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               MAX(SURFACE_STATE_PADDED_SIZE_GEN6, SURFACE_STATE_PADDED_SIZE_GEN7)
  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       3
  55 #define VME_BATCHBUFFER         2
  56
  57 #define CURBE_ALLOCATION_SIZE   37              /* in 256-bit */
  58 #define CURBE_TOTAL_DATA_LENGTH (4 * 32)        /* in byte, it should be less than or equal to CURBE_ALLOCATION_SIZE * 32 */
  59 #define CURBE_URB_ENTRY_LENGTH  4               /* in 256-bit, it should be less than or equal to CURBE_TOTAL_DATA_LENGTH / 32 */
  60
  61 #define VME_MSG_LENGTH          32
  62
  63 static const uint32_t gen8_vme_intra_frame[][4] = {
  64 #include "shaders/vme/intra_frame_haswell.g75b"
  65 };
  66
  67 static const uint32_t gen8_vme_inter_frame[][4] = {
  68 #include "shaders/vme/inter_frame_haswell.g75b"
  69 };
  70
  71 static const uint32_t gen8_vme_inter_bframe[][4] = {
  72 #include "shaders/vme/inter_bframe_haswell.g75b"
  73 };
  74
  75 static const uint32_t gen8_vme_batchbuffer[][4] = {
  76 #include "shaders/vme/batchbuffer.g75b"
  77 };
  78
  79 static struct i965_kernel gen8_vme_kernels[] = {
  80     {
  81         "VME Intra Frame",
  82         VME_INTRA_SHADER, /*index*/
  83         gen8_vme_intra_frame,
  84         sizeof(gen8_vme_intra_frame),
  85         NULL
  86     },
  87     {
  88         "VME inter Frame",
  89         VME_INTER_SHADER,
  90         gen8_vme_inter_frame,
  91         sizeof(gen8_vme_inter_frame),
  92         NULL
  93     },
  94     {
  95         "VME BATCHBUFFER",
  96         VME_BATCHBUFFER,
  97         gen8_vme_batchbuffer,
  98         sizeof(gen8_vme_batchbuffer),
  99         NULL
 100     },
 101     {
 102         "VME inter BFrame",
 103         VME_BINTER_SHADER,
 104         gen8_vme_inter_bframe,
 105         sizeof(gen8_vme_inter_bframe),
 106         NULL
 107     }
 108 };
 109
 110 static const uint32_t gen8_vme_mpeg2_intra_frame[][4] = {
 111 #include "shaders/vme/intra_frame_haswell.g75b"
 112 };
 113
 114 static const uint32_t gen8_vme_mpeg2_inter_frame[][4] = {
 115 #include "shaders/vme/mpeg2_inter_haswell.g75b"
 116 };
 117
 118 static const uint32_t gen8_vme_mpeg2_batchbuffer[][4] = {
 119 #include "shaders/vme/batchbuffer.g75b"
 120 };
 121
 122 static struct i965_kernel gen8_vme_mpeg2_kernels[] = {
 123     {
 124         "VME Intra Frame",
 125         VME_INTRA_SHADER, /*index*/
 126         gen8_vme_mpeg2_intra_frame,
 127         sizeof(gen8_vme_mpeg2_intra_frame),
 128         NULL
 129     },
 130     {
 131         "VME inter Frame",
 132         VME_INTER_SHADER,
 133         gen8_vme_mpeg2_inter_frame,
 134         sizeof(gen8_vme_mpeg2_inter_frame),
 135         NULL
 136     },
 137     {
 138         "VME BATCHBUFFER",
 139         VME_BATCHBUFFER,
 140         gen8_vme_mpeg2_batchbuffer,
 141         sizeof(gen8_vme_mpeg2_batchbuffer),
 142         NULL
 143     },
 144 };
 145
 146 /* only used for VME source surface state */
 147 static void
 148 gen8_vme_source_surface_state(VADriverContextP ctx,
 149                                int index,
 150                                struct object_surface *obj_surface,
 151                                struct intel_encoder_context *encoder_context)
 152 {
 153     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 154
 155     vme_context->vme_surface2_setup(ctx,
 156                                     &vme_context->gpe_context,
 157                                     obj_surface,
 158                                     BINDING_TABLE_OFFSET(index),
 159                                     SURFACE_STATE_OFFSET(index));
 160 }
 161
 162 static void
 163 gen8_vme_media_source_surface_state(VADriverContextP ctx,
 164                                      int index,
 165                                      struct object_surface *obj_surface,
 166                                      struct intel_encoder_context *encoder_context)
 167 {
 168     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 169
 170     vme_context->vme_media_rw_surface_setup(ctx,
 171                                             &vme_context->gpe_context,
 172                                             obj_surface,
 173                                             BINDING_TABLE_OFFSET(index),
