src/gen7_vme.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  *
  27  */
  28
  29 #include <stdio.h>
  30 #include <stdlib.h>
  31 #include <stdbool.h>
  32 #include <string.h>
  33 #include <assert.h>
  34
  35 #include "intel_batchbuffer.h"
  36 #include "intel_driver.h"
  37
  38 #include "i965_defines.h"
  39 #include "i965_drv_video.h"
  40 #include "i965_encoder.h"
  41 #include "gen6_vme.h"
  42 #include "gen6_mfc.h"
  43 #ifdef SURFACE_STATE_PADDED_SIZE
  44 #undef SURFACE_STATE_PADDED_SIZE
  45 #endif
  46
  47 #define VME_MSG_LENGTH          32
  48
  49 #define SURFACE_STATE_PADDED_SIZE               SURFACE_STATE_PADDED_SIZE_GEN7
  50 #define SURFACE_STATE_OFFSET(index)             (SURFACE_STATE_PADDED_SIZE * index)
  51 #define BINDING_TABLE_OFFSET(index)             (SURFACE_STATE_OFFSET(MAX_MEDIA_SURFACES_GEN6) + sizeof(unsigned int) * index)
  52
  53 #define CURBE_ALLOCATION_SIZE   37              /* in 256-bit */
  54 #define CURBE_TOTAL_DATA_LENGTH (4 * 32)        /* in byte, it should be less than or equal to CURBE_ALLOCATION_SIZE * 32 */
  55 #define CURBE_URB_ENTRY_LENGTH  4               /* in 256-bit, it should be less than or equal to CURBE_TOTAL_DATA_LENGTH / 32 */
  56
  57 enum VIDEO_CODING_TYPE{
  58     VIDEO_CODING_AVC = 0,
  59     VIDEO_CODING_MPEG2,
  60     VIDEO_CODING_SUM
  61 };
  62
  63 enum AVC_VME_KERNEL_TYPE{
  64     AVC_VME_INTRA_SHADER = 0,
  65     AVC_VME_INTER_SHADER,
  66     AVC_VME_BATCHBUFFER,
  67     AVC_VME_BINTER_SHADER,
  68     AVC_VME_KERNEL_SUM
  69 };
  70
  71 enum MPEG2_VME_KERNEL_TYPE{
  72     MPEG2_VME_INTER_SHADER = 0,
  73     MPEG2_VME_BATCHBUFFER,
  74     MPEG2_VME_KERNEL_SUM
  75 };
  76
  77
  78 static const uint32_t gen7_vme_intra_frame[][4] = {
  79 #include "shaders/vme/intra_frame_ivb.g7b"
  80 };
  81
  82 static const uint32_t gen7_vme_inter_frame[][4] = {
  83 #include "shaders/vme/inter_frame_ivb.g7b"
  84 };
  85
  86 static const uint32_t gen7_vme_batchbuffer[][4] = {
  87 #include "shaders/vme/batchbuffer.g7b"
  88 };
  89
  90 static const uint32_t gen7_vme_binter_frame[][4] = {
  91 #include "shaders/vme/inter_bframe_ivb.g7b"
  92 };
  93
  94 static struct i965_kernel gen7_vme_kernels[] = {
  95     {
  96         "AVC VME Intra Frame",
  97         AVC_VME_INTRA_SHADER,                   /*index*/
  98         gen7_vme_intra_frame,
  99         sizeof(gen7_vme_intra_frame),
 100         NULL
 101     },
 102     {
 103         "AVC VME inter Frame",
 104         AVC_VME_INTER_SHADER,
 105         gen7_vme_inter_frame,
 106         sizeof(gen7_vme_inter_frame),
 107         NULL
 108     },
 109     {
 110         "AVC VME BATCHBUFFER",
 111         AVC_VME_BATCHBUFFER,
 112         gen7_vme_batchbuffer,
 113         sizeof(gen7_vme_batchbuffer),
 114         NULL
 115     },
 116     {
 117         "AVC VME binter Frame",
 118         AVC_VME_BINTER_SHADER,
 119         gen7_vme_binter_frame,
 120         sizeof(gen7_vme_binter_frame),
 121         NULL
 122     }
 123 };
 124
 125 static const uint32_t gen7_vme_mpeg2_inter_frame[][4] = {
 126 #include "shaders/vme/mpeg2_inter_ivb.g7b"
 127 };
 128
 129 static const uint32_t gen7_vme_mpeg2_batchbuffer[][4] = {
 130 #include "shaders/vme/batchbuffer.g7b"
 131 };
 132
 133 static struct i965_kernel gen7_vme_mpeg2_kernels[] = {
 134     {
 135         "MPEG2 VME inter Frame",
 136         MPEG2_VME_INTER_SHADER,
 137         gen7_vme_mpeg2_inter_frame,
 138         sizeof(gen7_vme_mpeg2_inter_frame),
 139         NULL
 140     },
 141     {
 142         "MPEG2 VME BATCHBUFFER",
 143         MPEG2_VME_BATCHBUFFER,
 144         gen7_vme_mpeg2_batchbuffer,
 145         sizeof(gen7_vme_mpeg2_batchbuffer),
 146         NULL
 147     },
 148 };
 149
 150 /* only used for VME source surface state */
 151 static void
 152 gen7_vme_source_surface_state(VADriverContextP ctx,
 153                               int index,
 154                               struct object_surface *obj_surface,
 155                               struct intel_encoder_context *encoder_context)
 156 {
 157     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 158
 159     vme_context->vme_surface2_setup(ctx,
 160                                     &vme_context->gpe_context,
 161                                     obj_surface,
 162                                     BINDING_TABLE_OFFSET(index),
 163                                     SURFACE_STATE_OFFSET(index));
 164 }
 165
 166 static void
 167 gen7_vme_media_source_surface_state(VADriverContextP ctx,
 168                                     int index,
 169                                     struct object_surface *obj_surface,
 170                                     struct intel_encoder_context *encoder_context)
 171 {
 172     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 173
 174     vme_context->vme_media_rw_surface_setup(ctx,
 175                                             &vme_context->gpe_context,
 176                                             obj_surface,
 177                                             BINDING_TABLE_OFFSET(index),
 178                                             SURFACE_STATE_OFFSET(index),
 179                                             0);
 180 }
 181
 182 static void
 183 gen7_vme_output_buffer_setup(VADriverContextP ctx,
 184                              struct encode_state *encode_state,
 185                              int index,
 186                              struct intel_encoder_context *encoder_context)
 187
 188 {
 189     struct i965_driver_data *i965 = i965_driver_data(ctx);
 190     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 191     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 192     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 193     int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
 194     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 195     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 196
 197     vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
