• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * HW decode acceleration through NVDEC
3  *
4  * Copyright (c) 2016 Anton Khirnov
5  *
6  * This file is part of FFmpeg.
7  *
8  * FFmpeg is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU Lesser General Public
10  * License as published by the Free Software Foundation; either
11  * version 2.1 of the License, or (at your option) any later version.
12  *
13  * FFmpeg is distributed in the hope that it will be useful,
14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * Lesser General Public License for more details.
17  *
18  * You should have received a copy of the GNU Lesser General Public
19  * License along with FFmpeg; if not, write to the Free Software
20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21  */
22 
23 #include "config.h"
24 
25 #include "libavutil/common.h"
26 #include "libavutil/error.h"
27 #include "libavutil/hwcontext.h"
28 #include "libavutil/hwcontext_cuda_internal.h"
29 #include "libavutil/cuda_check.h"
30 #include "libavutil/pixdesc.h"
31 #include "libavutil/pixfmt.h"
32 
33 #include "avcodec.h"
34 #include "decode.h"
35 #include "nvdec.h"
36 #include "internal.h"
37 
38 #if !NVDECAPI_CHECK_VERSION(9, 0)
39 #define cudaVideoSurfaceFormat_YUV444 2
40 #define cudaVideoSurfaceFormat_YUV444_16Bit 3
41 #endif
42 
43 typedef struct NVDECDecoder {
44     CUvideodecoder decoder;
45 
46     AVBufferRef *hw_device_ref;
47     AVBufferRef *real_hw_frames_ref;
48     CUcontext    cuda_ctx;
49     CUstream     stream;
50 
51     CudaFunctions *cudl;
52     CuvidFunctions *cvdl;
53 } NVDECDecoder;
54 
55 typedef struct NVDECFramePool {
56     unsigned int dpb_size;
57     unsigned int nb_allocated;
58 } NVDECFramePool;
59 
60 #define CHECK_CU(x) FF_CUDA_CHECK_DL(logctx, decoder->cudl, x)
61 
map_avcodec_id(enum AVCodecID id)62 static int map_avcodec_id(enum AVCodecID id)
63 {
64     switch (id) {
65 #if CONFIG_AV1_NVDEC_HWACCEL
66     case AV_CODEC_ID_AV1:        return cudaVideoCodec_AV1;
67 #endif
68     case AV_CODEC_ID_H264:       return cudaVideoCodec_H264;
69     case AV_CODEC_ID_HEVC:       return cudaVideoCodec_HEVC;
70     case AV_CODEC_ID_MJPEG:      return cudaVideoCodec_JPEG;
71     case AV_CODEC_ID_MPEG1VIDEO: return cudaVideoCodec_MPEG1;
72     case AV_CODEC_ID_MPEG2VIDEO: return cudaVideoCodec_MPEG2;
73     case AV_CODEC_ID_MPEG4:      return cudaVideoCodec_MPEG4;
74     case AV_CODEC_ID_VC1:        return cudaVideoCodec_VC1;
75     case AV_CODEC_ID_VP8:        return cudaVideoCodec_VP8;
76     case AV_CODEC_ID_VP9:        return cudaVideoCodec_VP9;
77     case AV_CODEC_ID_WMV3:       return cudaVideoCodec_VC1;
78     }
79     return -1;
80 }
81 
map_chroma_format(enum AVPixelFormat pix_fmt)82 static int map_chroma_format(enum AVPixelFormat pix_fmt)
83 {
84     int shift_h = 0, shift_v = 0;
85 
86     if (av_pix_fmt_count_planes(pix_fmt) == 1)
87         return cudaVideoChromaFormat_Monochrome;
88 
89     av_pix_fmt_get_chroma_sub_sample(pix_fmt, &shift_h, &shift_v);
90 
91     if (shift_h == 1 && shift_v == 1)
92         return cudaVideoChromaFormat_420;
93     else if (shift_h == 1 && shift_v == 0)
94         return cudaVideoChromaFormat_422;
95     else if (shift_h == 0 && shift_v == 0)
96         return cudaVideoChromaFormat_444;
97 
98     return -1;
99 }
100 
