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
 244     /*Setup surfaces state*/
 245     /* current picture for encoding */
 246     obj_surface = encode_state->input_yuv_object;
 247     gen8_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 248     gen8_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 249     gen8_vme_media_chroma_source_surface_state(ctx, 6, obj_surface, encoder_context);
 250
 251     if (!is_intra) {
 252         VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 253         int slice_type;
 254
 255         slice_type = intel_avc_enc_slice_type_fixup(slice_param->slice_type);
 256         assert(slice_type != SLICE_TYPE_I && slice_type != SLICE_TYPE_SI);
 257
 258         intel_avc_vme_reference_state(ctx, encode_state, encoder_context, 0, 1, gen8_vme_source_surface_state);
 259
 260         if (slice_type == SLICE_TYPE_B)
 261             intel_avc_vme_reference_state(ctx, encode_state, encoder_context, 1, 2, gen8_vme_source_surface_state);
 262     }
 263
 264     /* VME output */
 265     gen8_vme_output_buffer_setup(ctx, encode_state, 3, encoder_context);
 266     gen8_vme_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 267
 268     return VA_STATUS_SUCCESS;
 269 }
 270
 271 static VAStatus gen8_vme_interface_setup(VADriverContextP ctx,
 272                                          struct encode_state *encode_state,
 273                                          struct intel_encoder_context *encoder_context)
 274 {
 275     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 276     struct gen8_interface_descriptor_data *desc;
 277     int i;
 278     dri_bo *bo;
 279
 280     bo = vme_context->gpe_context.idrt.bo;
 281     dri_bo_map(bo, 1);
 282     assert(bo->virtual);
 283     desc = bo->virtual;
 284
 285     for (i = 0; i < vme_context->vme_kernel_sum; i++) {
 286         struct i965_kernel *kernel;
 287         kernel = &vme_context->gpe_context.kernels[i];
 288         assert(sizeof(*desc) == 32);
 289         /*Setup the descritor table*/
 290         memset(desc, 0, sizeof(*desc));
 291         desc->desc0.kernel_start_pointer = (kernel->bo->offset >> 6);
 292         desc->desc3.sampler_count = 0; /* FIXME: */
 293         desc->desc3.sampler_state_pointer = 0;
 294         desc->desc4.binding_table_entry_count = 1; /* FIXME: */
 295         desc->desc4.binding_table_pointer = (BINDING_TABLE_OFFSET(0) >> 5);
 296         desc->desc5.constant_urb_entry_read_offset = 0;
 297         desc->desc5.constant_urb_entry_read_length = CURBE_URB_ENTRY_LENGTH;
 298
 299
 300         /*kernel start*/
 301         dri_bo_emit_reloc(bo,
 302                           I915_GEM_DOMAIN_INSTRUCTION, 0,
 303                           0,
 304                           i * sizeof(*desc) + offsetof(struct gen8_interface_descriptor_data, desc0),
 305                           kernel->bo);
 306         desc++;
 307     }
 308     dri_bo_unmap(bo);
 309
 310     return VA_STATUS_SUCCESS;
 311 }
 312
 313 static VAStatus gen8_vme_constant_setup(VADriverContextP ctx,
 314                                         struct encode_state *encode_state,
 315                                         struct intel_encoder_context *encoder_context)
 316 {
 317     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 318     unsigned char *constant_buffer;
 319     unsigned int *vme_state_message;
 320     int mv_num = 32;
 321
 322     vme_state_message = (unsigned int *)vme_context->vme_state_message;
 323
 324     if (encoder_context->codec == CODEC_H264) {
 325         if (vme_context->h264_level >= 30) {
 326             mv_num = 16;
 327
 328             if (vme_context->h264_level >= 31)
 329                 mv_num = 8;
 330         }
 331     } else if (encoder_context->codec == CODEC_MPEG2) {
 332         mv_num = 2;
 333     }
 334
 335     vme_state_message[31] = mv_num;
 336
 337     dri_bo_map(vme_context->gpe_context.curbe.bo, 1);
 338     assert(vme_context->gpe_context.curbe.bo->virtual);
 339     constant_buffer = vme_context->gpe_context.curbe.bo->virtual;
 340
 341     /* VME MV/Mb cost table is passed by using const buffer */
 342     /* Now it uses the fixed search path. So it is constructed directly
 343      * in the GPU shader.