 198     vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
 199
 200     if (is_intra)
 201         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES;
 202     else
 203         vme_context->vme_output.size_block = INTER_VME_OUTPUT_IN_BYTES;
 204
 205     vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
 206                                               "VME output buffer",
 207                                               vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
 208                                               0x1000);
 209     assert(vme_context->vme_output.bo);
 210     vme_context->vme_buffer_suface_setup(ctx,
 211                                          &vme_context->gpe_context,
 212                                          &vme_context->vme_output,
 213                                          BINDING_TABLE_OFFSET(index),
 214                                          SURFACE_STATE_OFFSET(index));
 215 }
 216
 217 static void
 218 gen7_vme_output_vme_batchbuffer_setup(VADriverContextP ctx,
 219                                       struct encode_state *encode_state,
 220                                       int index,
 221                                       struct intel_encoder_context *encoder_context)
 222
 223 {
 224     struct i965_driver_data *i965 = i965_driver_data(ctx);
 225     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 226     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 227     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 228     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 229
 230     vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
 231     vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
 232     vme_context->vme_batchbuffer.pitch = 16;
 233     vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
 234                                                    "VME batchbuffer",
 235                                                    vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
 236                                                    0x1000);
 237     vme_context->vme_buffer_suface_setup(ctx,
 238                                          &vme_context->gpe_context,
 239                                          &vme_context->vme_batchbuffer,
 240                                          BINDING_TABLE_OFFSET(index),
 241                                          SURFACE_STATE_OFFSET(index));
 242 }
 243
 244 static VAStatus
 245 gen7_vme_surface_setup(VADriverContextP ctx,
 246                        struct encode_state *encode_state,
 247                        int is_intra,
 248                        struct intel_encoder_context *encoder_context)
 249 {
 250     struct object_surface *obj_surface;
 251
 252     /*Setup surfaces state*/
 253     /* current picture for encoding */
 254     obj_surface = encode_state->input_yuv_object;
 255     gen7_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 256     gen7_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 257
 258     if (!is_intra) {
 259         VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 260         int slice_type;
 261
 262         slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
 263         assert(slice_type != SLICE_TYPE_I && slice_type != SLICE_TYPE_SI);
 264
 265         intel_avc_vme_reference_state(ctx, encode_state, encoder_context, 0, 1, gen7_vme_source_surface_state);
 266
 267         if (slice_type == SLICE_TYPE_B)
 268             intel_avc_vme_reference_state(ctx, encode_state, encoder_context, 1, 2, gen7_vme_source_surface_state);
 269     }
 270
 271     /* VME output */
 272     gen7_vme_output_buffer_setup(ctx, encode_state, 3, encoder_context);
 273     gen7_vme_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 274     intel_h264_setup_cost_surface(ctx, encode_state, encoder_context,
 275                                   BINDING_TABLE_OFFSET(INTEL_COST_TABLE_OFFSET),
 276                                   SURFACE_STATE_OFFSET(INTEL_COST_TABLE_OFFSET));
 277
 278     return VA_STATUS_SUCCESS;
 279 }
 280
 281 static VAStatus gen7_vme_interface_setup(VADriverContextP ctx,
 282                                          struct encode_state *encode_state,
 283                                          struct intel_encoder_context *encoder_context)
 284 {
 285     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 286     struct gen6_interface_descriptor_data *desc;
 287     int i;
 288     dri_bo *bo;
 289
 290     bo = vme_context->gpe_context.idrt.bo;
 291     dri_bo_map(bo, 1);
 292     assert(bo->virtual);
 293     desc = bo->virtual;
 294
 295     for (i = 0; i < vme_context->vme_kernel_sum; i++) {
 296         struct i965_kernel *kernel;
 297         kernel = &vme_context->gpe_context.kernels[i];
 298         assert(sizeof(*desc) == 32);
 299         /*Setup the descritor table*/
 300         memset(desc, 0, sizeof(*desc));
 301         desc->desc0.kernel_start_pointer = (kernel->bo->offset >> 6);
 302         desc->desc2.sampler_count = 1; /* FIXME: */
 303         desc->desc2.sampler_state_pointer = (vme_context->vme_state.bo->offset >> 5);
 304         desc->desc3.binding_table_entry_count = 1; /* FIXME: */
 305         desc->desc3.binding_table_pointer = (BINDING_TABLE_OFFSET(0) >> 5);
 306         desc->desc4.constant_urb_entry_read_offset = 0;
 307         desc->desc4.constant_urb_entry_read_length = CURBE_URB_ENTRY_LENGTH;
 308
 309         /*kernel start*/
 310         dri_bo_emit_reloc(bo,
 311                           I915_GEM_DOMAIN_INSTRUCTION, 0,
 312                           0,
 313                           i * sizeof(*desc) + offsetof(struct gen6_interface_descriptor_data, desc0),
 314                           kernel->bo);
 315         /*Sampler State(VME state pointer)*/
 316         dri_bo_emit_reloc(bo,
 317                           I915_GEM_DOMAIN_INSTRUCTION, 0,
 318                           (1 << 2),                                                                     //
