src/gen8_vme.c

   1 /*
   2  * Copyright © 2012 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 #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
  44 #ifdef SURFACE_STATE_PADDED_SIZE
  45 #undef SURFACE_STATE_PADDED_SIZE
  46 #endif
  47
  48 #define SURFACE_STATE_PADDED_SIZE               SURFACE_STATE_PADDED_SIZE_GEN8
  49 #define SURFACE_STATE_OFFSET(index)             (SURFACE_STATE_PADDED_SIZE * index)
  50 #define BINDING_TABLE_OFFSET(index)             (SURFACE_STATE_OFFSET(MAX_MEDIA_SURFACES_GEN6) + sizeof(unsigned int) * index)
  51
  52 #define VME_INTRA_SHADER        0
  53 #define VME_INTER_SHADER        1
  54 #define VME_BINTER_SHADER       2
  55
  56 #define CURBE_ALLOCATION_SIZE   37              /* in 256-bit */
  57 #define CURBE_TOTAL_DATA_LENGTH (4 * 32)        /* in byte, it should be less than or equal to CURBE_ALLOCATION_SIZE * 32 */
  58 #define CURBE_URB_ENTRY_LENGTH  4               /* in 256-bit, it should be less than or equal to CURBE_TOTAL_DATA_LENGTH / 32 */
  59
  60 #define VME_MSG_LENGTH          32
  61
  62 static const uint32_t gen8_vme_intra_frame[][4] = {
  63 #include "shaders/vme/intra_frame_gen8.g8b"
  64 };
  65
  66 static const uint32_t gen8_vme_inter_frame[][4] = {
  67 #include "shaders/vme/inter_frame_gen8.g8b"
  68 };
  69
  70 static const uint32_t gen8_vme_inter_bframe[][4] = {
  71 #include "shaders/vme/inter_bframe_gen8.g8b"
  72 };
  73
  74 static struct i965_kernel gen8_vme_kernels[] = {
  75     {
  76         "VME Intra Frame",
  77         VME_INTRA_SHADER, /*index*/
  78         gen8_vme_intra_frame,
  79         sizeof(gen8_vme_intra_frame),
  80         NULL
  81     },
  82     {
  83         "VME inter Frame",
  84         VME_INTER_SHADER,
  85         gen8_vme_inter_frame,
  86         sizeof(gen8_vme_inter_frame),
  87         NULL
  88     },
  89     {
  90         "VME inter BFrame",
  91         VME_BINTER_SHADER,
  92         gen8_vme_inter_bframe,
  93         sizeof(gen8_vme_inter_bframe),
  94         NULL
  95     }
  96 };
  97
  98 static const uint32_t gen8_vme_mpeg2_intra_frame[][4] = {
  99 #include "shaders/vme/intra_frame_gen8.g8b"
 100 };
 101
 102 static const uint32_t gen8_vme_mpeg2_inter_frame[][4] = {
 103 #include "shaders/vme/mpeg2_inter_gen8.g8b"
 104 };
 105
 106 static struct i965_kernel gen8_vme_mpeg2_kernels[] = {
 107     {
 108         "VME Intra Frame",
 109         VME_INTRA_SHADER, /*index*/
 110         gen8_vme_mpeg2_intra_frame,
 111         sizeof(gen8_vme_mpeg2_intra_frame),
 112         NULL
 113     },
 114     {
 115         "VME inter Frame",
 116         VME_INTER_SHADER,
 117         gen8_vme_mpeg2_inter_frame,
 118         sizeof(gen8_vme_mpeg2_inter_frame),
 119         NULL
 120     },
 121 };
 122
 123 /* only used for VME source surface state */
 124 static void
 125 gen8_vme_source_surface_state(VADriverContextP ctx,
 126                               int index,
 127                               struct object_surface *obj_surface,
 128                               struct intel_encoder_context *encoder_context)
 129 {
 130     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 131
 132     vme_context->vme_surface2_setup(ctx,
 133                                     &vme_context->gpe_context,
 134                                     obj_surface,
 135                                     BINDING_TABLE_OFFSET(index),
 136                                     SURFACE_STATE_OFFSET(index));
 137 }
 138
 139 static void
 140 gen8_vme_media_source_surface_state(VADriverContextP ctx,
 141                                     int index,
 142                                     struct object_surface *obj_surface,
 143                                     struct intel_encoder_context *encoder_context)
 144 {
 145     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 146
 147     vme_context->vme_media_rw_surface_setup(ctx,
 148                                             &vme_context->gpe_context,
 149                                             obj_surface,
 150                                             BINDING_TABLE_OFFSET(index),
 151                                             SURFACE_STATE_OFFSET(index));
 152 }
 153
 154 static void
 155 gen8_vme_media_chroma_source_surface_state(VADriverContextP ctx,
 156                                            int index,
 157                                            struct object_surface *obj_surface,
 158                                            struct intel_encoder_context *encoder_context)
 159 {
 160     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 161
 162     vme_context->vme_media_chroma_surface_setup(ctx,
 163                                                 &vme_context->gpe_context,