 344      */
 345     memcpy(constant_buffer, (char *)vme_context->vme_state_message, 128);
 346
 347     dri_bo_unmap(vme_context->gpe_context.curbe.bo);
 348
 349     return VA_STATUS_SUCCESS;
 350 }
 351
 352
 353 static void
 354 gen8_vme_fill_vme_batchbuffer(VADriverContextP ctx,
 355                               struct encode_state *encode_state,
 356                               int mb_width, int mb_height,
 357                               int kernel,
 358                               int transform_8x8_mode_flag,
 359                               struct intel_encoder_context *encoder_context)
 360 {
 361     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 362     int mb_x = 0, mb_y = 0;
 363     int i, s;
 364     unsigned int *command_ptr;
 365
 366     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 367     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 368
 369     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 370         VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 371         int slice_mb_begin = pSliceParameter->macroblock_address;
 372         int slice_mb_number = pSliceParameter->num_macroblocks;
 373         unsigned int mb_intra_ub;
 374         int slice_mb_x = pSliceParameter->macroblock_address % mb_width;
 375         for (i = 0; i < slice_mb_number;  ) {
 376             int mb_count = i + slice_mb_begin;
 377             mb_x = mb_count % mb_width;
 378             mb_y = mb_count / mb_width;
 379             mb_intra_ub = 0;
 380             if (mb_x != 0) {
 381                 mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 382             }
 383             if (mb_y != 0) {
 384                 mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 385                 if (mb_x != 0)
 386                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 387                 if (mb_x != (mb_width -1))
 388                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 389             }
 390             if (i < mb_width) {
 391                 if (i == 0)
 392                     mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_AE);
 393                 mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_BCD_MASK);
 394                 if ((i == (mb_width - 1)) && slice_mb_x) {
 395                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 396                 }
 397             }
 398
 399             if ((i == mb_width) && slice_mb_x) {
 400                 mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_D);
 401             }
 402             *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 403             *command_ptr++ = kernel;
 404             *command_ptr++ = 0;
 405             *command_ptr++ = 0;
 406             *command_ptr++ = 0;
 407             *command_ptr++ = 0;
 408
 409             /*inline data */
 410             *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 411             *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 412
 413             i += 1;
 414         }
 415     }
 416
 417     *command_ptr++ = 0;
 418     *command_ptr++ = MI_BATCH_BUFFER_END;
 419
 420     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 421 }
 422
 423 static void gen8_vme_media_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 424 {
 425     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 426
 427     i965_gpe_context_init(ctx, &vme_context->gpe_context);
 428
 429     /* VME output buffer */
 430     dri_bo_unreference(vme_context->vme_output.bo);
 431     vme_context->vme_output.bo = NULL;
 432
 433     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 434     vme_context->vme_batchbuffer.bo = NULL;
 435
 436     /* VME state */
 437     dri_bo_unreference(vme_context->vme_state.bo);
 438     vme_context->vme_state.bo = NULL;
 439 }
 440
 441 static void gen8_vme_pipeline_programing(VADriverContextP ctx,
 442                                          struct encode_state *encode_state,
 443                                          struct intel_encoder_context *encoder_context)
 444 {
 445     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 446     struct intel_batchbuffer *batch = encoder_context->base.batch;
 447     VAEncPictureParameterBufferH264 *pPicParameter = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
 448     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 449     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 450     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 451     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 452     int kernel_shader;
 453     bool allow_hwscore = true;
 454     int s;
 455
 456     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 457         pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 458         if ((pSliceParameter->macroblock_address % width_in_mbs)) {
 459             allow_hwscore = false;
 460             break;
 461         }
 462     }
 463     if ((pSliceParameter->slice_type == SLICE_TYPE_I) ||
 464         (pSliceParameter->slice_type == SLICE_TYPE_I)) {
 465         kernel_shader = VME_INTRA_SHADER;
 466     } else if ((pSliceParameter->slice_type == SLICE_TYPE_P) ||
