1 #include "vaapi_jpeg_decoder.h"
3 #include "jpeg_frame.h"
4 #include "memcpy_interleaved.h"
20 #include <va/va_drm.h>
21 #include <va/va_x11.h>
25 static unique_ptr<VADisplayWithCleanup> va_dpy;
26 static VAConfigID config_id;
27 static VAImageFormat uyvy_format;
28 bool vaapi_jpeg_decoding_usable = false;
31 unsigned width, height;
36 static list<VAResources> va_resources_freelist;
37 static mutex va_resources_mutex;
39 #define CHECK_VASTATUS(va_status, func) \
40 if (va_status != VA_STATUS_SUCCESS) { \
41 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
45 #define CHECK_VASTATUS_RET(va_status, func) \
46 if (va_status != VA_STATUS_SUCCESS) { \
47 fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
51 VAResources get_va_resources(unsigned width, unsigned height)
54 lock_guard<mutex> lock(va_resources_mutex);
55 for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
56 if (it->width == width && it->height == height) {
57 VAResources ret = *it;
58 va_resources_freelist.erase(it);
69 VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
71 &ret.surface, 1, nullptr, 0);
72 CHECK_VASTATUS(va_status, "vaCreateSurfaces");
74 va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
75 CHECK_VASTATUS(va_status, "vaCreateContext");
77 va_status = vaCreateImage(va_dpy->va_dpy, &uyvy_format, width, height, &ret.image);
78 CHECK_VASTATUS(va_status, "vaCreateImage");
83 void release_va_resources(VAResources resources)
85 lock_guard<mutex> lock(va_resources_mutex);
86 if (va_resources_freelist.size() > 10) {
87 auto it = va_resources_freelist.end();
90 VAStatus va_status = vaDestroyImage(va_dpy->va_dpy, it->image.image_id);
91 CHECK_VASTATUS(va_status, "vaDestroyImage");
93 va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
94 CHECK_VASTATUS(va_status, "vaDestroyContext");
96 va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
97 CHECK_VASTATUS(va_status, "vaDestroySurfaces");
99 va_resources_freelist.erase(it);
102 va_resources_freelist.push_front(resources);
105 // RAII wrapper to release VAResources on return (even on error).
106 class ReleaseVAResources {
108 ReleaseVAResources(const VAResources &resources)
109 : resources(resources) {}
110 ~ReleaseVAResources()
113 release_va_resources(resources);
117 void commit() { committed = true; }
120 const VAResources &resources;
121 bool committed = false;
124 VADisplayWithCleanup::~VADisplayWithCleanup()
126 if (va_dpy != nullptr) {
129 if (x11_display != nullptr) {
130 XCloseDisplay(x11_display);
137 unique_ptr<VADisplayWithCleanup> va_open_display(const string &va_display)
139 if (va_display.empty() || va_display[0] != '/') { // An X display.
140 Display *x11_display = XOpenDisplay(va_display.empty() ? nullptr : va_display.c_str());
141 if (x11_display == nullptr) {
142 fprintf(stderr, "error: can't connect to X server!\n");
146 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
147 ret->x11_display = x11_display;
148 ret->va_dpy = vaGetDisplay(x11_display);
149 if (ret->va_dpy == nullptr) {
153 } else { // A DRM node on the filesystem (e.g. /dev/dri/renderD128).
154 int drm_fd = open(va_display.c_str(), O_RDWR);
156 perror(va_display.c_str());
159 unique_ptr<VADisplayWithCleanup> ret(new VADisplayWithCleanup);
160 ret->drm_fd = drm_fd;
161 ret->va_dpy = vaGetDisplayDRM(drm_fd);
162 if (ret->va_dpy == nullptr) {
169 unique_ptr<VADisplayWithCleanup> try_open_va(const string &va_display, string *error)
171 unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
172 if (va_dpy == nullptr) {
174 *error = "Opening VA display failed";
177 int major_ver, minor_ver;
178 VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
179 if (va_status != VA_STATUS_SUCCESS) {
181 snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
182 if (error != nullptr)
187 int num_entrypoints = vaMaxNumEntrypoints(va_dpy->va_dpy);
188 unique_ptr<VAEntrypoint[]> entrypoints(new VAEntrypoint[num_entrypoints]);
189 if (entrypoints == nullptr) {
190 if (error != nullptr)
191 *error = "Failed to allocate memory for VA entry points";
195 vaQueryConfigEntrypoints(va_dpy->va_dpy, VAProfileJPEGBaseline, entrypoints.get(), &num_entrypoints);
196 for (int slice_entrypoint = 0; slice_entrypoint < num_entrypoints; slice_entrypoint++) {
197 if (entrypoints[slice_entrypoint] != VAEntrypointVLD) {
201 // We found a usable decode, so return it.
