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 != 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 != 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[1];
 286             }
 287
 288             if (obj_surface && obj_surface->bo)
 289                 gen8_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
 290         }
 291     }
 292
 293     /* VME output */
 294     gen8_vme_output_buffer_setup(ctx, encode_state, 3, encoder_context);
 295     gen8_vme_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 296
 297     return VA_STATUS_SUCCESS;
 298 }
 299
 300 static VAStatus gen8_vme_interface_setup(VADriverContextP ctx,
 301                                          struct encode_state *encode_state,
 302                                          struct intel_encoder_context *encoder_context)
 303 {
 304     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 305     struct gen8_interface_descriptor_data *desc;
 306     int i;
 307     dri_bo *bo;
 308
 309     bo = vme_context->gpe_context.idrt.bo;
 310     dri_bo_map(bo, 1);
 311     assert(bo->virtual);
 312     desc = bo->virtual;
 313
 314     for (i = 0; i < vme_context->vme_kernel_sum; i++) {
 315         struct i965_kernel *kernel;
 316         kernel = &vme_context->gpe_context.kernels[i];
 317         assert(sizeof(*desc) == 32);
 318         /*Setup the descritor table*/
 319         memset(desc, 0, sizeof(*desc));
 320         desc->desc0.kernel_start_pointer = (kernel->bo->offset >> 6);
 321         desc->desc3.sampler_count = 0; /* FIXME: */
 322         desc->desc3.sampler_state_pointer = 0;
 323         desc->desc4.binding_table_entry_count = 1; /* FIXME: */
 324         desc->desc4.binding_table_pointer = (BINDING_TABLE_OFFSET(0) >> 5);
 325         desc->desc5.constant_urb_entry_read_offset = 0;
 326         desc->desc5.constant_urb_entry_read_length = CURBE_URB_ENTRY_LENGTH;
 327
 328
 329         /*kernel start*/
 330         dri_bo_emit_reloc(bo,
 331                           I915_GEM_DOMAIN_INSTRUCTION, 0,
 332                           0,
 333                           i * sizeof(*desc) + offsetof(struct gen8_interface_descriptor_data, desc0),
 334                           kernel->bo);
 335         desc++;
 336     }
 337     dri_bo_unmap(bo);
 338
 339     return VA_STATUS_SUCCESS;
 340 }
 341
 342 static VAStatus gen8_vme_constant_setup(VADriverContextP ctx,
 343                                         struct encode_state *encode_state,
 344                                         struct intel_encoder_context *encoder_context)
 345 {
 346     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 347     unsigned char *constant_buffer;
 348     unsigned int *vme_state_message;
 349     int mv_num = 32;
 350
 351     vme_state_message = (unsigned int *)vme_context->vme_state_message;
 352
 353     if (encoder_context->codec == CODEC_H264) {
 354         if (vme_context->h264_level >= 30) {
 355             mv_num = 16;
 356
 357             if (vme_context->h264_level >= 31)
 358                 mv_num = 8;
 359         }
 360     } else if (encoder_context->codec == CODEC_MPEG2) {
 361         mv_num = 2;
 362     }
 363
 364     vme_state_message[31] = mv_num;
 365
 366     dri_bo_map(vme_context->gpe_context.curbe.bo, 1);
 367     assert(vme_context->gpe_context.curbe.bo->virtual);
 368     constant_buffer = vme_context->gpe_context.curbe.bo->virtual;
 369
 370     /* VME MV/Mb cost table is passed by using const buffer */
 371     /* Now it uses the fixed search path. So it is constructed directly
 372      * in the GPU shader.
