]> git.sesse.net Git - ffmpeg/blob - libavcodec/dxva2.c
avcodec/x86: allow future 8-bit simple idct to have "DC only hack"
[ffmpeg] / libavcodec / dxva2.c
1 /*
2  * DXVA2 HW acceleration.
3  *
4  * copyright (c) 2010 Laurent Aimar
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 <assert.h>
24 #include <string.h>
25 #include <initguid.h>
26
27 #include "libavutil/common.h"
28 #include "libavutil/log.h"
29 #include "libavutil/time.h"
30
31 #include "avcodec.h"
32 #include "dxva2_internal.h"
33
34 /* define all the GUIDs used directly here,
35  to avoid problems with inconsistent dxva2api.h versions in mingw-w64 and different MSVC version */
36 DEFINE_GUID(ff_DXVA2_ModeMPEG2_VLD,      0xee27417f, 0x5e28,0x4e65,0xbe,0xea,0x1d,0x26,0xb5,0x08,0xad,0xc9);
37 DEFINE_GUID(ff_DXVA2_ModeMPEG2and1_VLD,  0x86695f12, 0x340e,0x4f04,0x9f,0xd3,0x92,0x53,0xdd,0x32,0x74,0x60);
38 DEFINE_GUID(ff_DXVA2_ModeH264_E,         0x1b81be68, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
39 DEFINE_GUID(ff_DXVA2_ModeH264_F,         0x1b81be69, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
40 DEFINE_GUID(ff_DXVADDI_Intel_ModeH264_E, 0x604F8E68, 0x4951,0x4C54,0x88,0xFE,0xAB,0xD2,0x5C,0x15,0xB3,0xD6);
41 DEFINE_GUID(ff_DXVA2_ModeVC1_D,          0x1b81beA3, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
42 DEFINE_GUID(ff_DXVA2_ModeVC1_D2010,      0x1b81beA4, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
43 DEFINE_GUID(ff_DXVA2_ModeHEVC_VLD_Main,  0x5b11d51b, 0x2f4c,0x4452,0xbc,0xc3,0x09,0xf2,0xa1,0x16,0x0c,0xc0);
44 DEFINE_GUID(ff_DXVA2_ModeHEVC_VLD_Main10,0x107af0e0, 0xef1a,0x4d19,0xab,0xa8,0x67,0xa1,0x63,0x07,0x3d,0x13);
45 DEFINE_GUID(ff_DXVA2_ModeVP9_VLD_Profile0,0x463707f8,0xa1d0,0x4585,0x87,0x6d,0x83,0xaa,0x6d,0x60,0xb8,0x9e);
46 DEFINE_GUID(ff_DXVA2_NoEncrypt,          0x1b81beD0, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
47 DEFINE_GUID(ff_GUID_NULL,                0x00000000, 0x0000,0x0000,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00);
48 DEFINE_GUID(ff_IID_IDirectXVideoDecoderService, 0xfc51a551,0xd5e7,0x11d9,0xaf,0x55,0x00,0x05,0x4e,0x43,0xff,0x02);
49
50 typedef struct dxva_mode {
51     const GUID     *guid;
52     enum AVCodecID codec;
53     // List of supported profiles, terminated by a FF_PROFILE_UNKNOWN entry.
54     // If NULL, don't check profile.
55     const int      *profiles;
56 } dxva_mode;
57
58 static const int prof_mpeg2_main[]   = {FF_PROFILE_MPEG2_SIMPLE,
59                                         FF_PROFILE_MPEG2_MAIN,
60                                         FF_PROFILE_UNKNOWN};
61 static const int prof_h264_high[]    = {FF_PROFILE_H264_CONSTRAINED_BASELINE,
62                                         FF_PROFILE_H264_MAIN,
63                                         FF_PROFILE_H264_HIGH,
64                                         FF_PROFILE_UNKNOWN};
65 static const int prof_hevc_main[]    = {FF_PROFILE_HEVC_MAIN,
66                                         FF_PROFILE_UNKNOWN};
67 static const int prof_hevc_main10[]  = {FF_PROFILE_HEVC_MAIN,
68                                         FF_PROFILE_HEVC_MAIN_10,
69                                         FF_PROFILE_UNKNOWN};
70
71 static const dxva_mode dxva_modes[] = {
72     /* MPEG-2 */
73     { &ff_DXVA2_ModeMPEG2_VLD,       AV_CODEC_ID_MPEG2VIDEO, prof_mpeg2_main },
74     { &ff_DXVA2_ModeMPEG2and1_VLD,   AV_CODEC_ID_MPEG2VIDEO, prof_mpeg2_main },
75
76     /* H.264 */
77     { &ff_DXVA2_ModeH264_F,          AV_CODEC_ID_H264, prof_h264_high },
78     { &ff_DXVA2_ModeH264_E,          AV_CODEC_ID_H264, prof_h264_high },
79     /* Intel specific H.264 mode */
80     { &ff_DXVADDI_Intel_ModeH264_E,  AV_CODEC_ID_H264, prof_h264_high },
81
82     /* VC-1 / WMV3 */
83     { &ff_DXVA2_ModeVC1_D2010,       AV_CODEC_ID_VC1 },
84     { &ff_DXVA2_ModeVC1_D2010,       AV_CODEC_ID_WMV3 },
85     { &ff_DXVA2_ModeVC1_D,           AV_CODEC_ID_VC1 },
86     { &ff_DXVA2_ModeVC1_D,           AV_CODEC_ID_WMV3 },
87
88     /* HEVC/H.265 */
89     { &ff_DXVA2_ModeHEVC_VLD_Main10, AV_CODEC_ID_HEVC, prof_hevc_main10 },
90     { &ff_DXVA2_ModeHEVC_VLD_Main,   AV_CODEC_ID_HEVC, prof_hevc_main },
91
92     /* VP8/9 */
93     { &ff_DXVA2_ModeVP9_VLD_Profile0,AV_CODEC_ID_VP9 },
94
95     { NULL,                          0 },
96 };
97
98 static int dxva_get_decoder_configuration(AVCodecContext *avctx,
99                                           const void *cfg_list,
100                                           unsigned cfg_count)
101 {
102     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
103     unsigned i, best_score = 0;