 467                (pSliceParameter->slice_type == SLICE_TYPE_SP)) {
 468         kernel_shader = VME_INTER_SHADER;
 469     } else {
 470         kernel_shader = VME_BINTER_SHADER;
 471         if (!allow_hwscore)
 472             kernel_shader = VME_INTER_SHADER;
 473     }
 474     if (allow_hwscore)
 475         gen7_vme_walker_fill_vme_batchbuffer(ctx,
 476                                              encode_state,
 477                                              width_in_mbs, height_in_mbs,
 478                                              kernel_shader,
 479                                              pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 480                                              encoder_context);
 481     else
 482         gen8_vme_fill_vme_batchbuffer(ctx,
 483                                       encode_state,
 484                                       width_in_mbs, height_in_mbs,
 485                                       kernel_shader,
 486                                       pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 487                                       encoder_context);
 488
 489     intel_batchbuffer_start_atomic(batch, 0x1000);
 490     gen8_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 491     BEGIN_BATCH(batch, 3);
 492     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (1 << 8) | (1 << 0));
 493     OUT_RELOC(batch,
 494               vme_context->vme_batchbuffer.bo,
 495               I915_GEM_DOMAIN_COMMAND, 0,
 496               0);
 497     OUT_BATCH(batch, 0);
 498     ADVANCE_BATCH(batch);
 499
 500     intel_batchbuffer_end_atomic(batch);
 501 }
 502
 503 static VAStatus gen8_vme_prepare(VADriverContextP ctx,
 504                                  struct encode_state *encode_state,
 505                                  struct intel_encoder_context *encoder_context)
 506 {
 507     VAStatus vaStatus = VA_STATUS_SUCCESS;
 508     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 509     int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
 510     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 511     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 512
 513     if (!vme_context->h264_level ||
 514         (vme_context->h264_level != pSequenceParameter->level_idc)) {
 515         vme_context->h264_level = pSequenceParameter->level_idc;
 516     }
 517
 518     intel_vme_update_mbmv_cost(ctx, encode_state, encoder_context);
 519
 520     /*Setup all the memory object*/
 521     gen8_vme_surface_setup(ctx, encode_state, is_intra, encoder_context);
 522     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
 523     //gen8_vme_vme_state_setup(ctx, encode_state, is_intra, encoder_context);
 524     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 525
 526     /*Programing media pipeline*/
 527     gen8_vme_pipeline_programing(ctx, encode_state, encoder_context);
 528
 529     return vaStatus;
 530 }
 531
 532 static VAStatus gen8_vme_run(VADriverContextP ctx,
 533                              struct encode_state *encode_state,
 534                              struct intel_encoder_context *encoder_context)
 535 {
 536     struct intel_batchbuffer *batch = encoder_context->base.batch;
 537
 538     intel_batchbuffer_flush(batch);
 539
 540     return VA_STATUS_SUCCESS;
 541 }
 542
 543 static VAStatus gen8_vme_stop(VADriverContextP ctx,
 544                               struct encode_state *encode_state,
 545                               struct intel_encoder_context *encoder_context)
 546 {
 547     return VA_STATUS_SUCCESS;
 548 }
 549
 550 static VAStatus
 551 gen8_vme_pipeline(VADriverContextP ctx,
 552                   VAProfile profile,
 553                   struct encode_state *encode_state,
 554                   struct intel_encoder_context *encoder_context)
 555 {
 556     gen8_vme_media_init(ctx, encoder_context);
 557     gen8_vme_prepare(ctx, encode_state, encoder_context);
 558     gen8_vme_run(ctx, encode_state, encoder_context);
 559     gen8_vme_stop(ctx, encode_state, encoder_context);
 560
 561     return VA_STATUS_SUCCESS;
 562 }
 563
 564 static void
 565 gen8_vme_mpeg2_output_buffer_setup(VADriverContextP ctx,
 566                                    struct encode_state *encode_state,
 567                                    int index,
 568                                    int is_intra,
 569                                    struct intel_encoder_context *encoder_context)
 570
 571 {
 572     struct i965_driver_data *i965 = i965_driver_data(ctx);
 573     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 574     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 575     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 576     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 577
 578     vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
 579     vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
 580
 581     if (is_intra)
 582         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 2;
 583     else
 584         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 24;
 585     /*
 586      * Inter MV . 32-byte Intra search + 16 IME info + 128 IME MV + 32 IME Ref
 587      * + 16 FBR Info + 128 FBR MV + 32 FBR Ref.