 373      */
 374     memcpy(constant_buffer, (char *)vme_context->vme_state_message, 128);
 375
 376     dri_bo_unmap(vme_context->gpe_context.curbe.bo);
 377
 378     return VA_STATUS_SUCCESS;
 379 }
 380
 381
 382 static void
 383 gen8_vme_fill_vme_batchbuffer(VADriverContextP ctx,
 384                               struct encode_state *encode_state,
 385                               int mb_width, int mb_height,
 386                               int kernel,
 387                               int transform_8x8_mode_flag,
 388                               struct intel_encoder_context *encoder_context)
 389 {
 390     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 391     int mb_x = 0, mb_y = 0;
 392     int i, s;
 393     unsigned int *command_ptr;
 394
 395     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 396     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 397
 398     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 399         VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 400         int slice_mb_begin = pSliceParameter->macroblock_address;
 401         int slice_mb_number = pSliceParameter->num_macroblocks;
 402         unsigned int mb_intra_ub;
 403         int slice_mb_x = pSliceParameter->macroblock_address % mb_width;
 404         for (i = 0; i < slice_mb_number;  ) {
 405             int mb_count = i + slice_mb_begin;
 406             mb_x = mb_count % mb_width;
 407             mb_y = mb_count / mb_width;
 408             mb_intra_ub = 0;
 409             if (mb_x != 0) {
 410                 mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 411             }
 412             if (mb_y != 0) {
 413                 mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 414                 if (mb_x != 0)
 415                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 416                 if (mb_x != (mb_width -1))
 417                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 418             }
 419             if (i < mb_width) {
 420                 if (i == 0)
 421                     mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_AE);
 422                 mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_BCD_MASK);
 423                 if ((i == (mb_width - 1)) && slice_mb_x) {
 424                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 425                 }
 426             }
 427
 428             if ((i == mb_width) && slice_mb_x) {
 429                 mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_D);
 430             }
 431             *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 432             *command_ptr++ = kernel;
 433             *command_ptr++ = 0;
 434             *command_ptr++ = 0;
 435             *command_ptr++ = 0;
 436             *command_ptr++ = 0;
 437
 438             /*inline data */
 439             *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 440             *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 441
 442             i += 1;
 443         }
 444     }
 445
 446     *command_ptr++ = 0;
 447     *command_ptr++ = MI_BATCH_BUFFER_END;
 448
 449     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 450 }
 451
 452 static void gen8_vme_media_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 453 {
 454     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 455
 456     i965_gpe_context_init(ctx, &vme_context->gpe_context);
 457
 458     /* VME output buffer */
 459     dri_bo_unreference(vme_context->vme_output.bo);
 460     vme_context->vme_output.bo = NULL;
 461
 462     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 463     vme_context->vme_batchbuffer.bo = NULL;
 464
 465     /* VME state */
 466     dri_bo_unreference(vme_context->vme_state.bo);
 467     vme_context->vme_state.bo = NULL;
 468 }
 469
 470 static void gen8_vme_pipeline_programing(VADriverContextP ctx,
 471                                          struct encode_state *encode_state,
 472                                          struct intel_encoder_context *encoder_context)
 473 {
 474     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 475     struct intel_batchbuffer *batch = encoder_context->base.batch;
 476     VAEncPictureParameterBufferH264 *pPicParameter = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
 477     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 478     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 479     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 480     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 481     int kernel_shader;
 482     bool allow_hwscore = true;
 483     int s;
 484
 485     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 486         pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 487         if ((pSliceParameter->macroblock_address % width_in_mbs)) {
 488             allow_hwscore = false;
 489             break;
 490         }
 491     }
 492     if ((pSliceParameter->slice_type == SLICE_TYPE_I) ||
 493         (pSliceParameter->slice_type == SLICE_TYPE_I)) {
 494         kernel_shader = VME_INTRA_SHADER;
 495     } else if ((pSliceParameter->slice_type == SLICE_TYPE_P) ||
 496                (pSliceParameter->slice_type == SLICE_TYPE_SP)) {
