src/gen75_mfc.c

   1 /*
   2  * Copyright © 2010-2011 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the
   6  * "Software"), to deal in the Software without restriction, including
   7  * without limitation the rights to use, copy, modify, merge, publish,
   8  * distribute, sub license, and/or sell copies of the Software, and to
   9  * permit persons to whom the Software is furnished to do so, subject to
  10  * the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the
  13  * next paragraph) shall be included in all copies or substantial portions
  14  * of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  17  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  18  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  19  * IN NO EVENT SHALL PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR
  20  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  21  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  22  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *    Zhao Yakui <yakui.zhao@intel.com>
  26  *    Xiang Haihao <haihao.xiang@intel.com>
  27  *
  28  */
  29
  30 #include <stdio.h>
  31 #include <stdlib.h>
  32 #include <string.h>
  33 #include <assert.h>
  34
  35 #include "assert.h"
  36 #include "intel_batchbuffer.h"
  37 #include "i965_defines.h"
  38 #include "i965_structs.h"
  39 #include "i965_drv_video.h"
  40 #include "i965_encoder.h"
  41
  42 #define B0_STEP_REV             2
  43 #define IS_STEPPING_BPLUS(i965) ((i965->intel.revision) >= B0_STEP_REV)
  44
  45 static void
  46 gen75_mfc_pipe_mode_select(VADriverContextP ctx,
  47                            int standard_select,
  48                            struct gen6_encoder_context *gen6_encoder_context,
  49                            struct intel_batchbuffer *batch)
  50 {
  51     if (batch == NULL)
  52         batch = gen6_encoder_context->base.batch;
  53
  54     assert(standard_select == MFX_FORMAT_MPEG2 ||
  55            standard_select == MFX_FORMAT_AVC);
  56
  57     BEGIN_BCS_BATCH(batch, 5);
  58     OUT_BCS_BATCH(batch, MFX_PIPE_MODE_SELECT | (5 - 2));
  59     OUT_BCS_BATCH(batch,
  60                   (MFX_LONG_MODE << 17) | /* Must be long format for encoder */
  61                   (MFD_MODE_VLD << 15) | /* VLD mode */
  62                   (0 << 10) | /* disable Stream-Out */
  63                   (1 << 9)  | /* Post Deblocking Output */
  64                   (0 << 8)  | /* Pre Deblocking Output */
  65                   (0 << 5)  | /* not in stitch mode */
  66                   (1 << 4)  | /* encoding mode */
  67                   (standard_select << 0));  /* standard select: avc or mpeg2 */
  68     OUT_BCS_BATCH(batch,
  69                   (0 << 7)  | /* expand NOA bus flag */
  70                   (0 << 6)  | /* disable slice-level clock gating */
  71                   (0 << 5)  | /* disable clock gating for NOA */
  72                   (0 << 4)  | /* terminate if AVC motion and POC table error occurs */
  73                   (0 << 3)  | /* terminate if AVC mbdata error occurs */
  74                   (0 << 2)  | /* terminate if AVC CABAC/CAVLC decode error occurs */
  75                   (0 << 1)  |
  76                   (0 << 0));
  77     OUT_BCS_BATCH(batch, 0);
  78     OUT_BCS_BATCH(batch, 0);
  79
  80     ADVANCE_BCS_BATCH(batch);
  81 }
  82
  83 #define         INTER_MODE_MASK         0x03
  84 #define         INTER_8X8               0x03
  85 #define         SUBMB_SHAPE_MASK        0x00FF00
  86
  87 #define         INTER_MV8               (4 << 20)
  88 #define         INTER_MV32              (6 << 20)
  89
  90
  91 static void
  92 gen75_mfc_surface_state(VADriverContextP ctx,
  93                         struct gen6_encoder_context *gen6_encoder_context,
  94                         struct intel_batchbuffer *batch)
  95 {
  96     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
  97
  98     if (batch == NULL)
  99         batch = gen6_encoder_context->base.batch;
 100
 101     BEGIN_BCS_BATCH(batch, 6);
 102
 103     OUT_BCS_BATCH(batch, MFX_SURFACE_STATE | (6 - 2));
 104     OUT_BCS_BATCH(batch, 0);
 105     OUT_BCS_BATCH(batch,
 106                   ((mfc_context->surface_state.height - 1) << 18) |
 107                   ((mfc_context->surface_state.width - 1) << 4));
 108     OUT_BCS_BATCH(batch,
 109                   (MFX_SURFACE_PLANAR_420_8 << 28) | /* 420 planar YUV surface */
 110                   (1 << 27) | /* must be 1 for interleave U/V, hardware requirement */
 111                   (0 << 22) | /* surface object control state, FIXME??? */
 112                   ((mfc_context->surface_state.w_pitch - 1) << 3) | /* pitch */
 113                   (0 << 2)  | /* must be 0 for interleave U/V */
 114                   (1 << 1)  | /* must be tiled */
 115                   (I965_TILEWALK_YMAJOR << 0));  /* tile walk, TILEWALK_YMAJOR */
 116     OUT_BCS_BATCH(batch,
 117                   (0 << 16) |                                                           /* must be 0 for interleave U/V */
 118                   (mfc_context->surface_state.h_pitch));                /* y offset for U(cb) */
 119     OUT_BCS_BATCH(batch, 0);
 120     ADVANCE_BCS_BATCH(batch);
 121 }
 122
 123 static void
 124 gen75_mfc_pipe_buf_addr_state_bplus(VADriverContextP ctx,
 125                                     struct gen6_encoder_context *gen6_encoder_context,
 126                                     struct intel_batchbuffer *batch)
 127 {
 128     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 129     int i;
 130
 131     if (batch == NULL)
 132         batch = gen6_encoder_context->base.batch;
 133
 134     BEGIN_BCS_BATCH(batch, 61);
 135
 136     OUT_BCS_BATCH(batch, MFX_PIPE_BUF_ADDR_STATE | (61 - 2));
 137
 138     /* the DW1-3 is for pre_deblocking */
 139         OUT_BCS_BATCH(batch, 0);                                                                                        /* pre output addr   */
 140
 141         OUT_BCS_BATCH(batch, 0);
 142         OUT_BCS_BATCH(batch, 0);
 143      /* the DW4-6 is for the post_deblocking */
 144
 145     if (mfc_context->post_deblocking_output.bo)
 146         OUT_BCS_RELOC(batch, mfc_context->post_deblocking_output.bo,
 147                       I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 148                       0);                                                                                       /* post output addr  */