 588      * 16 * (2 + 2 * (1 + 8 + 2))= 16 * 24.
 589      */
 590
 591     vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
 592                                               "VME output buffer",
 593                                               vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
 594                                               0x1000);
 595     assert(vme_context->vme_output.bo);
 596     vme_context->vme_buffer_suface_setup(ctx,
 597                                          &vme_context->gpe_context,
 598                                          &vme_context->vme_output,
 599                                          BINDING_TABLE_OFFSET(index),
 600                                          SURFACE_STATE_OFFSET(index));
 601 }
 602
 603 static void
 604 gen8_vme_mpeg2_output_vme_batchbuffer_setup(VADriverContextP ctx,
 605                                             struct encode_state *encode_state,
 606                                             int index,
 607                                             struct intel_encoder_context *encoder_context)
 608
 609 {
 610     struct i965_driver_data *i965 = i965_driver_data(ctx);
 611     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 612     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 613     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 614     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 615
 616     vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
 617     vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
 618     vme_context->vme_batchbuffer.pitch = 16;
 619     vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
 620                                                    "VME batchbuffer",
 621                                                    vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
 622                                                    0x1000);
 623     vme_context->vme_buffer_suface_setup(ctx,
 624                                          &vme_context->gpe_context,
 625                                          &vme_context->vme_batchbuffer,
 626                                          BINDING_TABLE_OFFSET(index),
 627                                          SURFACE_STATE_OFFSET(index));
 628 }
 629
 630 static VAStatus
 631 gen8_vme_mpeg2_surface_setup(VADriverContextP ctx,
 632                              struct encode_state *encode_state,
 633                              int is_intra,
 634                              struct intel_encoder_context *encoder_context)
 635 {
 636     struct object_surface *obj_surface;
 637
 638     /*Setup surfaces state*/
 639     /* current picture for encoding */
 640     obj_surface = encode_state->input_yuv_object;
 641     gen8_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 642     gen8_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 643     gen8_vme_media_chroma_source_surface_state(ctx, 6, obj_surface, encoder_context);
 644
 645     if (!is_intra) {
 646         /* reference 0 */
 647         obj_surface = encode_state->reference_objects[0];
 648
 649         if (obj_surface->bo != NULL)
 650             gen8_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
 651
 652         /* reference 1 */
 653         obj_surface = encode_state->reference_objects[1];
 654
 655         if (obj_surface && obj_surface->bo != NULL)
 656             gen8_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
 657     }
 658
 659     /* VME output */
 660     gen8_vme_mpeg2_output_buffer_setup(ctx, encode_state, 3, is_intra, encoder_context);
 661     gen8_vme_mpeg2_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 662
 663     return VA_STATUS_SUCCESS;
 664 }
 665
 666 static void
 667 gen8_vme_mpeg2_fill_vme_batchbuffer(VADriverContextP ctx,
 668                                     struct encode_state *encode_state,
 669                                     int mb_width, int mb_height,
 670                                     int kernel,
 671                                     int transform_8x8_mode_flag,
 672                                     struct intel_encoder_context *encoder_context)
 673 {
 674     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 675     int mb_x = 0, mb_y = 0;
 676     int i, s, j;
 677     unsigned int *command_ptr;
 678
 679
 680     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 681     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 682
 683     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 684         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 685
 686         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 687             int slice_mb_begin = slice_param->macroblock_address;
 688             int slice_mb_number = slice_param->num_macroblocks;