 164                                                 obj_surface,
 165                                                 BINDING_TABLE_OFFSET(index),
 166                                                 SURFACE_STATE_OFFSET(index));
 167 }
 168
 169 static void
 170 gen8_vme_output_buffer_setup(VADriverContextP ctx,
 171                              struct encode_state *encode_state,
 172                              int index,
 173                              struct intel_encoder_context *encoder_context)
 174
 175 {
 176     struct i965_driver_data *i965 = i965_driver_data(ctx);
 177     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 178     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 179     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 180     int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
 181     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 182     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 183
 184     vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
 185     vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
 186
 187     if (is_intra)
 188         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 2;
 189     else
 190         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 24;
 191     /*
 192      * Inter MV . 32-byte Intra search + 16 IME info + 128 IME MV + 32 IME Ref
 193      * + 16 FBR Info + 128 FBR MV + 32 FBR Ref.
 194      * 16 * (2 + 2 * (1 + 8 + 2))= 16 * 24.
 195      */
 196
 197     vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
 198                                               "VME output buffer",
 199                                               vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
 200                                               0x1000);
 201     assert(vme_context->vme_output.bo);
 202     vme_context->vme_buffer_suface_setup(ctx,
 203                                          &vme_context->gpe_context,
 204                                          &vme_context->vme_output,
 205                                          BINDING_TABLE_OFFSET(index),
 206                                          SURFACE_STATE_OFFSET(index));
 207 }
 208
 209 static void
 210 gen8_vme_output_vme_batchbuffer_setup(VADriverContextP ctx,
 211                                       struct encode_state *encode_state,
 212                                       int index,
 213                                       struct intel_encoder_context *encoder_context)
 214
 215 {
 216     struct i965_driver_data *i965 = i965_driver_data(ctx);
 217     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 218     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 219     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 220     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 221
 222     vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
 223     vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
 224     vme_context->vme_batchbuffer.pitch = 16;
 225     vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
 226                                                    "VME batchbuffer",
 227                                                    vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
 228                                                    0x1000);
 229     vme_context->vme_buffer_suface_setup(ctx,
 230                                          &vme_context->gpe_context,
 231                                          &vme_context->vme_batchbuffer,
 232                                          BINDING_TABLE_OFFSET(index),
 233                                          SURFACE_STATE_OFFSET(index));
 234 }
 235
 236 static VAStatus
 237 gen8_vme_surface_setup(VADriverContextP ctx,
 238                        struct encode_state *encode_state,
 239                        int is_intra,
 240                        struct intel_encoder_context *encoder_context)
 241 {
 242     struct object_surface *obj_surface;
 243     struct i965_driver_data *i965 = i965_driver_data(ctx);
 244
 245     /*Setup surfaces state*/
 246     /* current picture for encoding */
 247     obj_surface = encode_state->input_yuv_object;
 248     gen8_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 249     gen8_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 250     gen8_vme_media_chroma_source_surface_state(ctx, 6, obj_surface, encoder_context);
 251
 252     if (!is_intra) {
 253         VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 254         int slice_type;
 255         struct object_surface *slice_obj_surface;
 256         int ref_surface_id;
 257
 258         slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
 259
 260         if (slice_type == SLICE_TYPE_P || slice_type == SLICE_TYPE_B) {