 149     else
 150         OUT_BCS_BATCH(batch, 0);
 151
 152         OUT_BCS_BATCH(batch, 0);
 153         OUT_BCS_BATCH(batch, 0);
 154
 155      /* the DW7-9 is for the uncompressed_picture */
 156     OUT_BCS_RELOC(batch, mfc_context->uncompressed_picture_source.bo,
 157                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 158                   0); /* uncompressed data */
 159
 160         OUT_BCS_BATCH(batch, 0);
 161         OUT_BCS_BATCH(batch, 0);
 162
 163      /* the DW10-12 is for the mb status */
 164         OUT_BCS_BATCH(batch, 0);
 165         OUT_BCS_BATCH(batch, 0);
 166         OUT_BCS_BATCH(batch, 0);
 167
 168      /* the DW13-15 is for the intra_row_store_scratch */
 169     OUT_BCS_RELOC(batch, mfc_context->intra_row_store_scratch_buffer.bo,
 170                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 171                   0);
 172         OUT_BCS_BATCH(batch, 0);
 173         OUT_BCS_BATCH(batch, 0);
 174
 175      /* the DW16-18 is for the deblocking filter */
 176     OUT_BCS_RELOC(batch, mfc_context->deblocking_filter_row_store_scratch_buffer.bo,
 177                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 178                   0);
 179         OUT_BCS_BATCH(batch, 0);
 180         OUT_BCS_BATCH(batch, 0);
 181
 182     /* the DW 19-50 is for Reference pictures*/
 183     for (i = 0; i < ARRAY_ELEMS(mfc_context->reference_surfaces); i++) {
 184         if ( mfc_context->reference_surfaces[i].bo != NULL) {
 185             OUT_BCS_RELOC(batch, mfc_context->reference_surfaces[i].bo,
 186                           I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 187                           0);
 188         } else {
 189             OUT_BCS_BATCH(batch, 0);
 190         }
 191         OUT_BCS_BATCH(batch, 0);
 192     }
 193         OUT_BCS_BATCH(batch, 0);
 194
 195         /* The DW 52-54 is for the MB status buffer */
 196         OUT_BCS_BATCH(batch, 0);
 197
 198         OUT_BCS_BATCH(batch, 0);
 199         OUT_BCS_BATCH(batch, 0);
 200
 201         /* the DW 55-57 is the ILDB buffer */
 202         OUT_BCS_BATCH(batch, 0);
 203         OUT_BCS_BATCH(batch, 0);
 204         OUT_BCS_BATCH(batch, 0);
 205
 206         /* the DW 58-60 is the second ILDB buffer */
 207         OUT_BCS_BATCH(batch, 0);
 208         OUT_BCS_BATCH(batch, 0);
 209         OUT_BCS_BATCH(batch, 0);
 210     ADVANCE_BCS_BATCH(batch);
 211 }
 212
 213 static void
 214 gen75_mfc_pipe_buf_addr_state(VADriverContextP ctx,
 215                               struct gen6_encoder_context *gen6_encoder_context,
 216                               struct intel_batchbuffer *batch)
 217 {
 218     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 219     int i;
 220     struct i965_driver_data *i965 = i965_driver_data(ctx);
 221
 222     if (batch == NULL)
 223         batch = gen6_encoder_context->base.batch;
 224
 225     if (IS_STEPPING_BPLUS(i965)) {
 226         gen75_mfc_pipe_buf_addr_state_bplus(ctx, gen6_encoder_context, batch);
 227         return;
 228     }
 229
 230     BEGIN_BCS_BATCH(batch, 25);
 231
 232     OUT_BCS_BATCH(batch, MFX_PIPE_BUF_ADDR_STATE | (25 - 2));
 233
 234     OUT_BCS_BATCH(batch, 0);                                                                                    /* pre output addr   */
 235
 236     OUT_BCS_RELOC(batch, mfc_context->post_deblocking_output.bo,
 237                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 238                   0);                                                                                   /* post output addr  */
 239
 240     OUT_BCS_RELOC(batch, mfc_context->uncompressed_picture_source.bo,
 241                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 242                   0);                                                                                   /* uncompressed data */
 243
 244     OUT_BCS_BATCH(batch, 0);                                                                                    /* StreamOut data*/
 245     OUT_BCS_RELOC(batch, mfc_context->intra_row_store_scratch_buffer.bo,
 246                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 247                   0);
 248     OUT_BCS_RELOC(batch, mfc_context->deblocking_filter_row_store_scratch_buffer.bo,
 249                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 250                   0);
 251     /* 7..22 Reference pictures*/
 252     for (i = 0; i < ARRAY_ELEMS(mfc_context->reference_surfaces); i++) {
 253         if ( mfc_context->reference_surfaces[i].bo != NULL) {
 254             OUT_BCS_RELOC(batch, mfc_context->reference_surfaces[i].bo,
 255                           I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 256                           0);
 257         } else {
 258             OUT_BCS_BATCH(batch, 0);
 259         }
 260     }
 261     OUT_BCS_BATCH(batch, 0);                                                                                    /* no block status  */
 262
 263     OUT_BCS_BATCH(batch, 0);
 264
 265     ADVANCE_BCS_BATCH(batch);
 266 }
 267
 268
 269 static void
 270 gen75_mfc_ind_obj_base_addr_state_bplus(VADriverContextP ctx,
 271                                         struct gen6_encoder_context *gen6_encoder_context,
 272                                         struct intel_batchbuffer *batch)
 273 {
 274     struct gen6_vme_context *vme_context = &gen6_encoder_context->vme_context;
 275     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 276
 277     if (batch == NULL)
 278         batch = gen6_encoder_context->base.batch;
 279
 280     BEGIN_BCS_BATCH(batch, 26);
 281
 282     OUT_BCS_BATCH(batch, MFX_IND_OBJ_BASE_ADDR_STATE | (26 - 2));
 283         /* the DW1-3 is for the MFX indirect bistream offset */
 284     OUT_BCS_BATCH(batch, 0);
 285     OUT_BCS_BATCH(batch, 0);
 286     OUT_BCS_BATCH(batch, 0);
 287         /* the DW4-5 is the MFX upper bound */
 288     OUT_BCS_BATCH(batch, 0);
 289     OUT_BCS_BATCH(batch, 0);
 290
 291     /* the DW6-10 is for MFX Indirect MV Object Base Address */
 292     OUT_BCS_RELOC(batch, vme_context->vme_output.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0);
 293     OUT_BCS_BATCH(batch, 0);
 294     OUT_BCS_BATCH(batch, 0);