 319                           i * sizeof(*desc) + offsetof(struct gen6_interface_descriptor_data, desc2),
 320                           vme_context->vme_state.bo);
 321         desc++;
 322     }
 323     dri_bo_unmap(bo);
 324
 325     return VA_STATUS_SUCCESS;
 326 }
 327
 328 static VAStatus gen7_vme_constant_setup(VADriverContextP ctx,
 329                                         struct encode_state *encode_state,
 330                                         struct intel_encoder_context *encoder_context)
 331 {
 332     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 333     unsigned char *constant_buffer;
 334     unsigned int *vme_state_message;
 335     int mv_num;
 336
 337     vme_state_message = (unsigned int *)vme_context->vme_state_message;
 338     mv_num = 32;
 339
 340     if (encoder_context->codec == CODEC_H264) {
 341         if (vme_context->h264_level >= 30) {
 342             mv_num = 16;
 343
 344             if (vme_context->h264_level >= 31)
 345                 mv_num = 8;
 346         }
 347     } else if (encoder_context->codec == CODEC_MPEG2) {
 348         mv_num = 2;
 349     }
 350
 351
 352     vme_state_message[31] = mv_num;
 353
 354     dri_bo_map(vme_context->gpe_context.curbe.bo, 1);
 355     assert(vme_context->gpe_context.curbe.bo->virtual);
 356     constant_buffer = vme_context->gpe_context.curbe.bo->virtual;
 357
 358     /* Pass the required constant info into the constant buffer */
 359     memcpy(constant_buffer, (char *)vme_context->vme_state_message, 128);
 360
 361     dri_bo_unmap( vme_context->gpe_context.curbe.bo);
 362
 363     return VA_STATUS_SUCCESS;
 364 }
 365
 366
 367 static VAStatus gen7_vme_avc_state_setup(VADriverContextP ctx,
 368                                          struct encode_state *encode_state,
 369                                          int is_intra,
 370                                          struct intel_encoder_context *encoder_context)
 371 {
 372     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 373     unsigned int *vme_state_message;
 374     unsigned int *mb_cost_table;
 375     int i;
 376     VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 377     unsigned int is_low_quality = (encoder_context->quality_level == ENCODER_LOW_QUALITY);
 378     dri_bo *cost_bo;
 379     int slice_type;
 380     uint8_t *cost_ptr;
 381     int qp;
 382
 383     slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
 384
 385     if (slice_type == SLICE_TYPE_I) {
 386         cost_bo = vme_context->i_qp_cost_table;
 387     } else if (slice_type == SLICE_TYPE_P) {
 388         cost_bo = vme_context->p_qp_cost_table;
 389     } else {
 390         cost_bo = vme_context->b_qp_cost_table;
 391     }
 392
 393     mb_cost_table = (unsigned int *)vme_context->vme_state_message;
 394     dri_bo_map(vme_context->vme_state.bo, 1);
 395     dri_bo_map(cost_bo, 0);
 396     assert(vme_context->vme_state.bo->virtual);
 397     assert(cost_bo->virtual);
 398     vme_state_message = (unsigned int *)vme_context->vme_state.bo->virtual;
 399
 400     cost_ptr = (uint8_t *)cost_bo->virtual;
 401
 402     /* up to 8 VME_SEARCH_PATH_LUT is supported */
 403     /* Two subsequent qp will share the same mode/motion-vector cost table */
 404     /* the range is from 0-51 */
 405     for (i = 0; i < 8; i++)  {
 406
 407         vme_state_message = (unsigned int *)vme_context->vme_state.bo->virtual +
 408                              i * 32;
 409         if ((slice_type == SLICE_TYPE_P) && !is_low_quality) {
 410             vme_state_message[0] = 0x01010101;
 411             vme_state_message[1] = 0x10010101;
 412             vme_state_message[2] = 0x0F0F0F0F;
 413             vme_state_message[3] = 0x100F0F0F;
 414             vme_state_message[4] = 0x01010101;
 415             vme_state_message[5] = 0x10010101;
 416             vme_state_message[6] = 0x0F0F0F0F;
 417             vme_state_message[7] = 0x100F0F0F;
 418             vme_state_message[8] = 0x01010101;
 419             vme_state_message[9] = 0x10010101;
 420             vme_state_message[10] = 0x0F0F0F0F;
 421             vme_state_message[11] = 0x000F0F0F;
 422             vme_state_message[12] = 0x00;
 423             vme_state_message[13] = 0x00;
 424         } else {
 425             vme_state_message[0] = 0x10010101;
 426             vme_state_message[1] = 0x100F0F0F;
 427             vme_state_message[2] = 0x10010101;
 428             vme_state_message[3] = 0x000F0F0F;
 429             vme_state_message[4] = 0;
 430             vme_state_message[5] = 0;
 431             vme_state_message[6] = 0;
 432             vme_state_message[7] = 0;
 433             vme_state_message[8] = 0;
 434             vme_state_message[9] = 0;
 435             vme_state_message[10] = 0;
 436             vme_state_message[11] = 0;
 437             vme_state_message[12] = 0;
 438             vme_state_message[13] = 0;
 439         }
 440
 441         qp = 8 * i;
 442
 443         /* when qp is greater than 51, use the cost_table of qp=51 to fulfill */
 444         if (qp > 51) {
 445             qp = 51;
 446         }
 447         /* Setup the four LUT sets for MbMV cost */
 448         mb_cost_table = (unsigned int *)(cost_ptr + qp * 32);
 449         vme_state_message[14] = (mb_cost_table[2] & 0xFFFF);
 450         vme_state_message[16] = mb_cost_table[0];
 451         vme_state_message[17] = mb_cost_table[1];
 452         vme_state_message[18] = mb_cost_table[3];
 453         vme_state_message[19] = mb_cost_table[4];
 454
 455         qp += 2;
 456         if (qp > 51) {
 457             qp = 51;
 458         }
 459         mb_cost_table = (unsigned int *)(cost_ptr + qp * 32);
 460         vme_state_message[14] |= ((mb_cost_table[2] & 0xFFFF) << 16);
 461         vme_state_message[20] = mb_cost_table[0];
 462         vme_state_message[21] = mb_cost_table[1];
 463         vme_state_message[22] = mb_cost_table[3];
 464         vme_state_message[23] = mb_cost_table[4];
 465
 466         qp += 2;
 467         if (qp > 51) {
 468             qp = 51;