205 if (error != nullptr)
206 *error = "Can't find VAEntrypointVLD for the JPEG profile";
210 string get_usable_va_display()
212 // Reduce the amount of chatter while probing,
213 // unless the user has specified otherwise.
214 bool need_env_reset = false;
215 if (getenv("LIBVA_MESSAGING_LEVEL") == nullptr) {
216 setenv("LIBVA_MESSAGING_LEVEL", "0", true);
217 need_env_reset = true;
220 // First try the default (ie., whatever $DISPLAY is set to).
221 unique_ptr<VADisplayWithCleanup> va_dpy = try_open_va("", nullptr);
222 if (va_dpy != nullptr) {
223 if (need_env_reset) {
224 unsetenv("LIBVA_MESSAGING_LEVEL");
229 fprintf(stderr, "The X11 display did not expose a VA-API JPEG decoder.\n");
231 // Try all /dev/dri/render* in turn. TODO: Accept /dev/dri/card*, too?
233 int err = glob("/dev/dri/renderD*", 0, nullptr, &g);
235 fprintf(stderr, "Couldn't list render nodes (%s) when trying to autodetect a replacement.\n", strerror(errno));
237 for (size_t i = 0; i < g.gl_pathc; ++i) {
238 string path = g.gl_pathv[i];
239 va_dpy = try_open_va(path, nullptr);
240 if (va_dpy != nullptr) {
241 fprintf(stderr, "Autodetected %s as a suitable replacement; using it.\n",
244 if (need_env_reset) {
245 unsetenv("LIBVA_MESSAGING_LEVEL");
252 fprintf(stderr, "No suitable VA-API JPEG decoders were found in /dev/dri; giving up.\n");
253 fprintf(stderr, "Note that if you are using an Intel CPU with an external GPU,\n");
254 fprintf(stderr, "you may need to enable the integrated Intel GPU in your BIOS\n");
255 fprintf(stderr, "to expose Quick Sync.\n");
259 void init_jpeg_vaapi()
261 string dpy = get_usable_va_display();
266 va_dpy = try_open_va(dpy, nullptr);
267 if (va_dpy == nullptr) {
271 VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
273 VAStatus va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointVLD,
274 &attr, 1, &config_id);
275 CHECK_VASTATUS(va_status, "vaCreateConfig");
277 int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
278 assert(num_formats > 0);
280 unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
281 va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
282 CHECK_VASTATUS(va_status, "vaQueryImageFormats");
285 for (int i = 0; i < num_formats; ++i) {
286 // Seemingly VA_FOURCC_422H is no good for vaGetImage(). :-/
287 if (formats[i].fourcc == VA_FOURCC_UYVY) {
288 memcpy(&uyvy_format, &formats[i], sizeof(VAImageFormat));
297 fprintf(stderr, "VA-API JPEG decoding initialized.\n");
298 vaapi_jpeg_decoding_usable = true;
301 shared_ptr<Frame> decode_jpeg_vaapi(const string &filename)
303 jpeg_decompress_struct dinfo;
305 dinfo.err = jpeg_std_error(&jerr);
306 jpeg_create_decompress(&dinfo);
308 FILE *fp = fopen(filename.c_str(), "rb");
310 perror(filename.c_str());
313 jpeg_stdio_src(&dinfo, fp);
315 jpeg_read_header(&dinfo, true);
317 // Read the data that comes after the header. VA-API will destuff and all for us.
318 std::string str((const char *)dinfo.src->next_input_byte, dinfo.src->bytes_in_buffer);
321 size_t ret = fread(buf, 1, sizeof(buf), fp);
322 str.append(buf, ret);
326 if (dinfo.num_components != 3) {
327 fprintf(stderr, "Not a color JPEG. (%d components, Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
328 dinfo.num_components,
329 dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
330 dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
331 dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
334 if (dinfo.comp_info[0].h_samp_factor != 2 ||
335 dinfo.comp_info[1].h_samp_factor != 1 ||
336 dinfo.comp_info[1].v_samp_factor != dinfo.comp_info[0].v_samp_factor ||
337 dinfo.comp_info[2].h_samp_factor != 1 ||
338 dinfo.comp_info[2].v_samp_factor != dinfo.comp_info[0].v_samp_factor) {
339 fprintf(stderr, "Not 4:2:2. (Y=%dx%d, Cb=%dx%d, Cr=%dx%d)\n",
340 dinfo.comp_info[0].h_samp_factor, dinfo.comp_info[0].v_samp_factor,
341 dinfo.comp_info[1].h_samp_factor, dinfo.comp_info[1].v_samp_factor,
342 dinfo.comp_info[2].h_samp_factor, dinfo.comp_info[2].v_samp_factor);
346 // Picture parameters.