 295     OUT_BCS_BATCH(batch, 0x80000000); /* must set, up to 2G */
 296     OUT_BCS_BATCH(batch, 0);
 297
 298      /* the DW11-15 is for MFX IT-COFF. Not used on encoder */
 299     OUT_BCS_BATCH(batch, 0);
 300     OUT_BCS_BATCH(batch, 0);
 301     OUT_BCS_BATCH(batch, 0);
 302     OUT_BCS_BATCH(batch, 0);
 303     OUT_BCS_BATCH(batch, 0);
 304
 305      /* the DW16-20 is for MFX indirect DBLK. Not used on encoder */
 306     OUT_BCS_BATCH(batch, 0);
 307     OUT_BCS_BATCH(batch, 0);
 308     OUT_BCS_BATCH(batch, 0);
 309     OUT_BCS_BATCH(batch, 0);
 310     OUT_BCS_BATCH(batch, 0);
 311
 312     /* the DW21-25 is for MFC Indirect PAK-BSE Object Base Address for Encoder*/
 313     OUT_BCS_RELOC(batch,
 314                   mfc_context->mfc_indirect_pak_bse_object.bo,
 315                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 316                   0);
 317     OUT_BCS_BATCH(batch, 0);
 318     OUT_BCS_BATCH(batch, 0);
 319
 320     OUT_BCS_BATCH(batch, 0);
 321     OUT_BCS_BATCH(batch, 0x00000000);
 322
 323     ADVANCE_BCS_BATCH(batch);
 324 }
 325
 326 static void
 327 gen75_mfc_ind_obj_base_addr_state(VADriverContextP ctx,
 328                                   struct gen6_encoder_context *gen6_encoder_context,
 329                                   struct intel_batchbuffer *batch)
 330 {
 331     struct gen6_vme_context *vme_context = &gen6_encoder_context->vme_context;
 332     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 333     struct i965_driver_data *i965 = i965_driver_data(ctx);
 334
 335     if (batch == NULL)
 336         batch = gen6_encoder_context->base.batch;
 337
 338     if (IS_STEPPING_BPLUS(i965)) {
 339         gen75_mfc_ind_obj_base_addr_state_bplus(ctx, gen6_encoder_context, batch);
 340         return;
 341     }
 342
 343     BEGIN_BCS_BATCH(batch, 11);
 344
 345     OUT_BCS_BATCH(batch, MFX_IND_OBJ_BASE_ADDR_STATE | (11 - 2));
 346     OUT_BCS_BATCH(batch, 0);
 347     OUT_BCS_BATCH(batch, 0);
 348     /* MFX Indirect MV Object Base Address */
 349     OUT_BCS_RELOC(batch, vme_context->vme_output.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0);
 350     OUT_BCS_BATCH(batch, 0x80000000); /* must set, up to 2G */
 351     OUT_BCS_BATCH(batch, 0);
 352     OUT_BCS_BATCH(batch, 0);
 353     OUT_BCS_BATCH(batch, 0);
 354     OUT_BCS_BATCH(batch, 0);
 355     /*MFC Indirect PAK-BSE Object Base Address for Encoder*/
 356     OUT_BCS_RELOC(batch,
 357                   mfc_context->mfc_indirect_pak_bse_object.bo,
 358                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 359                   0);
 360     OUT_BCS_BATCH(batch, 0x00000000); /* must set, up to 2G */
 361
 362     ADVANCE_BCS_BATCH(batch);
 363 }
 364
 365 static void
 366 gen75_mfc_bsp_buf_base_addr_state_bplus(VADriverContextP ctx,
 367                                         struct gen6_encoder_context *gen6_encoder_context,
 368                                         struct intel_batchbuffer *batch)
 369 {
 370     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 371
 372     if (batch == NULL)
 373         batch = gen6_encoder_context->base.batch;
 374
 375     BEGIN_BCS_BATCH(batch, 10);
 376
 377     OUT_BCS_BATCH(batch, MFX_BSP_BUF_BASE_ADDR_STATE | (10 - 2));
 378     OUT_BCS_RELOC(batch, mfc_context->bsd_mpc_row_store_scratch_buffer.bo,
 379                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 380                   0);
 381     OUT_BCS_BATCH(batch, 0);
 382     OUT_BCS_BATCH(batch, 0);
 383
 384         /* the DW4-6 is for MPR Row Store Scratch Buffer Base Address */
 385     OUT_BCS_BATCH(batch, 0);
 386     OUT_BCS_BATCH(batch, 0);
 387     OUT_BCS_BATCH(batch, 0);
 388
 389         /* the DW7-9 is for Bitplane Read Buffer Base Address */
 390     OUT_BCS_BATCH(batch, 0);
 391     OUT_BCS_BATCH(batch, 0);
 392     OUT_BCS_BATCH(batch, 0);
 393
 394     ADVANCE_BCS_BATCH(batch);
 395 }
 396
 397 static void
 398 gen75_mfc_bsp_buf_base_addr_state(VADriverContextP ctx,
 399                                   struct gen6_encoder_context *gen6_encoder_context,
 400                                   struct intel_batchbuffer *batch)
 401 {
 402     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 403     struct i965_driver_data *i965 = i965_driver_data(ctx);
 404
 405     if (batch == NULL)
 406         batch = gen6_encoder_context->base.batch;
 407
 408     if (IS_STEPPING_BPLUS(i965)) {
 409         gen75_mfc_bsp_buf_base_addr_state_bplus(ctx, gen6_encoder_context, batch);
 410         return;
 411     }
 412
 413
 414     BEGIN_BCS_BATCH(batch, 4);
 415
 416     OUT_BCS_BATCH(batch, MFX_BSP_BUF_BASE_ADDR_STATE | (4 - 2));
 417     OUT_BCS_RELOC(batch, mfc_context->bsd_mpc_row_store_scratch_buffer.bo,
 418                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 419                   0);
 420     OUT_BCS_BATCH(batch, 0);
 421     OUT_BCS_BATCH(batch, 0);
 422
 423     ADVANCE_BCS_BATCH(batch);
 424 }
 425
 426 static void
 427 gen75_mfc_avc_img_state(VADriverContextP ctx,
 428                         struct gen6_encoder_context *gen6_encoder_context,
 429                         struct intel_batchbuffer *batch)
 430 {
 431     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 432     int width_in_mbs = (mfc_context->surface_state.width + 15) / 16;
 433     int height_in_mbs = (mfc_context->surface_state.height + 15) / 16;
 434
 435     if (batch == NULL)
 436         batch = gen6_encoder_context->base.batch;
 437
 438     BEGIN_BCS_BATCH(batch, 16);
 439     OUT_BCS_BATCH(batch, MFX_AVC_IMG_STATE | (16 - 2));
 440     OUT_BCS_BATCH(batch,
 441                   ((width_in_mbs * height_in_mbs) & 0xFFFF));
 442     OUT_BCS_BATCH(batch,
 443                   ((height_in_mbs - 1) << 16) |
 444                   ((width_in_mbs - 1) << 0));
 445     OUT_BCS_BATCH(batch,
 446                   (0 << 24) |   /* Second Chroma QP Offset */
 447                   (0 << 16) |   /* Chroma QP Offset */
 448                   (0 << 14) |   /* Max-bit conformance Intra flag */
 449                   (0 << 13) |   /* Max Macroblock size conformance Inter flag */
 450                   (0 << 12) |   /* FIXME: Weighted_Pred_Flag */
 451                   (0 << 10) |   /* FIXME: Weighted_BiPred_Idc */
 452                   (0 << 8)  |   /* FIXME: Image Structure */
 453                   (0 << 0) );   /* Current Decoed Image Frame Store ID, reserved in Encode mode */
