git.sesse.net Git - vlc/blob - modules/codec/avcodec/dxva2.c

   1 /*****************************************************************************
   2  * va.c: Video Acceleration helpers
   3  *****************************************************************************
   4  * Copyright (C) 2009 Geoffroy Couprie
   5  * Copyright (C) 2009 Laurent Aimar
   6  * $Id$
   7  *
   8  * Authors: Geoffroy Couprie <geal@videolan.org>
   9  *          Laurent Aimar <fenrir _AT_ videolan _DOT_ org>
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2 of the License, or
  14  * (at your option) any later version.
  15  *
  16  * This program is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  19  * GNU General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU General Public License
  22  * along with this program; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
  24  *****************************************************************************/
  25
  26 #ifdef HAVE_CONFIG_H
  27 # include "config.h"
  28 #endif
  29
  30 #include <vlc_common.h>
  31 #include <vlc_picture.h>
  32 #include <vlc_fourcc.h>
  33 #include <vlc_cpu.h>
  34 #include <assert.h>
  35
  36 #ifdef HAVE_LIBAVCODEC_AVCODEC_H
  37 #   include <libavcodec/avcodec.h>
  38 #   ifdef HAVE_AVCODEC_DXVA2
  39 #       define DXVA2API_USE_BITFIELDS
  40 #       include <libavcodec/dxva2.h>
  41 #   endif
  42 #elif defined(HAVE_FFMPEG_AVCODEC_H)
  43 #   include <ffmpeg/avcodec.h>
  44 #else
  45 #   include <avcodec.h>
  46 #endif
  47
  48 #include "avcodec.h"
  49 #include "va.h"
  50
  51 #ifdef HAVE_AVCODEC_DXVA2
  52
  53 #include <windows.h>
  54 #include <windowsx.h>
  55 #include <ole2.h>
  56 #include <commctrl.h>
  57 #include <shlwapi.h>
  58 #include <d3d9.h>
  59
  60 /* FIXME */
  61 #define CoTaskMemFree(x)
  62
  63 /* */
  64 #define DXVA2_E_NOT_INITIALIZED     MAKE_HRESULT(1, 4, 4096)
  65 #define DXVA2_E_NEW_VIDEO_DEVICE    MAKE_HRESULT(1, 4, 4097)
  66 #define DXVA2_E_VIDEO_DEVICE_LOCKED MAKE_HRESULT(1, 4, 4098)
  67 #define DXVA2_E_NOT_AVAILABLE       MAKE_HRESULT(1, 4, 4099)
  68
  69 static const GUID DXVA2_ModeMPEG2_MoComp = {
  70     0xe6a9f44b, 0x61b0, 0x4563, {0x9e,0xa4,0x63,0xd2,0xa3,0xc6,0xfe,0x66}
  71 };
  72 static const GUID DXVA2_ModeMPEG2_IDCT = {
  73   0xbf22ad00, 0x03ea, 0x4690, {0x80,0x77,0x47,0x33,0x46,0x20,0x9b,0x7e}
  74 };
  75 static const GUID DXVA2_ModeMPEG2_VLD = {
  76    0xee27417f, 0x5e28, 0x4e65, {0xbe,0xea,0x1d,0x26,0xb5,0x08,0xad,0xc9}
  77 };
  78
  79 static const GUID DXVA2_ModeH264_A = {
  80     0x1b81be64, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
  81 };
  82 static const GUID DXVA2_ModeH264_B = {
  83     0x1b81be65, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
  84 };
  85 static const GUID DXVA2_ModeH264_C = {
  86     0x1b81be66, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
  87 };
  88 static const GUID DXVA2_ModeH264_D = {
  89     0x1b81be67, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
  90 };
  91 static const GUID DXVA2_ModeH264_E = {
  92     0x1b81be68, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
  93 };
  94 static const GUID DXVA2_ModeH264_F = {
  95     0x1b81be69, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
  96 };
  97 static const GUID DXVADDI_Intel_ModeH264_A = {
  98     0x604F8E64, 0x4951,0x4c54, {0x88,0xFE,0xAB,0xD2,0x5C,0x15,0xB3,0xD6}
  99 };
 100 static const GUID DXVADDI_Intel_ModeH264_C = {
 101     0x604F8E66,0x4951, 0x4c54, {0x88,0xFE,0xAB,0xD2,0x5C,0x15,0xB3,0xD6}
 102 };
 103 static const GUID DXVADDI_Intel_ModeH264_E = {
 104     0x604F8E68,0x4951, 0x4c54, {0x88,0xFE,0xAB,0xD2,0x5C,0x15,0xB3,0xD6}
 105 };
 106 static const GUID DXVA2_ModeWMV8_A = {
 107     0x1b81be80, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 108 };
 109 static const GUID DXVA2_ModeWMV8_B = {
 110     0x1b81be81, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 111 };
 112 static const GUID DXVA2_ModeWMV9_A = {
 113     0x1b81be90, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 114 };
 115 static const GUID DXVA2_ModeWMV9_B = {
 116     0x1b81be91, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 117 };
 118 static const GUID DXVA2_ModeWMV9_C = {
 119     0x1b81be94, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 120 };
 121
 122 static const GUID DXVA2_ModeVC1_A = {
 123     0x1b81beA0, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 124 };
 125 static const GUID DXVA2_ModeVC1_B = {
 126     0x1b81beA1, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 127 };
 128 static const GUID DXVA2_ModeVC1_C = {
 129     0x1b81beA2, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 130 };
 131 static const GUID DXVA2_ModeVC1_D = {
 132     0x1b81beA3, 0xa0c7,0x11d3, {0xb9,0x84,0x00,0xc0,0x4f,0x2e,0x73,0xc5}
 133 };
 134
 135 /* */
 136 typedef struct {
 137     const char   *name;
 138     const GUID   *guid;
 139     int          codec;
 140 } dxva2_mode_t;
 141 /* XXX Prefered modes must come first */
 142 static const dxva2_mode_t dxva2_modes[] = {
 143     { "DXVA2_ModeMPEG2_VLD",    &DXVA2_ModeMPEG2_VLD,     0 },
 144     { "DXVA2_ModeMPEG2_MoComp", &DXVA2_ModeMPEG2_MoComp,  0 },
 145     { "DXVA2_ModeMPEG2_IDCT",   &DXVA2_ModeMPEG2_IDCT,    0 },
 146