 469         }
 470         vme_state_message[15] = (mb_cost_table[2] & 0xFFFF);
 471         vme_state_message[24] = mb_cost_table[0];
 472         vme_state_message[25] = mb_cost_table[1];
 473         vme_state_message[26] = mb_cost_table[3];
 474         vme_state_message[27] = mb_cost_table[4];
 475
 476         qp += 2;
 477         if (qp > 51) {
 478             qp = 51;
 479         }
 480         mb_cost_table = (unsigned int *)(cost_ptr + qp * 32);
 481         vme_state_message[15] |= ((mb_cost_table[2] & 0xFFFF) << 16);
 482         vme_state_message[28] = mb_cost_table[0];
 483         vme_state_message[29] = mb_cost_table[1];
 484         vme_state_message[30] = mb_cost_table[3];
 485         vme_state_message[31] = mb_cost_table[4];
 486     }
 487
 488     dri_bo_unmap(cost_bo);
 489     dri_bo_unmap( vme_context->vme_state.bo);
 490     return VA_STATUS_SUCCESS;
 491 }
 492
 493 static VAStatus gen7_vme_mpeg2_state_setup(VADriverContextP ctx,
 494                                            struct encode_state *encode_state,
 495                                            int is_intra,
 496                                            struct intel_encoder_context *encoder_context)
 497 {
 498     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 499     unsigned int *vme_state_message;
 500     int i;
 501     unsigned int *mb_cost_table;
 502
 503     mb_cost_table = (unsigned int *)vme_context->vme_state_message;
 504
 505     //building VME state message
 506     dri_bo_map(vme_context->vme_state.bo, 1);
 507     assert(vme_context->vme_state.bo->virtual);
 508     vme_state_message = (unsigned int *)vme_context->vme_state.bo->virtual;
 509
 510     vme_state_message[0] = 0x01010101;
 511     vme_state_message[1] = 0x10010101;
 512     vme_state_message[2] = 0x0F0F0F0F;
 513     vme_state_message[3] = 0x100F0F0F;
 514     vme_state_message[4] = 0x01010101;
 515     vme_state_message[5] = 0x10010101;
 516     vme_state_message[6] = 0x0F0F0F0F;
 517     vme_state_message[7] = 0x100F0F0F;
 518     vme_state_message[8] = 0x01010101;
 519     vme_state_message[9] = 0x10010101;
 520     vme_state_message[10] = 0x0F0F0F0F;
 521     vme_state_message[11] = 0x000F0F0F;
 522     vme_state_message[12] = 0x00;
 523     vme_state_message[13] = 0x00;
 524
 525     vme_state_message[14] = (mb_cost_table[2] & 0xFFFF);
 526     vme_state_message[15] = 0;
 527     vme_state_message[16] = mb_cost_table[0];
 528     vme_state_message[17] = 0;
 529     vme_state_message[18] = mb_cost_table[3];
 530     vme_state_message[19] = mb_cost_table[4];
 531
 532     for(i = 20; i < 32; i++) {
 533         vme_state_message[i] = 0;
 534     }
 535     //vme_state_message[16] = 0x42424242;                       //cost function LUT set 0 for Intra
 536
 537     dri_bo_unmap( vme_context->vme_state.bo);
 538     return VA_STATUS_SUCCESS;
 539 }
 540
 541 static void
 542 gen7_vme_fill_vme_batchbuffer(VADriverContextP ctx,
 543                               struct encode_state *encode_state,
 544                               int mb_width, int mb_height,
 545                               int kernel,
 546                               int transform_8x8_mode_flag,
 547                               struct intel_encoder_context *encoder_context)
 548 {
 549     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 550     int mb_x = 0, mb_y = 0;
 551     int i, s, j;
 552     unsigned int *command_ptr;
 553     struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
 554     VAEncPictureParameterBufferH264 *pic_param = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
 555     VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 556     int qp;
 557     int slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
 558     int qp_mb, qp_index;
 559
 560     if (encoder_context->rate_control_mode == VA_RC_CQP)
 561         qp = pic_param->pic_init_qp + slice_param->slice_qp_delta;
 562     else
 563         qp = mfc_context->brc.qp_prime_y[encoder_context->layer.curr_frame_layer_id][slice_type];
 564
 565     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 566     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 567
 568     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 569         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 570
 571         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 572             int slice_mb_begin = slice_param->macroblock_address;
 573             int slice_mb_number = slice_param->num_macroblocks;
 574             unsigned int mb_intra_ub;
 575             int slice_mb_x = slice_param->macroblock_address % mb_width;
 576
 577             for (i = 0; i < slice_mb_number;) {
 578                 int mb_count = i + slice_mb_begin;
 579
 580                 mb_x = mb_count % mb_width;
 581                 mb_y = mb_count / mb_width;
 582                 mb_intra_ub = 0;
 583
 584                 if (mb_x != 0) {
 585                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 586                 }
 587
 588                 if (mb_y != 0) {
 589                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 590
 591                     if (mb_x != 0)
 592                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 593
 594                     if (mb_x != (mb_width -1))
 595                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 596                 }
 597
 598                 if (i < mb_width) {
 599                     if (i == 0)
 600                         mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_AE);
 601
 602                     mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_BCD_MASK);
 603
 604                     if ((i == (mb_width - 1)) && slice_mb_x) {
 605                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 606                     }
 607                 }
 608
 609                 if ((i == mb_width) && slice_mb_x) {
 610                     mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_D);