nvdec_test_capabilities(NVDECDecoder * decoder,CUVIDDECODECREATEINFO * params,void * logctx)101 static int nvdec_test_capabilities(NVDECDecoder *decoder,
102                                    CUVIDDECODECREATEINFO *params, void *logctx)
103 {
104     int ret;
105     CUVIDDECODECAPS caps = { 0 };
106 
107     caps.eCodecType      = params->CodecType;
108     caps.eChromaFormat   = params->ChromaFormat;
109     caps.nBitDepthMinus8 = params->bitDepthMinus8;
110 
111     if (!decoder->cvdl->cuvidGetDecoderCaps) {
112         av_log(logctx, AV_LOG_WARNING, "Used Nvidia driver is too old to perform a capability check.\n");
113         av_log(logctx, AV_LOG_WARNING, "The minimum required version is "
114 #if defined(_WIN32) || defined(__CYGWIN__)
115             "378.66"
116 #else
117             "378.13"
118 #endif
119             ". Continuing blind.\n");
120         return 0;
121     }
122 
123     ret = CHECK_CU(decoder->cvdl->cuvidGetDecoderCaps(&caps));
124     if (ret < 0)
125         return ret;
126 
127     av_log(logctx, AV_LOG_VERBOSE, "NVDEC capabilities:\n");
128     av_log(logctx, AV_LOG_VERBOSE, "format supported: %s, max_mb_count: %d\n",
129            caps.bIsSupported ? "yes" : "no", caps.nMaxMBCount);
130     av_log(logctx, AV_LOG_VERBOSE, "min_width: %d, max_width: %d\n",
131            caps.nMinWidth, caps.nMaxWidth);
132     av_log(logctx, AV_LOG_VERBOSE, "min_height: %d, max_height: %d\n",
133            caps.nMinHeight, caps.nMaxHeight);
134 
135     if (!caps.bIsSupported) {
136         av_log(logctx, AV_LOG_ERROR, "Hardware is lacking required capabilities\n");
137         return AVERROR(EINVAL);
138     }
139 
140     if (params->ulWidth > caps.nMaxWidth || params->ulWidth < caps.nMinWidth) {
141         av_log(logctx, AV_LOG_ERROR, "Video width %d not within range from %d to %d\n",
142                (int)params->ulWidth, caps.nMinWidth, caps.nMaxWidth);
143         return AVERROR(EINVAL);
144     }
145 
146     if (params->ulHeight > caps.nMaxHeight || params->ulHeight < caps.nMinHeight) {
147         av_log(logctx, AV_LOG_ERROR, "Video height %d not within range from %d to %d\n",
148                (int)params->ulHeight, caps.nMinHeight, caps.nMaxHeight);
149         return AVERROR(EINVAL);
150     }
151 
152     if ((params->ulWidth * params->ulHeight) / 256 > caps.nMaxMBCount) {
153         av_log(logctx, AV_LOG_ERROR, "Video macroblock count %d exceeds maximum of %d\n",
154                (int)(params->ulWidth * params->ulHeight) / 256, caps.nMaxMBCount);
155         return AVERROR(EINVAL);
156     }
157 
158     return 0;
159 }
160 
nvdec_decoder_free(void * opaque,uint8_t * data)161 static void nvdec_decoder_free(void *opaque, uint8_t *data)
162 {
163     NVDECDecoder *decoder = (NVDECDecoder*)data;
164 
165     if (decoder->decoder) {
166         void *logctx = decoder->hw_device_ref->data;
167         CUcontext dummy;
168         CHECK_CU(decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx));
169         CHECK_CU(decoder->cvdl->cuvidDestroyDecoder(decoder->decoder));
170         CHECK_CU(decoder->cudl->cuCtxPopCurrent(&dummy));
171     }
172 
173     av_buffer_unref(&decoder->real_hw_frames_ref);
174     av_buffer_unref(&decoder->hw_device_ref);
175 
176     cuvid_free_functions(&decoder->cvdl);
177 
178     av_freep(&decoder);
179 }
180 
