a2d6012fa1f9a23a70e33db0111f469f860347c5
[ffmpeg.git] / avconv_dxva2.c
1 /*
2  * This file is part of Libav.
3  *
4  * Libav is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Lesser General Public
6  * License as published by the Free Software Foundation; either
7  * version 2.1 of the License, or (at your option) any later version.
8  *
9  * Libav is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * Lesser General Public License for more details.
13  *
14  * You should have received a copy of the GNU Lesser General Public
15  * License along with Libav; if not, write to the Free Software
16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17  */
18
19 #include <windows.h>
20
21 #ifdef _WIN32_WINNT
22 #undef _WIN32_WINNT
23 #endif
24 #define _WIN32_WINNT 0x0600
25 #define DXVA2API_USE_BITFIELDS
26 #define COBJMACROS
27
28 #include <stdint.h>
29
30 #include <d3d9.h>
31 #include <dxva2api.h>
32
33 #include "avconv.h"
34
35 #include "libavcodec/dxva2.h"
36
37 #include "libavutil/avassert.h"
38 #include "libavutil/buffer.h"
39 #include "libavutil/frame.h"
40 #include "libavutil/imgutils.h"
41 #include "libavutil/pixfmt.h"
42
43 #include "libavutil/hwcontext.h"
44 #include "libavutil/hwcontext_dxva2.h"
45
46 /* define all the GUIDs used directly here,
47    to avoid problems with inconsistent dxva2api.h versions in mingw-w64 and different MSVC version */
48 #include <initguid.h>
49 DEFINE_GUID(IID_IDirectXVideoDecoderService, 0xfc51a551,0xd5e7,0x11d9,0xaf,0x55,0x00,0x05,0x4e,0x43,0xff,0x02);
50
51 DEFINE_GUID(DXVA2_ModeMPEG2_VLD,      0xee27417f, 0x5e28,0x4e65,0xbe,0xea,0x1d,0x26,0xb5,0x08,0xad,0xc9);
52 DEFINE_GUID(DXVA2_ModeMPEG2and1_VLD,  0x86695f12, 0x340e,0x4f04,0x9f,0xd3,0x92,0x53,0xdd,0x32,0x74,0x60);
53 DEFINE_GUID(DXVA2_ModeH264_E,         0x1b81be68, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
54 DEFINE_GUID(DXVA2_ModeH264_F,         0x1b81be69, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
55 DEFINE_GUID(DXVADDI_Intel_ModeH264_E, 0x604F8E68, 0x4951,0x4C54,0x88,0xFE,0xAB,0xD2,0x5C,0x15,0xB3,0xD6);
56 DEFINE_GUID(DXVA2_ModeVC1_D,          0x1b81beA3, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
57 DEFINE_GUID(DXVA2_ModeVC1_D2010,      0x1b81beA4, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
58 DEFINE_GUID(DXVA2_ModeHEVC_VLD_Main,  0x5b11d51b, 0x2f4c,0x4452,0xbc,0xc3,0x09,0xf2,0xa1,0x16,0x0c,0xc0);
59 DEFINE_GUID(DXVA2_NoEncrypt,          0x1b81beD0, 0xa0c7,0x11d3,0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5);
60 DEFINE_GUID(GUID_NULL,                0x00000000, 0x0000,0x0000,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00);
61
62 typedef IDirect3D9* WINAPI pDirect3DCreate9(UINT);
63 typedef HRESULT WINAPI pCreateDeviceManager9(UINT *, IDirect3DDeviceManager9 **);
64
65 typedef struct dxva2_mode {
66   const GUID     *guid;
67   enum AVCodecID codec;
68 } dxva2_mode;
69
70 static const dxva2_mode dxva2_modes[] = {
71     /* MPEG-2 */
72     { &DXVA2_ModeMPEG2_VLD,      AV_CODEC_ID_MPEG2VIDEO },
73     { &DXVA2_ModeMPEG2and1_VLD,  AV_CODEC_ID_MPEG2VIDEO },
74
75     /* H.264 */
76     { &DXVA2_ModeH264_F,         AV_CODEC_ID_H264 },
77     { &DXVA2_ModeH264_E,         AV_CODEC_ID_H264 },
78     /* Intel specific H.264 mode */
79     { &DXVADDI_Intel_ModeH264_E, AV_CODEC_ID_H264 },
80