 147     { "H.264 variable-length decoder (VLD), FGT",               &DXVA2_ModeH264_F,         CODEC_ID_H264 },
 148     { "H.264 VLD, no FGT",                                      &DXVA2_ModeH264_E,         CODEC_ID_H264 },
 149     { "H.264 VLD, no FGT (Intel)",                              &DXVADDI_Intel_ModeH264_E, CODEC_ID_H264 },
 150     { "H.264 IDCT, FGT",                                        &DXVA2_ModeH264_D,         0             },
 151     { "H.264 inverse discrete cosine transform (IDCT), no FGT", &DXVA2_ModeH264_C,         0             },
 152     { "H.264 inverse discrete cosine transform (IDCT), no FGT (Intel)", &DXVADDI_Intel_ModeH264_C, 0     },
 153     { "H.264 MoComp, FGT",                                      &DXVA2_ModeH264_B,         0             },
 154     { "H.264 motion compensation (MoComp), no FGT",             &DXVA2_ModeH264_A,         0             },
 155     { "H.264 motion compensation (MoComp), no FGT (Intel)",     &DXVADDI_Intel_ModeH264_A, 0             },
 156
 157     { "Windows Media Video 8 MoComp",           &DXVA2_ModeWMV8_B, 0 },
 158     { "Windows Media Video 8 post processing",  &DXVA2_ModeWMV8_A, 0 },
 159
 160     {  "Windows Media Video 9 IDCT",            &DXVA2_ModeWMV9_C, 0 },
 161     {  "Windows Media Video 9 MoComp",          &DXVA2_ModeWMV9_B, 0 },
 162     {  "Windows Media Video 9 post processing", &DXVA2_ModeWMV9_A, 0 },
 163
 164     { "VC-1 VLD",             &DXVA2_ModeVC1_D, CODEC_ID_VC1 },
 165     { "VC-1 VLD",             &DXVA2_ModeVC1_D, CODEC_ID_WMV3 },
 166     { "VC-1 IDCT",            &DXVA2_ModeVC1_C, 0 },
 167     { "VC-1 MoComp",          &DXVA2_ModeVC1_B, 0 },
 168     { "VC-1 post processing", &DXVA2_ModeVC1_A, 0 },
 169
 170     { NULL, NULL, 0 }
 171 };
 172
 173 static const dxva2_mode_t *Dxva2FindMode(const GUID *guid)
 174 {
 175     for (unsigned i = 0; dxva2_modes[i].name; i++) {
 176         if (IsEqualGUID(dxva2_modes[i].guid, guid))
 177             return &dxva2_modes[i];
 178     }
 179     return NULL;
 180 }
 181
 182 /* */
 183 typedef struct {
 184     const char   *name;
 185     D3DFORMAT    format;
 186     vlc_fourcc_t codec;
 187 } d3d_format_t;
 188 /* XXX Prefered format must come first */
 189 static const d3d_format_t d3d_formats[] = {
 190     { "YV12",   MAKEFOURCC('Y','V','1','2'),    VLC_CODEC_YV12 },
 191     { "NV12",   MAKEFOURCC('N','V','1','2'),    VLC_CODEC_NV12 },
 192
 193     { NULL, 0, 0 }
 194 };
 195
 196 static const d3d_format_t *D3dFindFormat(D3DFORMAT format)
 197 {
 198     for (unsigned i = 0; d3d_formats[i].name; i++) {
 199         if (d3d_formats[i].format == format)
 200             return &d3d_formats[i];
 201     }
 202     return NULL;
 203 }
 204
 205 static const GUID IID_IDirectXVideoDecoderService = {
 206     0xfc51a551, 0xd5e7, 0x11d9, {0xaf,0x55,0x00,0x05,0x4e,0x43,0xff,0x02}
 207 };
 208 static const GUID IID_IDirectXVideoAccelerationService = {
 209     0xfc51a550, 0xd5e7, 0x11d9, {0xaf,0x55,0x00,0x05,0x4e,0x43,0xff,0x02}
 210 };
 211
 212 /* */
 213 typedef struct {
 214     LPDIRECT3DSURFACE9 d3d;
 215     int                refcount;
 216     unsigned int       order;
 217 } vlc_va_surface_t;
 218
 219 #define VA_DXVA2_MAX_SURFACE_COUNT (64)
 220 typedef struct
 221 {
 222     /* */
 223     vlc_va_t va;
 224
 225     /* */
 226     vlc_object_t *log;
 227     int          codec_id;
 228
 229     /* DLL */
 230         HINSTANCE             hd3d9_dll;
 231     HINSTANCE             hdxva2_dll;
 232
 233     /* Direct3D */
 234     D3DPRESENT_PARAMETERS  d3dpp;
 235     LPDIRECT3D9            d3dobj;
 236     D3DADAPTER_IDENTIFIER9 d3dai;
 237     LPDIRECT3DDEVICE9      d3ddev;
 238
 239     /* Device manager */
 240     UINT                     token;
 241     IDirect3DDeviceManager9  *devmng;
 242     HANDLE                   device;
 243
 244     /* Video service */
 245     IDirectXVideoDecoderService  *vs;
 246     GUID                         input;
 247     D3DFORMAT                    render;
 248
 249     /* Video decoder */
 250     DXVA2_ConfigPictureDecode    cfg;
 251     IDirectXVideoDecoder         *decoder;
 252
 253     /* Option conversion */
 254     D3DFORMAT                    output;
 255     uint8_t                      *surface_cache_base;
 256     uint8_t                      *surface_cache;
 257     size_t                       surface_cache_size;
 258
 259     /* */
 260     struct dxva_context hw;
 261
 262     /* */
 263     unsigned     surface_count;
 264     unsigned     surface_order;
 265     int          surface_width;
 266     int          surface_height;
 267     vlc_fourcc_t surface_chroma;
 268
 269     vlc_va_surface_t surface[VA_DXVA2_MAX_SURFACE_COUNT];
 270     LPDIRECT3DSURFACE9 hw_surface[VA_DXVA2_MAX_SURFACE_COUNT];
 271 } vlc_va_dxva2_t;
 272
 273 /* */
 274 static vlc_va_dxva2_t *vlc_va_dxva2_Get(void *external)
 275 {
 276     assert(external == (void*)(&((vlc_va_dxva2_t*)external)->va));
 277     return external;
 278 }
 279
 280 /* */
 281 static int D3dCreateDevice(vlc_va_dxva2_t *);
 282 static void D3dDestroyDevice(vlc_va_dxva2_t *);
 283 static char *DxDescribe(vlc_va_dxva2_t *);
 284
 285 static int D3dCreateDeviceManager(vlc_va_dxva2_t *);
 286 static void D3dDestroyDeviceManager(vlc_va_dxva2_t *);
 287