 261             slice_obj_surface = NULL;
 262             ref_surface_id = slice_param->RefPicList0[0].picture_id;
 263             if (ref_surface_id != 0 && ref_surface_id != VA_INVALID_SURFACE) {
 264                 slice_obj_surface = SURFACE(ref_surface_id);
 265             }
 266             if (slice_obj_surface && slice_obj_surface->bo) {
 267                 obj_surface = slice_obj_surface;
 268             } else {
 269                 obj_surface = encode_state->reference_objects[0];
 270             }
 271             /* reference 0 */
 272             if (obj_surface && obj_surface->bo)
 273                 gen8_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
 274         }
 275         if (slice_type == SLICE_TYPE_B) {
 276             /* reference 1 */
 277             slice_obj_surface = NULL;
 278             ref_surface_id = slice_param->RefPicList1[0].picture_id;
 279             if (ref_surface_id != 0 && ref_surface_id != VA_INVALID_SURFACE) {
 280                 slice_obj_surface = SURFACE(ref_surface_id);
 281             }
 282             if (slice_obj_surface && slice_obj_surface->bo) {
 283                 obj_surface = slice_obj_surface;
 284             } else {
 285                 obj_surface = encode_state->reference_objects[0];
 286             }
 287
 288             obj_surface = encode_state->reference_objects[1];
 289             if (obj_surface && obj_surface->bo)
 290                 gen8_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
 291         }
 292     }
 293
 294     /* VME output */
 295     gen8_vme_output_buffer_setup(ctx, encode_state, 3, encoder_context);
 296     gen8_vme_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 297
 298     return VA_STATUS_SUCCESS;
 299 }
 300
 301 static VAStatus gen8_vme_interface_setup(VADriverContextP ctx,
 302                                          struct encode_state *encode_state,
 303                                          struct intel_encoder_context *encoder_context)
 304 {
 305     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 306     struct gen8_interface_descriptor_data *desc;
 307     int i;
 308     dri_bo *bo;
 309
 310     bo = vme_context->gpe_context.idrt.bo;
 311     dri_bo_map(bo, 1);
 312     assert(bo->virtual);
 313     desc = bo->virtual;
 314
 315     for (i = 0; i < vme_context->vme_kernel_sum; i++) {
 316         struct i965_kernel *kernel;
 317         kernel = &vme_context->gpe_context.kernels[i];
 318         assert(sizeof(*desc) == 32);
 319         /*Setup the descritor table*/
 320         memset(desc, 0, sizeof(*desc));
 321         desc->desc0.kernel_start_pointer = (kernel->bo->offset >> 6);
 322         desc->desc3.sampler_count = 0; /* FIXME: */
 323         desc->desc3.sampler_state_pointer = 0;
 324         desc->desc4.binding_table_entry_count = 1; /* FIXME: */
 325         desc->desc4.binding_table_pointer = (BINDING_TABLE_OFFSET(0) >> 5);
 326         desc->desc5.constant_urb_entry_read_offset = 0;
 327         desc->desc5.constant_urb_entry_read_length = CURBE_URB_ENTRY_LENGTH;
 328
 329
 330         /*kernel start*/
 331         dri_bo_emit_reloc(bo,
 332                           I915_GEM_DOMAIN_INSTRUCTION, 0,
 333                           0,
 334                           i * sizeof(*desc) + offsetof(struct gen8_interface_descriptor_data, desc0),
 335                           kernel->bo);
 336         desc++;
 337     }
 338     dri_bo_unmap(bo);
 339
 340     return VA_STATUS_SUCCESS;
 341 }
 342
 343 static VAStatus gen8_vme_constant_setup(VADriverContextP ctx,
 344                                         struct encode_state *encode_state,
 345                                         struct intel_encoder_context *encoder_context)
 346 {
 347     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 348     unsigned char *constant_buffer;
 349     unsigned int *vme_state_message;
 350     int mv_num = 32;
 351
 352     vme_state_message = (unsigned int *)vme_context->vme_state_message;
 353
 354     if (encoder_context->codec == CODEC_H264) {
 355         if (vme_context->h264_level >= 30) {
 356             mv_num = 16;
 357
 358             if (vme_context->h264_level >= 31)
 359                 mv_num = 8;
 360         }
 361     } else if (encoder_context->codec == CODEC_MPEG2) {
 362         mv_num = 2;
 363     }
 364
 365     vme_state_message[31] = mv_num;
 366
 367     dri_bo_map(vme_context->gpe_context.curbe.bo, 1);
 368     assert(vme_context->gpe_context.curbe.bo->virtual);
 369     constant_buffer = vme_context->gpe_context.curbe.bo->virtual;
 370
 371     /* VME MV/Mb cost table is passed by using const buffer */
 372     /* Now it uses the fixed search path. So it is constructed directly
 373      * in the GPU shader.
