.image_height, resources.image.image_id);
491 CHECK_VASTATUS_RET(va_status, "vaGetImage");
495 va_status = vaMapBuffer(va_dpy->va_dpy, resources.image.buf, &mapped);
496 CHECK_VASTATUS_RET(va_status, "vaMapBuffer");
498 shared_ptr<Frame> frame(new Frame);
500 // 4:2:2 planar (for vaDeriveImage).
501 frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
502 frame->cb.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
503 frame->cr.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
504 for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
507 if (component_idx == 0) {
508 dptr = frame->y.get();
509 width = dinfo.image_width;
510 } else if (component_idx == 1) {
511 dptr = frame->cb.get();
512 width = dinfo.image_width / 2;
513 } else if (component_idx == 2) {
514 dptr = frame->cr.get();
515 width = dinfo.image_width / 2;
519 const uint8_t *sptr = (const uint8_t *)mapped + image.offsets[component_idx];
520 size_t spitch = image.pitches[component_idx];
521 for (size_t y = 0; y < dinfo.image_height; ++y) {
522 memcpy(dptr + y * width, sptr + y * spitch, width);
526 // Convert Y'CbCr to separate Y' and CbCr.
527 frame->is_semiplanar = true;
528 frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
529 frame->cbcr.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
530 const uint8_t *src = (const uint8_t *)mapped + resources.image.offsets[0];
531 if (resources.image.pitches[0] == dinfo.image_width * 2) {
532 memcpy_interleaved(frame->cbcr.get(), frame->y.get(), src, dinfo.image_width * dinfo.image_height * 2);
534 for (unsigned y = 0; y < dinfo.image_height; ++y) {
535 memcpy_interleaved(frame->cbcr.get() + y * dinfo.image_width, frame->y.get() + y * dinfo.image_width,
536 src + y * resources.image.pitches[0], dinfo.image_width * 2);
540 frame->width = dinfo.image_width;
541 frame->height = dinfo.image_height;
542 frame->chroma_subsampling_x = 2;
543 frame->chroma_subsampling_y = 1;
544 frame->pitch_y = dinfo.image_width;
545 frame->pitch_chroma = dinfo.image_width / 2;
547 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
548 CHECK_VASTATUS_RET(va_status, "vaUnmapBuffer");