 288 static int DxCreateVideoService(vlc_va_dxva2_t *);
 289 static void DxDestroyVideoService(vlc_va_dxva2_t *);
 290 static int DxFindVideoServiceConversion(vlc_va_dxva2_t *, GUID *input, D3DFORMAT *output);
 291
 292 static int DxCreateVideoDecoder(vlc_va_dxva2_t *,
 293                                 int codec_id, const video_format_t *);
 294 static void DxDestroyVideoDecoder(vlc_va_dxva2_t *);
 295 static int DxResetVideoDecoder(vlc_va_dxva2_t *);
 296
 297 static void DxCreateVideoConversion(vlc_va_dxva2_t *);
 298 static void DxDestroyVideoConversion(vlc_va_dxva2_t *);
 299
 300 static void CopyFromNv12(picture_t *dst, const D3DLOCKED_RECT *src,
 301                          uint8_t *cache, size_t cache_size,
 302                          unsigned width, unsigned height);
 303 static void CopyFromYv12(picture_t *dst, const D3DLOCKED_RECT *src,
 304                          uint8_t *cache, size_t cache_size,
 305                          unsigned width, unsigned height);
 306
 307 /* */
 308 static int Setup(vlc_va_t *external, void **hw, vlc_fourcc_t *chroma,
 309                  int width, int height)
 310 {
 311     vlc_va_dxva2_t *va = vlc_va_dxva2_Get(external);
 312
 313     if (va->surface_width  == width &&
 314         va->surface_height == height)
 315         goto ok;
 316
 317     /* */
 318     DxDestroyVideoConversion(va);
 319     DxDestroyVideoDecoder(va);
 320
 321     *hw = NULL;
 322     *chroma = 0;
 323     if (width <= 0 || height <= 0)
 324         return VLC_EGENERIC;
 325
 326     /* FIXME transmit a video_format_t by VaSetup directly */
 327     video_format_t fmt;
 328     memset(&fmt, 0, sizeof(fmt));
 329     fmt.i_width = width;
 330     fmt.i_height = height;
 331
 332     if (DxCreateVideoDecoder(va, va->codec_id, &fmt))
 333         return VLC_EGENERIC;
 334     /* */
 335     va->hw.decoder = va->decoder;
 336     va->hw.cfg = &va->cfg;
 337     va->hw.surface_count = va->surface_count;
 338     va->hw.surface = va->hw_surface;
 339     for (unsigned i = 0; i < va->surface_count; i++)
 340         va->hw.surface[i] = va->surface[i].d3d;
 341
 342     /* */
 343     DxCreateVideoConversion(va);
 344
 345     /* */
 346 ok:
 347     *hw = &va->hw;
 348     const d3d_format_t *output = D3dFindFormat(va->output);
 349     *chroma = output->codec;
 350
 351     return VLC_SUCCESS;
 352 }
 353
 354 static int Extract(vlc_va_t *external, picture_t *picture, AVFrame *ff)
 355 {
 356     vlc_va_dxva2_t *va = vlc_va_dxva2_Get(external);
 357     LPDIRECT3DSURFACE9 d3d = (LPDIRECT3DSURFACE9)(uintptr_t)ff->data[3];
 358
 359     if (!va->surface_cache)
 360         return VLC_EGENERIC;
 361
 362     /* */
 363     assert(va->output == MAKEFOURCC('Y','V','1','2'));
 364
 365     /* */
 366     D3DLOCKED_RECT lock;
 367     if (FAILED(IDirect3DSurface9_LockRect(d3d, &lock, NULL, D3DLOCK_READONLY))) {
 368         msg_Err(va->log, "Failed to lock surface");
 369         return VLC_EGENERIC;
 370     }
 371
 372     if (va->render == MAKEFOURCC('Y','V','1','2')) {
 373         CopyFromYv12(picture, &lock,
 374                      va->surface_cache, va->surface_cache_size,
 375                      va->surface_width, va->surface_height);
 376     } else {
 377         assert(va->render == MAKEFOURCC('N','V','1','2'));
 378         CopyFromNv12(picture, &lock,
 379                      va->surface_cache, va->surface_cache_size,
 380                      va->surface_width, va->surface_height);
 381     }
 382
 383     /* */
 384     IDirect3DSurface9_UnlockRect(d3d);
 385     return VLC_SUCCESS;
 386 }
 387 /* FIXME it is nearly common with VAAPI */
 388 static int Get(vlc_va_t *external, AVFrame *ff)
 389 {
 390     vlc_va_dxva2_t *va = vlc_va_dxva2_Get(external);
 391
 392     /* Check the device */
 393     HRESULT hr = IDirect3DDeviceManager9_TestDevice(va->devmng, va->device);
 394     if (hr == DXVA2_E_NEW_VIDEO_DEVICE) {
 395         if (DxResetVideoDecoder(va))
 396             return VLC_EGENERIC;
 397     } else if (FAILED(hr)) {
 398         msg_Err(va->log, "IDirect3DDeviceManager9_TestDevice %u", (unsigned)hr);
 399         return VLC_EGENERIC;
 400     }
 401
 402     /* Grab an unused surface, in case none are, try the oldest
 403      * XXX using the oldest is a workaround in case a problem happens with ffmpeg */
 404     unsigned i, old;
 405     for (i = 0, old = 0; i < va->surface_count; i++) {
 406         vlc_va_surface_t *surface = &va->surface[i];
 407
 408         if (!surface->refcount)
 409             break;
 410
 411         if (surface->order < va->surface[old].order)
 412             old = i;
 413     }
 414     if (i >= va->surface_count)
 415         i = old;
 416
 417     vlc_va_surface_t *surface = &va->surface[i];
 418
 419     surface->refcount = 1;
 420     surface->order = va->surface_order++;
 421
 422     /* */
 423     for (int i = 0; i < 4; i++) {
 424         ff->data[i] = NULL;
 425         ff->linesize[i] = 0;
 426
 427         if (i == 0 || i == 3)
 428             ff->data[i] = (void*)surface->d3d;/* Yummie */
 429     }
 430     return VLC_SUCCESS;
 431 }
 432 static void Release(vlc_va_t *external, AVFrame *ff)
 433 {
 434     vlc_va_dxva2_t *va = vlc_va_dxva2_Get(external);
 435     LPDIRECT3DSURFACE9 d3d = (LPDIRECT3DSURFACE9)(uintptr_t)ff->data[3];