104     int best_cfg = -1;
105
106     for (i = 0; i < cfg_count; i++) {
107         unsigned score;
108         UINT ConfigBitstreamRaw;
109         GUID guidConfigBitstreamEncryption;
110
111 #if CONFIG_D3D11VA
112         if (sctx->pix_fmt == AV_PIX_FMT_D3D11) {
113             D3D11_VIDEO_DECODER_CONFIG *cfg = &((D3D11_VIDEO_DECODER_CONFIG *)cfg_list)[i];
114             ConfigBitstreamRaw = cfg->ConfigBitstreamRaw;
115             guidConfigBitstreamEncryption = cfg->guidConfigBitstreamEncryption;
116         }
117 #endif
118 #if CONFIG_DXVA2
119         if (sctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
120             DXVA2_ConfigPictureDecode *cfg = &((DXVA2_ConfigPictureDecode *)cfg_list)[i];
121             ConfigBitstreamRaw = cfg->ConfigBitstreamRaw;
122             guidConfigBitstreamEncryption = cfg->guidConfigBitstreamEncryption;
123         }
124 #endif
125
126         if (ConfigBitstreamRaw == 1)
127             score = 1;
128         else if (avctx->codec_id == AV_CODEC_ID_H264 && ConfigBitstreamRaw == 2)
129             score = 2;
130         else
131             continue;
132         if (IsEqualGUID(&guidConfigBitstreamEncryption, &ff_DXVA2_NoEncrypt))
133             score += 16;
134         if (score > best_score) {
135             best_score = score;
136             best_cfg = i;
137         }
138     }
139
140     if (!best_score) {
141         av_log(avctx, AV_LOG_VERBOSE, "No valid decoder configuration available\n");
142         return AVERROR(EINVAL);
143     }
144
145     return best_cfg;
146 }
147
148 #if CONFIG_D3D11VA
149 static int d3d11va_validate_output(void *service, GUID guid, const void *surface_format)
150 {
151     HRESULT hr;
152     BOOL is_supported = FALSE;
153     hr = ID3D11VideoDevice_CheckVideoDecoderFormat((ID3D11VideoDevice *)service,
154                                                    &guid,
155                                                    *(DXGI_FORMAT *)surface_format,
156                                                    &is_supported);
157     return SUCCEEDED(hr) && is_supported;
158 }
159 #endif
160
161 #if CONFIG_DXVA2
162 static int dxva2_validate_output(void *decoder_service, GUID guid, const void *surface_format)
163 {
164     HRESULT hr;
165     int ret = 0;
166     unsigned j, target_count;
167     D3DFORMAT *target_list;
168     hr = IDirectXVideoDecoderService_GetDecoderRenderTargets((IDirectXVideoDecoderService *)decoder_service, &guid, &target_count, &target_list);
169     if (SUCCEEDED(hr)) {
170         for (j = 0; j < target_count; j++) {
171             const D3DFORMAT format = target_list[j];
172             if (format == *(D3DFORMAT *)surface_format) {
173                 ret = 1;
174                 break;
175             }
176         }
177         CoTaskMemFree(target_list);
178     }
179     return ret;
180 }
181 #endif
182
183 static int dxva_check_codec_compatibility(AVCodecContext *avctx, const dxva_mode *mode)
184 {
185     if (mode->codec != avctx->codec_id)
186             return 0;
187
188     if (mode->profiles && !(avctx->hwaccel_flags & AV_HWACCEL_FLAG_ALLOW_PROFILE_MISMATCH)) {
189         int i, found = 0;
190         for (i = 0; mode->profiles[i] != FF_PROFILE_UNKNOWN; i++) {
191             if (avctx->profile == mode->profiles[i]) {
192                 found = 1;
193                 break;
194             }
195         }
196         if (!found)
197             return 0;
198     }
199
200     return 1;
201 }
202
203 static void dxva_list_guids_debug(AVCodecContext *avctx, void *service,
204                                  unsigned guid_count, const GUID *guid_list)
205 {
206     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
207     int i;
208
209     av_log(avctx, AV_LOG_VERBOSE, "Decoder GUIDs reported as supported:\n");
210
211     for (i = 0; i < guid_count; i++) {
212         const GUID *guid = &guid_list[i];
213
214         av_log(avctx, AV_LOG_VERBOSE,
215              "{%8.8x-%4.4x-%4.4x-%2.2x%2.2x-%2.2x%2.2x%2.2x%2.2x%2.2x%2.2x}",
216              (unsigned) guid->Data1, guid->Data2, guid->Data3,
217              guid->Data4[0], guid->Data4[1],
218              guid->Data4[2], guid->Data4[3],
219              guid->Data4[4], guid->Data4[5],
220              guid->Data4[6], guid->Data4[7]);
221
222 #if CONFIG_D3D11VA
223         if (sctx->pix_fmt == AV_PIX_FMT_D3D11) {
224             DXGI_FORMAT format;
225             // We don't know the maximum valid DXGI_FORMAT, so use 200 as
226             // arbitrary upper bound (that could become outdated).