 611                 }
 612
 613                 *command_ptr++ = (CMD_MEDIA_OBJECT | (9 - 2));
 614                 *command_ptr++ = kernel;
 615                 *command_ptr++ = 0;
 616                 *command_ptr++ = 0;
 617                 *command_ptr++ = 0;
 618                 *command_ptr++ = 0;
 619
 620                 /*inline data */
 621                 *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 622                 *command_ptr++ = ((encoder_context->quality_level << 24) | (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 623
 624                 if (vme_context->roi_enabled) {
 625                     qp_index = mb_y * mb_width + mb_x;
 626                     qp_mb = *(vme_context->qp_per_mb + qp_index);
 627                 } else
 628                     qp_mb = qp;
 629                 *command_ptr++ = qp_mb;
 630
 631                 i += 1;
 632             }
 633
 634             slice_param++;
 635         }
 636     }
 637
 638     *command_ptr++ = 0;
 639     *command_ptr++ = MI_BATCH_BUFFER_END;
 640
 641     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 642 }
 643
 644
 645 static void gen7_vme_media_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 646 {
 647     struct i965_driver_data *i965 = i965_driver_data(ctx);
 648     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 649     dri_bo *bo;
 650
 651     i965_gpe_context_init(ctx, &vme_context->gpe_context);
 652
 653     /* VME output buffer */
 654     dri_bo_unreference(vme_context->vme_output.bo);
 655     vme_context->vme_output.bo = NULL;
 656
 657     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 658     vme_context->vme_batchbuffer.bo = NULL;
 659
 660     /* VME state */
 661     dri_bo_unreference(vme_context->vme_state.bo);
 662     bo = dri_bo_alloc(i965->intel.bufmgr,
 663                       "Buffer",
 664                       1024*16, 64);
 665     assert(bo);
 666     vme_context->vme_state.bo = bo;
 667 }
 668
 669 static void gen7_vme_pipeline_programing(VADriverContextP ctx,
 670                                          struct encode_state *encode_state,
 671                                          struct intel_encoder_context *encoder_context)
 672 {
 673     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 674     struct intel_batchbuffer *batch = encoder_context->base.batch;
 675     VAEncPictureParameterBufferH264 *pPicParameter = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
 676     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 677     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 678     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 679     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 680     int s;
 681     bool allow_hwscore = true;
 682     int kernel_shader;
 683     unsigned int is_low_quality = (encoder_context->quality_level == ENCODER_LOW_QUALITY);
 684
 685     if (is_low_quality)
 686         allow_hwscore = false;
 687     else {
 688         for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 689             pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 690             if ((pSliceParameter->macroblock_address % width_in_mbs)) {
 691                 allow_hwscore = false;
 692                 break;
 693             }
 694         }
 695     }
 696
 697     if ((pSliceParameter->slice_type == SLICE_TYPE_I) ||
 698         (pSliceParameter->slice_type == SLICE_TYPE_SI)) {
 699         kernel_shader = AVC_VME_INTRA_SHADER;
 700     } else if ((pSliceParameter->slice_type == SLICE_TYPE_P) ||
 701                (pSliceParameter->slice_type == SLICE_TYPE_SP)) {
 702         kernel_shader = AVC_VME_INTER_SHADER;
 703     } else {
 704         kernel_shader = AVC_VME_BINTER_SHADER;
 705         if (!allow_hwscore)
 706             kernel_shader = AVC_VME_INTER_SHADER;
 707     }
 708
 709     if (allow_hwscore)
 710         gen7_vme_walker_fill_vme_batchbuffer(ctx,
 711                                              encode_state,
 712                                              width_in_mbs, height_in_mbs,
 713                                              kernel_shader,
 714                                              pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 715                                              encoder_context);
 716
 717     else
 718         gen7_vme_fill_vme_batchbuffer(ctx,
 719                                       encode_state,
 720                                       width_in_mbs, height_in_mbs,
 721                                       kernel_shader,
 722                                       pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 723                                       encoder_context);
 724
 725     intel_batchbuffer_start_atomic(batch, 0x1000);
 726     gen6_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 727     BEGIN_BATCH(batch, 2);
 728     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (1 << 8));
 729     OUT_RELOC(batch,
 730               vme_context->vme_batchbuffer.bo,
 731               I915_GEM_DOMAIN_COMMAND, 0,
 732               0);
 733     ADVANCE_BATCH(batch);
 734
 735     intel_batchbuffer_end_atomic(batch);
 736 }
 737
 738 static VAStatus gen7_vme_prepare(VADriverContextP ctx,
 739                                  struct encode_state *encode_state,
 740                                  struct intel_encoder_context *encoder_context)
 741 {
 742     VAStatus vaStatus = VA_STATUS_SUCCESS;
 743     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 744     int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
 745     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 746     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 747
 748     if (!vme_context->h264_level ||
 749         (vme_context->h264_level != pSequenceParameter->level_idc)) {
 750         vme_context->h264_level = pSequenceParameter->level_idc;
 751     }
 752
 753     intel_vme_update_mbmv_cost(ctx, encode_state, encoder_context);