 436
 437     for (unsigned i = 0; i < va->surface_count; i++) {
 438         vlc_va_surface_t *surface = &va->surface[i];
 439
 440         if (surface->d3d == d3d)
 441             surface->refcount--;
 442     }
 443 }
 444 static void Close(vlc_va_t *external)
 445 {
 446     vlc_va_dxva2_t *va = vlc_va_dxva2_Get(external);
 447
 448     DxDestroyVideoConversion(va);
 449     DxDestroyVideoDecoder(va);
 450     DxDestroyVideoService(va);
 451     D3dDestroyDeviceManager(va);
 452     D3dDestroyDevice(va);
 453
 454     if (va->hdxva2_dll)
 455         FreeLibrary(va->hdxva2_dll);
 456     if (va->hd3d9_dll)
 457         FreeLibrary(va->hd3d9_dll);
 458
 459     free(va->va.description);
 460     free(va);
 461 }
 462
 463 vlc_va_t *vlc_va_NewDxva2(vlc_object_t *log, int codec_id)
 464 {
 465     vlc_va_dxva2_t *va = calloc(1, sizeof(*va));
 466     if (!va)
 467         return NULL;
 468
 469     /* */
 470     va->log = log;
 471     va->codec_id = codec_id;
 472
 473     /* Load dll*/
 474     va->hd3d9_dll = LoadLibrary(TEXT("D3D9.DLL"));
 475     if (!va->hd3d9_dll) {
 476         msg_Warn(va->log, "cannot load d3d9.dll");
 477         goto error;
 478     }
 479     va->hdxva2_dll = LoadLibrary(TEXT("DXVA2.DLL"));
 480     if (!va->hdxva2_dll) {
 481         msg_Warn(va->log, "cannot load dxva2.dll");
 482         goto error;
 483     }
 484     msg_Dbg(va->log, "DLLs loaded");
 485
 486     /* */
 487     if (D3dCreateDevice(va)) {
 488         msg_Err(va->log, "Failed to create Direct3D device");
 489         goto error;
 490     }
 491     msg_Dbg(va->log, "D3dCreateDevice succeed");
 492
 493     if (D3dCreateDeviceManager(va)) {
 494         msg_Err(va->log, "D3dCreateDeviceManager failed");
 495         goto error;
 496     }
 497
 498     if (DxCreateVideoService(va)) {
 499         msg_Err(va->log, "DxCreateVideoService failed");
 500         goto error;
 501     }
 502
 503     /* */
 504     if (DxFindVideoServiceConversion(va, &va->input, &va->render)) {
 505         msg_Err(va->log, "DxFindVideoServiceConversion failed");
 506         goto error;
 507     }
 508
 509     /* TODO print the hardware name/vendor for debugging purposes */
 510     va->va.description = DxDescribe(va);
 511     va->va.setup   = Setup;
 512     va->va.get     = Get;
 513     va->va.release = Release;
 514     va->va.extract = Extract;
 515     va->va.close   = Close;
 516     return &va->va;
 517
 518 error:
 519     Close(&va->va);
 520     return NULL;
 521 }
 522 /* */
 523
 524 /**
 525  * It creates a Direct3D device usable for DXVA 2
 526  */
 527 static int D3dCreateDevice(vlc_va_dxva2_t *va)
 528 {
 529     /* */
 530     LPDIRECT3D9 (WINAPI *Create9)(UINT SDKVersion);
 531     Create9 = (void *)GetProcAddress(va->hd3d9_dll,
 532                                      TEXT("Direct3DCreate9"));
 533     if (!Create9) {
 534         msg_Err(va->log, "Cannot locate reference to Direct3DCreate9 ABI in DLL");
 535         return VLC_EGENERIC;
 536     }
 537
 538     /* */
 539     LPDIRECT3D9 d3dobj;
 540     d3dobj = Create9(D3D_SDK_VERSION);
 541     if (!d3dobj) {
 542         msg_Err(va->log, "Direct3DCreate9 failed");
 543         return VLC_EGENERIC;
 544     }
 545     va->d3dobj = d3dobj;
 546
 547     /* */
 548     D3DADAPTER_IDENTIFIER9 *d3dai = &va->d3dai;
 549     if (FAILED(IDirect3D9_GetAdapterIdentifier(va->d3dobj,
 550                                                D3DADAPTER_DEFAULT, 0, d3dai))) {
 551         msg_Warn(va->log, "IDirect3D9_GetAdapterIdentifier failed");
 552         ZeroMemory(d3dai, sizeof(*d3dai));
 553     }
 554
 555     /* */
 556     D3DPRESENT_PARAMETERS *d3dpp = &va->d3dpp;
 557     ZeroMemory(d3dpp, sizeof(*d3dpp));
 558     d3dpp->Flags                  = D3DPRESENTFLAG_VIDEO;
 559     d3dpp->Windowed               = TRUE;
 560     d3dpp->hDeviceWindow          = NULL;
 561     d3dpp->SwapEffect             = D3DSWAPEFFECT_DISCARD;
 562     d3dpp->MultiSampleType        = D3DMULTISAMPLE_NONE;
 563     d3dpp->PresentationInterval   = D3DPRESENT_INTERVAL_DEFAULT;
 564     d3dpp->BackBufferCount        = 0;                  /* FIXME what to put here */
 565     d3dpp->BackBufferFormat       = D3DFMT_X8R8G8B8;    /* FIXME what to put here */
 566     d3dpp->BackBufferWidth        = 0;
 567     d3dpp->BackBufferHeight       = 0;
 568     d3dpp->EnableAutoDepthStencil = FALSE;
 569
 570     /* Direct3D needs a HWND to create a device, even without using ::Present
 571     this HWND is used to alert Direct3D when there's a change of focus window.
 572     For now, use GetShellWindow, as it looks harmless */
 573     LPDIRECT3DDEVICE9 d3ddev;
 574     if (FAILED(IDirect3D9_CreateDevice(d3dobj, D3DADAPTER_DEFAULT,
 575                                        D3DDEVTYPE_HAL, GetShellWindow(),
 576                                        D3DCREATE_SOFTWARE_VERTEXPROCESSING |
 577                                        D3DCREATE_MULTITHREADED,
 578                                        d3dpp, &d3ddev))) {
 579         msg_Err(va->log, "IDirect3D9_CreateDevice failed");
 580         return VLC_EGENERIC;
 581     }
 582     va->d3ddev = d3ddev;
 583
 584     return VLC_SUCCESS;
 585 }
 586 /**
 587  * It releases a Direct3D device and its resources.