 497         kernel_shader = VME_INTER_SHADER;
 498     } else {
 499         kernel_shader = VME_BINTER_SHADER;
 500         if (!allow_hwscore)
 501             kernel_shader = VME_INTER_SHADER;
 502     }
 503     if (allow_hwscore)
 504         gen7_vme_walker_fill_vme_batchbuffer(ctx,
 505                                              encode_state,
 506                                              width_in_mbs, height_in_mbs,
 507                                              kernel_shader,
 508                                              pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 509                                              encoder_context);
 510     else
 511         gen8_vme_fill_vme_batchbuffer(ctx,
 512                                       encode_state,
 513                                       width_in_mbs, height_in_mbs,
 514                                       kernel_shader,
 515                                       pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 516                                       encoder_context);
 517
 518     intel_batchbuffer_start_atomic(batch, 0x1000);
 519     gen8_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 520     BEGIN_BATCH(batch, 3);
 521     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (1 << 8) | (1 << 0));
 522     OUT_RELOC(batch,
 523               vme_context->vme_batchbuffer.bo,
 524               I915_GEM_DOMAIN_COMMAND, 0,
 525               0);
 526     OUT_BATCH(batch, 0);
 527     ADVANCE_BATCH(batch);
 528
 529     intel_batchbuffer_end_atomic(batch);
 530 }
 531
 532 static VAStatus gen8_vme_prepare(VADriverContextP ctx,
 533                                  struct encode_state *encode_state,
 534                                  struct intel_encoder_context *encoder_context)
 535 {
 536     VAStatus vaStatus = VA_STATUS_SUCCESS;
 537     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 538     int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
 539     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 540     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 541
 542     if (!vme_context->h264_level ||
 543         (vme_context->h264_level != pSequenceParameter->level_idc)) {
 544         vme_context->h264_level = pSequenceParameter->level_idc;
 545     }
 546
 547     intel_vme_update_mbmv_cost(ctx, encode_state, encoder_context);
 548
 549     /*Setup all the memory object*/
 550     gen8_vme_surface_setup(ctx, encode_state, is_intra, encoder_context);
 551     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
 552     //gen8_vme_vme_state_setup(ctx, encode_state, is_intra, encoder_context);
 553     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 554
 555     /*Programing media pipeline*/
 556     gen8_vme_pipeline_programing(ctx, encode_state, encoder_context);
 557
 558     return vaStatus;
 559 }
 560
 561 static VAStatus gen8_vme_run(VADriverContextP ctx,
 562                              struct encode_state *encode_state,
 563                              struct intel_encoder_context *encoder_context)
 564 {
 565     struct intel_batchbuffer *batch = encoder_context->base.batch;
 566
 567     intel_batchbuffer_flush(batch);
 568
 569     return VA_STATUS_SUCCESS;
 570 }
 571
 572 static VAStatus gen8_vme_stop(VADriverContextP ctx,
 573                               struct encode_state *encode_state,
 574                               struct intel_encoder_context *encoder_context)
 575 {
 576     return VA_STATUS_SUCCESS;
 577 }
 578
 579 static VAStatus
 580 gen8_vme_pipeline(VADriverContextP ctx,
 581                   VAProfile profile,
 582                   struct encode_state *encode_state,
 583                   struct intel_encoder_context *encoder_context)
 584 {
 585     gen8_vme_media_init(ctx, encoder_context);
 586     gen8_vme_prepare(ctx, encode_state, encoder_context);
 587     gen8_vme_run(ctx, encode_state, encoder_context);
 588     gen8_vme_stop(ctx, encode_state, encoder_context);
 589
 590     return VA_STATUS_SUCCESS;
 591 }
 592
 593 static void
 594 gen8_vme_mpeg2_output_buffer_setup(VADriverContextP ctx,
 595                                    struct encode_state *encode_state,
 596                                    int index,
 597                                    int is_intra,
 598                                    struct intel_encoder_context *encoder_context)
 599
 600 {
 601     struct i965_driver_data *i965 = i965_driver_data(ctx);
 602     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 603     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 604     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 605     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 606
 607     vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
 608     vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
 609
 610     if (is_intra)
 611         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 2;
 612     else
 613         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 24;
 614     /*
 615      * Inter MV . 32-byte Intra search + 16 IME info + 128 IME MV + 32 IME Ref
 616      * + 16 FBR Info + 128 FBR MV + 32 FBR Ref.
 617      * 16 * (2 + 2 * (1 + 8 + 2))= 16 * 24.