227             for (format = 0; format < 200; format++) {
228                 if (d3d11va_validate_output(service, *guid, &format))
229                     av_log(avctx, AV_LOG_VERBOSE, " %d", (int)format);
230             }
231         }
232 #endif
233 #if CONFIG_DXVA2
234         if (sctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
235             const D3DFORMAT formats[] = {MKTAG('N', 'V', '1', '2'),
236                                          MKTAG('P', '0', '1', '0')};
237             int i;
238             for (i = 0; i < FF_ARRAY_ELEMS(formats); i++) {
239                 if (dxva2_validate_output(service, *guid, &formats[i]))
240                     av_log(avctx, AV_LOG_VERBOSE, " %d", i);
241             }
242         }
243 #endif
244         av_log(avctx, AV_LOG_VERBOSE, "\n");
245     }
246 }
247
248 static int dxva_get_decoder_guid(AVCodecContext *avctx, void *service, void *surface_format,
249                                  unsigned guid_count, const GUID *guid_list, GUID *decoder_guid)
250 {
251     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
252     unsigned i, j;
253
254     dxva_list_guids_debug(avctx, service, guid_count, guid_list);
255
256     *decoder_guid = ff_GUID_NULL;
257     for (i = 0; dxva_modes[i].guid; i++) {
258         const dxva_mode *mode = &dxva_modes[i];
259         int validate;
260         if (!dxva_check_codec_compatibility(avctx, mode))
261             continue;
262
263         for (j = 0; j < guid_count; j++) {
264             if (IsEqualGUID(mode->guid, &guid_list[j]))
265                 break;
266         }
267         if (j == guid_count)
268             continue;
269
270 #if CONFIG_D3D11VA
271         if (sctx->pix_fmt == AV_PIX_FMT_D3D11)
272             validate = d3d11va_validate_output(service, *mode->guid, surface_format);
273 #endif
274 #if CONFIG_DXVA2
275         if (sctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
276             validate = dxva2_validate_output(service, *mode->guid, surface_format);
277 #endif
278         if (validate) {
279             *decoder_guid = *mode->guid;
280             break;
281         }
282     }
283
284     if (IsEqualGUID(decoder_guid, &ff_GUID_NULL)) {
285         av_log(avctx, AV_LOG_VERBOSE, "No decoder device for codec found\n");
286         return AVERROR(EINVAL);
287     }
288
289     if (IsEqualGUID(decoder_guid, &ff_DXVADDI_Intel_ModeH264_E))
290         sctx->workaround |= FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO;
291
292     return 0;
293 }
294
295 static void bufref_free_interface(void *opaque, uint8_t *data)
296 {
297     IUnknown_Release((IUnknown *)opaque);
298 }
299
300 static AVBufferRef *bufref_wrap_interface(IUnknown *iface)
301 {
302     return av_buffer_create((uint8_t*)iface, 1, bufref_free_interface, iface, 0);
303 }
304
305 #if CONFIG_DXVA2
306
307 static int dxva2_get_decoder_configuration(AVCodecContext *avctx, const GUID *device_guid,
308                                            const DXVA2_VideoDesc *desc,
309                                            DXVA2_ConfigPictureDecode *config)
310 {
311     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
312     unsigned cfg_count;
313     DXVA2_ConfigPictureDecode *cfg_list;
314     HRESULT hr;
315     int ret;
316
317     hr = IDirectXVideoDecoderService_GetDecoderConfigurations(sctx->dxva2_service, device_guid, desc, NULL, &cfg_count, &cfg_list);
318     if (FAILED(hr)) {
319         av_log(avctx, AV_LOG_ERROR, "Unable to retrieve decoder configurations\n");
320         return AVERROR(EINVAL);
321     }
322
323     ret = dxva_get_decoder_configuration(avctx, cfg_list, cfg_count);
324     if (ret >= 0)
325         *config = cfg_list[ret];
326     CoTaskMemFree(cfg_list);
327     return ret;
328 }
329
330 static int dxva2_create_decoder(AVCodecContext *avctx)
331 {
332     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
333     GUID *guid_list;
334     unsigned guid_count;
335     GUID device_guid;
336     D3DFORMAT surface_format = avctx->sw_pix_fmt == AV_PIX_FMT_YUV420P10 ?
337                                MKTAG('P', '0', '1', '0') : MKTAG('N', 'V', '1', '2');
338     DXVA2_VideoDesc desc = { 0 };
339     DXVA2_ConfigPictureDecode config;
340     HRESULT hr;
341     int ret;
342     HANDLE device_handle;
343     AVHWFramesContext *frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
344     AVDXVA2FramesContext *frames_hwctx = frames_ctx->hwctx;
345     AVDXVA2DeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
346
347     hr = IDirect3DDeviceManager9_OpenDeviceHandle(device_hwctx->devmgr,
348                                                   &device_handle);
349     if (FAILED(hr)) {
350         av_log(avctx, AV_LOG_ERROR, "Failed to open a device handle\n");
351         goto fail;
352     }
353
354     hr = IDirect3DDeviceManager9_GetVideoService(device_hwctx->devmgr, device_handle,
355                                                  &ff_IID_IDirectXVideoDecoderService,
356                                                  (void **)&sctx->dxva2_service);
357     IDirect3DDeviceManager9_CloseDeviceHandle(device_hwctx->devmgr, device_handle);