 588  */
 589 static void D3dDestroyDevice(vlc_va_dxva2_t *va)
 590 {
 591     if (va->d3ddev)
 592         IDirect3DDevice9_Release(va->d3ddev);
 593     if (va->d3dobj)
 594         IDirect3D9_Release(va->d3dobj);
 595 }
 596 /**
 597  * It describes our Direct3D object
 598  */
 599 static char *DxDescribe(vlc_va_dxva2_t *va)
 600 {
 601     static const struct {
 602         unsigned id;
 603         char     name[32];
 604     } vendors [] = {
 605         { 0x1002, "ATI" },
 606         { 0x10DE, "NVIDIA" },
 607         { 0x8086, "Intel" },
 608         { 0x5333, "S3 Graphics" },
 609         { 0, "" }
 610     };
 611     D3DADAPTER_IDENTIFIER9 *id = &va->d3dai;
 612
 613     const char *vendor = "Unknown";
 614     for (int i = 0; vendors[i].id != 0; i++) {
 615         if (vendors[i].id == id->VendorId) {
 616             vendor = vendors[i].name;
 617             break;
 618         }
 619     }
 620
 621     char *description;
 622     if (asprintf(&description, "DXVA2 (%.*s, vendor %d(%s), device %d, revision %d)",
 623                  sizeof(id->Description), id->Description,
 624                  id->VendorId, vendor, id->DeviceId, id->Revision) < 0)
 625         return NULL;
 626     return description;
 627 }
 628
 629 /**
 630  * It creates a Direct3D device manager
 631  */
 632 static int D3dCreateDeviceManager(vlc_va_dxva2_t *va)
 633 {
 634     HRESULT (WINAPI *CreateDeviceManager9)(UINT *pResetToken,
 635                                            IDirect3DDeviceManager9 **);
 636     CreateDeviceManager9 =
 637       (void *)GetProcAddress(va->hdxva2_dll,
 638                              TEXT("DXVA2CreateDirect3DDeviceManager9"));
 639
 640     if (!CreateDeviceManager9) {
 641         msg_Err(va->log, "cannot load function\n");
 642         return VLC_EGENERIC;
 643     }
 644     msg_Dbg(va->log, "OurDirect3DCreateDeviceManager9 Success!");
 645
 646     UINT token;
 647     IDirect3DDeviceManager9 *devmng;
 648     if (FAILED(CreateDeviceManager9(&token, &devmng))) {
 649         msg_Err(va->log, " OurDirect3DCreateDeviceManager9 failed");
 650         return VLC_EGENERIC;
 651     }
 652     va->token  = token;
 653     va->devmng = devmng;
 654     msg_Info(va->log, "obtained IDirect3DDeviceManager9");
 655
 656     HRESULT hr = IDirect3DDeviceManager9_ResetDevice(devmng, va->d3ddev, token);
 657     if (FAILED(hr)) {
 658         msg_Err(va->log, "IDirect3DDeviceManager9_ResetDevice failed: %08x", (unsigned)hr);
 659         return VLC_EGENERIC;
 660     }
 661     return VLC_SUCCESS;
 662 }
 663 /**
 664  * It destroys a Direct3D device manager
 665  */
 666 static void D3dDestroyDeviceManager(vlc_va_dxva2_t *va)
 667 {
 668     if (va->devmng)
 669         IDirect3DDeviceManager9_Release(va->devmng);
 670 }
 671
 672 /**
 673  * It creates a DirectX video service
 674  */
 675 static int DxCreateVideoService(vlc_va_dxva2_t *va)
 676 {
 677     HRESULT (WINAPI *CreateVideoService)(IDirect3DDevice9 *,
 678                                          REFIID riid,
 679                                          void **ppService);
 680     CreateVideoService =
 681       (void *)GetProcAddress(va->hdxva2_dll,
 682                              TEXT("DXVA2CreateVideoService"));
 683
 684     if (!CreateVideoService) {
 685         msg_Err(va->log, "cannot load function\n");
 686         return 4;
 687     }
 688     msg_Info(va->log, "DXVA2CreateVideoService Success!");
 689
 690     HRESULT hr;
 691
 692     HANDLE device;
 693     hr = IDirect3DDeviceManager9_OpenDeviceHandle(va->devmng, &device);
 694     if (FAILED(hr)) {
 695         msg_Err(va->log, "OpenDeviceHandle failed");
 696         return VLC_EGENERIC;
 697     }
 698     va->device = device;
 699
 700     IDirectXVideoDecoderService *vs;
 701     hr = IDirect3DDeviceManager9_GetVideoService(va->devmng, device,
 702                                                  &IID_IDirectXVideoDecoderService,
 703                                                  &vs);
 704     if (FAILED(hr)) {
 705         msg_Err(va->log, "GetVideoService failed");
 706         return VLC_EGENERIC;
 707     }
 708     va->vs = vs;
 709
 710     return VLC_SUCCESS;
 711 }
 712 /**
 713  * It destroys a DirectX video service
 714  */
 715 static void DxDestroyVideoService(vlc_va_dxva2_t *va)
 716 {
 717     if (va->device)
 718         IDirect3DDeviceManager9_CloseDeviceHandle(va->devmng, va->device);
 719     if (va->vs)
 720         IDirectXVideoDecoderService_Release(va->vs);
 721 }
 722 /**
 723  * Find the best suited decoder mode GUID and render format.
 724  */
 725 static int DxFindVideoServiceConversion(vlc_va_dxva2_t *va, GUID *input, D3DFORMAT *output)
 726 {
 727     /* Retreive supported modes from the decoder service */
 728     UINT input_count = 0;
 729     GUID *input_list = NULL;
 730     if (FAILED(IDirectXVideoDecoderService_GetDecoderDeviceGuids(va->vs,
 731                                                                  &input_count,
 732                                                                  &input_list))) {
 733         msg_Err(va->log, "IDirectXVideoDecoderService_GetDecoderDeviceGuids failed");
 734         return VLC_EGENERIC;
 735     }
 736     for (unsigned i = 0; i < input_count; i++) {
 737         const GUID *g = &input_list[i];
 738         const dxva2_mode_t *mode = Dxva2FindMode(g);
 739         if (mode) {
 740             msg_Dbg(va->log, "- '%s' is supported by hardware", mode->name);
 741         } else {
 742             msg_Warn(va->log, "- Unknown GUID = %08X-%04x-%04x-XXXX",
 743                      (unsigned)g->Data1, g->Data2, g->Data3);
 744         }
 745     }
 746
 747     /* Try all supported mode by our priority */
 748     for (unsigned i = 0; dxva2_modes[i].name; i++) {
 749         const dxva2_mode_t *mode = &dxva2_modes[i];
 750         if (!mode->codec || mode->codec != va->codec_id)
 751             continue;
 752
 753         /* */
 754         bool is_suported = false;
 755         for (const GUID *g = &input_list[0]; !is_suported && g < &input_list[input_count]; g++) {
 756             is_suported = IsEqualGUID(mode->guid, g);
 757         }
 758         if (!is_suported)
 759             continue;
 760
 761         /* */
 762         msg_Dbg(va->log, "Trying to use '%s' as input", mode->name);
 763         UINT      output_count = 0;
 764         D3DFORMAT *output_list = NULL;
 765         if (FAILED(IDirectXVideoDecoderService_GetDecoderRenderTargets(va->vs, mode->guid,
 766                                                                        &output_count,
 767                                                                        &output_list))) {
 768             msg_Err(va->log, "IDirectXVideoDecoderService_GetDecoderRenderTargets failed");
 769             continue;
 770         }
 771         for (unsigned j = 0; j < output_count; j++) {
 772             const D3DFORMAT f = output_list[j];
 773             const d3d_format_t *format = D3dFindFormat(f);
 774             if (format) {
 775                 msg_Dbg(va->log, "%s is supported for output", format->name);
 776             } else {
 777                 msg_Dbg(va->log, "%d is supported for output (%4.4s)", f, (const char*)&f);
 778             }
 779         }
 780
 781         /* */
 782         for (unsigned j = 0; d3d_formats[j].name; j++) {
 783             const d3d_format_t *format = &d3d_formats[j];
 784
 785             /* */
 786             bool is_suported = false;
 787             for (unsigned k = 0; !is_suported && k < output_count; k++) {
 788                 is_suported = format->format == output_list[k];
 789             }
 790             if (!is_suported)
 791                 continue;
 792
 793             /* We have our solution */
 794             msg_Dbg(va->log, "Using '%s' to decode to '%s'", mode->name, format->name);