 689             unsigned int mb_intra_ub;
 690
 691             for (i = 0; i < slice_mb_number;) {
 692                 int mb_count = i + slice_mb_begin;
 693
 694                 mb_x = mb_count % mb_width;
 695                 mb_y = mb_count / mb_width;
 696                 mb_intra_ub = 0;
 697
 698                 if (mb_x != 0) {
 699                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 700                 }
 701
 702                 if (mb_y != 0) {
 703                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 704
 705                     if (mb_x != 0)
 706                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 707
 708                     if (mb_x != (mb_width -1))
 709                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 710                 }
 711
 712                 *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 713                 *command_ptr++ = kernel;
 714                 *command_ptr++ = 0;
 715                 *command_ptr++ = 0;
 716                 *command_ptr++ = 0;
 717                 *command_ptr++ = 0;
 718
 719                 /*inline data */
 720                 *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 721                 *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 722
 723                 i += 1;
 724             }
 725
 726             slice_param++;
 727         }
 728     }
 729
 730     *command_ptr++ = 0;
 731     *command_ptr++ = MI_BATCH_BUFFER_END;
 732
 733     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 734 }
 735
 736 static void
 737 gen8_vme_mpeg2_pipeline_programing(VADriverContextP ctx,
 738                                    struct encode_state *encode_state,
 739                                    int is_intra,
 740                                    struct intel_encoder_context *encoder_context)
 741 {
 742     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 743     struct intel_batchbuffer *batch = encoder_context->base.batch;
 744     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 745     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 746     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 747     bool allow_hwscore = true;
 748     int s;
 749     int kernel_shader;
 750     VAEncPictureParameterBufferMPEG2 *pic_param = NULL;
 751
 752     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 753         int j;
 754         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 755
 756         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 757             if (slice_param->macroblock_address % width_in_mbs) {
 758                 allow_hwscore = false;
 759                 break;
 760             }
 761         }
 762     }
 763
 764     pic_param = (VAEncPictureParameterBufferMPEG2 *)encode_state->pic_param_ext->buffer;
 765     if (pic_param->picture_type == VAEncPictureTypeIntra) {
 766         allow_hwscore = false;
 767         kernel_shader = VME_INTRA_SHADER;
 768     } else {
 769         kernel_shader = VME_INTER_SHADER;
 770     }
 771
 772     if (allow_hwscore)
 773         gen7_vme_mpeg2_walker_fill_vme_batchbuffer(ctx,
 774                                                    encode_state,
 775                                                    width_in_mbs, height_in_mbs,
 776                                                    kernel_shader,
 777                                                    encoder_context);
 778     else
 779         gen8_vme_mpeg2_fill_vme_batchbuffer(ctx,
 780                                             encode_state,
 781                                             width_in_mbs, height_in_mbs,
 782                                             is_intra ? VME_INTRA_SHADER : VME_INTER_SHADER,
 783                                             0,
 784                                             encoder_context);
 785
 786     intel_batchbuffer_start_atomic(batch, 0x1000);
 787     gen8_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 788     BEGIN_BATCH(batch, 2);
 789     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (2 << 6));
 790     OUT_RELOC(batch,
 791               vme_context->vme_batchbuffer.bo,
 792               I915_GEM_DOMAIN_COMMAND, 0,
 793               0);
 794     ADVANCE_BATCH(batch);
 795
 796     intel_batchbuffer_end_atomic(batch);
 797 }
 798
 799 static VAStatus
 800 gen8_vme_mpeg2_prepare(VADriverContextP ctx,
 801                        struct encode_state *encode_state,
 802                        struct intel_encoder_context *encoder_context)
 803 {
 804     VAStatus vaStatus = VA_STATUS_SUCCESS;
 805     VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[0]->buffer;
 806     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 807     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 808