 618      */
 619
 620     vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
 621                                               "VME output buffer",
 622                                               vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
 623                                               0x1000);
 624     assert(vme_context->vme_output.bo);
 625     vme_context->vme_buffer_suface_setup(ctx,
 626                                          &vme_context->gpe_context,
 627                                          &vme_context->vme_output,
 628                                          BINDING_TABLE_OFFSET(index),
 629                                          SURFACE_STATE_OFFSET(index));
 630 }
 631
 632 static void
 633 gen8_vme_mpeg2_output_vme_batchbuffer_setup(VADriverContextP ctx,
 634                                             struct encode_state *encode_state,
 635                                             int index,
 636                                             struct intel_encoder_context *encoder_context)
 637
 638 {
 639     struct i965_driver_data *i965 = i965_driver_data(ctx);
 640     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 641     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 642     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 643     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 644
 645     vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
 646     vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
 647     vme_context->vme_batchbuffer.pitch = 16;
 648     vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
 649                                                    "VME batchbuffer",
 650                                                    vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
 651                                                    0x1000);
 652     vme_context->vme_buffer_suface_setup(ctx,
 653                                          &vme_context->gpe_context,
 654                                          &vme_context->vme_batchbuffer,
 655                                          BINDING_TABLE_OFFSET(index),
 656                                          SURFACE_STATE_OFFSET(index));
 657 }
 658
 659 static VAStatus
 660 gen8_vme_mpeg2_surface_setup(VADriverContextP ctx,
 661                              struct encode_state *encode_state,
 662                              int is_intra,
 663                              struct intel_encoder_context *encoder_context)
 664 {
 665     struct object_surface *obj_surface;
 666
 667     /*Setup surfaces state*/
 668     /* current picture for encoding */
 669     obj_surface = encode_state->input_yuv_object;
 670     gen8_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 671     gen8_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 672     gen8_vme_media_chroma_source_surface_state(ctx, 6, obj_surface, encoder_context);
 673
 674     if (!is_intra) {
 675         /* reference 0 */
 676         obj_surface = encode_state->reference_objects[0];
 677
 678         if (obj_surface->bo != NULL)
 679             gen8_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
 680
 681         /* reference 1 */
 682         obj_surface = encode_state->reference_objects[1];
 683
 684         if (obj_surface && obj_surface->bo != NULL)
 685             gen8_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
 686     }
 687
 688     /* VME output */
 689     gen8_vme_mpeg2_output_buffer_setup(ctx, encode_state, 3, is_intra, encoder_context);
 690     gen8_vme_mpeg2_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 691
 692     return VA_STATUS_SUCCESS;
 693 }
 694
 695 static void
 696 gen8_vme_mpeg2_fill_vme_batchbuffer(VADriverContextP ctx,
 697                                     struct encode_state *encode_state,
 698                                     int mb_width, int mb_height,
 699                                     int kernel,
 700                                     int transform_8x8_mode_flag,
 701                                     struct intel_encoder_context *encoder_context)
 702 {
 703     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 704     int mb_x = 0, mb_y = 0;
 705     int i, s, j;
 706     unsigned int *command_ptr;
 707
 708
 709     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 710     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 711
 712     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 713         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 714
 715         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 716             int slice_mb_begin = slice_param->macroblock_address;
 717             int slice_mb_number = slice_param->num_macroblocks;
 718             unsigned int mb_intra_ub;