 795             *input  = *mode->guid;
 796             *output = format->format;
 797             CoTaskMemFree(output_list);
 798             CoTaskMemFree(input_list);
 799             return VLC_SUCCESS;
 800         }
 801         CoTaskMemFree(output_list);
 802     }
 803     CoTaskMemFree(input_list);
 804     return VLC_EGENERIC;
 805 }
 806
 807 /**
 808  * It creates a DXVA2 decoder using the given video format
 809  */
 810 static int DxCreateVideoDecoder(vlc_va_dxva2_t *va,
 811                                 int codec_id, const video_format_t *fmt)
 812 {
 813     /* */
 814     msg_Dbg(va->log, "DxCreateVideoDecoder id %d %dx%d",
 815             codec_id, fmt->i_width, fmt->i_height);
 816
 817     /* Allocates all surfaces needed for the decoder */
 818     switch (codec_id) {
 819     case CODEC_ID_H264:
 820         va->surface_count = 16 + 1;
 821         break;
 822     default:
 823         va->surface_count = 2 + 1;
 824         break;
 825     }
 826     LPDIRECT3DSURFACE9 surface_list[VA_DXVA2_MAX_SURFACE_COUNT];
 827     if (FAILED(IDirectXVideoDecoderService_CreateSurface(va->vs,
 828                                                          fmt->i_width,
 829                                                          fmt->i_height,
 830                                                          va->surface_count - 1,
 831                                                          va->render,
 832                                                          D3DPOOL_DEFAULT,
 833                                                          0,
 834                                                          DXVA2_VideoDecoderRenderTarget,
 835                                                          surface_list,
 836                                                          NULL))) {
 837         msg_Err(va->log, "IDirectXVideoAccelerationService_CreateSurface failed\n");
 838         va->surface_count = 0;
 839         return VLC_EGENERIC;
 840     }
 841     for (unsigned i = 0; i < va->surface_count; i++) {
 842         vlc_va_surface_t *surface = &va->surface[i];
 843         surface->d3d = surface_list[i];
 844         surface->refcount = 0;
 845         surface->order = 0;
 846     }
 847     va->surface_width  = fmt->i_width;
 848     va->surface_height = fmt->i_height;
 849     msg_Dbg(va->log, "IDirectXVideoAccelerationService_CreateSurface succeed with %d surfaces (%dx%d)",
 850             va->surface_count, fmt->i_width, fmt->i_height);
 851
 852     /* */
 853     DXVA2_VideoDesc dsc;
 854     ZeroMemory(&dsc, sizeof(dsc));
 855     dsc.SampleWidth     = fmt->i_width;
 856     dsc.SampleHeight    = fmt->i_height;
 857     dsc.Format          = va->render;
 858     if (fmt->i_frame_rate > 0 && fmt->i_frame_rate_base > 0) {
 859         dsc.InputSampleFreq.Numerator   = fmt->i_frame_rate;
 860         dsc.InputSampleFreq.Denominator = fmt->i_frame_rate_base;
 861     } else {
 862         dsc.InputSampleFreq.Numerator   = 0;
 863         dsc.InputSampleFreq.Denominator = 0;
 864     }
 865     dsc.OutputFrameFreq = dsc.InputSampleFreq;
 866     dsc.UABProtectionLevel = FALSE;
 867     dsc.Reserved = 0;
 868
 869     /* FIXME I am unsure we can let unknown everywhere */
 870     DXVA2_ExtendedFormat *ext = &dsc.SampleFormat;
 871     ext->SampleFormat = 0;//DXVA2_SampleUnknown;
 872     ext->VideoChromaSubsampling = 0;//DXVA2_VideoChromaSubsampling_Unknown;
 873     ext->NominalRange = 0;//DXVA2_NominalRange_Unknown;
 874     ext->VideoTransferMatrix = 0;//DXVA2_VideoTransferMatrix_Unknown;
 875     ext->VideoLighting = 0;//DXVA2_VideoLighting_Unknown;
 876     ext->VideoPrimaries = 0;//DXVA2_VideoPrimaries_Unknown;
 877     ext->VideoTransferFunction = 0;//DXVA2_VideoTransFunc_Unknown;
 878
 879     /* List all configurations available for the decoder */
 880     UINT                      cfg_count = 0;
 881     DXVA2_ConfigPictureDecode *cfg_list = NULL;
 882     if (FAILED(IDirectXVideoDecoderService_GetDecoderConfigurations(va->vs,
 883                                                                     &va->input,
 884                                                                     &dsc,
 885                                                                     NULL,
 886                                                                     &cfg_count,
 887                                                                     &cfg_list))) {
 888         msg_Err(va->log, "IDirectXVideoDecoderService_GetDecoderConfigurations failed\n");
 889         return VLC_EGENERIC;
 890     }
 891     msg_Dbg(va->log, "we got %d decoder configurations", cfg_count);
 892
 893     /* Select the best decoder configuration */
 894     bool has_cfg = false;
 895     for (unsigned i = 0; i < cfg_count; i++) {
 896         const DXVA2_ConfigPictureDecode *cfg = &cfg_list[i];
 897
 898         /* */
 899         msg_Dbg(va->log, "configuration[%d] ConfigBitstreamRaw %d",
 900                 i, cfg->ConfigBitstreamRaw);
 901
 902         /* */
 903         if ((!has_cfg && cfg->ConfigBitstreamRaw == 1) ||
 904             (codec_id == CODEC_ID_H264 && cfg->ConfigBitstreamRaw == 2)) {
 905             va->cfg = *cfg;
 906             has_cfg = true;
 907         }
 908     }
 909     CoTaskMemFree(cfg_list);
 910     if (!has_cfg) {
 911         msg_Err(va->log, "Failed to find a supported decoder configuration");
 912         return VLC_EGENERIC;
 913     }
 914
 915     /* Create the decoder */
 916     IDirectXVideoDecoder *decoder;
 917     if (FAILED(IDirectXVideoDecoderService_CreateVideoDecoder(va->vs,
 918                                                               &va->input,
 919                                                               &dsc,
 920                                                               &va->cfg,
 921                                                               surface_list,
 922                                                               va->surface_count,
 923                                                               &decoder))) {
 924         msg_Err(va->log, "IDirectXVideoDecoderService_CreateVideoDecoder failed\n");
 925         return VLC_EGENERIC;
 926     }
 927     va->decoder = decoder;
 928     msg_Dbg(va->log, "IDirectXVideoDecoderService_CreateVideoDecoder succeed");
 929     return VLC_SUCCESS;
 930 }
 931 static void DxDestroyVideoDecoder(vlc_va_dxva2_t *va)
 932 {
 933     if (va->decoder)
 934         IDirectXVideoDecoder_Release(va->decoder);
 935     va->decoder = NULL;
 936
 937     for (unsigned i = 0; i < va->surface_count; i++)
 938         IDirect3DSurface9_Release(va->surface[i].d3d);
 939     va->surface_count = 0;
 940 }
 941 static int DxResetVideoDecoder(vlc_va_dxva2_t *va)
 942 {
 943     msg_Err(va->log, "DxResetVideoDecoder unimplemented");
 944     return VLC_EGENERIC;
 945 }
 946
 947 static void DxCreateVideoConversion(vlc_va_dxva2_t *va)
 948 {
 949     switch (va->render) {
 950     case MAKEFOURCC('N','V','1','2'):
 951         va->output = MAKEFOURCC('Y','V','1','2');
 952         break;
 953     default:
 954         va->output = va->render;
 955         break;
 956     }
 957     va->surface_cache_size = __MAX((va->surface_width + 0x0f) & ~ 0x0f, 4096);
 958     va->surface_cache_base = malloc(16 + va->surface_cache_size);
 959     va->surface_cache      = &va->surface_cache_base[16 - ((intptr_t)va->surface_cache_base & 0x0f)];
 960 }
 961 static void DxDestroyVideoConversion(vlc_va_dxva2_t *va)
 962 {
 963     free(va->surface_cache_base);
 964     va->surface_cache_base = NULL;
 965     va->surface_cache      = NULL;
 966     va->surface_cache_size = 0;
 967 }
 968
 969 /* Copy 64 bytes from srcp to dsp loading data with the SSE>=2 instruction load and
 970  * storing data with the SSE>=2 instruction store.