358     if (FAILED(hr)) {
359         av_log(avctx, AV_LOG_ERROR, "Failed to create IDirectXVideoDecoderService\n");
360         goto fail;
361     }
362
363     hr = IDirectXVideoDecoderService_GetDecoderDeviceGuids(sctx->dxva2_service, &guid_count, &guid_list);
364     if (FAILED(hr)) {
365         av_log(avctx, AV_LOG_ERROR, "Failed to retrieve decoder device GUIDs\n");
366         goto fail;
367     }
368
369     ret = dxva_get_decoder_guid(avctx, sctx->dxva2_service, &surface_format,
370                                 guid_count, guid_list, &device_guid);
371     CoTaskMemFree(guid_list);
372     if (ret < 0) {
373         goto fail;
374     }
375
376     desc.SampleWidth  = avctx->coded_width;
377     desc.SampleHeight = avctx->coded_height;
378     desc.Format       = surface_format;
379
380     ret = dxva2_get_decoder_configuration(avctx, &device_guid, &desc, &config);
381     if (ret < 0) {
382         goto fail;
383     }
384
385     hr = IDirectXVideoDecoderService_CreateVideoDecoder(sctx->dxva2_service, &device_guid,
386                                                         &desc, &config, frames_hwctx->surfaces,
387                                                         frames_hwctx->nb_surfaces, &sctx->dxva2_decoder);
388     if (FAILED(hr)) {
389         av_log(avctx, AV_LOG_ERROR, "Failed to create DXVA2 video decoder\n");
390         goto fail;
391     }
392
393     sctx->dxva2_config = config;
394
395     sctx->decoder_ref = bufref_wrap_interface((IUnknown *)sctx->dxva2_decoder);
396     if (!sctx->decoder_ref)
397         return AVERROR(ENOMEM);
398
399     return 0;
400 fail:
401     return AVERROR(EINVAL);
402 }
403
404 #endif
405
406 #if CONFIG_D3D11VA
407
408 static int d3d11va_get_decoder_configuration(AVCodecContext *avctx,
409                                              ID3D11VideoDevice *video_device,
410                                              const D3D11_VIDEO_DECODER_DESC *desc,
411                                              D3D11_VIDEO_DECODER_CONFIG *config)
412 {
413     unsigned cfg_count = 0;
414     D3D11_VIDEO_DECODER_CONFIG *cfg_list = NULL;
415     HRESULT hr;
416     int i, ret;
417
418     hr = ID3D11VideoDevice_GetVideoDecoderConfigCount(video_device, desc, &cfg_count);
419     if (FAILED(hr)) {
420         av_log(avctx, AV_LOG_ERROR, "Unable to retrieve decoder configurations\n");
421         return AVERROR(EINVAL);
422     }
423
424     cfg_list = av_malloc_array(cfg_count, sizeof(D3D11_VIDEO_DECODER_CONFIG));
425     if (cfg_list == NULL)
426         return AVERROR(ENOMEM);
427     for (i = 0; i < cfg_count; i++) {
428         hr = ID3D11VideoDevice_GetVideoDecoderConfig(video_device, desc, i, &cfg_list[i]);
429         if (FAILED(hr)) {
430             av_log(avctx, AV_LOG_ERROR, "Unable to retrieve decoder configurations. (hr=0x%lX)\n", hr);
431             av_free(cfg_list);
432             return AVERROR(EINVAL);
433         }
434     }
435
436     ret = dxva_get_decoder_configuration(avctx, cfg_list, cfg_count);
437     if (ret >= 0)
438         *config = cfg_list[ret];
439     av_free(cfg_list);
440     return ret;
441 }
442
443 static DXGI_FORMAT d3d11va_map_sw_to_hw_format(enum AVPixelFormat pix_fmt)
444 {
445     switch (pix_fmt) {
446     case AV_PIX_FMT_NV12:       return DXGI_FORMAT_NV12;
447     case AV_PIX_FMT_P010:       return DXGI_FORMAT_P010;
448     case AV_PIX_FMT_YUV420P:    return DXGI_FORMAT_420_OPAQUE;
449     default:                    return DXGI_FORMAT_UNKNOWN;
450     }
451 }
452
453 static int d3d11va_create_decoder(AVCodecContext *avctx)
454 {
455     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
456     GUID *guid_list;
457     unsigned guid_count, i;
458     GUID decoder_guid;
459     D3D11_VIDEO_DECODER_DESC desc = { 0 };
460     D3D11_VIDEO_DECODER_CONFIG config;
461     AVHWFramesContext *frames_ctx = (AVHWFramesContext *)avctx->hw_frames_ctx->data;
462     AVD3D11VADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
463     AVD3D11VAFramesContext *frames_hwctx = frames_ctx->hwctx;
464     DXGI_FORMAT surface_format = d3d11va_map_sw_to_hw_format(frames_ctx->sw_format);
465     D3D11_TEXTURE2D_DESC texdesc;
466     HRESULT hr;
467     int ret;
468
469     if (!frames_hwctx->texture) {
470         av_log(avctx, AV_LOG_ERROR, "AVD3D11VAFramesContext.texture not set.\n");
471         return AVERROR(EINVAL);
472     }
473     ID3D11Texture2D_GetDesc(frames_hwctx->texture, &texdesc);
474
475     guid_count = ID3D11VideoDevice_GetVideoDecoderProfileCount(device_hwctx->video_device);
476     guid_list = av_malloc_array(guid_count, sizeof(*guid_list));
477     if (guid_list == NULL || guid_count == 0) {
478         av_log(avctx, AV_LOG_ERROR, "Failed to get the decoder GUIDs\n");
479         av_free(guid_list);
480         return AVERROR(EINVAL);
481     }
482     for (i = 0; i < guid_count; i++) {
483         hr = ID3D11VideoDevice_GetVideoDecoderProfile(device_hwctx->video_device, i, &guid_list[i]);
484         if (FAILED(hr)) {
485             av_log(avctx, AV_LOG_ERROR, "Failed to retrieve decoder GUID %d\n", i);
486             av_free(guid_list);