347 VAPictureParameterBufferJPEGBaseline pic_param;
348 memset(&pic_param, 0, sizeof(pic_param));
349 pic_param.picture_width = dinfo.image_width;
350 pic_param.picture_height = dinfo.image_height;
351 for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
352 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
353 pic_param.components[component_idx].component_id = comp->component_id;
354 pic_param.components[component_idx].h_sampling_factor = comp->h_samp_factor;
355 pic_param.components[component_idx].v_sampling_factor = comp->v_samp_factor;
356 pic_param.components[component_idx].quantiser_table_selector = comp->quant_tbl_no;
358 pic_param.num_components = dinfo.num_components;
359 pic_param.color_space = 0; // YUV.
360 pic_param.rotation = VA_ROTATION_NONE;
362 VABufferID pic_param_buffer;
363 VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAPictureParameterBufferType, sizeof(pic_param), 1, &pic_param, &pic_param_buffer);
364 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
366 // Quantization matrices.
367 VAIQMatrixBufferJPEGBaseline iq;
368 memset(&iq, 0, sizeof(iq));
370 for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
371 const JQUANT_TBL *qtbl = dinfo.quant_tbl_ptrs[quant_tbl_idx];
372 if (qtbl == nullptr) {
373 iq.load_quantiser_table[quant_tbl_idx] = 0;
375 iq.load_quantiser_table[quant_tbl_idx] = 1;
376 for (int i = 0; i < 64; ++i) {
377 if (qtbl->quantval[i] > 255) {
378 fprintf(stderr, "Baseline JPEG only!\n");
381 iq.quantiser_table[quant_tbl_idx][i] = qtbl->quantval[i];
386 VABufferID iq_buffer;
387 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAIQMatrixBufferType, sizeof(iq), 1, &iq, &iq_buffer);
388 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
390 // Huffman tables (arithmetic is not supported).
391 VAHuffmanTableBufferJPEGBaseline huff;
392 memset(&huff, 0, sizeof(huff));
394 for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
395 const JHUFF_TBL *ac_hufftbl = dinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
396 const JHUFF_TBL *dc_hufftbl = dinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
397 if (ac_hufftbl == nullptr) {
398 assert(dc_hufftbl == nullptr);
399 huff.load_huffman_table[huff_tbl_idx] = 0;
401 assert(dc_hufftbl != nullptr);
402 huff.load_huffman_table[huff_tbl_idx] = 1;
404 for (int i = 0; i < 16; ++i) {
405 huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
407 for (int i = 0; i < 12; ++i) {
408 huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
410 for (int i = 0; i < 16; ++i) {
411 huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
413 for (int i = 0; i < 162; ++i) {
414 huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
419 VABufferID huff_buffer;
420 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(huff), 1, &huff, &huff_buffer);
421 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
423 // Slice parameters (metadata about the slice).