nvdec_decoder_create(AVBufferRef ** out,AVBufferRef * hw_device_ref,CUVIDDECODECREATEINFO * params,void * logctx)181 static int nvdec_decoder_create(AVBufferRef **out, AVBufferRef *hw_device_ref,
182                                 CUVIDDECODECREATEINFO *params, void *logctx)
183 {
184     AVHWDeviceContext  *hw_device_ctx = (AVHWDeviceContext*)hw_device_ref->data;
185     AVCUDADeviceContext *device_hwctx = hw_device_ctx->hwctx;
186 
187     AVBufferRef *decoder_ref;
188     NVDECDecoder *decoder;
189 
190     CUcontext dummy;
191     int ret;
192 
193     decoder = av_mallocz(sizeof(*decoder));
194     if (!decoder)
195         return AVERROR(ENOMEM);
196 
197     decoder_ref = av_buffer_create((uint8_t*)decoder, sizeof(*decoder),
198                                    nvdec_decoder_free, NULL, AV_BUFFER_FLAG_READONLY);
199     if (!decoder_ref) {
200         av_freep(&decoder);
201         return AVERROR(ENOMEM);
202     }
203 
204     decoder->hw_device_ref = av_buffer_ref(hw_device_ref);
205     if (!decoder->hw_device_ref) {
206         ret = AVERROR(ENOMEM);
207         goto fail;
208     }
209     decoder->cuda_ctx = device_hwctx->cuda_ctx;
210     decoder->cudl = device_hwctx->internal->cuda_dl;
211     decoder->stream = device_hwctx->stream;
212 
213     ret = cuvid_load_functions(&decoder->cvdl, logctx);
214     if (ret < 0) {
215         av_log(logctx, AV_LOG_ERROR, "Failed loading nvcuvid.\n");
216         goto fail;
217     }
218 
219     ret = CHECK_CU(decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx));
220     if (ret < 0)
221         goto fail;
222 
223     ret = nvdec_test_capabilities(decoder, params, logctx);
224     if (ret < 0) {
225         CHECK_CU(decoder->cudl->cuCtxPopCurrent(&dummy));
226         goto fail;
227     }
228 
229     ret = CHECK_CU(decoder->cvdl->cuvidCreateDecoder(&decoder->decoder, params));
230 
231     CHECK_CU(decoder->cudl->cuCtxPopCurrent(&dummy));
232 
233     if (ret < 0) {
234         goto fail;
235     }
236 
237     *out = decoder_ref;
238 
239     return 0;
240 fail:
241     av_buffer_unref(&decoder_ref);
242     return ret;
243 }
244 
nvdec_decoder_frame_alloc(void * opaque,buffer_size_t size)245 static AVBufferRef *nvdec_decoder_frame_alloc(void *opaque, buffer_size_t size)
246 {
247     NVDECFramePool *pool = opaque;
248     AVBufferRef *ret;
249 
250     if (pool->nb_allocated >= pool->dpb_size)
251         return NULL;
252 
253     ret = av_buffer_alloc(sizeof(unsigned int));
254     if (!ret)
255         return NULL;
256 
257     *(unsigned int*)ret->data = pool->nb_allocated++;
258 
259     return ret;
260 }
261 
ff_nvdec_decode_uninit(AVCodecContext * avctx)262 int ff_nvdec_decode_uninit(AVCodecContext *avctx)
263 {
264     NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
265 
266     av_freep(&ctx->bitstream);
267     av_freep(&ctx->bitstream_internal);
268     ctx->bitstream_len       = 0;
269     ctx->bitstream_allocated = 0;
270 
271     av_freep(&ctx->slice_offsets);
272     ctx->nb_slices               = 0;
273     ctx->slice_offsets_allocated = 0;
274 
275     av_buffer_unref(&ctx->decoder_ref);
276     av_buffer_pool_uninit(&ctx->decoder_pool);
277 
278     return 0;
279 }
280 
nvdec_free_dummy(struct AVHWFramesContext * ctx)281 static void nvdec_free_dummy(struct AVHWFramesContext *ctx)
282 {
283     av_buffer_pool_uninit(&ctx->pool);
284 }
285 