487             return AVERROR(EINVAL);
488         }
489     }
490
491     ret = dxva_get_decoder_guid(avctx, device_hwctx->video_device, &surface_format,
492                                 guid_count, guid_list, &decoder_guid);
493     av_free(guid_list);
494     if (ret < 0)
495         return AVERROR(EINVAL);
496
497     desc.SampleWidth  = avctx->coded_width;
498     desc.SampleHeight = avctx->coded_height;
499     desc.OutputFormat = surface_format;
500     desc.Guid         = decoder_guid;
501
502     ret = d3d11va_get_decoder_configuration(avctx, device_hwctx->video_device, &desc, &config);
503     if (ret < 0)
504         return AVERROR(EINVAL);
505
506     sctx->d3d11_views = av_mallocz_array(texdesc.ArraySize, sizeof(sctx->d3d11_views[0]));
507     if (!sctx->d3d11_views)
508         return AVERROR(ENOMEM);
509     sctx->nb_d3d11_views = texdesc.ArraySize;
510
511     for (i = 0; i < sctx->nb_d3d11_views; i++) {
512         D3D11_VIDEO_DECODER_OUTPUT_VIEW_DESC viewDesc = {
513             .DecodeProfile = decoder_guid,
514             .ViewDimension = D3D11_VDOV_DIMENSION_TEXTURE2D,
515             .Texture2D = {
516                 .ArraySlice = i,
517             }
518         };
519         hr = ID3D11VideoDevice_CreateVideoDecoderOutputView(device_hwctx->video_device,
520                                                             (ID3D11Resource*) frames_hwctx->texture,
521                                                             &viewDesc,
522                                                             (ID3D11VideoDecoderOutputView**) &sctx->d3d11_views[i]);
523         if (FAILED(hr)) {
524             av_log(avctx, AV_LOG_ERROR, "Could not create the decoder output view %d\n", i);
525             return AVERROR_UNKNOWN;
526         }
527     }
528
529     hr = ID3D11VideoDevice_CreateVideoDecoder(device_hwctx->video_device, &desc,
530                                               &config, &sctx->d3d11_decoder);
531     if (FAILED(hr)) {
532         av_log(avctx, AV_LOG_ERROR, "Failed to create D3D11VA video decoder\n");
533         return AVERROR(EINVAL);
534     }
535
536     sctx->d3d11_config = config;
537     sctx->d3d11_texture = frames_hwctx->texture;
538
539     sctx->decoder_ref = bufref_wrap_interface((IUnknown *)sctx->d3d11_decoder);
540     if (!sctx->decoder_ref)
541         return AVERROR(ENOMEM);
542
543     return 0;
544 }
545
546 #endif
547
548 static void ff_dxva2_lock(AVCodecContext *avctx)
549 {
550 #if CONFIG_D3D11VA
551     if (ff_dxva2_is_d3d11(avctx)) {
552         FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
553         AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
554         if (D3D11VA_CONTEXT(ctx)->context_mutex != INVALID_HANDLE_VALUE)
555             WaitForSingleObjectEx(D3D11VA_CONTEXT(ctx)->context_mutex, INFINITE, FALSE);
556         if (sctx->device_ctx) {
557             AVD3D11VADeviceContext *hwctx = sctx->device_ctx->hwctx;
558             hwctx->lock(hwctx->lock_ctx);
559         }
560     }
561 #endif
562 }
563
564 static void ff_dxva2_unlock(AVCodecContext *avctx)
565 {
566 #if CONFIG_D3D11VA
567     if (ff_dxva2_is_d3d11(avctx)) {
568         FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
569         AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
570         if (D3D11VA_CONTEXT(ctx)->context_mutex != INVALID_HANDLE_VALUE)
571             ReleaseMutex(D3D11VA_CONTEXT(ctx)->context_mutex);
572         if (sctx->device_ctx) {
573             AVD3D11VADeviceContext *hwctx = sctx->device_ctx->hwctx;
574             hwctx->unlock(hwctx->lock_ctx);
575         }
576     }
577 #endif
578 }
579
580 // This must work before the decoder is created.
581 // This somehow needs to be exported to the user.
582 static void dxva_adjust_hwframes(AVCodecContext *avctx, AVHWFramesContext *frames_ctx)
583 {
584     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
585     int surface_alignment, num_surfaces;
586
587     frames_ctx->format = sctx->pix_fmt;
588
589     /* decoding MPEG-2 requires additional alignment on some Intel GPUs,
590     but it causes issues for H.264 on certain AMD GPUs..... */
591     if (avctx->codec_id == AV_CODEC_ID_MPEG2VIDEO)
592         surface_alignment = 32;
593     /* the HEVC DXVA2 spec asks for 128 pixel aligned surfaces to ensure
594     all coding features have enough room to work with */
595     else if (avctx->codec_id == AV_CODEC_ID_HEVC)
596         surface_alignment = 128;
597     else
598         surface_alignment = 16;
599
600     /* 4 base work surfaces */
601     num_surfaces = 4;
602
603     /* add surfaces based on number of possible refs */
604     if (avctx->codec_id == AV_CODEC_ID_H264 || avctx->codec_id == AV_CODEC_ID_HEVC)
605         num_surfaces += 16;
606     else if (avctx->codec_id == AV_CODEC_ID_VP9)
607         num_surfaces += 8;
608     else
609         num_surfaces += 2;
610
611     /* add extra surfaces for frame threading */
612     if (avctx->active_thread_type & FF_THREAD_FRAME)
613         num_surfaces += avctx->thread_count;
614
615     frames_ctx->sw_format = avctx->sw_pix_fmt == AV_PIX_FMT_YUV420P10 ?