81     /* VC-1 / WMV3 */
82     { &DXVA2_ModeVC1_D2010,      AV_CODEC_ID_VC1  },
83     { &DXVA2_ModeVC1_D2010,      AV_CODEC_ID_WMV3 },
84     { &DXVA2_ModeVC1_D,          AV_CODEC_ID_VC1  },
85     { &DXVA2_ModeVC1_D,          AV_CODEC_ID_WMV3 },
86
87     /* HEVC/H.265 */
88     { &DXVA2_ModeHEVC_VLD_Main,  AV_CODEC_ID_HEVC },
89
90     { NULL,                      0 },
91 };
92
93 typedef struct DXVA2DevicePriv {
94     HMODULE d3dlib;
95     HMODULE dxva2lib;
96
97     HANDLE  deviceHandle;
98
99     IDirect3D9                  *d3d9;
100     IDirect3DDevice9            *d3d9device;
101 } DXVA2DevicePriv;
102
103 typedef struct DXVA2Context {
104     IDirectXVideoDecoder        *decoder;
105
106     GUID                        decoder_guid;
107     DXVA2_ConfigPictureDecode   decoder_config;
108     IDirectXVideoDecoderService *decoder_service;
109
110     AVFrame                     *tmp_frame;
111
112     AVBufferRef                 *hw_device_ctx;
113     AVBufferRef                 *hw_frames_ctx;
114 } DXVA2Context;
115
116 static void dxva2_device_uninit(AVHWDeviceContext *ctx)
117 {
118     AVDXVA2DeviceContext *hwctx = ctx->hwctx;
119     DXVA2DevicePriv       *priv = ctx->user_opaque;
120
121     if (hwctx->devmgr && priv->deviceHandle != INVALID_HANDLE_VALUE)
122         IDirect3DDeviceManager9_CloseDeviceHandle(hwctx->devmgr, priv->deviceHandle);
123
124     if (hwctx->devmgr)
125         IDirect3DDeviceManager9_Release(hwctx->devmgr);
126
127     if (priv->d3d9device)
128         IDirect3DDevice9_Release(priv->d3d9device);
129
130     if (priv->d3d9)
131         IDirect3D9_Release(priv->d3d9);
132
133     if (priv->d3dlib)
134         FreeLibrary(priv->d3dlib);
135
136     if (priv->dxva2lib)
137         FreeLibrary(priv->dxva2lib);
138
139     av_freep(&ctx->user_opaque);
140 }
141
142 static void dxva2_uninit(AVCodecContext *s)
143 {
144     InputStream  *ist = s->opaque;
145     DXVA2Context *ctx = ist->hwaccel_ctx;
146
147     ist->hwaccel_uninit        = NULL;
148     ist->hwaccel_get_buffer    = NULL;
149     ist->hwaccel_retrieve_data = NULL;
150
151     if (ctx->decoder_service)
152         IDirectXVideoDecoderService_Release(ctx->decoder_service);
153
154     av_buffer_unref(&ctx->hw_frames_ctx);
155     av_buffer_unref(&ctx->hw_device_ctx);
156
157     av_frame_free(&ctx->tmp_frame);
158
159     av_freep(&ist->hwaccel_ctx);
160     av_freep(&s->hwaccel_context);
161 }
162
163 static int dxva2_get_buffer(AVCodecContext *s, AVFrame *frame, int flags)
164 {
165     InputStream  *ist = s->opaque;
166     DXVA2Context *ctx = ist->hwaccel_ctx;
167
168     return av_hwframe_get_buffer(ctx->hw_frames_ctx, frame, 0);
169 }
170
171 static int dxva2_retrieve_data(AVCodecContext *s, AVFrame *frame)
172 {
173     InputStream        *ist = s->opaque;
174     DXVA2Context       *ctx = ist->hwaccel_ctx;
175     int                ret;
176
177     ret = av_hwframe_transfer_data(ctx->tmp_frame, frame, 0);
178     if (ret < 0)
179         return ret;
180
181     ret = av_frame_copy_props(ctx->tmp_frame, frame);
182     if (ret < 0) {
183         av_frame_unref(ctx->tmp_frame);
184         return ret;
185     }
186
187     av_frame_unref(frame);
188     av_frame_move_ref(frame, ctx->tmp_frame);
189
190     return 0;
191 }
192