 719
 720             for (i = 0; i < slice_mb_number;) {
 721                 int mb_count = i + slice_mb_begin;
 722
 723                 mb_x = mb_count % mb_width;
 724                 mb_y = mb_count / mb_width;
 725                 mb_intra_ub = 0;
 726
 727                 if (mb_x != 0) {
 728                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 729                 }
 730
 731                 if (mb_y != 0) {
 732                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 733
 734                     if (mb_x != 0)
 735                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 736
 737                     if (mb_x != (mb_width -1))
 738                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 739                 }
 740
 741                 *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 742                 *command_ptr++ = kernel;
 743                 *command_ptr++ = 0;
 744                 *command_ptr++ = 0;
 745                 *command_ptr++ = 0;
 746                 *command_ptr++ = 0;
 747
 748                 /*inline data */
 749                 *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 750                 *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 751
 752                 i += 1;
 753             }
 754
 755             slice_param++;
 756         }
 757     }
 758
 759     *command_ptr++ = 0;
 760     *command_ptr++ = MI_BATCH_BUFFER_END;
 761
 762     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 763 }
 764
 765 static void
 766 gen8_vme_mpeg2_pipeline_programing(VADriverContextP ctx,
 767                                    struct encode_state *encode_state,
 768                                    int is_intra,
 769                                    struct intel_encoder_context *encoder_context)
 770 {
 771     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 772     struct intel_batchbuffer *batch = encoder_context->base.batch;
 773     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 774     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 775     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 776     bool allow_hwscore = true;
 777     int s;
 778     int kernel_shader;
 779     VAEncPictureParameterBufferMPEG2 *pic_param = NULL;
 780
 781     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 782         int j;
 783         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 784
 785         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 786             if (slice_param->macroblock_address % width_in_mbs) {
 787                 allow_hwscore = false;
 788                 break;
 789             }
 790         }
 791     }
 792
 793     pic_param = (VAEncPictureParameterBufferMPEG2 *)encode_state->pic_param_ext->buffer;
 794     if (pic_param->picture_type == VAEncPictureTypeIntra) {
 795         allow_hwscore = false;
 796         kernel_shader = VME_INTRA_SHADER;
 797     } else {
 798         kernel_shader = VME_INTER_SHADER;
 799     }
 800
 801     if (allow_hwscore)
 802         gen7_vme_mpeg2_walker_fill_vme_batchbuffer(ctx,
 803                                                    encode_state,
 804                                                    width_in_mbs, height_in_mbs,
 805                                                    kernel_shader,
 806                                                    encoder_context);
 807     else
 808         gen8_vme_mpeg2_fill_vme_batchbuffer(ctx,
 809                                             encode_state,
 810                                             width_in_mbs, height_in_mbs,
 811                                             is_intra ? VME_INTRA_SHADER : VME_INTER_SHADER,
 812                                             0,
 813                                             encoder_context);
 814
 815     intel_batchbuffer_start_atomic(batch, 0x1000);
 816     gen8_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 817     BEGIN_BATCH(batch, 2);
 818     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (2 << 6));
 819     OUT_RELOC(batch,
 820               vme_context->vme_batchbuffer.bo,
 821               I915_GEM_DOMAIN_COMMAND, 0,
 822               0);
 823     ADVANCE_BATCH(batch);
 824
 825     intel_batchbuffer_end_atomic(batch);
 826 }
 827
 828 static VAStatus
 829 gen8_vme_mpeg2_prepare(VADriverContextP ctx,
 830                        struct encode_state *encode_state,
 831                        struct intel_encoder_context *encoder_context)
 832 {
 833     VAStatus vaStatus = VA_STATUS_SUCCESS;
 834     VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[0]->buffer;
 835     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 836     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 837