nvdec_alloc_dummy(buffer_size_t size)286 static AVBufferRef *nvdec_alloc_dummy(buffer_size_t size)
287 {
288     return av_buffer_create(NULL, 0, NULL, NULL, 0);
289 }
290 
nvdec_init_hwframes(AVCodecContext * avctx,AVBufferRef ** out_frames_ref,int dummy)291 static int nvdec_init_hwframes(AVCodecContext *avctx, AVBufferRef **out_frames_ref, int dummy)
292 {
293     AVHWFramesContext *frames_ctx;
294     int ret;
295 
296     ret = avcodec_get_hw_frames_parameters(avctx,
297                                            avctx->hw_device_ctx,
298                                            avctx->hwaccel->pix_fmt,
299                                            out_frames_ref);
300     if (ret < 0)
301         return ret;
302 
303     frames_ctx = (AVHWFramesContext*)(*out_frames_ref)->data;
304 
305     if (dummy) {
306         // Copied from ff_decode_get_hw_frames_ctx for compatibility
307         frames_ctx->initial_pool_size += 3;
308 
309         frames_ctx->free = nvdec_free_dummy;
310         frames_ctx->pool = av_buffer_pool_init(0, nvdec_alloc_dummy);
311 
312         if (!frames_ctx->pool) {
313             av_buffer_unref(out_frames_ref);
314             return AVERROR(ENOMEM);
315         }
316     } else {
317         // This is normally not used to actually allocate frames from
318         frames_ctx->initial_pool_size = 0;
319     }
320 
321     ret = av_hwframe_ctx_init(*out_frames_ref);
322     if (ret < 0) {
323         av_buffer_unref(out_frames_ref);
324         return ret;
325     }
326 
327     return 0;
328 }
329 
ff_nvdec_decode_init(AVCodecContext * avctx)330 int ff_nvdec_decode_init(AVCodecContext *avctx)
331 {
332     NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
333 
334     NVDECDecoder        *decoder;
335     AVBufferRef         *real_hw_frames_ref;
336     NVDECFramePool      *pool;
337     AVHWFramesContext   *frames_ctx;
338     const AVPixFmtDescriptor *sw_desc;
339 
340     CUVIDDECODECREATEINFO params = { 0 };
341 
342     cudaVideoSurfaceFormat output_format;
343     int cuvid_codec_type, cuvid_chroma_format, chroma_444;
344     int ret = 0;
345 
346     sw_desc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
347     if (!sw_desc)
348         return AVERROR_BUG;
349 
350     cuvid_codec_type = map_avcodec_id(avctx->codec_id);
351     if (cuvid_codec_type < 0) {
352         av_log(avctx, AV_LOG_ERROR, "Unsupported codec ID\n");
353         return AVERROR_BUG;
354     }
355 
356     cuvid_chroma_format = map_chroma_format(avctx->sw_pix_fmt);
357     if (cuvid_chroma_format < 0) {
358         av_log(avctx, AV_LOG_ERROR, "Unsupported chroma format\n");
359         return AVERROR(ENOSYS);
360     }
361     chroma_444 = ctx->supports_444 && cuvid_chroma_format == cudaVideoChromaFormat_444;
362 
363     if (!avctx->hw_frames_ctx) {
364         ret = nvdec_init_hwframes(avctx, &avctx->hw_frames_ctx, 1);
365         if (ret < 0)
366             return ret;
367 
368         ret = nvdec_init_hwframes(avctx, &real_hw_frames_ref, 0);
369         if (ret < 0)
370             return ret;
371     } else {
372         real_hw_frames_ref = av_buffer_ref(avctx->hw_frames_ctx);
373         if (!real_hw_frames_ref)
374             return AVERROR(ENOMEM);
375     }
376 
377     switch (sw_desc->comp[0].depth) {
378     case 8:
379         output_format = chroma_444 ? cudaVideoSurfaceFormat_YUV444 :