193 static int dxva2_alloc(AVCodecContext *s)
194 {
195     InputStream  *ist = s->opaque;
196     int loglevel = (ist->hwaccel_id == HWACCEL_AUTO) ? AV_LOG_VERBOSE : AV_LOG_ERROR;
197     DXVA2Context *ctx;
198     pDirect3DCreate9      *createD3D = NULL;
199     pCreateDeviceManager9 *createDeviceManager = NULL;
200     HRESULT hr;
201     D3DPRESENT_PARAMETERS d3dpp = {0};
202     D3DDISPLAYMODE        d3ddm;
203     unsigned resetToken = 0;
204     UINT adapter = D3DADAPTER_DEFAULT;
205
206     AVHWDeviceContext    *device_ctx;
207     AVDXVA2DeviceContext *device_hwctx;
208     DXVA2DevicePriv      *device_priv;
209     int ret;
210
211     ctx = av_mallocz(sizeof(*ctx));
212     if (!ctx)
213         return AVERROR(ENOMEM);
214
215     ist->hwaccel_ctx           = ctx;
216     ist->hwaccel_uninit        = dxva2_uninit;
217     ist->hwaccel_get_buffer    = dxva2_get_buffer;
218     ist->hwaccel_retrieve_data = dxva2_retrieve_data;
219
220     ctx->hw_device_ctx = av_hwdevice_ctx_alloc(AV_HWDEVICE_TYPE_DXVA2);
221     if (!ctx->hw_device_ctx)
222         goto fail;
223
224     device_ctx   = (AVHWDeviceContext*)ctx->hw_device_ctx->data;
225     device_hwctx = device_ctx->hwctx;
226
227     device_priv = av_mallocz(sizeof(*device_priv));
228     if (!device_priv)
229         goto fail;
230
231     device_ctx->user_opaque = device_priv;
232     device_ctx->free        = dxva2_device_uninit;
233
234     device_priv->deviceHandle = INVALID_HANDLE_VALUE;
235
236     device_priv->d3dlib = LoadLibrary("d3d9.dll");
237     if (!device_priv->d3dlib) {
238         av_log(NULL, loglevel, "Failed to load D3D9 library\n");
239         goto fail;
240     }
241     device_priv->dxva2lib = LoadLibrary("dxva2.dll");
242     if (!device_priv->dxva2lib) {
243         av_log(NULL, loglevel, "Failed to load DXVA2 library\n");
244         goto fail;
245     }
246
247     createD3D = (pDirect3DCreate9 *)GetProcAddress(device_priv->d3dlib, "Direct3DCreate9");
248     if (!createD3D) {
249         av_log(NULL, loglevel, "Failed to locate Direct3DCreate9\n");
250         goto fail;
251     }
252     createDeviceManager = (pCreateDeviceManager9 *)GetProcAddress(device_priv->dxva2lib, "DXVA2CreateDirect3DDeviceManager9");
253     if (!createDeviceManager) {
254         av_log(NULL, loglevel, "Failed to locate DXVA2CreateDirect3DDeviceManager9\n");
255         goto fail;
256     }
257
258     device_priv->d3d9 = createD3D(D3D_SDK_VERSION);
259     if (!device_priv->d3d9) {
260         av_log(NULL, loglevel, "Failed to create IDirect3D object\n");
261         goto fail;
262     }
263
264     if (ist->hwaccel_device) {
265         adapter = atoi(ist->hwaccel_device);
266         av_log(NULL, AV_LOG_INFO, "Using HWAccel device %d\n", adapter);
267     }
268
269     IDirect3D9_GetAdapterDisplayMode(device_priv->d3d9, adapter, &d3ddm);
270     d3dpp.Windowed         = TRUE;
271     d3dpp.BackBufferWidth  = 640;
272     d3dpp.BackBufferHeight = 480;
273     d3dpp.BackBufferCount  = 0;
274     d3dpp.BackBufferFormat = d3ddm.Format;
275     d3dpp.SwapEffect       = D3DSWAPEFFECT_DISCARD;
276     d3dpp.Flags            = D3DPRESENTFLAG_VIDEO;
277
278     hr = IDirect3D9_CreateDevice(device_priv->d3d9, adapter, D3DDEVTYPE_HAL, GetShellWindow(),