 971  */
 972 #define COPY64(dstp, srcp, load, store) \
 973     asm volatile (                      \
 974         load "  0(%[src]), %%xmm1\n"    \
 975         load " 16(%[src]), %%xmm2\n"    \
 976         load " 32(%[src]), %%xmm3\n"    \
 977         load " 48(%[src]), %%xmm4\n"    \
 978         store " %%xmm1,    0(%[dst])\n" \
 979         store " %%xmm2,   16(%[dst])\n" \
 980         store " %%xmm3,   32(%[dst])\n" \
 981         store " %%xmm4,   48(%[dst])\n" \
 982         : : [dst]"r"(dstp), [src]"r"(srcp) : "memory")
 983
 984 /* Execute the instruction op only if SSE2 is supported. */
 985 #ifdef CAN_COMPILE_SSE2
 986 #   define ASM_SSE2(cpu, op) do {          \
 987         if (cpu & CPU_CAPABILITY_SSE2)  \
 988             asm volatile (op);    \
 989     } while (0)
 990 #else
 991 #   define ASM_SSE2(cpu, op)
 992 #endif
 993
 994 /* Optimized copy from "Uncacheable Speculative Write Combining" memory
 995  * as used by some video surface.
 996  * XXX It is really efficient only when SSE4.1 is available.
 997  */
 998 static void CopyFromUswc(uint8_t *dst, size_t dst_pitch,
 999                          const uint8_t *src, size_t src_pitch,
1000                          unsigned unaligned,
1001                          unsigned width, unsigned height,
1002                          unsigned cpu)
1003 {
1004     assert(((intptr_t)dst & 0x0f) == 0 && (dst_pitch & 0x0f) == 0);
1005
1006     ASM_SSE2(cpu, "mfence");
1007     for (unsigned y = 0; y < height; y++) {
1008         unsigned x;
1009
1010         for (x = 0; x < unaligned; x++)
1011             dst[x] = src[x];
1012
1013 #ifdef CAN_COMPILE_SSE4_1
1014         if (cpu & CPU_CAPABILITY_SSE4_1) {
1015             if (!unaligned) {
1016                 for (; x+63 < width; x += 64)
1017                     COPY64(&dst[x], &src[x], "movntdqa", "movdqa");
1018             } else {
1019                 for (; x+63 < width; x += 64)
1020                     COPY64(&dst[x], &src[x], "movntdqa", "movdqu");
1021             }
1022         } else
1023 #endif
1024 #ifdef CAN_COMPILE_SSE2
1025         if (cpu & CPU_CAPABILITY_SSE2) {
1026             if (!unaligned) {
1027                 for (; x+63 < width; x += 64)
1028                     COPY64(&dst[x], &src[x], "movdqa", "movdqa");
1029             } else {
1030                 for (; x+63 < width; x += 64)
1031                     COPY64(&dst[x], &src[x], "movdqa", "movdqu");
1032             }
1033         }
1034 #endif
1035
1036         for (; x < width; x++)
1037             dst[x] = src[x];
1038
1039         src += src_pitch;
1040         dst += dst_pitch;
1041     }
1042 }
1043
1044 static void Copy2d(uint8_t *dst, size_t dst_pitch,
1045                    const uint8_t *src, size_t src_pitch,
1046                    unsigned width, unsigned height,
1047                    unsigned cpu)
1048 {
1049     assert(((intptr_t)src & 0x0f) == 0 && (src_pitch & 0x0f) == 0);
1050
1051     ASM_SSE2(cpu, "mfence");
1052
1053     for (unsigned y = 0; y < height; y++) {
1054         unsigned x = 0;
1055         bool unaligned = ((intptr_t)dst & 0x0f) != 0;
1056
1057 #ifdef CAN_COMPILE_SSE2
1058         if (cpu & CPU_CAPABILITY_SSE2) {
1059             if (!unaligned) {
1060                 for (; x+63 < width; x += 64)
1061                     COPY64(&dst[x], &src[x], "movdqa", "movntdq");
1062             } else {
1063                 for (; x+63 < width; x += 64)
1064                     COPY64(&dst[x], &src[x], "movdqa", "movdqu");
1065             }
1066         }
1067 #endif
1068
1069         for (; x < width; x++)
1070             dst[x] = src[x];
1071
1072         src += src_pitch;
1073         dst += dst_pitch;
1074     }
1075 }
1076
1077 static void SplitUV(uint8_t *dstu, size_t dstu_pitch,
1078                     uint8_t *dstv, size_t dstv_pitch,
1079                     const uint8_t *src, size_t src_pitch,
1080                     unsigned width, unsigned height, unsigned cpu)
1081 {
1082     const uint8_t shuffle[] = { 0, 2, 4, 6, 8, 10, 12, 14,
1083                                 1, 3, 5, 7, 9, 11, 13, 15 };
1084     const uint8_t mask[] = { 0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00,
1085                              0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00 };
1086
1087     assert(((intptr_t)src & 0x0f) == 0 && (src_pitch & 0x0f) == 0);
1088
1089     ASM_SSE2(cpu, "mfence");
1090
1091     for (unsigned y = 0; y < height; y++) {
1092         unsigned x = 0;
1093
1094 #define LOAD64 \
1095     "movdqa  0(%[src]), %%xmm0\n" \
1096     "movdqa 16(%[src]), %%xmm1\n" \
1097     "movdqa 32(%[src]), %%xmm2\n" \
1098     "movdqa 48(%[src]), %%xmm3\n"
1099
1100 #define STORE2X32 \
1101     "movq   %%xmm0,   0(%[dst1])\n" \
1102     "movq   %%xmm1,   8(%[dst1])\n" \
1103     "movhpd %%xmm0,   0(%[dst2])\n" \
1104     "movhpd %%xmm1,   8(%[dst2])\n" \
1105     "movq   %%xmm2,  16(%[dst1])\n" \
1106     "movq   %%xmm3,  24(%[dst1])\n" \
1107     "movhpd %%xmm2,  16(%[dst2])\n" \
1108     "movhpd %%xmm3,  24(%[dst2])\n"
1109
1110 #ifdef CAN_COMPILE_SSSE3
1111         if (cpu & CPU_CAPABILITY_SSSE3) {
1112             for (x = 0; x < (width & ~31); x += 32) {
1113                 asm volatile (
1114                     "movdqu (%[shuffle]), %%xmm7\n"
1115                     LOAD64
1116                     "pshufb  %%xmm7, %%xmm0\n"
1117                     "pshufb  %%xmm7, %%xmm1\n"
1118                     "pshufb  %%xmm7, %%xmm2\n"
1119                     "pshufb  %%xmm7, %%xmm3\n"
1120                     STORE2X32
1121                     : : [dst1]"r"(&dstu[x]), [dst2]"r"(&dstv[x]), [src]"r"(&src[2*x]), [shuffle]"r"(shuffle) : "memory");
1122             }
1123         } else