380                                      cudaVideoSurfaceFormat_NV12;
381         break;
382     case 10:
383     case 12:
384         output_format = chroma_444 ? cudaVideoSurfaceFormat_YUV444_16Bit :
385                                      cudaVideoSurfaceFormat_P016;
386         break;
387     default:
388         av_log(avctx, AV_LOG_ERROR, "Unsupported bit depth\n");
389         av_buffer_unref(&real_hw_frames_ref);
390         return AVERROR(ENOSYS);
391     }
392 
393     frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
394 
395     params.ulWidth             = avctx->coded_width;
396     params.ulHeight            = avctx->coded_height;
397     params.ulTargetWidth       = avctx->coded_width;
398     params.ulTargetHeight      = avctx->coded_height;
399     params.bitDepthMinus8      = sw_desc->comp[0].depth - 8;
400     params.OutputFormat        = output_format;
401     params.CodecType           = cuvid_codec_type;
402     params.ChromaFormat        = cuvid_chroma_format;
403     params.ulNumDecodeSurfaces = frames_ctx->initial_pool_size;
404     params.ulNumOutputSurfaces = frames_ctx->initial_pool_size;
405 
406     ret = nvdec_decoder_create(&ctx->decoder_ref, frames_ctx->device_ref, &params, avctx);
407     if (ret < 0) {
408         if (params.ulNumDecodeSurfaces > 32) {
409             av_log(avctx, AV_LOG_WARNING, "Using more than 32 (%d) decode surfaces might cause nvdec to fail.\n",
410                    (int)params.ulNumDecodeSurfaces);
411             av_log(avctx, AV_LOG_WARNING, "Try lowering the amount of threads. Using %d right now.\n",
412                    avctx->thread_count);
413         }
414         av_buffer_unref(&real_hw_frames_ref);
415         return ret;
416     }
417 
418     decoder = (NVDECDecoder*)ctx->decoder_ref->data;
419     decoder->real_hw_frames_ref = real_hw_frames_ref;
420     real_hw_frames_ref = NULL;
421 
422     pool = av_mallocz(sizeof(*pool));
423     if (!pool) {
424         ret = AVERROR(ENOMEM);
425         goto fail;
426     }
427     pool->dpb_size = frames_ctx->initial_pool_size;
428 
429     ctx->decoder_pool = av_buffer_pool_init2(sizeof(int), pool,
430                                              nvdec_decoder_frame_alloc, av_free);
431     if (!ctx->decoder_pool) {
432         ret = AVERROR(ENOMEM);
433         goto fail;
434     }
435 
436     return 0;
437 fail:
438     ff_nvdec_decode_uninit(avctx);
439     return ret;
440 }
441 
nvdec_fdd_priv_free(void * priv)442 static void nvdec_fdd_priv_free(void *priv)
443 {
444     NVDECFrame *cf = priv;
445 
446     if (!cf)
447         return;
448 
449     av_buffer_unref(&cf->idx_ref);
450     av_buffer_unref(&cf->decoder_ref);
451     av_buffer_unref(&cf->ref_idx_ref);
452 
453     av_freep(&priv);
454 }
455 
nvdec_unmap_mapped_frame(void * opaque,uint8_t * data)456 static void nvdec_unmap_mapped_frame(void *opaque, uint8_t *data)
457 {
458     NVDECFrame *unmap_data = (NVDECFrame*)data;
459     NVDECDecoder *decoder = (NVDECDecoder*)unmap_data->decoder_ref->data;
460     void *logctx = decoder->hw_device_ref->data;
461     CUdeviceptr devptr = (CUdeviceptr)opaque;
462     int ret;
463     CUcontext dummy;
464 
465     ret = CHECK_CU(decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx));
466     if (ret < 0)
467         goto finish;
468 