279                                  D3DCREATE_SOFTWARE_VERTEXPROCESSING | D3DCREATE_MULTITHREADED | D3DCREATE_FPU_PRESERVE,
280                                  &d3dpp, &device_priv->d3d9device);
281     if (FAILED(hr)) {
282         av_log(NULL, loglevel, "Failed to create Direct3D device\n");
283         goto fail;
284     }
285
286     hr = createDeviceManager(&resetToken, &device_hwctx->devmgr);
287     if (FAILED(hr)) {
288         av_log(NULL, loglevel, "Failed to create Direct3D device manager\n");
289         goto fail;
290     }
291
292     hr = IDirect3DDeviceManager9_ResetDevice(device_hwctx->devmgr, device_priv->d3d9device, resetToken);
293     if (FAILED(hr)) {
294         av_log(NULL, loglevel, "Failed to bind Direct3D device to device manager\n");
295         goto fail;
296     }
297
298     hr = IDirect3DDeviceManager9_OpenDeviceHandle(device_hwctx->devmgr, &device_priv->deviceHandle);
299     if (FAILED(hr)) {
300         av_log(NULL, loglevel, "Failed to open device handle\n");
301         goto fail;
302     }
303
304     hr = IDirect3DDeviceManager9_GetVideoService(device_hwctx->devmgr, device_priv->deviceHandle, &IID_IDirectXVideoDecoderService, (void **)&ctx->decoder_service);
305     if (FAILED(hr)) {
306         av_log(NULL, loglevel, "Failed to create IDirectXVideoDecoderService\n");
307         goto fail;
308     }
309
310     ret = av_hwdevice_ctx_init(ctx->hw_device_ctx);
311     if (ret < 0) {
312         av_log(NULL, loglevel, "Failed to initialize the HW device context\n");
313         goto fail;
314     }
315
316     ctx->tmp_frame = av_frame_alloc();
317     if (!ctx->tmp_frame)
318         goto fail;
319
320     s->hwaccel_context = av_mallocz(sizeof(struct dxva_context));
321     if (!s->hwaccel_context)
322         goto fail;
323
324     return 0;
325 fail:
326     dxva2_uninit(s);
327     return AVERROR(EINVAL);
328 }
329
330 static int dxva2_get_decoder_configuration(AVCodecContext *s, const GUID *device_guid,
331                                            const DXVA2_VideoDesc *desc,
332                                            DXVA2_ConfigPictureDecode *config)
333 {
334     InputStream  *ist = s->opaque;
335     int loglevel = (ist->hwaccel_id == HWACCEL_AUTO) ? AV_LOG_VERBOSE : AV_LOG_ERROR;
336     DXVA2Context *ctx = ist->hwaccel_ctx;
337     unsigned cfg_count = 0, best_score = 0;
338     DXVA2_ConfigPictureDecode *cfg_list = NULL;
339     DXVA2_ConfigPictureDecode best_cfg = {{0}};
340     HRESULT hr;
341     int i;
342
343     hr = IDirectXVideoDecoderService_GetDecoderConfigurations(ctx->decoder_service, device_guid, desc, NULL, &cfg_count, &cfg_list);
344     if (FAILED(hr)) {
345         av_log(NULL, loglevel, "Unable to retrieve decoder configurations\n");
346         return AVERROR(EINVAL);
347     }
348
349     for (i = 0; i < cfg_count; i++) {
350         DXVA2_ConfigPictureDecode *cfg = &cfg_list[i];
351
352         unsigned score;
353         if (cfg->ConfigBitstreamRaw == 1)
354             score = 1;
355         else if (s->codec_id == AV_CODEC_ID_H264 && cfg->ConfigBitstreamRaw == 2)
356             score = 2;
357         else
358             continue;
359         if (IsEqualGUID(&cfg->guidConfigBitstreamEncryption, &DXVA2_NoEncrypt))
360             score += 16;
361         if (score > best_score) {
362             best_score = score;