 454     OUT_BCS_BATCH(batch,
 455                   (0 << 16) |   /* Mininum Frame size */
 456                   (0 << 15) |   /* Disable reading of Macroblock Status Buffer */
 457                   (0 << 14) |   /* Load BitStream Pointer only once, 1 slic 1 frame */
 458                   (0 << 13) |   /* CABAC 0 word insertion test enable */
 459                   (1 << 12) |   /* MVUnpackedEnable,compliant to DXVA */
 460                   (1 << 10) |   /* Chroma Format IDC, 4:2:0 */
 461                   (0 << 8)  |   /* FIXME: MbMvFormatFlag */
 462                   (1 << 7)  |   /* 0:CAVLC encoding mode,1:CABAC */
 463                   (0 << 6)  |   /* Only valid for VLD decoding mode */
 464                   (0 << 5)  |   /* Constrained Intra Predition Flag, from PPS */
 465                   (0 << 4)  |   /* Direct 8x8 inference flag */
 466                   (0 << 3)  |   /* Only 8x8 IDCT Transform Mode Flag */
 467                   (1 << 2)  |   /* Frame MB only flag */
 468                   (0 << 1)  |   /* MBAFF mode is in active */
 469                   (0 << 0));    /* Field picture flag */
 470     OUT_BCS_BATCH(batch, 0);    /* Mainly about MB rate control and debug, just ignoring */
 471     OUT_BCS_BATCH(batch,        /* Inter and Intra Conformance Max size limit */
 472                   (0xBB8 << 16) |       /* InterMbMaxSz */
 473                   (0xEE8) );            /* IntraMbMaxSz */
 474     OUT_BCS_BATCH(batch, 0);            /* Reserved */
 475     OUT_BCS_BATCH(batch, 0);            /* Slice QP Delta for bitrate control */
 476     OUT_BCS_BATCH(batch, 0);            /* Slice QP Delta for bitrate control */
 477     OUT_BCS_BATCH(batch, 0x8C000000);
 478     OUT_BCS_BATCH(batch, 0x00010000);
 479     OUT_BCS_BATCH(batch, 0);
 480     OUT_BCS_BATCH(batch, 0);
 481     OUT_BCS_BATCH(batch, 0);
 482     OUT_BCS_BATCH(batch, 0);
 483
 484     ADVANCE_BCS_BATCH(batch);
 485 }
 486
 487
 488 static void
 489 gen75_mfc_avc_directmode_state_bplus(VADriverContextP ctx,
 490                                      struct gen6_encoder_context *gen6_encoder_context,
 491                                      struct intel_batchbuffer *batch)
 492 {
 493     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 494     int i;
 495
 496     if (batch == NULL)
 497         batch = gen6_encoder_context->base.batch;
 498
 499     BEGIN_BCS_BATCH(batch, 71);
 500
 501     OUT_BCS_BATCH(batch, MFX_AVC_DIRECTMODE_STATE | (71 - 2));
 502
 503     /* Reference frames and Current frames */
 504     /* the DW1-32 is for the direct MV for reference */
 505     for(i = 0; i < NUM_MFC_DMV_BUFFERS - 2; i += 2) {
 506         if ( mfc_context->direct_mv_buffers[i].bo != NULL) {
 507             OUT_BCS_RELOC(batch, mfc_context->direct_mv_buffers[i].bo,
 508                           I915_GEM_DOMAIN_INSTRUCTION, 0,
 509                           0);
 510             OUT_BCS_BATCH(batch, 0);
 511         } else {
 512             OUT_BCS_BATCH(batch, 0);
 513             OUT_BCS_BATCH(batch, 0);
 514         }
 515     }
 516         OUT_BCS_BATCH(batch, 0);
 517
 518         /* the DW34-36 is the MV for the current reference */
 519         OUT_BCS_RELOC(batch, mfc_context->direct_mv_buffers[NUM_MFC_DMV_BUFFERS - 2].bo,
 520                           I915_GEM_DOMAIN_INSTRUCTION, 0,
 521                           0);
 522
 523         OUT_BCS_BATCH(batch, 0);
 524         OUT_BCS_BATCH(batch, 0);
 525
 526     /* POL list */
 527     for(i = 0; i < 32; i++) {
 528         OUT_BCS_BATCH(batch, i/2);
 529     }
 530     OUT_BCS_BATCH(batch, 0);
 531     OUT_BCS_BATCH(batch, 0);
 532
 533     ADVANCE_BCS_BATCH(batch);
 534 }
 535
 536 static void gen75_mfc_avc_directmode_state(VADriverContextP ctx,
 537                                            struct gen6_encoder_context *gen6_encoder_context,
 538                                            struct intel_batchbuffer *batch)
 539 {
 540     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 541     int i;
 542     struct i965_driver_data *i965 = i965_driver_data(ctx);
 543
 544     if (batch == NULL)
 545         batch = gen6_encoder_context->base.batch;
 546
 547     if (IS_STEPPING_BPLUS(i965)) {
 548         gen75_mfc_avc_directmode_state_bplus(ctx, gen6_encoder_context, batch);
 549         return;
 550     }
 551
 552     BEGIN_BCS_BATCH(batch, 69);
 553
 554     OUT_BCS_BATCH(batch, MFX_AVC_DIRECTMODE_STATE | (69 - 2));
 555     //TODO: reference DMV
 556     for (i = 0; i < NUM_MFC_DMV_BUFFERS - 2; i++){
 557         if (mfc_context->direct_mv_buffers[i].bo)
 558                 OUT_BCS_RELOC(batch, mfc_context->direct_mv_buffers[i].bo,
 559                           I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 560                   0);
 561         else
 562                 OUT_BCS_BATCH(batch, 0);
 563     }
 564
 565     //TODO: current DMV just for test
 566 #if 0
 567     OUT_BCS_RELOC(batch, mfc_context->direct_mv_buffers[0].bo,
 568                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 569                   0);
 570 #else
 571     //drm_intel_bo_pin(mfc_context->direct_mv_buffers[0].bo, 0x1000);
 572     //OUT_BCS_BATCH(batch, mfc_context->direct_mv_buffers[0].bo->offset);
 573     OUT_BCS_RELOC(batch, mfc_context->direct_mv_buffers[NUM_MFC_DMV_BUFFERS - 2].bo,
 574                   I915_GEM_DOMAIN_INSTRUCTION, I915_GEM_DOMAIN_INSTRUCTION,
 575                   0);
 576 #endif
 577
 578
 579     OUT_BCS_BATCH(batch, 0);
 580
 581     //TODO: POL list
 582     for(i = 0; i < 34; i++) {
 583         OUT_BCS_BATCH(batch, 0);
 584     }
 585
 586     ADVANCE_BCS_BATCH(batch);
 587 }
 588
 589 static void gen75_mfc_avc_slice_state(VADriverContextP ctx,
 590                                       int intra_slice,
 591                                       struct gen6_encoder_context *gen6_encoder_context,
 592                                       struct intel_batchbuffer *batch)
 593 {
 594     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 595
 596     if (batch == NULL)
 597         batch = gen6_encoder_context->base.batch;
 598
 599     BEGIN_BCS_BATCH(batch, 11);;
 600
 601     OUT_BCS_BATCH(batch, MFX_AVC_SLICE_STATE | (11 - 2) );
 602
 603     if ( intra_slice )
 604         OUT_BCS_BATCH(batch, 2);                        /*Slice Type: I Slice*/