 374      */
 375     memcpy(constant_buffer, (char *)vme_context->vme_state_message, 128);
 376
 377     dri_bo_unmap(vme_context->gpe_context.curbe.bo);
 378
 379     return VA_STATUS_SUCCESS;
 380 }
 381
 382
 383 static void
 384 gen8_vme_fill_vme_batchbuffer(VADriverContextP ctx,
 385                               struct encode_state *encode_state,
 386                               int mb_width, int mb_height,
 387                               int kernel,
 388                               int transform_8x8_mode_flag,
 389                               struct intel_encoder_context *encoder_context)
 390 {
 391     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 392     int mb_x = 0, mb_y = 0;
 393     int i, s;
 394     unsigned int *command_ptr;
 395
 396     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 397     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 398
 399     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 400         VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 401         int slice_mb_begin = pSliceParameter->macroblock_address;
 402         int slice_mb_number = pSliceParameter->num_macroblocks;
 403         unsigned int mb_intra_ub;
 404         int slice_mb_x = pSliceParameter->macroblock_address % mb_width;
 405         for (i = 0; i < slice_mb_number;  ) {
 406             int mb_count = i + slice_mb_begin;
 407             mb_x = mb_count % mb_width;
 408             mb_y = mb_count / mb_width;
 409             mb_intra_ub = 0;
 410             if (mb_x != 0) {
 411                 mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 412             }
 413             if (mb_y != 0) {
 414                 mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 415                 if (mb_x != 0)
 416                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 417                 if (mb_x != (mb_width -1))
 418                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 419             }
 420             if (i < mb_width) {
 421                 if (i == 0)
 422                     mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_AE);
 423                 mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_BCD_MASK);
 424                 if ((i == (mb_width - 1)) && slice_mb_x) {
 425                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 426                 }
 427             }
 428
 429             if ((i == mb_width) && slice_mb_x) {
 430                 mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_D);
 431             }
 432             *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 433             *command_ptr++ = kernel;
 434             *command_ptr++ = 0;
 435             *command_ptr++ = 0;
 436             *command_ptr++ = 0;
 437             *command_ptr++ = 0;
 438
 439             /*inline data */
 440             *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 441             *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 442
 443             i += 1;
 444         }
 445     }
 446
 447     *command_ptr++ = 0;
 448     *command_ptr++ = MI_BATCH_BUFFER_END;
 449
 450     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 451 }
 452
 453 static void gen8_vme_media_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 454 {
 455     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 456
 457     i965_gpe_context_init(ctx, &vme_context->gpe_context);
 458
 459     /* VME output buffer */
 460     dri_bo_unreference(vme_context->vme_output.bo);
 461     vme_context->vme_output.bo = NULL;
 462
 463     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 464     vme_context->vme_batchbuffer.bo = NULL;
 465
 466     /* VME state */
 467     dri_bo_unreference(vme_context->vme_state.bo);
 468     vme_context->vme_state.bo = NULL;
 469 }
 470
 471 static void gen8_vme_pipeline_programing(VADriverContextP ctx,
 472                                          struct encode_state *encode_state,
 473                                          struct intel_encoder_context *encoder_context)
 474 {
 475     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 476     struct intel_batchbuffer *batch = encoder_context->base.batch;
 477     VAEncPictureParameterBufferH264 *pPicParameter = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
 478     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 479     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 480     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 481     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 482     int kernel_shader;
 483     bool allow_hwscore = true;
 484     int s;
 485
 486     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 487         pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 488         if ((pSliceParameter->macroblock_address % width_in_mbs)) {
 489             allow_hwscore = false;
 490             break;
 491         }
 492     }
 493     if ((pSliceParameter->slice_type == SLICE_TYPE_I) ||
 494         (pSliceParameter->slice_type == SLICE_TYPE_I)) {
 495         kernel_shader = VME_INTRA_SHADER;
 496     } else if ((pSliceParameter->slice_type == SLICE_TYPE_P) ||