 838     if ((!vme_context->mpeg2_level) ||
 839         (vme_context->mpeg2_level != (seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK))) {
 840         vme_context->mpeg2_level = seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK;
 841     }
 842
 843
 844     /*Setup all the memory object*/
 845     gen8_vme_mpeg2_surface_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 846     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
 847     //gen8_vme_vme_state_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 848     intel_vme_mpeg2_state_setup(ctx, encode_state, encoder_context);
 849     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 850
 851     /*Programing media pipeline*/
 852     gen8_vme_mpeg2_pipeline_programing(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 853
 854     return vaStatus;
 855 }
 856
 857 static VAStatus
 858 gen8_vme_mpeg2_pipeline(VADriverContextP ctx,
 859                         VAProfile profile,
 860                         struct encode_state *encode_state,
 861                         struct intel_encoder_context *encoder_context)
 862 {
 863     gen8_vme_media_init(ctx, encoder_context);
 864     gen8_vme_mpeg2_prepare(ctx, encode_state, encoder_context);
 865     gen8_vme_run(ctx, encode_state, encoder_context);
 866     gen8_vme_stop(ctx, encode_state, encoder_context);
 867
 868     return VA_STATUS_SUCCESS;
 869 }
 870
 871 static void
 872 gen8_vme_context_destroy(void *context)
 873 {
 874     struct gen6_vme_context *vme_context = context;
 875
 876     i965_gpe_context_destroy(&vme_context->gpe_context);
 877
 878     dri_bo_unreference(vme_context->vme_output.bo);
 879     vme_context->vme_output.bo = NULL;
 880
 881     dri_bo_unreference(vme_context->vme_state.bo);
 882     vme_context->vme_state.bo = NULL;
 883
 884     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 885     vme_context->vme_batchbuffer.bo = NULL;
 886
 887     if (vme_context->vme_state_message) {
 888         free(vme_context->vme_state_message);
 889         vme_context->vme_state_message = NULL;
 890     }
 891
 892     free(vme_context);
 893 }
 894
 895 Bool gen8_vme_context_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 896 {
 897     struct gen6_vme_context *vme_context = calloc(1, sizeof(struct gen6_vme_context));
 898     struct i965_kernel *vme_kernel_list = NULL;
 899     int i965_kernel_num;
 900
 901     switch (encoder_context->codec) {
 902     case CODEC_H264:
 903         vme_kernel_list = gen8_vme_kernels;
 904         encoder_context->vme_pipeline = gen8_vme_pipeline;
 905         i965_kernel_num = sizeof(gen8_vme_kernels) / sizeof(struct i965_kernel);
 906         break;
 907
 908     case CODEC_MPEG2:
 909         vme_kernel_list = gen8_vme_mpeg2_kernels;
 910         encoder_context->vme_pipeline = gen8_vme_mpeg2_pipeline;
 911         i965_kernel_num = sizeof(gen8_vme_mpeg2_kernels) / sizeof(struct i965_kernel);
 912
 913         break;
 914
 915     default:
 916         /* never get here */
 917         assert(0);
 918
 919         break;
 920     }
 921     vme_context->vme_kernel_sum = i965_kernel_num;
 922     vme_context->gpe_context.surface_state_binding_table.length = (SURFACE_STATE_PADDED_SIZE + sizeof(unsigned int)) * MAX_MEDIA_SURFACES_GEN6;
 923
 924     vme_context->gpe_context.idrt.max_entries = MAX_INTERFACE_DESC_GEN6;
 925     vme_context->gpe_context.idrt.entry_size = sizeof(struct gen8_interface_descriptor_data);
 926
 927     vme_context->gpe_context.curbe.length = CURBE_TOTAL_DATA_LENGTH;
 928
 929     vme_context->gpe_context.vfe_state.max_num_threads = 60 - 1;
 930     vme_context->gpe_context.vfe_state.num_urb_entries = 16;
 931     vme_context->gpe_context.vfe_state.gpgpu_mode = 0;
 932     vme_context->gpe_context.vfe_state.urb_entry_size = 59 - 1;
 933     vme_context->gpe_context.vfe_state.curbe_allocation_size = CURBE_ALLOCATION_SIZE - 1;
 934
 935     gen7_vme_scoreboard_init(ctx, vme_context);
 936
 937     i965_gpe_load_kernels(ctx,
 938                           &vme_context->gpe_context,
 939                           vme_kernel_list,
 940                           i965_kernel_num);
 941     vme_context->vme_surface2_setup = gen8_gpe_surface2_setup;
 942     vme_context->vme_media_rw_surface_setup = gen8_gpe_media_rw_surface_setup;
 943     vme_context->vme_buffer_suface_setup = gen8_gpe_buffer_suface_setup;
 944     vme_context->vme_media_chroma_surface_setup = gen8_gpe_media_chroma_surface_setup;
 945
 946     encoder_context->vme_context = vme_context;
 947     encoder_context->vme_context_destroy = gen8_vme_context_destroy;
 948
 949     vme_context->vme_state_message = malloc(VME_MSG_LENGTH * sizeof(int));
 950
 951     return True;
 952 }