 605     else
 606         OUT_BCS_BATCH(batch, 0);                        /*Slice Type: P Slice*/
 607
 608     if ( intra_slice )
 609         OUT_BCS_BATCH(batch, 0);                        /*no reference frames and pred_weight_table*/
 610     else
 611         OUT_BCS_BATCH(batch, 0x00010000);       /*1 reference frame*/
 612
 613     OUT_BCS_BATCH(batch, (0<<24) |                /*Enable deblocking operation*/
 614                   (26<<16) |                    /*Slice Quantization Parameter*/
 615                   0x0202 );
 616     OUT_BCS_BATCH(batch, 0);                    /*First MB X&Y , the postion of current slice*/
 617     OUT_BCS_BATCH(batch, ( ((mfc_context->surface_state.height+15)/16) << 16) );
 618
 619     OUT_BCS_BATCH(batch,
 620                   (0<<31) |             /*RateControlCounterEnable = disable*/
 621                   (1<<30) |             /*ResetRateControlCounter*/
 622                   (2<<28) |             /*RC Triggle Mode = Loose Rate Control*/
 623                   (1<<19) |             /*IsLastSlice*/
 624                   (0<<18) |             /*BitstreamOutputFlag Compressed BitStream Output Disable Flag 0:enable 1:disable*/
 625                   (0<<17) |             /*HeaderPresentFlag*/
 626                   (1<<16) |             /*SliceData PresentFlag*/
 627                   (0<<15) |             /*TailPresentFlag*/
 628                   (1<<13) |             /*RBSP NAL TYPE*/
 629                   (0<<12) );            /*CabacZeroWordInsertionEnable*/
 630
 631
 632     OUT_BCS_BATCH(batch, mfc_context->mfc_indirect_pak_bse_object.offset);
 633
 634     OUT_BCS_BATCH(batch, 0);
 635     OUT_BCS_BATCH(batch, 0);
 636     OUT_BCS_BATCH(batch, 0);
 637
 638     ADVANCE_BCS_BATCH(batch);
 639 }
 640
 641 static void
 642 gen75_mfc_qm_state(VADriverContextP ctx,
 643                    int qm_type,
 644                    unsigned int *qm,
 645                    int qm_length,
 646                    struct gen6_encoder_context *gen6_encoder_context,
 647                    struct intel_batchbuffer *batch)
 648 {
 649     unsigned int qm_buffer[16];
 650
 651     if (batch == NULL)
 652         batch = gen6_encoder_context->base.batch;
 653
 654     assert(qm_length <= 16);
 655     assert(sizeof(*qm) == 4);
 656     memcpy(qm_buffer, qm, qm_length * 4);
 657
 658     BEGIN_BCS_BATCH(batch, 18);
 659     OUT_BCS_BATCH(batch, MFX_QM_STATE | (18 - 2));
 660     OUT_BCS_BATCH(batch, qm_type << 0);
 661     intel_batchbuffer_data(batch, qm_buffer, 16 * 4);
 662     ADVANCE_BCS_BATCH(batch);
 663 }
 664
 665 static void gen75_mfc_avc_qm_state(VADriverContextP ctx,
 666                                    struct gen6_encoder_context *gen6_encoder_context,
 667                                    struct intel_batchbuffer *batch)
 668 {
 669     unsigned int qm[16] = {
 670         0x10101010, 0x10101010, 0x10101010, 0x10101010,
 671         0x10101010, 0x10101010, 0x10101010, 0x10101010,
 672         0x10101010, 0x10101010, 0x10101010, 0x10101010,
 673         0x10101010, 0x10101010, 0x10101010, 0x10101010
 674     };
 675
 676     gen75_mfc_qm_state(ctx, MFX_QM_AVC_4X4_INTRA_MATRIX, qm, 12, gen6_encoder_context, batch);
 677     gen75_mfc_qm_state(ctx, MFX_QM_AVC_4X4_INTER_MATRIX, qm, 12, gen6_encoder_context, batch);
 678     gen75_mfc_qm_state(ctx, MFX_QM_AVC_8x8_INTRA_MATRIX, qm, 16, gen6_encoder_context, batch);
 679     gen75_mfc_qm_state(ctx, MFX_QM_AVC_8x8_INTER_MATRIX, qm, 16, gen6_encoder_context, batch);
 680 }
 681
 682 static void
 683 gen75_mfc_fqm_state(VADriverContextP ctx,
 684                     int fqm_type,
 685                     unsigned int *fqm,
 686                     int fqm_length,
 687                     struct gen6_encoder_context *gen6_encoder_context,
 688                     struct intel_batchbuffer *batch)
 689 {
 690     unsigned int fqm_buffer[32];
 691
 692     if (batch == NULL)
 693         batch = gen6_encoder_context->base.batch;
 694
 695     assert(fqm_length <= 32);
 696     assert(sizeof(*fqm) == 4);
 697     memcpy(fqm_buffer, fqm, fqm_length * 4);
 698
 699     BEGIN_BCS_BATCH(batch, 34);
 700     OUT_BCS_BATCH(batch, MFX_FQM_STATE | (34 - 2));
 701     OUT_BCS_BATCH(batch, fqm_type << 0);
 702     intel_batchbuffer_data(batch, fqm_buffer, 32 * 4);
 703     ADVANCE_BCS_BATCH(batch);
 704 }
 705
 706 static void gen75_mfc_avc_fqm_state(VADriverContextP ctx,
 707                                     struct gen6_encoder_context *gen6_encoder_context,
 708                                     struct intel_batchbuffer *batch)
 709 {
 710     unsigned int qm[32] = {
 711         0x10001000, 0x10001000, 0x10001000, 0x10001000,
 712         0x10001000, 0x10001000, 0x10001000, 0x10001000,
 713         0x10001000, 0x10001000, 0x10001000, 0x10001000,
 714         0x10001000, 0x10001000, 0x10001000, 0x10001000,
 715         0x10001000, 0x10001000, 0x10001000, 0x10001000,
 716         0x10001000, 0x10001000, 0x10001000, 0x10001000,
 717         0x10001000, 0x10001000, 0x10001000, 0x10001000,
 718         0x10001000, 0x10001000, 0x10001000, 0x10001000
 719     };
 720
 721     gen75_mfc_fqm_state(ctx, MFX_QM_AVC_4X4_INTRA_MATRIX, qm, 24, gen6_encoder_context, batch);
 722     gen75_mfc_fqm_state(ctx, MFX_QM_AVC_4X4_INTER_MATRIX, qm, 24, gen6_encoder_context, batch);
 723     gen75_mfc_fqm_state(ctx, MFX_QM_AVC_8x8_INTRA_MATRIX, qm, 32, gen6_encoder_context, batch);
 724     gen75_mfc_fqm_state(ctx, MFX_QM_AVC_8x8_INTER_MATRIX, qm, 32, gen6_encoder_context, batch);
 725 }
 726
 727 static void gen75_mfc_avc_ref_idx_state(VADriverContextP ctx,
 728                                         struct gen6_encoder_context *gen6_encoder_context,
 729                                         struct intel_batchbuffer *batch)
 730 {
 731     int i;
 732
 733     if (batch == NULL)
 734         batch = gen6_encoder_context->base.batch;
 735
 736     BEGIN_BCS_BATCH(batch, 10);
 737
 738     OUT_BCS_BATCH(batch, MFX_AVC_REF_IDX_STATE | 8);
 739     OUT_BCS_BATCH(batch, 0);                  //Select L0
 740
 741     OUT_BCS_BATCH(batch, 0x80808000);         //Only 1 reference
 742     for(i = 0; i < 7; i++) {
 743         OUT_BCS_BATCH(batch, 0x80808080);
 744     }
 745
 746     ADVANCE_BCS_BATCH(batch);
 747 }
 748
 749 static void
 750 gen75_mfc_avc_insert_object(VADriverContextP ctx, int flush_data,