 497                (pSliceParameter->slice_type == SLICE_TYPE_SP)) {
 498         kernel_shader = VME_INTER_SHADER;
 499     } else {
 500         kernel_shader = VME_BINTER_SHADER;
 501         if (!allow_hwscore)
 502             kernel_shader = VME_INTER_SHADER;
 503     }
 504     if (allow_hwscore)
 505         gen7_vme_walker_fill_vme_batchbuffer(ctx,
 506                                              encode_state,
 507                                              width_in_mbs, height_in_mbs,
 508                                              kernel_shader,
 509                                              pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 510                                              encoder_context);
 511     else
 512         gen8_vme_fill_vme_batchbuffer(ctx,
 513                                       encode_state,
 514                                       width_in_mbs, height_in_mbs,
 515                                       kernel_shader,
 516                                       pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 517                                       encoder_context);
 518
 519     intel_batchbuffer_start_atomic(batch, 0x1000);
 520     gen8_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 521     BEGIN_BATCH(batch, 3);
 522     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (1 << 8) | (1 << 0));
 523     OUT_RELOC(batch,
 524               vme_context->vme_batchbuffer.bo,
 525               I915_GEM_DOMAIN_COMMAND, 0,
 526               0);
 527     OUT_BATCH(batch, 0);
 528     ADVANCE_BATCH(batch);
 529
 530     intel_batchbuffer_end_atomic(batch);
 531 }
 532
 533 static VAStatus gen8_vme_prepare(VADriverContextP ctx,
 534                                  struct encode_state *encode_state,
 535                                  struct intel_encoder_context *encoder_context)
 536 {
 537     VAStatus vaStatus = VA_STATUS_SUCCESS;
 538     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 539     int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
 540     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 541     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 542
 543     if (!vme_context->h264_level ||
 544         (vme_context->h264_level != pSequenceParameter->level_idc)) {
 545         vme_context->h264_level = pSequenceParameter->level_idc;
 546     }
 547
 548     intel_vme_update_mbmv_cost(ctx, encode_state, encoder_context);
 549
 550     /*Setup all the memory object*/
 551     gen8_vme_surface_setup(ctx, encode_state, is_intra, encoder_context);
 552     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
 553     //gen8_vme_vme_state_setup(ctx, encode_state, is_intra, encoder_context);
 554     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 555
 556     /*Programing media pipeline*/
 557     gen8_vme_pipeline_programing(ctx, encode_state, encoder_context);
 558
 559     return vaStatus;
 560 }
 561
 562 static VAStatus gen8_vme_run(VADriverContextP ctx,
 563                              struct encode_state *encode_state,
 564                              struct intel_encoder_context *encoder_context)
 565 {
 566     struct intel_batchbuffer *batch = encoder_context->base.batch;
 567
 568     intel_batchbuffer_flush(batch);
 569
 570     return VA_STATUS_SUCCESS;
 571 }
 572
 573 static VAStatus gen8_vme_stop(VADriverContextP ctx,
 574                               struct encode_state *encode_state,
 575                               struct intel_encoder_context *encoder_context)
 576 {
 577     return VA_STATUS_SUCCESS;
 578 }
 579
 580 static VAStatus
 581 gen8_vme_pipeline(VADriverContextP ctx,
 582                   VAProfile profile,
 583                   struct encode_state *encode_state,
 584                   struct intel_encoder_context *encoder_context)
 585 {
 586     gen8_vme_media_init(ctx, encoder_context);
 587     gen8_vme_prepare(ctx, encode_state, encoder_context);
 588     gen8_vme_run(ctx, encode_state, encoder_context);
 589     gen8_vme_stop(ctx, encode_state, encoder_context);
 590
 591     return VA_STATUS_SUCCESS;
 592 }
 593
 594 static void
 595 gen8_vme_mpeg2_output_buffer_setup(VADriverContextP ctx,
 596                                    struct encode_state *encode_state,
 597                                    int index,
 598                                    int is_intra,
 599                                    struct intel_encoder_context *encoder_context)
 600
 601 {
 602     struct i965_driver_data *i965 = i965_driver_data(ctx);
 603     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 604     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 605     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 606     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 607
 608     vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
 609     vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
 610
 611     if (is_intra)
 612         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 2;
 613     else
 614         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 24;
 615     /*
 616      * Inter MV . 32-byte Intra search + 16 IME info + 128 IME MV + 32 IME Ref
 617      * + 16 FBR Info + 128 FBR MV + 32 FBR Ref.