616                             AV_PIX_FMT_P010 : AV_PIX_FMT_NV12;
617     frames_ctx->width = FFALIGN(avctx->coded_width, surface_alignment);
618     frames_ctx->height = FFALIGN(avctx->coded_height, surface_alignment);
619     frames_ctx->initial_pool_size = num_surfaces;
620
621
622 #if CONFIG_DXVA2
623     if (frames_ctx->format == AV_PIX_FMT_DXVA2_VLD) {
624         AVDXVA2FramesContext *frames_hwctx = frames_ctx->hwctx;
625
626         frames_hwctx->surface_type = DXVA2_VideoDecoderRenderTarget;
627     }
628 #endif
629
630 #if CONFIG_D3D11VA
631     if (frames_ctx->format == AV_PIX_FMT_D3D11) {
632         AVD3D11VAFramesContext *frames_hwctx = frames_ctx->hwctx;
633
634         frames_hwctx->BindFlags |= D3D11_BIND_DECODER;
635     }
636 #endif
637 }
638
639 int ff_dxva2_decode_init(AVCodecContext *avctx)
640 {
641     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
642     AVHWFramesContext *frames_ctx = NULL;
643     int ret = 0;
644
645     // Old API.
646     if (avctx->hwaccel_context)
647         return 0;
648
649     // (avctx->pix_fmt is not updated yet at this point)
650     sctx->pix_fmt = avctx->hwaccel->pix_fmt;
651
652     if (!avctx->hw_frames_ctx && !avctx->hw_device_ctx) {
653         av_log(avctx, AV_LOG_ERROR, "Either a hw_frames_ctx or a hw_device_ctx needs to be set for hardware decoding.\n");
654         return AVERROR(EINVAL);
655     }
656
657     if (avctx->hw_frames_ctx) {
658         frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
659     } else {
660         avctx->hw_frames_ctx = av_hwframe_ctx_alloc(avctx->hw_device_ctx);
661         if (!avctx->hw_frames_ctx)
662             return AVERROR(ENOMEM);
663
664         frames_ctx = (AVHWFramesContext*)avctx->hw_frames_ctx->data;
665
666         dxva_adjust_hwframes(avctx, frames_ctx);
667
668         ret = av_hwframe_ctx_init(avctx->hw_frames_ctx);
669         if (ret < 0)
670             goto fail;
671     }
672
673     sctx->device_ctx = frames_ctx->device_ctx;
674
675     if (frames_ctx->format != sctx->pix_fmt ||
676         !((sctx->pix_fmt == AV_PIX_FMT_D3D11 && CONFIG_D3D11VA) ||
677           (sctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD && CONFIG_DXVA2))) {
678         av_log(avctx, AV_LOG_ERROR, "Invalid pixfmt for hwaccel!\n");
679         ret = AVERROR(EINVAL);
680         goto fail;
681     }
682
683 #if CONFIG_D3D11VA
684     if (sctx->pix_fmt == AV_PIX_FMT_D3D11) {
685         AVD3D11VADeviceContext *device_hwctx = frames_ctx->device_ctx->hwctx;
686         AVD3D11VAContext *d3d11_ctx = &sctx->ctx.d3d11va;
687
688         ff_dxva2_lock(avctx);
689         ret = d3d11va_create_decoder(avctx);
690         ff_dxva2_unlock(avctx);
691         if (ret < 0)
692             goto fail;
693
694         d3d11_ctx->decoder       = sctx->d3d11_decoder;
695         d3d11_ctx->video_context = device_hwctx->video_context;
696         d3d11_ctx->cfg           = &sctx->d3d11_config;
697         d3d11_ctx->surface_count = sctx->nb_d3d11_views;
698         d3d11_ctx->surface       = sctx->d3d11_views;
699         d3d11_ctx->workaround    = sctx->workaround;
700         d3d11_ctx->context_mutex = INVALID_HANDLE_VALUE;
701     }
702 #endif
703
704 #if CONFIG_DXVA2
705     if (sctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
706         AVDXVA2FramesContext *frames_hwctx = frames_ctx->hwctx;
707         struct dxva_context *dxva_ctx = &sctx->ctx.dxva2;
708
709         ff_dxva2_lock(avctx);
710         ret = dxva2_create_decoder(avctx);
711         ff_dxva2_unlock(avctx);
712         if (ret < 0)
713             goto fail;
714
715         dxva_ctx->decoder       = sctx->dxva2_decoder;
716         dxva_ctx->cfg           = &sctx->dxva2_config;
717         dxva_ctx->surface       = frames_hwctx->surfaces;
718         dxva_ctx->surface_count = frames_hwctx->nb_surfaces;
719         dxva_ctx->workaround    = sctx->workaround;
720     }
721 #endif
722
723     return 0;
724
725 fail:
726     ff_dxva2_decode_uninit(avctx);
727     return ret;
728 }
729
730 int ff_dxva2_decode_uninit(AVCodecContext *avctx)
731 {
732     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
733     int i;
734
735     av_buffer_unref(&sctx->decoder_ref);
736
737 #if CONFIG_D3D11VA
738     for (i = 0; i < sctx->nb_d3d11_views; i++) {
739         if (sctx->d3d11_views[i])
740             ID3D11VideoDecoderOutputView_Release(sctx->d3d11_views[i]);
741     }
742     av_freep(&sctx->d3d11_views);
743 #endif
744
745 #if CONFIG_DXVA2
746     if (sctx->dxva2_service)
747         IDirectXVideoDecoderService_Release(sctx->dxva2_service);
748 #endif
749
750     return 0;
751 }
752
753 static void *get_surface(const AVCodecContext *avctx, const AVFrame *frame)
754 {
755 #if CONFIG_D3D11VA
756     if (frame->format == AV_PIX_FMT_D3D11) {