 751                             struct gen6_encoder_context *gen6_encoder_context,
 752                             struct intel_batchbuffer *batch)
 753 {
 754     if (batch == NULL)
 755         batch = gen6_encoder_context->base.batch;
 756
 757     BEGIN_BCS_BATCH(batch, 4);
 758
 759     OUT_BCS_BATCH(batch, MFC_AVC_INSERT_OBJECT | (4 -2 ) );
 760     OUT_BCS_BATCH(batch, (32<<8) |
 761                   (1 << 3) |
 762                   (1 << 2) |
 763                   (flush_data << 1) |
 764                   (1<<0) );
 765     OUT_BCS_BATCH(batch, 0x00000003);
 766     OUT_BCS_BATCH(batch, 0xABCD1234);
 767
 768     ADVANCE_BCS_BATCH(batch);
 769 }
 770
 771 static int
 772 gen75_mfc_avc_pak_object_intra(VADriverContextP ctx, int x, int y, int end_mb, int qp,unsigned int *msg,
 773                                struct gen6_encoder_context *gen6_encoder_context,
 774                                struct intel_batchbuffer *batch)
 775 {
 776     int len_in_dwords = 12;
 777
 778     unsigned int intra_msg;
 779 #define         INTRA_MSG_FLAG          (1 << 13)
 780 #define         INTRA_MBTYPE_MASK       (0x1F0000)
 781
 782     if (batch == NULL)
 783         batch = gen6_encoder_context->base.batch;
 784
 785     BEGIN_BCS_BATCH(batch, len_in_dwords);
 786
 787     intra_msg = msg[0] & 0xC0FF;
 788     intra_msg |= INTRA_MSG_FLAG;
 789     intra_msg |= ((msg[0] & INTRA_MBTYPE_MASK) >> 8);
 790     OUT_BCS_BATCH(batch, MFC_AVC_PAK_OBJECT | (len_in_dwords - 2));
 791     OUT_BCS_BATCH(batch, 0);
 792     OUT_BCS_BATCH(batch, 0);
 793     OUT_BCS_BATCH(batch,
 794                   (0 << 24) |           /* PackedMvNum, Debug*/
 795                   (0 << 20) |           /* No motion vector */
 796                   (1 << 19) |           /* CbpDcY */
 797                   (1 << 18) |           /* CbpDcU */
 798                   (1 << 17) |           /* CbpDcV */
 799                   intra_msg);
 800
 801     OUT_BCS_BATCH(batch, (0xFFFF<<16) | (y << 8) | x);          /* Code Block Pattern for Y*/
 802     OUT_BCS_BATCH(batch, 0x000F000F);                                                   /* Code Block Pattern */
 803     OUT_BCS_BATCH(batch, (0 << 27) | (end_mb << 26) | qp);      /* Last MB */
 804
 805     /*Stuff for Intra MB*/
 806     OUT_BCS_BATCH(batch, msg[1]);                       /* We using Intra16x16 no 4x4 predmode*/
 807     OUT_BCS_BATCH(batch, msg[2]);
 808     OUT_BCS_BATCH(batch, msg[3]&0xFC);
 809
 810     OUT_BCS_BATCH(batch, 0x00000);      /*MaxSizeInWord and TargetSzieInWord*/
 811         OUT_BCS_BATCH(batch, 0);
 812
 813     ADVANCE_BCS_BATCH(batch);
 814
 815     return len_in_dwords;
 816 }
 817
 818 static int gen75_mfc_avc_pak_object_inter(VADriverContextP ctx, int x, int y, int end_mb, int qp,
 819                                           unsigned int offset, unsigned int *msg, struct gen6_encoder_context *gen6_encoder_context,
 820                                           struct intel_batchbuffer *batch)
 821 {
 822     int len_in_dwords = 12;
 823     unsigned int inter_msg;
 824
 825     if (batch == NULL)
 826         batch = gen6_encoder_context->base.batch;
 827
 828     BEGIN_BCS_BATCH(batch, len_in_dwords);
 829
 830     OUT_BCS_BATCH(batch, MFC_AVC_PAK_OBJECT | (len_in_dwords - 2));
 831
 832         inter_msg = 32;
 833         /* MV quantity */
 834         if ((msg[0] & INTER_MODE_MASK) == INTER_8X8) {
 835                 if (msg[1] & SUBMB_SHAPE_MASK)
 836                         inter_msg = 128;
 837         }
 838     OUT_BCS_BATCH(batch, inter_msg);         /* 32 MV*/
 839
 840     OUT_BCS_BATCH(batch, offset);
 841         inter_msg = msg[0] & (0x1F00FFFF);
 842         inter_msg |= INTER_MV8;
 843         if (((msg[0] & INTER_MODE_MASK) == INTER_8X8) &&
 844                         (msg[1] & SUBMB_SHAPE_MASK)) {
 845                 inter_msg |= INTER_MV32;
 846         }
 847
 848     OUT_BCS_BATCH(batch, inter_msg);
 849
 850     OUT_BCS_BATCH(batch, (0xFFFF<<16) | (y << 8) | x);        /* Code Block Pattern for Y*/
 851     OUT_BCS_BATCH(batch, 0x000F000F);                         /* Code Block Pattern */
 852     OUT_BCS_BATCH(batch, (0 << 27) | (end_mb << 26) | qp);    /* Last MB */
 853
 854     /*Stuff for Inter MB*/
 855         inter_msg = msg[1] >> 8;
 856     OUT_BCS_BATCH(batch, inter_msg);
 857     OUT_BCS_BATCH(batch, 0x0);
 858     OUT_BCS_BATCH(batch, 0x0);
 859
 860     OUT_BCS_BATCH(batch, 0x00000000); /*MaxSizeInWord and TargetSzieInWord*/
 861
 862     OUT_BCS_BATCH(batch, 0x0);
 863
 864     ADVANCE_BCS_BATCH(batch);
 865
 866     return len_in_dwords;
 867 }
 868
 869 static void gen75_mfc_init(VADriverContextP ctx,
 870                            struct encode_state *encode_state,
 871                            struct gen6_encoder_context *gen6_encoder_context)
 872 {
 873     struct i965_driver_data *i965 = i965_driver_data(ctx);
 874     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 875     dri_bo *bo;
 876     int i;
 877     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param->buffer;
 878     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 879     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 880
 881     /*Encode common setup for MFC*/
 882     dri_bo_unreference(mfc_context->post_deblocking_output.bo);
 883     mfc_context->post_deblocking_output.bo = NULL;
 884
 885     dri_bo_unreference(mfc_context->pre_deblocking_output.bo);
 886     mfc_context->pre_deblocking_output.bo = NULL;
 887
 888     dri_bo_unreference(mfc_context->uncompressed_picture_source.bo);
 889     mfc_context->uncompressed_picture_source.bo = NULL;
 890
 891     dri_bo_unreference(mfc_context->mfc_indirect_pak_bse_object.bo);
 892     mfc_context->mfc_indirect_pak_bse_object.bo = NULL;
 893
 894     for (i = 0; i < MAX_MFC_REFERENCE_SURFACES; i++){
 895         if (mfc_context->reference_surfaces[i].bo != NULL)
 896             dri_bo_unreference(mfc_context->reference_surfaces[i].bo);
 897         mfc_context->reference_surfaces[i].bo = NULL;
 898     }
 899
 900     dri_bo_unreference(mfc_context->intra_row_store_scratch_buffer.bo);
 901     bo = dri_bo_alloc(i965->intel.bufmgr,
 902                       "Buffer",
 903                       width_in_mbs * 64,
 904                       64);
 905     assert(bo);