469     CHECK_CU(decoder->cvdl->cuvidUnmapVideoFrame(decoder->decoder, devptr));
470 
471     CHECK_CU(decoder->cudl->cuCtxPopCurrent(&dummy));
472 
473 finish:
474     av_buffer_unref(&unmap_data->idx_ref);
475     av_buffer_unref(&unmap_data->decoder_ref);
476     av_buffer_unref(&unmap_data->ref_idx_ref);
477     av_free(unmap_data);
478 }
479 
nvdec_retrieve_data(void * logctx,AVFrame * frame)480 static int nvdec_retrieve_data(void *logctx, AVFrame *frame)
481 {
482     FrameDecodeData  *fdd = (FrameDecodeData*)frame->private_ref->data;
483     NVDECFrame        *cf = (NVDECFrame*)fdd->hwaccel_priv;
484     NVDECDecoder *decoder = (NVDECDecoder*)cf->decoder_ref->data;
485 
486     AVHWFramesContext *hwctx = (AVHWFramesContext *)frame->hw_frames_ctx->data;
487 
488     CUVIDPROCPARAMS vpp = { 0 };
489     NVDECFrame *unmap_data = NULL;
490 
491     CUcontext dummy;
492     CUdeviceptr devptr;
493 
494     unsigned int pitch, i;
495     unsigned int offset = 0;
496     int shift_h = 0, shift_v = 0;
497     int ret = 0;
498 
499     vpp.progressive_frame = 1;
500     vpp.output_stream = decoder->stream;
501 
502     ret = CHECK_CU(decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx));
503     if (ret < 0)
504         return ret;
505 
506     ret = CHECK_CU(decoder->cvdl->cuvidMapVideoFrame(decoder->decoder,
507                                                      cf->idx, &devptr,
508                                                      &pitch, &vpp));
509     if (ret < 0)
510         goto finish;
511 
512     unmap_data = av_mallocz(sizeof(*unmap_data));
513     if (!unmap_data) {
514         ret = AVERROR(ENOMEM);
515         goto copy_fail;
516     }
517 
518     frame->buf[1] = av_buffer_create((uint8_t *)unmap_data, sizeof(*unmap_data),
519                                      nvdec_unmap_mapped_frame, (void*)devptr,
520                                      AV_BUFFER_FLAG_READONLY);
521     if (!frame->buf[1]) {
522         ret = AVERROR(ENOMEM);
523         goto copy_fail;
524     }
525 
526     av_buffer_unref(&frame->hw_frames_ctx);
527     frame->hw_frames_ctx = av_buffer_ref(decoder->real_hw_frames_ref);
528     if (!frame->hw_frames_ctx) {
529         ret = AVERROR(ENOMEM);
530         goto copy_fail;
531     }
532 
533     unmap_data->idx = cf->idx;
534     unmap_data->idx_ref = av_buffer_ref(cf->idx_ref);
535     unmap_data->decoder_ref = av_buffer_ref(cf->decoder_ref);
536 
537     av_pix_fmt_get_chroma_sub_sample(hwctx->sw_format, &shift_h, &shift_v);
538     for (i = 0; frame->linesize[i]; i++) {
539         frame->data[i] = (uint8_t*)(devptr + offset);
540         frame->linesize[i] = pitch;
541         offset += pitch * (frame->height >> (i ? shift_v : 0));
542     }
543 
544     goto finish;
545 
546 copy_fail:
547     if (!frame->buf[1]) {
548         CHECK_CU(decoder->cvdl->cuvidUnmapVideoFrame(decoder->decoder, devptr));
549         av_freep(&unmap_data);
550     } else {
551         av_buffer_unref(&frame->buf[1]);
552     }
553 
554 finish:
555     CHECK_CU(decoder->cudl->cuCtxPopCurrent(&dummy));
556     return ret;
557 }
558 
ff_nvdec_start_frame(AVCodecContext * avctx,AVFrame * frame)559 int ff_nvdec_start_frame(AVCodecContext *avctx, AVFrame *frame)
560 {
561     NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
562     FrameDecodeData *fdd = (FrameDecodeData*)frame->private_ref->data;