757         FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
758         intptr_t index = (intptr_t)frame->data[1];
759         if (index < 0 || index >= sctx->nb_d3d11_views ||
760             sctx->d3d11_texture != (ID3D11Texture2D *)frame->data[0]) {
761             av_log((void *)avctx, AV_LOG_ERROR, "get_buffer frame is invalid!\n");
762             return NULL;
763         }
764         return sctx->d3d11_views[index];
765     }
766 #endif
767     return frame->data[3];
768 }
769
770 unsigned ff_dxva2_get_surface_index(const AVCodecContext *avctx,
771                                     const AVDXVAContext *ctx,
772                                     const AVFrame *frame)
773 {
774     void *surface = get_surface(avctx, frame);
775     unsigned i;
776
777 #if CONFIG_D3D11VA
778     if (avctx->pix_fmt == AV_PIX_FMT_D3D11)
779         return (intptr_t)frame->data[1];
780     if (avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD) {
781         D3D11_VIDEO_DECODER_OUTPUT_VIEW_DESC viewDesc;
782         ID3D11VideoDecoderOutputView_GetDesc((ID3D11VideoDecoderOutputView*) surface, &viewDesc);
783         return viewDesc.Texture2D.ArraySlice;
784     }
785 #endif
786 #if CONFIG_DXVA2
787     for (i = 0; i < DXVA_CONTEXT_COUNT(avctx, ctx); i++) {
788         if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD && ctx->dxva2.surface[i] == surface)
789             return i;
790     }
791 #endif
792
793     assert(0);
794     return 0;
795 }
796
797 int ff_dxva2_commit_buffer(AVCodecContext *avctx,
798                            AVDXVAContext *ctx,
799                            DECODER_BUFFER_DESC *dsc,
800                            unsigned type, const void *data, unsigned size,
801                            unsigned mb_count)
802 {
803     void     *dxva_data;
804     unsigned dxva_size;
805     int      result;
806     HRESULT hr = 0;
807
808 #if CONFIG_D3D11VA
809     if (ff_dxva2_is_d3d11(avctx))
810         hr = ID3D11VideoContext_GetDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context,
811                                                  D3D11VA_CONTEXT(ctx)->decoder,
812                                                  type,
813                                                  &dxva_size, &dxva_data);
814 #endif
815 #if CONFIG_DXVA2
816     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
817         hr = IDirectXVideoDecoder_GetBuffer(DXVA2_CONTEXT(ctx)->decoder, type,
818                                             &dxva_data, &dxva_size);
819 #endif
820     if (FAILED(hr)) {
821         av_log(avctx, AV_LOG_ERROR, "Failed to get a buffer for %u: 0x%x\n",
822                type, (unsigned)hr);
823         return -1;
824     }
825     if (size <= dxva_size) {
826         memcpy(dxva_data, data, size);
827
828 #if CONFIG_D3D11VA
829         if (ff_dxva2_is_d3d11(avctx)) {
830             D3D11_VIDEO_DECODER_BUFFER_DESC *dsc11 = dsc;
831             memset(dsc11, 0, sizeof(*dsc11));
832             dsc11->BufferType           = type;
833             dsc11->DataSize             = size;
834             dsc11->NumMBsInBuffer       = mb_count;
835         }
836 #endif
837 #if CONFIG_DXVA2
838         if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
839             DXVA2_DecodeBufferDesc *dsc2 = dsc;
840             memset(dsc2, 0, sizeof(*dsc2));
841             dsc2->CompressedBufferType = type;
842             dsc2->DataSize             = size;
843             dsc2->NumMBsInBuffer       = mb_count;
844         }
845 #endif
846
847         result = 0;
848     } else {
849         av_log(avctx, AV_LOG_ERROR, "Buffer for type %u was too small\n", type);
850         result = -1;
851     }
852
853 #if CONFIG_D3D11VA
854     if (ff_dxva2_is_d3d11(avctx))
855         hr = ID3D11VideoContext_ReleaseDecoderBuffer(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder, type);
856 #endif
857 #if CONFIG_DXVA2
858     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
859         hr = IDirectXVideoDecoder_ReleaseBuffer(DXVA2_CONTEXT(ctx)->decoder, type);
860 #endif
861     if (FAILED(hr)) {
862         av_log(avctx, AV_LOG_ERROR,
863                "Failed to release buffer type %u: 0x%x\n",
864                type, (unsigned)hr);
865         result = -1;
866     }
867     return result;
868 }
869
870 static int frame_add_buf(AVFrame *frame, AVBufferRef *ref)
871 {
872     int i;
873
874     for (i = 0; i < AV_NUM_DATA_POINTERS; i++) {
875         if (!frame->buf[i]) {
876             frame->buf[i] = av_buffer_ref(ref);
877             return frame->buf[i] ? 0 : AVERROR(ENOMEM);
878         }
879     }
880
881     // For now we expect that the caller does not use more than
882     // AV_NUM_DATA_POINTERS-1 buffers if the user uses a custom pool.