 906     mfc_context->intra_row_store_scratch_buffer.bo = bo;
 907
 908     dri_bo_unreference(mfc_context->deblocking_filter_row_store_scratch_buffer.bo);
 909     bo = dri_bo_alloc(i965->intel.bufmgr,
 910                       "Buffer",
 911                       4 * width_in_mbs * 64,  /* 4 * width_in_mbs * 64 */
 912                       64);
 913     assert(bo);
 914     mfc_context->deblocking_filter_row_store_scratch_buffer.bo = bo;
 915
 916     dri_bo_unreference(mfc_context->bsd_mpc_row_store_scratch_buffer.bo);
 917     bo = dri_bo_alloc(i965->intel.bufmgr,
 918                       "Buffer",
 919                       2 * width_in_mbs * 64, /* 2 * width_in_mbs * 64 */
 920                       0x1000);
 921     assert(bo);
 922     mfc_context->bsd_mpc_row_store_scratch_buffer.bo = bo;
 923 }
 924
 925 #define         INTRA_RDO_OFFSET        4
 926 #define         INTER_RDO_OFFSET        54
 927 #define         INTER_MSG_OFFSET        52
 928 #define         INTER_MV_OFFSET         224
 929 #define         RDO_MASK                0xFFFF
 930
 931 static void gen75_mfc_avc_pipeline_programing(VADriverContextP ctx,
 932                                       struct encode_state *encode_state,
 933                                       struct gen6_encoder_context *gen6_encoder_context)
 934 {
 935     struct i965_driver_data *i965 = i965_driver_data(ctx);
 936     struct intel_batchbuffer *batch = gen6_encoder_context->base.batch;
 937     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
 938     struct gen6_vme_context *vme_context = &gen6_encoder_context->vme_context;
 939     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param->buffer;
 940     VAEncSliceParameterBuffer *pSliceParameter = (VAEncSliceParameterBuffer *)encode_state->slice_params[0]->buffer; /* FIXME: multi slices */
 941     unsigned int *msg = NULL, offset = 0;
 942     unsigned char *msg_ptr = NULL;
 943     int emit_new_state = 1, object_len_in_bytes;
 944     int is_intra = pSliceParameter->slice_flags.bits.is_intra;
 945     int width_in_mbs = (mfc_context->surface_state.width + 15) / 16;
 946     int height_in_mbs = (mfc_context->surface_state.height + 15) / 16;
 947     int x,y, mb_index;
 948     int inter_rdo, intra_rdo;
 949
 950     intel_batchbuffer_start_atomic_bcs(batch, 0x1000);
 951
 952     dri_bo_map(vme_context->vme_output.bo , 1);
 953     msg_ptr = (unsigned char *)vme_context->vme_output.bo->virtual;
 954     if (is_intra) {
 955         msg = (unsigned int *) (msg_ptr + 0 * vme_context->vme_output.size_block);
 956     } else {
 957         msg = (unsigned int *) (msg_ptr + 0 * vme_context->vme_output.size_block);
 958         offset = 0;
 959     }
 960
 961     for (y = 0; y < height_in_mbs; y++) {
 962         for (x = 0; x < width_in_mbs; x++) {
 963             int last_mb = (y == (height_in_mbs-1)) && ( x == (width_in_mbs-1) );
 964             int qp = pSequenceParameter->initial_qp;
 965              mb_index = (y * width_in_mbs) + x;
 966             if (emit_new_state) {
 967                 intel_batchbuffer_emit_mi_flush(batch);
 968
 969                 gen75_mfc_pipe_mode_select(ctx, MFX_FORMAT_AVC, gen6_encoder_context, batch);
 970                 gen75_mfc_surface_state(ctx, gen6_encoder_context, batch);
 971                 gen75_mfc_ind_obj_base_addr_state(ctx, gen6_encoder_context, batch);
 972
 973                 gen75_mfc_pipe_buf_addr_state(ctx, gen6_encoder_context, batch);
 974                 gen75_mfc_bsp_buf_base_addr_state(ctx, gen6_encoder_context, batch);
 975
 976                 gen75_mfc_avc_img_state(ctx, gen6_encoder_context, batch);
 977                 gen75_mfc_avc_qm_state(ctx, gen6_encoder_context, batch);
 978                 gen75_mfc_avc_fqm_state(ctx, gen6_encoder_context, batch);
 979                 gen75_mfc_avc_directmode_state(ctx, gen6_encoder_context, batch);
 980
 981                 gen75_mfc_avc_ref_idx_state(ctx, gen6_encoder_context, batch);
 982                 gen75_mfc_avc_slice_state(ctx, is_intra, gen6_encoder_context, batch);
 983                 emit_new_state = 0;
 984             }
 985
 986             msg = (unsigned int *) (msg_ptr + mb_index * vme_context->vme_output.size_block);
 987             if (is_intra) {
 988                 object_len_in_bytes = gen75_mfc_avc_pak_object_intra(ctx, x, y, last_mb, qp, msg, gen6_encoder_context, batch);
 989             } else {
 990                 inter_rdo = msg[INTER_RDO_OFFSET] & RDO_MASK;
 991                 intra_rdo = msg[INTRA_RDO_OFFSET] & RDO_MASK;
 992                 if (intra_rdo < inter_rdo) {
 993                     object_len_in_bytes = gen75_mfc_avc_pak_object_intra(ctx, x, y, last_mb, qp, msg, gen6_encoder_context, batch);
 994                 } else {
 995                     msg += INTER_MSG_OFFSET;
 996                     offset = mb_index * vme_context->vme_output.size_block + INTER_MV_OFFSET;
 997                     object_len_in_bytes = gen75_mfc_avc_pak_object_inter(ctx, x, y, last_mb, qp, offset, msg, gen6_encoder_context, batch);
 998                 }
 999             }
1000             if (intel_batchbuffer_check_free_space(batch, object_len_in_bytes) == 0) {
1001                 intel_batchbuffer_end_atomic(batch);
1002                 intel_batchbuffer_flush(batch);
1003                 emit_new_state = 1;
1004                 intel_batchbuffer_start_atomic_bcs(batch, 0x1000);
1005             }
1006         }
1007     }
1008
1009     dri_bo_unmap(vme_context->vme_output.bo);
1010
1011     intel_batchbuffer_end_atomic(batch);
1012 }
1013
1014 static VAStatus gen75_mfc_avc_prepare(VADriverContextP ctx,
1015                                      struct encode_state *encode_state,
1016                                      struct gen6_encoder_context *gen6_encoder_context)
1017 {
1018     struct i965_driver_data *i965 = i965_driver_data(ctx);
1019     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
1020     struct object_surface *obj_surface;
1021     struct object_buffer *obj_buffer;
1022     dri_bo *bo;
1023     VAEncPictureParameterBufferH264 *pPicParameter = (VAEncPictureParameterBufferH264 *)encode_state->pic_param->buffer;
1024     VAStatus vaStatus = VA_STATUS_SUCCESS;
1025
1026     /*Setup all the input&output object*/
1027     obj_surface = SURFACE(pPicParameter->reconstructed_picture);
1028     assert(obj_surface);