 618      * 16 * (2 + 2 * (1 + 8 + 2))= 16 * 24.
 619      */
 620
 621     vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
 622                                               "VME output buffer",
 623                                               vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
 624                                               0x1000);
 625     assert(vme_context->vme_output.bo);
 626     vme_context->vme_buffer_suface_setup(ctx,
 627                                          &vme_context->gpe_context,
 628                                          &vme_context->vme_output,
 629                                          BINDING_TABLE_OFFSET(index),
 630                                          SURFACE_STATE_OFFSET(index));
 631 }
 632
 633 static void
 634 gen8_vme_mpeg2_output_vme_batchbuffer_setup(VADriverContextP ctx,
 635                                             struct encode_state *encode_state,
 636                                             int index,
 637                                             struct intel_encoder_context *encoder_context)
 638
 639 {
 640     struct i965_driver_data *i965 = i965_driver_data(ctx);
 641     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 642     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 643     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 644     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 645
 646     vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
 647     vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
 648     vme_context->vme_batchbuffer.pitch = 16;
 649     vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
 650                                                    "VME batchbuffer",
 651                                                    vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
 652                                                    0x1000);
 653     vme_context->vme_buffer_suface_setup(ctx,
 654                                          &vme_context->gpe_context,
 655                                          &vme_context->vme_batchbuffer,
 656                                          BINDING_TABLE_OFFSET(index),
 657                                          SURFACE_STATE_OFFSET(index));
 658 }
 659
 660 static VAStatus
 661 gen8_vme_mpeg2_surface_setup(VADriverContextP ctx,
 662                              struct encode_state *encode_state,
 663                              int is_intra,
 664                              struct intel_encoder_context *encoder_context)
 665 {
 666     struct object_surface *obj_surface;
 667
 668     /*Setup surfaces state*/
 669     /* current picture for encoding */
 670     obj_surface = encode_state->input_yuv_object;
 671     gen8_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 672     gen8_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 673     gen8_vme_media_chroma_source_surface_state(ctx, 6, obj_surface, encoder_context);
 674
 675     if (!is_intra) {
 676         /* reference 0 */
 677         obj_surface = encode_state->reference_objects[0];
 678
 679         if (obj_surface->bo != NULL)
 680             gen8_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
 681
 682         /* reference 1 */
 683         obj_surface = encode_state->reference_objects[1];
 684
 685         if (obj_surface && obj_surface->bo != NULL)
 686             gen8_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
 687     }
 688
 689     /* VME output */
 690     gen8_vme_mpeg2_output_buffer_setup(ctx, encode_state, 3, is_intra, encoder_context);
 691     gen8_vme_mpeg2_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 692
 693     return VA_STATUS_SUCCESS;
 694 }
 695
 696 static void
 697 gen8_vme_mpeg2_fill_vme_batchbuffer(VADriverContextP ctx,
 698                                     struct encode_state *encode_state,
 699                                     int mb_width, int mb_height,
 700                                     int kernel,
 701                                     int transform_8x8_mode_flag,
 702                                     struct intel_encoder_context *encoder_context)
 703 {
 704     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 705     int mb_x = 0, mb_y = 0;
 706     int i, s, j;
 707     unsigned int *command_ptr;
 708
 709
 710     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 711     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 712
 713     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 714         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 715
 716         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 717             int slice_mb_begin = slice_param->macroblock_address;
 718             int slice_mb_number = slice_param->num_macroblocks;