424 VASliceParameterBufferJPEGBaseline parms;
425 memset(&parms, 0, sizeof(parms));
426 parms.slice_data_size = str.size();
427 parms.slice_data_offset = 0;
428 parms.slice_data_flag = VA_SLICE_DATA_FLAG_ALL;
429 parms.slice_horizontal_position = 0;
430 parms.slice_vertical_position = 0;
431 for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
432 const jpeg_component_info *comp = &dinfo.comp_info[component_idx];
433 parms.components[component_idx].component_selector = comp->component_id;
434 parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
435 parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
436 if (parms.components[component_idx].dc_table_selector > 1 ||
437 parms.components[component_idx].ac_table_selector > 1) {
438 fprintf(stderr, "Uses too many Huffman tables\n");
442 parms.num_components = dinfo.num_components;
443 parms.restart_interval = dinfo.restart_interval;
444 int horiz_mcus = (dinfo.image_width + (DCTSIZE * 2) - 1) / (DCTSIZE * 2);
445 int vert_mcus = (dinfo.image_height + DCTSIZE - 1) / DCTSIZE;
446 parms.num_mcus = horiz_mcus * vert_mcus;
448 VABufferID slice_param_buffer;
449 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceParameterBufferType, sizeof(parms), 1, &parms, &slice_param_buffer);
450 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
453 VABufferID data_buffer;
454 va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VASliceDataBufferType, str.size(), 1, &str[0], &data_buffer);
455 CHECK_VASTATUS_RET(va_status, "vaCreateBuffer");
457 VAResources resources = get_va_resources(dinfo.image_width, dinfo.image_height);
458 ReleaseVAResources release(resources);
460 va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
461 CHECK_VASTATUS_RET(va_status, "vaBeginPicture");
462 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
463 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(pic_param)");
464 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &iq_buffer, 1);
465 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(iq)");
466 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
467 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(huff)");
468 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
469 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(slice_param)");
470 va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &data_buffer, 1);
471 CHECK_VASTATUS_RET(va_status, "vaRenderPicture(data)");
472 va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
473 CHECK_VASTATUS_RET(va_status, "vaEndPicture");
475 // vaDeriveImage() works, but the resulting image seems to live in
476 // uncached memory, which makes copying data out from it very, very slow.
477 // Thanks to FFmpeg for the observation that you can vaGetImage() the
478 // surface onto your own image (although then, it can't be planar, which
479 // is unfortunate for us).
482 va_status = vaDeriveImage(va_dpy->va_dpy, surf, &image);
483 CHECK_VASTATUS_RET(va_status, "vaDeriveImage");
485 va_status = vaSyncSurface(va_dpy->va_dpy, resources.surface);
486 CHECK_VASTATUS_RET(va_status, "vaSyncSurface");
488 va_status = vaGetImage(va_dpy->va_dpy, resources.surface, 0, 0, dinfo.image_width, dinfo.image_height, resources.image.image_id);
489 CHECK_VASTATUS_RET(va_status, "vaGetImage");
493 va_status = vaMapBuffer(va_dpy->va_dpy, resources.image.buf, &mapped);
494 CHECK_VASTATUS_RET(va_status, "vaMapBuffer");
496 shared_ptr<Frame> frame(new Frame);
498 // 4:2:2 planar (for vaDeriveImage).
499 frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
500 frame->cb.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
501 frame->cr.reset(new uint8_t[(dinfo.image_width / 2) * dinfo.image_height]);
502 for (int component_idx = 0; component_idx < dinfo.num_components; ++component_idx) {
505 if (component_idx == 0) {
506 dptr = frame->y.get();
507 width = dinfo.image_width;
508 } else if (component_idx == 1) {
509 dptr = frame->cb.get();
510 width = dinfo.image_width / 2;
511 } else if (component_idx == 2) {
512 dptr = frame->cr.get();
513 width = dinfo.image_width / 2;
517 const uint8_t *sptr = (const uint8_t *)mapped + image.offsets[component_idx];
518 size_t spitch = image.pitches[component_idx];
519 for (size_t y = 0; y < dinfo.image_height; ++y) {
520 memcpy(dptr + y * width, sptr + y * spitch, width);
524 // Convert Y'CbCr to separate Y' and CbCr.
525 frame->is_semiplanar = true;
526 frame->y.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
527 frame->cbcr.reset(new uint8_t[dinfo.image_width * dinfo.image_height]);
528 const uint8_t *src = (const uint8_t *)mapped + resources.image.offsets[0];
529 if (resources.image.pitches[0] == dinfo.image_width * 2) {
530 memcpy_interleaved(frame->cbcr.get(), frame->y.get(), src, dinfo.image_width * dinfo.image_height * 2);
532 for (unsigned y = 0; y < dinfo.image_height; ++y) {
533 memcpy_interleaved(frame->cbcr.get() + y * dinfo.image_width, frame->y.get() + y * dinfo.image_width,
534 src + y * resources.image.pitches[0], dinfo.image_width * 2);
538 frame->width = dinfo.image_width;
539 frame->height = dinfo.image_height;
540 frame->chroma_subsampling_x = 2;
541 frame->chroma_subsampling_y = 1;
542 frame->pitch_y = dinfo.image_width;
543 frame->pitch_chroma = dinfo.image_width / 2;
545 va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.image.buf);
546 CHECK_VASTATUS_RET(va_status, "vaUnmapBuffer");