1029     i965_check_alloc_surface_bo(ctx, obj_surface, 1, VA_FOURCC('N','V','1','2'), SUBSAMPLE_YUV420);
1030     mfc_context->post_deblocking_output.bo = obj_surface->bo;
1031     dri_bo_reference(mfc_context->post_deblocking_output.bo);
1032
1033     mfc_context->surface_state.width = obj_surface->orig_width;
1034     mfc_context->surface_state.height = obj_surface->orig_height;
1035     mfc_context->surface_state.w_pitch = obj_surface->width;
1036     mfc_context->surface_state.h_pitch = obj_surface->height;
1037
1038     obj_surface = SURFACE(pPicParameter->reference_picture);
1039     assert(obj_surface);
1040     if (obj_surface->bo != NULL) {
1041         mfc_context->reference_surfaces[0].bo = obj_surface->bo;
1042         dri_bo_reference(obj_surface->bo);
1043     }
1044
1045     obj_surface = SURFACE(encode_state->current_render_target);
1046     assert(obj_surface && obj_surface->bo);
1047     mfc_context->uncompressed_picture_source.bo = obj_surface->bo;
1048     dri_bo_reference(mfc_context->uncompressed_picture_source.bo);
1049
1050     obj_buffer = BUFFER (pPicParameter->coded_buf); /* FIXME: fix this later */
1051     bo = obj_buffer->buffer_store->bo;
1052     assert(bo);
1053     mfc_context->mfc_indirect_pak_bse_object.bo = bo;
1054     mfc_context->mfc_indirect_pak_bse_object.offset = ALIGN(sizeof(VACodedBufferSegment), 64);
1055     dri_bo_reference(mfc_context->mfc_indirect_pak_bse_object.bo);
1056
1057     /*Programing bcs pipeline*/
1058     gen75_mfc_avc_pipeline_programing(ctx, encode_state, gen6_encoder_context); //filling the pipeline
1059
1060     return vaStatus;
1061 }
1062
1063 static VAStatus gen75_mfc_run(VADriverContextP ctx,
1064                              struct encode_state *encode_state,
1065                              struct gen6_encoder_context *gen6_encoder_context)
1066 {
1067     struct intel_batchbuffer *batch = gen6_encoder_context->base.batch;
1068
1069     intel_batchbuffer_flush(batch);             //run the pipeline
1070
1071     return VA_STATUS_SUCCESS;
1072 }
1073
1074 static VAStatus gen75_mfc_stop(VADriverContextP ctx,
1075                               struct encode_state *encode_state,
1076                               struct gen6_encoder_context *gen6_encoder_context)
1077 {
1078 #if 0
1079     struct i965_driver_data *i965 = i965_driver_data(ctx);
1080     struct gen6_mfc_context *mfc_context = &gen6_encoder_context->mfc_context;
1081
1082     VAEncPictureParameterBufferH264 *pPicParameter = (VAEncPictureParameterBufferH264 *)encode_state->pic_param->buffer;
1083
1084     struct object_surface *obj_surface = SURFACE(pPicParameter->reconstructed_picture);
1085     //struct object_surface *obj_surface = SURFACE(pPicParameter->reference_picture[0]);
1086     //struct object_surface *obj_surface = SURFACE(encode_state->current_render_target);
1087     my_debug(obj_surface);
1088
1089 #endif
1090
1091     return VA_STATUS_SUCCESS;
1092 }
1093
1094 static VAStatus
1095 gen75_mfc_avc_encode_picture(VADriverContextP ctx,
1096                             struct encode_state *encode_state,
1097                             struct gen6_encoder_context *gen6_encoder_context)
1098 {
1099     gen75_mfc_init(ctx, encode_state, gen6_encoder_context);
1100     gen75_mfc_avc_prepare(ctx, encode_state, gen6_encoder_context);
1101     gen75_mfc_run(ctx, encode_state, gen6_encoder_context);
1102     gen75_mfc_stop(ctx, encode_state, gen6_encoder_context);
1103
1104     return VA_STATUS_SUCCESS;
1105 }
1106
1107 VAStatus
1108 gen75_mfc_pipeline(VADriverContextP ctx,
1109                   VAProfile profile,
1110                   struct encode_state *encode_state,
1111                   struct gen6_encoder_context *gen6_encoder_context)
1112 {
1113     VAStatus vaStatus;
1114
1115     switch (profile) {
1116     case VAProfileH264Baseline:
1117         vaStatus = gen75_mfc_avc_encode_picture(ctx, encode_state, gen6_encoder_context);
1118         break;
1119
1120         /* FIXME: add for other profile */
1121     default:
1122         vaStatus = VA_STATUS_ERROR_UNSUPPORTED_PROFILE;
1123         break;
1124     }
1125
1126     return vaStatus;
1127 }
1128
1129 Bool gen75_mfc_context_init(VADriverContextP ctx, struct gen6_mfc_context *mfc_context)
1130 {
1131     int i;
1132     struct i965_driver_data *i965 = i965_driver_data(ctx);
1133     dri_bo *bo;
1134
1135     for (i = 0; i < NUM_MFC_DMV_BUFFERS; i++){
1136         dri_bo_unreference(mfc_context->direct_mv_buffers[i].bo);
1137         mfc_context->direct_mv_buffers[i].bo = NULL;
1138     }
1139     bo = dri_bo_alloc(i965->intel.bufmgr,
1140                         "Buffer",
1141                          68*8192,
1142                          64);
1143     mfc_context->direct_mv_buffers[0].bo = bo;
1144     bo = dri_bo_alloc(i965->intel.bufmgr,
1145                         "Buffer",
1146                          68*8192,
1147                          64);
1148     mfc_context->direct_mv_buffers[NUM_MFC_DMV_BUFFERS - 2].bo = bo;
1149     return True;
1150 }
1151
1152 Bool gen75_mfc_context_destroy(struct gen6_mfc_context *mfc_context)
1153 {
1154     int i;
1155
1156     dri_bo_unreference(mfc_context->post_deblocking_output.bo);
1157     mfc_context->post_deblocking_output.bo = NULL;
1158
1159     dri_bo_unreference(mfc_context->pre_deblocking_output.bo);
1160     mfc_context->pre_deblocking_output.bo = NULL;
1161
1162     dri_bo_unreference(mfc_context->uncompressed_picture_source.bo);
1163     mfc_context->uncompressed_picture_source.bo = NULL;
1164
1165     dri_bo_unreference(mfc_context->mfc_indirect_pak_bse_object.bo);
1166     mfc_context->mfc_indirect_pak_bse_object.bo = NULL;
1167
1168     for (i = 0; i < NUM_MFC_DMV_BUFFERS; i++){
1169         dri_bo_unreference(mfc_context->direct_mv_buffers[i].bo);
1170         mfc_context->direct_mv_buffers[i].bo = NULL;
1171     }
1172
1173     dri_bo_unreference(mfc_context->intra_row_store_scratch_buffer.bo);
1174     mfc_context->intra_row_store_scratch_buffer.bo = NULL;
1175
1176     dri_bo_unreference(mfc_context->deblocking_filter_row_store_scratch_buffer.bo);
1177     mfc_context->deblocking_filter_row_store_scratch_buffer.bo = NULL;
1178
1179     dri_bo_unreference(mfc_context->bsd_mpc_row_store_scratch_buffer.bo);
1180     mfc_context->bsd_mpc_row_store_scratch_buffer.bo = NULL;
1181
1182     return True;
1183 }