 174                                             SURFACE_STATE_OFFSET(index));
 175 }
 176
 177 static void
 178 gen8_vme_media_chroma_source_surface_state(VADriverContextP ctx,
 179                                             int index,
 180                                             struct object_surface *obj_surface,
 181                                             struct intel_encoder_context *encoder_context)
 182 {
 183     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 184
 185     vme_context->vme_media_chroma_surface_setup(ctx,
 186                                                 &vme_context->gpe_context,
 187                                                 obj_surface,
 188                                                 BINDING_TABLE_OFFSET(index),
 189                                                 SURFACE_STATE_OFFSET(index));
 190 }
 191
 192 static void
 193 gen8_vme_output_buffer_setup(VADriverContextP ctx,
 194                               struct encode_state *encode_state,
 195                               int index,
 196                               struct intel_encoder_context *encoder_context)
 197
 198 {
 199     struct i965_driver_data *i965 = i965_driver_data(ctx);
 200     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 201     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 202     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 203     int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
 204     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 205     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 206
 207     vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
 208     vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
 209
 210     if (is_intra)
 211         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 2;
 212     else
 213         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 24;
 214     /*
 215      * Inter MV . 32-byte Intra search + 16 IME info + 128 IME MV + 32 IME Ref
 216      * + 16 FBR Info + 128 FBR MV + 32 FBR Ref.
 217      * 16 * (2 + 2 * (1 + 8 + 2))= 16 * 24.
 218      */
 219
 220     vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
 221                                               "VME output buffer",
 222                                               vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
 223                                               0x1000);
 224     assert(vme_context->vme_output.bo);
 225     vme_context->vme_buffer_suface_setup(ctx,
 226                                          &vme_context->gpe_context,
 227                                          &vme_context->vme_output,
 228                                          BINDING_TABLE_OFFSET(index),
 229                                          SURFACE_STATE_OFFSET(index));
 230 }
 231
 232 static void
 233 gen8_vme_output_vme_batchbuffer_setup(VADriverContextP ctx,
 234                                        struct encode_state *encode_state,
 235                                        int index,
 236                                        struct intel_encoder_context *encoder_context)
 237
 238 {
 239     struct i965_driver_data *i965 = i965_driver_data(ctx);
 240     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 241     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 242     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 243     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 244
 245     vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
 246     vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
 247     vme_context->vme_batchbuffer.pitch = 16;
 248     vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
 249                                                    "VME batchbuffer",
 250                                                    vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
 251                                                    0x1000);
 252     vme_context->vme_buffer_suface_setup(ctx,
 253                                          &vme_context->gpe_context,
 254                                          &vme_context->vme_batchbuffer,
 255                                          BINDING_TABLE_OFFSET(index),
 256                                          SURFACE_STATE_OFFSET(index));
 257 }
 258
 259 static VAStatus
 260 gen8_vme_surface_setup(VADriverContextP ctx,