 719             unsigned int mb_intra_ub;
 720
 721             for (i = 0; i < slice_mb_number;) {
 722                 int mb_count = i + slice_mb_begin;
 723
 724                 mb_x = mb_count % mb_width;
 725                 mb_y = mb_count / mb_width;
 726                 mb_intra_ub = 0;
 727
 728                 if (mb_x != 0) {
 729                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 730                 }
 731
 732                 if (mb_y != 0) {
 733                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 734
 735                     if (mb_x != 0)
 736                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 737
 738                     if (mb_x != (mb_width -1))
 739                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 740                 }
 741
 742                 *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 743                 *command_ptr++ = kernel;
 744                 *command_ptr++ = 0;
 745                 *command_ptr++ = 0;
 746                 *command_ptr++ = 0;
 747                 *command_ptr++ = 0;
 748
 749                 /*inline data */
 750                 *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 751                 *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 752
 753                 i += 1;
 754             }
 755
 756             slice_param++;
 757         }
 758     }
 759
 760     *command_ptr++ = 0;
 761     *command_ptr++ = MI_BATCH_BUFFER_END;
 762
 763     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 764 }
 765
 766 static void
 767 gen8_vme_mpeg2_pipeline_programing(VADriverContextP ctx,
 768                                    struct encode_state *encode_state,
 769                                    int is_intra,
 770                                    struct intel_encoder_context *encoder_context)
 771 {
 772     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 773     struct intel_batchbuffer *batch = encoder_context->base.batch;
 774     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 775     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 776     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 777     bool allow_hwscore = true;
 778     int s;
 779     int kernel_shader;
 780     VAEncPictureParameterBufferMPEG2 *pic_param = NULL;
 781
 782     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 783         int j;
 784         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 785
 786         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 787             if (slice_param->macroblock_address % width_in_mbs) {
 788                 allow_hwscore = false;
 789                 break;
 790             }
 791         }
 792     }
 793
 794     pic_param = (VAEncPictureParameterBufferMPEG2 *)encode_state->pic_param_ext->buffer;
 795     if (pic_param->picture_type == VAEncPictureTypeIntra) {
 796         allow_hwscore = false;
 797         kernel_shader = VME_INTRA_SHADER;
 798     } else {
 799         kernel_shader = VME_INTER_SHADER;
 800     }
 801
 802     if (allow_hwscore)
 803         gen7_vme_mpeg2_walker_fill_vme_batchbuffer(ctx,
 804                                                    encode_state,
 805                                                    width_in_mbs, height_in_mbs,
 806                                                    kernel_shader,
 807                                                    encoder_context);
 808     else
 809         gen8_vme_mpeg2_fill_vme_batchbuffer(ctx,
 810                                             encode_state,
 811                                             width_in_mbs, height_in_mbs,
 812                                             is_intra ? VME_INTRA_SHADER : VME_INTER_SHADER,
 813                                             0,
 814                                             encoder_context);
 815
 816     intel_batchbuffer_start_atomic(batch, 0x1000);
 817     gen8_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 818     BEGIN_BATCH(batch, 2);
 819     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (2 << 6));
 820     OUT_RELOC(batch,
 821               vme_context->vme_batchbuffer.bo,
 822               I915_GEM_DOMAIN_COMMAND, 0,
 823               0);
 824     ADVANCE_BATCH(batch);
 825
 826     intel_batchbuffer_end_atomic(batch);
 827 }
 828
 829 static VAStatus
 830 gen8_vme_mpeg2_prepare(VADriverContextP ctx,
 831                        struct encode_state *encode_state,
 832                        struct intel_encoder_context *encoder_context)
 833 {
 834     VAStatus vaStatus = VA_STATUS_SUCCESS;
 835     VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[0]->buffer;
 836     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 837     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 838