883     return AVERROR(EINVAL);
884 }
885
886 int ff_dxva2_common_end_frame(AVCodecContext *avctx, AVFrame *frame,
887                               const void *pp, unsigned pp_size,
888                               const void *qm, unsigned qm_size,
889                               int (*commit_bs_si)(AVCodecContext *,
890                                                   DECODER_BUFFER_DESC *bs,
891                                                   DECODER_BUFFER_DESC *slice))
892 {
893     AVDXVAContext *ctx = DXVA_CONTEXT(avctx);
894     unsigned               buffer_count = 0;
895 #if CONFIG_D3D11VA
896     D3D11_VIDEO_DECODER_BUFFER_DESC buffer11[4];
897 #endif
898 #if CONFIG_DXVA2
899     DXVA2_DecodeBufferDesc          buffer2[4];
900 #endif
901     DECODER_BUFFER_DESC             *buffer = NULL, *buffer_slice = NULL;
902     int result, runs = 0;
903     HRESULT hr;
904     unsigned type;
905     FFDXVASharedContext *sctx = DXVA_SHARED_CONTEXT(avctx);
906
907     if (sctx->decoder_ref) {
908         result = frame_add_buf(frame, sctx->decoder_ref);
909         if (result < 0)
910             return result;
911     }
912
913     do {
914         ff_dxva2_lock(avctx);
915 #if CONFIG_D3D11VA
916         if (ff_dxva2_is_d3d11(avctx))
917             hr = ID3D11VideoContext_DecoderBeginFrame(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder,
918                                                       get_surface(avctx, frame),
919                                                       0, NULL);
920 #endif
921 #if CONFIG_DXVA2
922         if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
923             hr = IDirectXVideoDecoder_BeginFrame(DXVA2_CONTEXT(ctx)->decoder,
924                                                  get_surface(avctx, frame),
925                                                  NULL);
926 #endif
927         if (hr != E_PENDING || ++runs > 50)
928             break;
929         ff_dxva2_unlock(avctx);
930         av_usleep(2000);
931     } while(1);
932
933     if (FAILED(hr)) {
934         av_log(avctx, AV_LOG_ERROR, "Failed to begin frame: 0x%x\n", (unsigned)hr);
935         ff_dxva2_unlock(avctx);
936         return -1;
937     }
938
939 #if CONFIG_D3D11VA
940     if (ff_dxva2_is_d3d11(avctx)) {
941         buffer = &buffer11[buffer_count];
942         type = D3D11_VIDEO_DECODER_BUFFER_PICTURE_PARAMETERS;
943     }
944 #endif
945 #if CONFIG_DXVA2
946     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
947         buffer = &buffer2[buffer_count];
948         type = DXVA2_PictureParametersBufferType;
949     }
950 #endif
951     result = ff_dxva2_commit_buffer(avctx, ctx, buffer,
952                                     type,
953                                     pp, pp_size, 0);
954     if (result) {
955         av_log(avctx, AV_LOG_ERROR,
956                "Failed to add picture parameter buffer\n");
957         goto end;
958     }
959     buffer_count++;
960
961     if (qm_size > 0) {
962 #if CONFIG_D3D11VA
963         if (ff_dxva2_is_d3d11(avctx)) {
964             buffer = &buffer11[buffer_count];
965             type = D3D11_VIDEO_DECODER_BUFFER_INVERSE_QUANTIZATION_MATRIX;
966         }
967 #endif
968 #if CONFIG_DXVA2
969         if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
970             buffer = &buffer2[buffer_count];
971             type = DXVA2_InverseQuantizationMatrixBufferType;
972         }
973 #endif
974         result = ff_dxva2_commit_buffer(avctx, ctx, buffer,
975                                         type,
976                                         qm, qm_size, 0);
977         if (result) {
978             av_log(avctx, AV_LOG_ERROR,
979                    "Failed to add inverse quantization matrix buffer\n");
980             goto end;
981         }
982         buffer_count++;
983     }
984
985 #if CONFIG_D3D11VA
986     if (ff_dxva2_is_d3d11(avctx)) {
987         buffer       = &buffer11[buffer_count + 0];
988         buffer_slice = &buffer11[buffer_count + 1];
989     }
990 #endif
991 #if CONFIG_DXVA2
992     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
993         buffer       = &buffer2[buffer_count + 0];
994         buffer_slice = &buffer2[buffer_count + 1];
995     }
996 #endif
997
998     result = commit_bs_si(avctx,
999                           buffer,
1000                           buffer_slice);
1001     if (result) {
1002         av_log(avctx, AV_LOG_ERROR,
1003                "Failed to add bitstream or slice control buffer\n");
1004         goto end;
1005     }
1006     buffer_count += 2;
1007
1008     /* TODO Film Grain when possible */
1009
1010     assert(buffer_count == 1 + (qm_size > 0) + 2);
1011
1012 #if CONFIG_D3D11VA
1013     if (ff_dxva2_is_d3d11(avctx))
1014         hr = ID3D11VideoContext_SubmitDecoderBuffers(D3D11VA_CONTEXT(ctx)->video_context,
1015                                                      D3D11VA_CONTEXT(ctx)->decoder,
1016                                                      buffer_count, buffer11);
1017 #endif
1018 #if CONFIG_DXVA2
1019     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD) {
1020         DXVA2_DecodeExecuteParams exec = {
1021             .NumCompBuffers     = buffer_count,
1022             .pCompressedBuffers = buffer2,
1023             .pExtensionData     = NULL,
1024         };
1025         hr = IDirectXVideoDecoder_Execute(DXVA2_CONTEXT(ctx)->decoder, &exec);
1026     }
1027 #endif
1028     if (FAILED(hr)) {
1029         av_log(avctx, AV_LOG_ERROR, "Failed to execute: 0x%x\n", (unsigned)hr);
1030         result = -1;
1031     }
1032
1033 end:
1034 #if CONFIG_D3D11VA
1035     if (ff_dxva2_is_d3d11(avctx))
1036         hr = ID3D11VideoContext_DecoderEndFrame(D3D11VA_CONTEXT(ctx)->video_context, D3D11VA_CONTEXT(ctx)->decoder);
1037 #endif
1038 #if CONFIG_DXVA2
1039     if (avctx->pix_fmt == AV_PIX_FMT_DXVA2_VLD)
1040         hr = IDirectXVideoDecoder_EndFrame(DXVA2_CONTEXT(ctx)->decoder, NULL);
1041 #endif
1042     ff_dxva2_unlock(avctx);
1043     if (FAILED(hr)) {
1044         av_log(avctx, AV_LOG_ERROR, "Failed to end frame: 0x%x\n", (unsigned)hr);
1045         result = -1;
1046     }
1047
1048     return result;
1049 }
1050
1051 int ff_dxva2_is_d3d11(const AVCodecContext *avctx)
1052 {
1053     if (CONFIG_D3D11VA)
1054         return avctx->pix_fmt == AV_PIX_FMT_D3D11VA_VLD ||
1055                avctx->pix_fmt == AV_PIX_FMT_D3D11;
1056     else
1057         return 0;
1058 }