 261                         struct encode_state *encode_state,
 262                         int is_intra,
 263                         struct intel_encoder_context *encoder_context)
 264 {
 265     struct object_surface *obj_surface;
 266
 267     /*Setup surfaces state*/
 268     /* current picture for encoding */
 269     obj_surface = encode_state->input_yuv_object;
 270     gen8_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 271     gen8_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 272     gen8_vme_media_chroma_source_surface_state(ctx, 6, obj_surface, encoder_context);
 273
 274     if (!is_intra) {
 275         /* reference 0 */
 276         obj_surface = encode_state->reference_objects[0];
 277
 278         if (obj_surface && obj_surface->bo)
 279             gen8_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
 280
 281         /* reference 1 */
 282         obj_surface = encode_state->reference_objects[1];
 283
 284         if (obj_surface && obj_surface->bo)
 285             gen8_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
 286     }
 287
 288     /* VME output */
 289     gen8_vme_output_buffer_setup(ctx, encode_state, 3, encoder_context);
 290     gen8_vme_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 291
 292     return VA_STATUS_SUCCESS;
 293 }
 294
 295 static VAStatus gen8_vme_interface_setup(VADriverContextP ctx,
 296                                           struct encode_state *encode_state,
 297                                           struct intel_encoder_context *encoder_context)
 298 {
 299     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 300     struct gen6_interface_descriptor_data *desc;
 301     int i;
 302     dri_bo *bo;
 303
 304     bo = vme_context->gpe_context.idrt.bo;
 305     dri_bo_map(bo, 1);
 306     assert(bo->virtual);
 307     desc = bo->virtual;
 308
 309     for (i = 0; i < vme_context->vme_kernel_sum; i++) {
 310         struct i965_kernel *kernel;
 311         kernel = &vme_context->gpe_context.kernels[i];
 312         assert(sizeof(*desc) == 32);
 313         /*Setup the descritor table*/
 314         memset(desc, 0, sizeof(*desc));
 315         desc->desc0.kernel_start_pointer = (kernel->bo->offset >> 6);
 316         desc->desc2.sampler_count = 0; /* FIXME: */
 317         desc->desc2.sampler_state_pointer = 0;
 318         desc->desc3.binding_table_entry_count = 1; /* FIXME: */
 319         desc->desc3.binding_table_pointer = (BINDING_TABLE_OFFSET(0) >> 5);
 320         desc->desc4.constant_urb_entry_read_offset = 0;
 321         desc->desc4.constant_urb_entry_read_length = CURBE_URB_ENTRY_LENGTH;
 322
 323         /*kernel start*/
 324         dri_bo_emit_reloc(bo,
 325                           I915_GEM_DOMAIN_INSTRUCTION, 0,
 326                           0,
 327                           i * sizeof(*desc) + offsetof(struct gen6_interface_descriptor_data, desc0),
 328                           kernel->bo);
 329         desc++;
 330     }
 331     dri_bo_unmap(bo);
 332
 333     return VA_STATUS_SUCCESS;
 334 }
 335
 336 static VAStatus gen8_vme_constant_setup(VADriverContextP ctx,
 337                                          struct encode_state *encode_state,
 338                                          struct intel_encoder_context *encoder_context)
 339 {
 340     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 341     unsigned char *constant_buffer;
 342     unsigned int *vme_state_message;
 343     int mv_num = 32;
 344
 345     vme_state_message = (unsigned int *)vme_context->vme_state_message;
 346
 347     if (encoder_context->codec == CODEC_H264) {
 348         if (vme_context->h264_level >= 30) {
 349             mv_num = 16;
 350
 351             if (vme_context->h264_level >= 31)
 352                 mv_num = 8;
 353         }
 354     } else if (encoder_context->codec == CODEC_MPEG2) {
 355         mv_num = 2;
 356     }
 357
 358     vme_state_message[31] = mv_num;
 359
 360     dri_bo_map(vme_context->gpe_context.curbe.bo, 1);
 361     assert(vme_context->gpe_context.curbe.bo->virtual);
 362     constant_buffer = vme_context->gpe_context.curbe.bo->virtual;
 363
 364     /* VME MV/Mb cost table is passed by using const buffer */
 365     /* Now it uses the fixed search path. So it is constructed directly
 366      * in the GPU shader.