1124 #endif
1125 #ifdef CAN_COMPILE_SSE2
1126         if (cpu & CPU_CAPABILITY_SSE2) {
1127             for (x = 0; x < (width & ~31); x += 32) {
1128                 asm volatile (
1129                     "movdqu (%[mask]), %%xmm7\n"
1130                     LOAD64
1131                     "movdqa   %%xmm0, %%xmm4\n"
1132                     "movdqa   %%xmm1, %%xmm5\n"
1133                     "movdqa   %%xmm2, %%xmm6\n"
1134                     "psrlw    $8,     %%xmm0\n"
1135                     "psrlw    $8,     %%xmm1\n"
1136                     "pand     %%xmm7, %%xmm4\n"
1137                     "pand     %%xmm7, %%xmm5\n"
1138                     "pand     %%xmm7, %%xmm6\n"
1139                     "packuswb %%xmm4, %%xmm0\n"
1140                     "packuswb %%xmm5, %%xmm1\n"
1141                     "pand     %%xmm3, %%xmm7\n"
1142                     "psrlw    $8,     %%xmm2\n"
1143                     "psrlw    $8,     %%xmm3\n"
1144                     "packuswb %%xmm6, %%xmm2\n"
1145                     "packuswb %%xmm7, %%xmm3\n"
1146                     STORE2X32
1147                     : : [dst2]"r"(&dstu[x]), [dst1]"r"(&dstv[x]), [src]"r"(&src[2*x]), [mask]"r"(mask) : "memory");
1148             }
1149         }
1150 #endif
1151 #undef STORE2X32
1152 #undef LOAD64
1153
1154         for (; x < width; x++) {
1155             dstu[x] = src[2*x+0];
1156             dstv[x] = src[2*x+1];
1157         }
1158         src  += src_pitch;
1159         dstu += dstu_pitch;
1160         dstv += dstv_pitch;
1161     }
1162 }
1163
1164 static void CopyPlane(uint8_t *dst, size_t dst_pitch, const uint8_t *src, size_t src_pitch,
1165                       uint8_t *cache, size_t cache_size,
1166                       unsigned width, unsigned height,
1167                       unsigned cpu)
1168 {
1169     const unsigned w16 = (width+15) & ~15;
1170     const unsigned hstep = cache_size / w16;
1171     assert(hstep > 0);
1172
1173     for (unsigned y = 0; y < height; y += hstep) {
1174         const unsigned unaligned = (intptr_t)src & 0x0f;
1175         const unsigned hblock =  __MIN(hstep, height - y);
1176
1177         /* Copy a bunch of line into our cache */
1178         CopyFromUswc(cache, w16,
1179                      src, src_pitch,
1180                      unaligned,
1181                      width, hblock, cpu);
1182
1183         /* Copy from our cache to the destination */
1184         Copy2d(dst, dst_pitch,
1185                cache, w16,
1186                width, hblock, cpu);
1187
1188         /* */
1189         src += src_pitch * hblock;
1190         dst += dst_pitch * hblock;
1191     }
1192
1193     ASM_SSE2(cpu, "mfence");
1194 }
1195 static void SplitPlanes(uint8_t *dstu, size_t dstu_pitch,
1196                         uint8_t *dstv, size_t dstv_pitch,
1197                         const uint8_t *src, size_t src_pitch,
1198                         uint8_t *cache, size_t cache_size,
1199                         unsigned width, unsigned height,
1200                         unsigned cpu)
1201 {
1202     const unsigned w2_16 = (2*width+15) & ~15;
1203     const unsigned hstep = cache_size / w2_16;
1204     assert(hstep > 0);
1205
1206     for (unsigned y = 0; y < height; y += hstep) {
1207         const unsigned unaligned = (intptr_t)src & 0x0f;
1208         const unsigned hblock =  __MIN(hstep, height - y);
1209
1210         /* Copy a bunch of line into our cache */
1211         CopyFromUswc(cache, w2_16,
1212                      src, src_pitch,
1213                      unaligned,
1214                      2*width, hblock, cpu);
1215
1216         /* Copy from our cache to the destination */
1217         SplitUV(dstu, dstu_pitch,
1218                 dstv, dstv_pitch,
1219                 cache, w2_16,
1220                 width, hblock, cpu);
1221
1222         /* */
1223         src  += src_pitch  * hblock;
1224         dstu += dstu_pitch * hblock;
1225         dstv += dstv_pitch * hblock;
1226     }
1227
1228     ASM_SSE2(cpu, "mfence");
1229 }
1230
1231 static void CopyFromNv12(picture_t *dst, const D3DLOCKED_RECT *src,
1232                          uint8_t *cache, size_t cache_size,
1233                          unsigned width, unsigned height)
1234 {
1235     const unsigned cpu = vlc_CPU();
1236
1237     /* */
1238     CopyPlane(dst->p[0].p_pixels, dst->p[0].i_pitch,
1239               src->pBits, src->Pitch,
1240               cache, cache_size,
1241               width, height, cpu);
1242     SplitPlanes(dst->p[2].p_pixels, dst->p[2].i_pitch,
1243                 dst->p[1].p_pixels, dst->p[1].i_pitch,
1244                 (const uint8_t*)src->pBits + src->Pitch * height, src->Pitch,
1245                 cache, cache_size,
1246                 width/2, height/2, cpu);
1247
1248     ASM_SSE2(cpu, "emms");
1249 }
1250 static void CopyFromYv12(picture_t *dst, const D3DLOCKED_RECT *src,
1251                          uint8_t *cache, size_t cache_size,
1252                          unsigned width, unsigned height)
1253 {
1254     const unsigned cpu = vlc_CPU();
1255
1256     /* */
1257     for (unsigned n = 0, offset = 0; n < 3; n++) {
1258         const unsigned d = n > 0 ? 2 : 1;
1259         CopyPlane(dst->p[n].p_pixels, dst->p[n].i_pitch,
1260                   (const uint8_t*)src->pBits + offset, src->Pitch/d,
1261                   cache, cache_size,
1262                   width/d, height/d, cpu);
1263         offset += (src->Pitch/d) * (height/d);
1264     }
1265     ASM_SSE2(cpu, "emms");
1266 }
1267 #undef ASM_SSE2
1268 #undef COPY64
1269
1270 #else
1271 vlc_va_t *vlc_va_NewDxva2(vlc_object_t *log, int codec_id)
1272 {
1273     (void)log;
1274     (void)codec_id;
1275     return NULL;
1276 }
1277 #endif