 809     if ((!vme_context->mpeg2_level) ||
 810         (vme_context->mpeg2_level != (seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK))) {
 811         vme_context->mpeg2_level = seq_param->sequence_extension.bits.profile_and_level_indication & MPEG2_LEVEL_MASK;
 812     }
 813
 814
 815     /*Setup all the memory object*/
 816     gen8_vme_mpeg2_surface_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 817     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
 818     //gen8_vme_vme_state_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 819     intel_vme_mpeg2_state_setup(ctx, encode_state, encoder_context);
 820     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 821
 822     /*Programing media pipeline*/
 823     gen8_vme_mpeg2_pipeline_programing(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 824
 825     return vaStatus;
 826 }
 827
 828 static VAStatus
 829 gen8_vme_mpeg2_pipeline(VADriverContextP ctx,
 830                         VAProfile profile,
 831                         struct encode_state *encode_state,
 832                         struct intel_encoder_context *encoder_context)
 833 {
 834     gen8_vme_media_init(ctx, encoder_context);
 835     gen8_vme_mpeg2_prepare(ctx, encode_state, encoder_context);
 836     gen8_vme_run(ctx, encode_state, encoder_context);
 837     gen8_vme_stop(ctx, encode_state, encoder_context);
 838
 839     return VA_STATUS_SUCCESS;
 840 }
 841
 842 static void
 843 gen8_vme_context_destroy(void *context)
 844 {
 845     struct gen6_vme_context *vme_context = context;
 846
 847     i965_gpe_context_destroy(&vme_context->gpe_context);
 848
 849     dri_bo_unreference(vme_context->vme_output.bo);
 850     vme_context->vme_output.bo = NULL;
 851
 852     dri_bo_unreference(vme_context->vme_state.bo);
 853     vme_context->vme_state.bo = NULL;
 854
 855     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 856     vme_context->vme_batchbuffer.bo = NULL;
 857
 858     if (vme_context->vme_state_message) {
 859         free(vme_context->vme_state_message);
 860         vme_context->vme_state_message = NULL;
 861     }
 862
 863     free(vme_context);
 864 }
 865
 866 Bool gen8_vme_context_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 867 {
 868     struct gen6_vme_context *vme_context = calloc(1, sizeof(struct gen6_vme_context));
 869     struct i965_kernel *vme_kernel_list = NULL;
 870     int i965_kernel_num;
 871
 872     switch (encoder_context->codec) {
 873     case CODEC_H264:
 874         vme_kernel_list = gen8_vme_kernels;
 875         encoder_context->vme_pipeline = gen8_vme_pipeline;
 876         i965_kernel_num = sizeof(gen8_vme_kernels) / sizeof(struct i965_kernel);
 877         break;
 878
 879     case CODEC_MPEG2:
 880         vme_kernel_list = gen8_vme_mpeg2_kernels;
 881         encoder_context->vme_pipeline = gen8_vme_mpeg2_pipeline;
 882         i965_kernel_num = sizeof(gen8_vme_mpeg2_kernels) / sizeof(struct i965_kernel);
 883
 884         break;
 885
 886     default:
 887         /* never get here */
 888         assert(0);
 889
 890         break;
 891     }
 892     vme_context->vme_kernel_sum = i965_kernel_num;
 893     vme_context->gpe_context.surface_state_binding_table.length = (SURFACE_STATE_PADDED_SIZE + sizeof(unsigned int)) * MAX_MEDIA_SURFACES_GEN6;
 894
 895     vme_context->gpe_context.idrt.max_entries = MAX_INTERFACE_DESC_GEN6;
 896     vme_context->gpe_context.idrt.entry_size = sizeof(struct gen8_interface_descriptor_data);
 897
 898     vme_context->gpe_context.curbe.length = CURBE_TOTAL_DATA_LENGTH;
 899
 900     vme_context->gpe_context.vfe_state.max_num_threads = 60 - 1;
 901     vme_context->gpe_context.vfe_state.num_urb_entries = 16;
 902     vme_context->gpe_context.vfe_state.gpgpu_mode = 0;
 903     vme_context->gpe_context.vfe_state.urb_entry_size = 59 - 1;
 904     vme_context->gpe_context.vfe_state.curbe_allocation_size = CURBE_ALLOCATION_SIZE - 1;
 905
 906     gen7_vme_scoreboard_init(ctx, vme_context);
 907
 908     i965_gpe_load_kernels(ctx,
 909                           &vme_context->gpe_context,
 910                           vme_kernel_list,
 911                           i965_kernel_num);
 912     vme_context->vme_surface2_setup = gen8_gpe_surface2_setup;
 913     vme_context->vme_media_rw_surface_setup = gen8_gpe_media_rw_surface_setup;
 914     vme_context->vme_buffer_suface_setup = gen8_gpe_buffer_suface_setup;
 915     vme_context->vme_media_chroma_surface_setup = gen8_gpe_media_chroma_surface_setup;
 916
 917     encoder_context->vme_context = vme_context;
 918     encoder_context->vme_context_destroy = gen8_vme_context_destroy;
 919
 920     vme_context->vme_state_message = malloc(VME_MSG_LENGTH * sizeof(int));
 921
 922     return True;
 923 }