563     NVDECFrame *cf = NULL;
564     int ret;
565 
566     ctx->bitstream_len = 0;
567     ctx->nb_slices     = 0;
568 
569     if (fdd->hwaccel_priv)
570         return 0;
571 
572     cf = av_mallocz(sizeof(*cf));
573     if (!cf)
574         return AVERROR(ENOMEM);
575 
576     cf->decoder_ref = av_buffer_ref(ctx->decoder_ref);
577     if (!cf->decoder_ref) {
578         ret = AVERROR(ENOMEM);
579         goto fail;
580     }
581 
582     cf->idx_ref = av_buffer_pool_get(ctx->decoder_pool);
583     if (!cf->idx_ref) {
584         av_log(avctx, AV_LOG_ERROR, "No decoder surfaces left\n");
585         ret = AVERROR(ENOMEM);
586         goto fail;
587     }
588     cf->ref_idx = cf->idx = *(unsigned int*)cf->idx_ref->data;
589 
590     fdd->hwaccel_priv      = cf;
591     fdd->hwaccel_priv_free = nvdec_fdd_priv_free;
592     fdd->post_process      = nvdec_retrieve_data;
593 
594     return 0;
595 fail:
596     nvdec_fdd_priv_free(cf);
597     return ret;
598 
599 }
600 
ff_nvdec_start_frame_sep_ref(AVCodecContext * avctx,AVFrame * frame,int has_sep_ref)601 int ff_nvdec_start_frame_sep_ref(AVCodecContext *avctx, AVFrame *frame, int has_sep_ref)
602 {
603     NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
604     FrameDecodeData *fdd = (FrameDecodeData*)frame->private_ref->data;
605     NVDECFrame *cf;
606     int ret;
607 
608     ret = ff_nvdec_start_frame(avctx, frame);
609     if (ret < 0)
610         return ret;
611 
612     cf = fdd->hwaccel_priv;
613 
614     if (has_sep_ref) {
615         if (!cf->ref_idx_ref) {
616             cf->ref_idx_ref = av_buffer_pool_get(ctx->decoder_pool);
617             if (!cf->ref_idx_ref) {
618                 av_log(avctx, AV_LOG_ERROR, "No decoder surfaces left\n");
619                 ret = AVERROR(ENOMEM);
620                 goto fail;
621             }
622         }
623         cf->ref_idx = *(unsigned int*)cf->ref_idx_ref->data;
624     } else {
625         av_buffer_unref(&cf->ref_idx_ref);
626         cf->ref_idx = cf->idx;
627     }
628 
629     return 0;
630 fail:
631     nvdec_fdd_priv_free(cf);
632     return ret;
633 }
634 
ff_nvdec_end_frame(AVCodecContext * avctx)635 int ff_nvdec_end_frame(AVCodecContext *avctx)
636 {
637     NVDECContext     *ctx = avctx->internal->hwaccel_priv_data;
638     NVDECDecoder *decoder = (NVDECDecoder*)ctx->decoder_ref->data;
639     void *logctx          = avctx;
640     CUVIDPICPARAMS    *pp = &ctx->pic_params;
641 
642     CUcontext dummy;
643 
644     int ret = 0;
645 
646     pp->nBitstreamDataLen = ctx->bitstream_len;
647     pp->pBitstreamData    = ctx->bitstream;
648     pp->nNumSlices        = ctx->nb_slices;
649     pp->pSliceDataOffsets = ctx->slice_offsets;
650 
651     ret = CHECK_CU(decoder->cudl->cuCtxPushCurrent(decoder->cuda_ctx));
652     if (ret < 0)
653         return ret;
654 
655     ret = CHECK_CU(decoder->cvdl->cuvidDecodePicture(decoder->decoder, &ctx->pic_params));
656     if (ret < 0)
657         goto finish;
658 
659 finish:
660     CHECK_CU(decoder->cudl->cuCtxPopCurrent(&dummy));
661 
662     return ret;
663 }
664 
ff_nvdec_simple_end_frame(AVCodecContext * avctx)665 int ff_nvdec_simple_end_frame(AVCodecContext *avctx)
666 {
667     NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
668     int ret = ff_nvdec_end_frame(avctx);