 839     if ((!vme_context->mpeg2_level) ||
 840         (vme_context->mpeg2_level != (seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK))) {
 841         vme_context->mpeg2_level = seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK;
 842     }
 843
 844
 845     /*Setup all the memory object*/
 846     gen8_vme_mpeg2_surface_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 847     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
 848     //gen8_vme_vme_state_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 849     intel_vme_mpeg2_state_setup(ctx, encode_state, encoder_context);
 850     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 851
 852     /*Programing media pipeline*/
 853     gen8_vme_mpeg2_pipeline_programing(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 854
 855     return vaStatus;
 856 }
 857
 858 static VAStatus
 859 gen8_vme_mpeg2_pipeline(VADriverContextP ctx,
 860                         VAProfile profile,
 861                         struct encode_state *encode_state,
 862                         struct intel_encoder_context *encoder_context)
 863 {
 864     gen8_vme_media_init(ctx, encoder_context);
 865     gen8_vme_mpeg2_prepare(ctx, encode_state, encoder_context);
 866     gen8_vme_run(ctx, encode_state, encoder_context);
 867     gen8_vme_stop(ctx, encode_state, encoder_context);
 868
 869     return VA_STATUS_SUCCESS;
 870 }
 871
 872 static void
 873 gen8_vme_context_destroy(void *context)
 874 {
 875     struct gen6_vme_context *vme_context = context;
 876
 877     i965_gpe_context_destroy(&vme_context->gpe_context);
 878
 879     dri_bo_unreference(vme_context->vme_output.bo);
 880     vme_context->vme_output.bo = NULL;
 881
 882     dri_bo_unreference(vme_context->vme_state.bo);
 883     vme_context->vme_state.bo = NULL;
 884
 885     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 886     vme_context->vme_batchbuffer.bo = NULL;
 887
 888     if (vme_context->vme_state_message) {
 889         free(vme_context->vme_state_message);
 890         vme_context->vme_state_message = NULL;
 891     }
 892
 893     free(vme_context);
 894 }
 895
 896 Bool gen8_vme_context_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 897 {
 898     struct gen6_vme_context *vme_context = calloc(1, sizeof(struct gen6_vme_context));
 899     struct i965_kernel *vme_kernel_list = NULL;
 900     int i965_kernel_num;
 901
 902     switch (encoder_context->codec) {
 903     case CODEC_H264:
 904         vme_kernel_list = gen8_vme_kernels;
 905         encoder_context->vme_pipeline = gen8_vme_pipeline;
 906         i965_kernel_num = sizeof(gen8_vme_kernels) / sizeof(struct i965_kernel);
 907         break;
 908
 909     case CODEC_MPEG2:
 910         vme_kernel_list = gen8_vme_mpeg2_kernels;
 911         encoder_context->vme_pipeline = gen8_vme_mpeg2_pipeline;
 912         i965_kernel_num = sizeof(gen8_vme_mpeg2_kernels) / sizeof(struct i965_kernel);
 913
 914         break;
 915
 916     default:
 917         /* never get here */
 918         assert(0);
 919
 920         break;
 921     }
 922     vme_context->vme_kernel_sum = i965_kernel_num;
 923     vme_context->gpe_context.surface_state_binding_table.length = (SURFACE_STATE_PADDED_SIZE + sizeof(unsigned int)) * MAX_MEDIA_SURFACES_GEN6;
 924
 925     vme_context->gpe_context.idrt.max_entries = MAX_INTERFACE_DESC_GEN6;
 926     vme_context->gpe_context.idrt.entry_size = sizeof(struct gen8_interface_descriptor_data);
 927
 928     vme_context->gpe_context.curbe.length = CURBE_TOTAL_DATA_LENGTH;
 929
 930     vme_context->gpe_context.vfe_state.max_num_threads = 60 - 1;
 931     vme_context->gpe_context.vfe_state.num_urb_entries = 16;
 932     vme_context->gpe_context.vfe_state.gpgpu_mode = 0;
 933     vme_context->gpe_context.vfe_state.urb_entry_size = 59 - 1;
 934     vme_context->gpe_context.vfe_state.curbe_allocation_size = CURBE_ALLOCATION_SIZE - 1;
 935
 936     gen7_vme_scoreboard_init(ctx, vme_context);
 937
 938     i965_gpe_load_kernels(ctx,
 939                           &vme_context->gpe_context,
 940                           vme_kernel_list,
 941                           i965_kernel_num);
 942     vme_context->vme_surface2_setup = gen8_gpe_surface2_setup;
 943     vme_context->vme_media_rw_surface_setup = gen8_gpe_media_rw_surface_setup;
 944     vme_context->vme_buffer_suface_setup = gen8_gpe_buffer_suface_setup;
 945     vme_context->vme_media_chroma_surface_setup = gen8_gpe_media_chroma_surface_setup;
 946
 947     encoder_context->vme_context = vme_context;
 948     encoder_context->vme_context_destroy = gen8_vme_context_destroy;
 949
 950     vme_context->vme_state_message = malloc(VME_MSG_LENGTH * sizeof(int));
 951
 952     return True;
 953 }