 754     intel_h264_initialize_mbmv_cost(ctx, encode_state, encoder_context);
 755
 756     /*Setup all the memory object*/
 757     gen7_vme_surface_setup(ctx, encode_state, is_intra, encoder_context);
 758     gen7_vme_interface_setup(ctx, encode_state, encoder_context);
 759     gen7_vme_constant_setup(ctx, encode_state, encoder_context);
 760     gen7_vme_avc_state_setup(ctx, encode_state, is_intra, encoder_context);
 761
 762     /*Programing media pipeline*/
 763     gen7_vme_pipeline_programing(ctx, encode_state, encoder_context);
 764
 765     return vaStatus;
 766 }
 767
 768 static VAStatus gen7_vme_run(VADriverContextP ctx,
 769                              struct encode_state *encode_state,
 770                              struct intel_encoder_context *encoder_context)
 771 {
 772     struct intel_batchbuffer *batch = encoder_context->base.batch;
 773
 774     intel_batchbuffer_flush(batch);
 775
 776     return VA_STATUS_SUCCESS;
 777 }
 778
 779 static VAStatus gen7_vme_stop(VADriverContextP ctx,
 780                               struct encode_state *encode_state,
 781                               struct intel_encoder_context *encoder_context)
 782 {
 783     return VA_STATUS_SUCCESS;
 784 }
 785
 786 static VAStatus
 787 gen7_vme_pipeline(VADriverContextP ctx,
 788                   VAProfile profile,
 789                   struct encode_state *encode_state,
 790                   struct intel_encoder_context *encoder_context)
 791 {
 792     gen7_vme_media_init(ctx, encoder_context);
 793     gen7_vme_prepare(ctx, encode_state, encoder_context);
 794     gen7_vme_run(ctx, encode_state, encoder_context);
 795     gen7_vme_stop(ctx, encode_state, encoder_context);
 796
 797     return VA_STATUS_SUCCESS;
 798 }
 799
 800 static void
 801 gen7_vme_mpeg2_output_buffer_setup(VADriverContextP ctx,
 802                                    struct encode_state *encode_state,
 803                                    int index,
 804                                    int is_intra,
 805                                    struct intel_encoder_context *encoder_context)
 806
 807 {
 808     struct i965_driver_data *i965 = i965_driver_data(ctx);
 809     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 810     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 811     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 812     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 813
 814     vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
 815     vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
 816
 817     if (is_intra)
 818         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES;
 819     else
 820         vme_context->vme_output.size_block = INTER_VME_OUTPUT_IN_BYTES;
 821
 822     vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
 823                                               "VME output buffer",
 824                                               vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
 825                                               0x1000);
 826     assert(vme_context->vme_output.bo);
 827     vme_context->vme_buffer_suface_setup(ctx,
 828                                          &vme_context->gpe_context,
 829                                          &vme_context->vme_output,
 830                                          BINDING_TABLE_OFFSET(index),
 831                                          SURFACE_STATE_OFFSET(index));
 832 }
 833
 834 static void
 835 gen7_vme_mpeg2_output_vme_batchbuffer_setup(VADriverContextP ctx,
 836                                             struct encode_state *encode_state,
 837                                             int index,
 838                                             struct intel_encoder_context *encoder_context)
 839
 840 {
 841     struct i965_driver_data *i965 = i965_driver_data(ctx);
 842     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 843     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 844     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 845     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 846
 847     vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
 848     vme_context->vme_batchbuffer.size_block = 32; /* 4 OWORDs */
 849     vme_context->vme_batchbuffer.pitch = 16;
 850     vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
 851                                                    "VME batchbuffer",
 852                                                    vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
 853                                                    0x1000);
 854     vme_context->vme_buffer_suface_setup(ctx,
 855                                          &vme_context->gpe_context,
 856                                          &vme_context->vme_batchbuffer,
 857                                          BINDING_TABLE_OFFSET(index),
 858                                          SURFACE_STATE_OFFSET(index));
 859 }
 860
 861 static VAStatus
 862 gen7_vme_mpeg2_surface_setup(VADriverContextP ctx,
 863                              struct encode_state *encode_state,
 864                              int is_intra,
 865                              struct intel_encoder_context *encoder_context)
 866 {
 867     struct object_surface *obj_surface;
 868
 869     /*Setup surfaces state*/
 870     /* current picture for encoding */
 871     obj_surface = encode_state->input_yuv_object;
 872     gen7_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 873     gen7_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 874
 875     if (!is_intra) {
 876         /* reference 0 */
 877         obj_surface = encode_state->reference_objects[0];
 878         if (obj_surface->bo != NULL)
 879             gen7_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
 880
 881         /* reference 1 */
 882         obj_surface = encode_state->reference_objects[1];
 883         if (obj_surface && obj_surface->bo != NULL)
 884             gen7_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
 885     }
 886