363             best_cfg   = *cfg;
364         }
365     }
366     CoTaskMemFree(cfg_list);
367
368     if (!best_score) {
369         av_log(NULL, loglevel, "No valid decoder configuration available\n");
370         return AVERROR(EINVAL);
371     }
372
373     *config = best_cfg;
374     return 0;
375 }
376
377 static int dxva2_create_decoder(AVCodecContext *s)
378 {
379     InputStream  *ist = s->opaque;
380     int loglevel = (ist->hwaccel_id == HWACCEL_AUTO) ? AV_LOG_VERBOSE : AV_LOG_ERROR;
381     DXVA2Context *ctx = ist->hwaccel_ctx;
382     struct dxva_context *dxva_ctx = s->hwaccel_context;
383     GUID *guid_list = NULL;
384     unsigned guid_count = 0, i, j;
385     GUID device_guid = GUID_NULL;
386     D3DFORMAT target_format = 0;
387     DXVA2_VideoDesc desc = { 0 };
388     DXVA2_ConfigPictureDecode config;
389     HRESULT hr;
390     int surface_alignment, num_surfaces;
391     int ret;
392
393     AVDXVA2FramesContext *frames_hwctx;
394     AVHWFramesContext *frames_ctx;
395
396     hr = IDirectXVideoDecoderService_GetDecoderDeviceGuids(ctx->decoder_service, &guid_count, &guid_list);
397     if (FAILED(hr)) {
398         av_log(NULL, loglevel, "Failed to retrieve decoder device GUIDs\n");
399         goto fail;
400     }
401
402     for (i = 0; dxva2_modes[i].guid; i++) {
403         D3DFORMAT *target_list = NULL;
404         unsigned target_count = 0;
405         const dxva2_mode *mode = &dxva2_modes[i];
406         if (mode->codec != s->codec_id)
407             continue;
408
409         for (j = 0; j < guid_count; j++) {
410             if (IsEqualGUID(mode->guid, &guid_list[j]))
411                 break;
412         }
413         if (j == guid_count)
414             continue;
415
416         hr = IDirectXVideoDecoderService_GetDecoderRenderTargets(ctx->decoder_service, mode->guid, &target_count, &target_list);
417         if (FAILED(hr)) {
418             continue;
419         }
420         for (j = 0; j < target_count; j++) {
421             const D3DFORMAT format = target_list[j];
422             if (format == MKTAG('N','V','1','2')) {
423                 target_format = format;
424                 break;
425             }
426         }
427         CoTaskMemFree(target_list);
428         if (target_format) {
429             device_guid = *mode->guid;
430             break;
431         }
432     }
433     CoTaskMemFree(guid_list);
434
435     if (IsEqualGUID(&device_guid, &GUID_NULL)) {
436         av_log(NULL, loglevel, "No decoder device for codec found\n");
437         goto fail;
438     }
439
440     desc.SampleWidth  = s->coded_width;
441     desc.SampleHeight = s->coded_height;
442     desc.Format       = target_format;
443
444     ret = dxva2_get_decoder_configuration(s, &device_guid, &desc, &config);
445     if (ret < 0) {
446         goto fail;
447     }
448
449     /* decoding MPEG-2 requires additional alignment on some Intel GPUs,
450        but it causes issues for H.264 on certain AMD GPUs..... */
451     if (s->codec_id == AV_CODEC_ID_MPEG2VIDEO)
452         surface_alignment = 32;
453     /* the HEVC DXVA2 spec asks for 128 pixel aligned surfaces to ensure
454        all coding features have enough room to work with */
455     else if  (s->codec_id == AV_CODEC_ID_HEVC)
456         surface_alignment = 128;
457     else