669     ctx->bitstream = NULL;
670     return ret;
671 }
672 
ff_nvdec_simple_decode_slice(AVCodecContext * avctx,const uint8_t * buffer,uint32_t size)673 int ff_nvdec_simple_decode_slice(AVCodecContext *avctx, const uint8_t *buffer,
674                                  uint32_t size)
675 {
676     NVDECContext *ctx = avctx->internal->hwaccel_priv_data;
677     void *tmp;
678 
679     tmp = av_fast_realloc(ctx->slice_offsets, &ctx->slice_offsets_allocated,
680                           (ctx->nb_slices + 1) * sizeof(*ctx->slice_offsets));
681     if (!tmp)
682         return AVERROR(ENOMEM);
683     ctx->slice_offsets = tmp;
684 
685     if (!ctx->bitstream)
686         ctx->bitstream = (uint8_t*)buffer;
687 
688     ctx->slice_offsets[ctx->nb_slices] = buffer - ctx->bitstream;
689     ctx->bitstream_len += size;
690     ctx->nb_slices++;
691 
692     return 0;
693 }
694 
ff_nvdec_frame_params(AVCodecContext * avctx,AVBufferRef * hw_frames_ctx,int dpb_size,int supports_444)695 int ff_nvdec_frame_params(AVCodecContext *avctx,
696                           AVBufferRef *hw_frames_ctx,
697                           int dpb_size,
698                           int supports_444)
699 {
700     AVHWFramesContext *frames_ctx = (AVHWFramesContext*)hw_frames_ctx->data;
701     const AVPixFmtDescriptor *sw_desc;
702     int cuvid_codec_type, cuvid_chroma_format, chroma_444;
703 
704     sw_desc = av_pix_fmt_desc_get(avctx->sw_pix_fmt);
705     if (!sw_desc)
706         return AVERROR_BUG;
707 
708     cuvid_codec_type = map_avcodec_id(avctx->codec_id);
709     if (cuvid_codec_type < 0) {
710         av_log(avctx, AV_LOG_ERROR, "Unsupported codec ID\n");
711         return AVERROR_BUG;
712     }
713 
714     cuvid_chroma_format = map_chroma_format(avctx->sw_pix_fmt);
715     if (cuvid_chroma_format < 0) {
716         av_log(avctx, AV_LOG_VERBOSE, "Unsupported chroma format\n");
717         return AVERROR(EINVAL);
718     }
719     chroma_444 = supports_444 && cuvid_chroma_format == cudaVideoChromaFormat_444;
720 
721     frames_ctx->format            = AV_PIX_FMT_CUDA;
722     frames_ctx->width             = (avctx->coded_width + 1) & ~1;
723     frames_ctx->height            = (avctx->coded_height + 1) & ~1;
724     /*
725      * We add two extra frames to the pool to account for deinterlacing filters
726      * holding onto their frames.
727      */
728     frames_ctx->initial_pool_size = dpb_size + 2;
729 
730     switch (sw_desc->comp[0].depth) {
731     case 8:
732         frames_ctx->sw_format = chroma_444 ? AV_PIX_FMT_YUV444P : AV_PIX_FMT_NV12;
733         break;
734     case 10:
735         frames_ctx->sw_format = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P010;
736         break;
737     case 12:
738         frames_ctx->sw_format = chroma_444 ? AV_PIX_FMT_YUV444P16 : AV_PIX_FMT_P016;
739         break;
740     default:
741         return AVERROR(EINVAL);
742     }
743 
744     return 0;
745 }
746 
ff_nvdec_get_ref_idx(AVFrame * frame)747 int ff_nvdec_get_ref_idx(AVFrame *frame)
748 {
749     FrameDecodeData *fdd;
750     NVDECFrame *cf;
751 
752     if (!frame || !frame->private_ref)
753         return -1;
754 
755     fdd = (FrameDecodeData*)frame->private_ref->data;
756     cf  = (NVDECFrame*)fdd->hwaccel_priv;
757     if (!cf)
758         return -1;
759 
760     return cf->ref_idx;
761 }
762