 887     /* VME output */
 888     gen7_vme_mpeg2_output_buffer_setup(ctx, encode_state, 3, is_intra, encoder_context);
 889     gen7_vme_mpeg2_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 890
 891     return VA_STATUS_SUCCESS;
 892 }
 893
 894 static void
 895 gen7_vme_mpeg2_fill_vme_batchbuffer(VADriverContextP ctx,
 896                                     struct encode_state *encode_state,
 897                                     int mb_width, int mb_height,
 898                                     int kernel,
 899                                     int transform_8x8_mode_flag,
 900                                     struct intel_encoder_context *encoder_context)
 901 {
 902     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 903     int mb_x = 0, mb_y = 0;
 904     int i, s, j;
 905     unsigned int *command_ptr;
 906
 907     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 908     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 909
 910     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 911         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 912
 913         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 914             int slice_mb_begin = slice_param->macroblock_address;
 915             int slice_mb_number = slice_param->num_macroblocks;
 916             unsigned int mb_intra_ub;
 917
 918             for (i = 0; i < slice_mb_number;) {
 919                 int mb_count = i + slice_mb_begin;
 920
 921                 mb_x = mb_count % mb_width;
 922                 mb_y = mb_count / mb_width;
 923                 mb_intra_ub = 0;
 924
 925                 if (mb_x != 0) {
 926                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 927                 }
 928
 929                 if (mb_y != 0) {
 930                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 931
 932                     if (mb_x != 0)
 933                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 934
 935                     if (mb_x != (mb_width -1))
 936                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 937                 }
 938
 939
 940
 941                 *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 942                 *command_ptr++ = kernel;
 943                 *command_ptr++ = 0;
 944                 *command_ptr++ = 0;
 945                 *command_ptr++ = 0;
 946                 *command_ptr++ = 0;
 947
 948                 /*inline data */
 949                 *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 950                 *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 951
 952                 i += 1;
 953             }
 954
 955             slice_param++;
 956         }
 957     }
 958
 959     *command_ptr++ = 0;
 960     *command_ptr++ = MI_BATCH_BUFFER_END;
 961
 962     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 963 }
 964
 965 static void
 966 gen7_vme_mpeg2_pipeline_programing(VADriverContextP ctx,
 967                                    struct encode_state *encode_state,
 968                                    int is_intra,
 969                                    struct intel_encoder_context *encoder_context)
 970 {
 971     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 972     struct intel_batchbuffer *batch = encoder_context->base.batch;
 973     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 974     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 975     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 976
 977     bool allow_hwscore = true;
 978     int s;
 979
 980     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 981         int j;
 982         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 983
 984         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 985             if (slice_param->macroblock_address % width_in_mbs) {
 986                 allow_hwscore = false;
 987                 break;
 988             }
 989         }
 990     }
 991
 992     if (allow_hwscore)
 993         gen7_vme_mpeg2_walker_fill_vme_batchbuffer(ctx,
 994                                                    encode_state,
 995                                                    width_in_mbs, height_in_mbs,
 996                                                    MPEG2_VME_INTER_SHADER,
 997                                                    encoder_context);
 998     else
 999         gen7_vme_mpeg2_fill_vme_batchbuffer(ctx,
1000                                             encode_state,
1001                                             width_in_mbs, height_in_mbs,
1002                                             MPEG2_VME_INTER_SHADER,
1003                                             0,
1004                                             encoder_context);
1005
1006     intel_batchbuffer_start_atomic(batch, 0x1000);
1007     gen6_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
1008     BEGIN_BATCH(batch, 2);
1009     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (1 << 8));
1010     OUT_RELOC(batch,
1011               vme_context->vme_batchbuffer.bo,
1012               I915_GEM_DOMAIN_COMMAND, 0,
1013               0);
1014     ADVANCE_BATCH(batch);
1015
1016     intel_batchbuffer_end_atomic(batch);
1017 }
1018
1019 static VAStatus
1020 gen7_vme_mpeg2_prepare(VADriverContextP ctx,
1021                        struct encode_state *encode_state,
1022                        struct intel_encoder_context *encoder_context)
1023 {
1024     VAStatus vaStatus = VA_STATUS_SUCCESS;
1025     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
1026     struct gen6_vme_context *vme_context = encoder_context->vme_context;
1027
1028     if ((!vme_context->mpeg2_level) ||
1029         (vme_context->mpeg2_level != (seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK))) {
1030         vme_context->mpeg2_level = seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK;
1031     }
1032
1033     /*Setup all the memory object*/
1034
1035     intel_vme_mpeg2_state_setup(ctx, encode_state, encoder_context);