458         surface_alignment = 16;
459
460     /* 4 base work surfaces */
461     num_surfaces = 4;
462
463     /* add surfaces based on number of possible refs */
464     if (s->codec_id == AV_CODEC_ID_H264 || s->codec_id == AV_CODEC_ID_HEVC)
465         num_surfaces += 16;
466     else
467         num_surfaces += 2;
468
469     /* add extra surfaces for frame threading */
470     if (s->active_thread_type & FF_THREAD_FRAME)
471         num_surfaces += s->thread_count;
472
473     ctx->hw_frames_ctx = av_hwframe_ctx_alloc(ctx->hw_device_ctx);
474     if (!ctx->hw_frames_ctx)
475         goto fail;
476     frames_ctx   = (AVHWFramesContext*)ctx->hw_frames_ctx->data;
477     frames_hwctx = frames_ctx->hwctx;
478
479     frames_ctx->format            = AV_PIX_FMT_DXVA2_VLD;
480     frames_ctx->sw_format         = AV_PIX_FMT_NV12;
481     frames_ctx->width             = FFALIGN(s->coded_width, surface_alignment);
482     frames_ctx->height            = FFALIGN(s->coded_height, surface_alignment);
483     frames_ctx->initial_pool_size = num_surfaces;
484
485     frames_hwctx->surface_type = DXVA2_VideoDecoderRenderTarget;
486
487     ret = av_hwframe_ctx_init(ctx->hw_frames_ctx);
488     if (ret < 0) {
489         av_log(NULL, loglevel, "Failed to initialize the HW frames context\n");
490         goto fail;
491     }
492
493     hr = IDirectXVideoDecoderService_CreateVideoDecoder(ctx->decoder_service, &device_guid,
494                                                         &desc, &config, frames_hwctx->surfaces,
495                                                         frames_hwctx->nb_surfaces, &frames_hwctx->decoder_to_release);
496     if (FAILED(hr)) {
497         av_log(NULL, loglevel, "Failed to create DXVA2 video decoder\n");
498         goto fail;
499     }
500
501     ctx->decoder_guid   = device_guid;
502     ctx->decoder_config = config;
503
504     dxva_ctx->cfg           = &ctx->decoder_config;
505     dxva_ctx->decoder       = frames_hwctx->decoder_to_release;
506     dxva_ctx->surface       = frames_hwctx->surfaces;
507     dxva_ctx->surface_count = frames_hwctx->nb_surfaces;
508
509     if (IsEqualGUID(&ctx->decoder_guid, &DXVADDI_Intel_ModeH264_E))
510         dxva_ctx->workaround |= FF_DXVA2_WORKAROUND_INTEL_CLEARVIDEO;
511
512     return 0;
513 fail:
514     av_buffer_unref(&ctx->hw_frames_ctx);
515     return AVERROR(EINVAL);
516 }
517
518 int dxva2_init(AVCodecContext *s)
519 {
520     InputStream *ist = s->opaque;
521     int loglevel = (ist->hwaccel_id == HWACCEL_AUTO) ? AV_LOG_VERBOSE : AV_LOG_ERROR;
522     DXVA2Context *ctx;
523     int ret;
524
525     if (!ist->hwaccel_ctx) {
526         ret = dxva2_alloc(s);
527         if (ret < 0)
528             return ret;
529     }
530     ctx = ist->hwaccel_ctx;
531
532     if (s->codec_id == AV_CODEC_ID_H264 &&
533         (s->profile & ~FF_PROFILE_H264_CONSTRAINED) > FF_PROFILE_H264_HIGH) {
534         av_log(NULL, loglevel, "Unsupported H.264 profile for DXVA2 HWAccel: %d\n", s->profile);
535         return AVERROR(EINVAL);
536     }
537
538     av_buffer_unref(&ctx->hw_frames_ctx);
539
540     ret = dxva2_create_decoder(s);
541     if (ret < 0) {
542         av_log(NULL, loglevel, "Error creating the DXVA2 decoder\n");
543         return ret;
544     }
545
546     return 0;
547 }