1036     gen7_vme_mpeg2_surface_setup(ctx, encode_state, 0, encoder_context);
1037     gen7_vme_interface_setup(ctx, encode_state, encoder_context);
1038     gen7_vme_constant_setup(ctx, encode_state, encoder_context);
1039     gen7_vme_mpeg2_state_setup(ctx, encode_state, 0, encoder_context);
1040
1041     /*Programing media pipeline*/
1042     gen7_vme_mpeg2_pipeline_programing(ctx, encode_state, 0, encoder_context);
1043
1044     return vaStatus;
1045 }
1046
1047 static VAStatus
1048 gen7_vme_mpeg2_pipeline(VADriverContextP ctx,
1049                         VAProfile profile,
1050                         struct encode_state *encode_state,
1051                         struct intel_encoder_context *encoder_context)
1052 {
1053     struct i965_driver_data *i965 = i965_driver_data(ctx);
1054     struct gen6_vme_context *vme_context = encoder_context->vme_context;
1055     VAEncSliceParameterBufferMPEG2 *slice_param =
1056         (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[0]->buffer;
1057     VAEncSequenceParameterBufferMPEG2 *seq_param =
1058         (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
1059
1060     /*No need of to exec VME for Intra slice */
1061     if (slice_param->is_intra_slice) {
1062         if(!vme_context->vme_output.bo) {
1063             int w_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
1064             int h_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
1065
1066             vme_context->vme_output.num_blocks = w_in_mbs * h_in_mbs;
1067             vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
1068             vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES;
1069             vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
1070                                                       "MPEG2 VME output buffer",
1071                                                       vme_context->vme_output.num_blocks
1072                                                       * vme_context->vme_output.size_block,
1073                                                       0x1000);
1074         }
1075
1076         return VA_STATUS_SUCCESS;
1077     }
1078
1079     gen7_vme_media_init(ctx, encoder_context);
1080     gen7_vme_mpeg2_prepare(ctx, encode_state, encoder_context);
1081     gen7_vme_run(ctx, encode_state, encoder_context);
1082     gen7_vme_stop(ctx, encode_state, encoder_context);
1083
1084     return VA_STATUS_SUCCESS;
1085 }
1086
1087 static void
1088 gen7_vme_context_destroy(void *context)
1089 {
1090     struct gen6_vme_context *vme_context = context;
1091
1092     i965_gpe_context_destroy(&vme_context->gpe_context);
1093
1094     dri_bo_unreference(vme_context->vme_output.bo);
1095     vme_context->vme_output.bo = NULL;
1096
1097     dri_bo_unreference(vme_context->vme_state.bo);
1098     vme_context->vme_state.bo = NULL;
1099
1100     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
1101     vme_context->vme_batchbuffer.bo = NULL;
1102
1103     free(vme_context->vme_state_message);
1104     vme_context->vme_state_message = NULL;
1105
1106     dri_bo_unreference(vme_context->i_qp_cost_table);
1107     vme_context->i_qp_cost_table = NULL;
1108
1109     dri_bo_unreference(vme_context->p_qp_cost_table);
1110     vme_context->p_qp_cost_table = NULL;
1111
1112     dri_bo_unreference(vme_context->b_qp_cost_table);
1113     vme_context->b_qp_cost_table = NULL;
1114
1115     free(vme_context->qp_per_mb);
1116     vme_context->qp_per_mb = NULL;
1117
1118     free(vme_context);
1119 }
1120
1121 Bool gen7_vme_context_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
1122 {
1123     struct gen6_vme_context *vme_context = calloc(1, sizeof(struct gen6_vme_context));
1124     struct i965_kernel *vme_kernel_list = NULL;
1125
1126     assert(vme_context);
1127     vme_context->gpe_context.surface_state_binding_table.length =
1128         (SURFACE_STATE_PADDED_SIZE + sizeof(unsigned int)) * MAX_MEDIA_SURFACES_GEN6;
1129
1130     vme_context->gpe_context.idrt.max_entries = MAX_INTERFACE_DESC_GEN6;
1131     vme_context->gpe_context.idrt.entry_size = sizeof(struct gen6_interface_descriptor_data);
1132     vme_context->gpe_context.curbe.length = CURBE_TOTAL_DATA_LENGTH;
1133
1134     vme_context->gpe_context.vfe_state.max_num_threads = 60 - 1;
1135     vme_context->gpe_context.vfe_state.num_urb_entries = 16;
1136     vme_context->gpe_context.vfe_state.gpgpu_mode = 0;
1137     vme_context->gpe_context.vfe_state.urb_entry_size = 59 - 1;
1138     vme_context->gpe_context.vfe_state.curbe_allocation_size = CURBE_ALLOCATION_SIZE - 1;
1139
1140     gen7_vme_scoreboard_init(ctx, vme_context);
1141
1142     if (encoder_context->codec == CODEC_H264) {
1143         vme_kernel_list = gen7_vme_kernels;
1144         vme_context->video_coding_type = VIDEO_CODING_AVC;
1145         vme_context->vme_kernel_sum = AVC_VME_KERNEL_SUM;
1146         encoder_context->vme_pipeline = gen7_vme_pipeline;
1147     } else if (encoder_context->codec == CODEC_MPEG2) {
1148         vme_kernel_list = gen7_vme_mpeg2_kernels;
1149         vme_context->video_coding_type = VIDEO_CODING_MPEG2;
1150         vme_context->vme_kernel_sum = MPEG2_VME_KERNEL_SUM;
1151         encoder_context->vme_pipeline = gen7_vme_mpeg2_pipeline;
1152     } else {
1153         /* Unsupported codec */
1154         assert(0);
1155     }
1156
1157     i965_gpe_load_kernels(ctx,
1158                           &vme_context->gpe_context,
1159                           vme_kernel_list,
1160                           vme_context->vme_kernel_sum);
1161
1162     vme_context->vme_surface2_setup = gen7_gpe_surface2_setup;
1163     vme_context->vme_media_rw_surface_setup = gen7_gpe_media_rw_surface_setup;
1164     vme_context->vme_buffer_suface_setup = gen7_gpe_buffer_suface_setup;
1165
1166     encoder_context->vme_context = vme_context;
1167     encoder_context->vme_context_destroy = gen7_vme_context_destroy;
1168     vme_context->vme_state_message = malloc(VME_MSG_LENGTH * sizeof(int));
1169
1170     return True;
1171 }