 367      */
 368     memcpy(constant_buffer, (char *)vme_context->vme_state_message, 128);
 369
 370     dri_bo_unmap(vme_context->gpe_context.curbe.bo);
 371
 372     return VA_STATUS_SUCCESS;
 373 }
 374
 375 static const unsigned int intra_mb_mode_cost_table[] = {
 376     0x31110001, // for qp0
 377     0x09110001, // for qp1
 378     0x15030001, // for qp2
 379     0x0b030001, // for qp3
 380     0x0d030011, // for qp4
 381     0x17210011, // for qp5
 382     0x41210011, // for qp6
 383     0x19210011, // for qp7
 384     0x25050003, // for qp8
 385     0x1b130003, // for qp9
 386     0x1d130003, // for qp10
 387     0x27070021, // for qp11
 388     0x51310021, // for qp12
 389     0x29090021, // for qp13
 390     0x35150005, // for qp14
 391     0x2b0b0013, // for qp15
 392     0x2d0d0013, // for qp16
 393     0x37170007, // for qp17
 394     0x61410031, // for qp18
 395     0x39190009, // for qp19
 396     0x45250015, // for qp20
 397     0x3b1b000b, // for qp21
 398     0x3d1d000d, // for qp22
 399     0x47270017, // for qp23
 400     0x71510041, // for qp24 ! center for qp=0..30
 401     0x49290019, // for qp25
 402     0x55350025, // for qp26
 403     0x4b2b001b, // for qp27
 404     0x4d2d001d, // for qp28
 405     0x57370027, // for qp29
 406     0x81610051, // for qp30
 407     0x57270017, // for qp31
 408     0x81510041, // for qp32 ! center for qp=31..51
 409     0x59290019, // for qp33
 410     0x65350025, // for qp34
 411     0x5b2b001b, // for qp35
 412     0x5d2d001d, // for qp36
 413     0x67370027, // for qp37
 414     0x91610051, // for qp38
 415     0x69390029, // for qp39
 416     0x75450035, // for qp40
 417     0x6b3b002b, // for qp41
 418     0x6d3d002d, // for qp42
 419     0x77470037, // for qp43
 420     0xa1710061, // for qp44
 421     0x79490039, // for qp45
 422     0x85550045, // for qp46
 423     0x7b4b003b, // for qp47
 424     0x7d4d003d, // for qp48
 425     0x87570047, // for qp49
 426     0xb1810071, // for qp50
 427     0x89590049  // for qp51
 428 };
 429
 430 static void gen8_vme_state_setup_fixup(VADriverContextP ctx,
 431                                         struct encode_state *encode_state,
 432                                         struct intel_encoder_context *encoder_context,
 433                                         unsigned int *vme_state_message)
 434 {
 435     struct gen6_mfc_context *mfc_context = encoder_context->mfc_context;
 436     VAEncPictureParameterBufferH264 *pic_param = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
 437     VAEncSliceParameterBufferH264 *slice_param = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 438
 439     if (slice_param->slice_type != SLICE_TYPE_I &&
 440         slice_param->slice_type != SLICE_TYPE_SI)
 441         return;
 442     if (encoder_context->rate_control_mode == VA_RC_CQP)
 443         vme_state_message[0] = intra_mb_mode_cost_table[pic_param->pic_init_qp + slice_param->slice_qp_delta];
 444     else
 445         vme_state_message[0] = intra_mb_mode_cost_table[mfc_context->bit_rate_control_context[slice_param->slice_type].QpPrimeY];
 446 }
 447
 448 static VAStatus gen8_vme_vme_state_setup(VADriverContextP ctx,
 449                                           struct encode_state *encode_state,
 450                                           int is_intra,
 451                                           struct intel_encoder_context *encoder_context)
 452 {
 453     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 454     unsigned int *vme_state_message;
 455     int i;
 456
 457     //pass the MV/Mb cost into VME message on HASWell
 458     assert(vme_context->vme_state_message);
 459     vme_state_message = (unsigned int *)vme_context->vme_state_message;
 460
 461     vme_state_message[0] = 0x4a4a4a4a;
 462     vme_state_message[1] = 0x4a4a4a4a;
 463     vme_state_message[2] = 0x4a4a4a4a;
 464     vme_state_message[3] = 0x22120200;
 465     vme_state_message[4] = 0x62524232;
 466
 467     for (i=5; i < 8; i++) {
 468         vme_state_message[i] = 0;
 469     }
 470
 471     switch (encoder_context->codec) {
 472     case CODEC_H264:
 473         gen8_vme_state_setup_fixup(ctx, encode_state, encoder_context, vme_state_message);
 474
 475         break;
 476
 477     default:
 478         /* no fixup */
 479         break;
 480     }
 481
 482     return VA_STATUS_SUCCESS;
 483 }
 484
 485
 486 static void
 487 gen8_vme_fill_vme_batchbuffer(VADriverContextP ctx,
 488                                struct encode_state *encode_state,
 489                                int mb_width, int mb_height,
 490                                int kernel,
 491                                int transform_8x8_mode_flag,
 492                                struct intel_encoder_context *encoder_context)
 493 {
 494     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 495     int mb_x = 0, mb_y = 0;
 496     int i, s;
 497     unsigned int *command_ptr;
 498
 499     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 500     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 501
 502     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 503         VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 504         int slice_mb_begin = pSliceParameter->macroblock_address;
 505         int slice_mb_number = pSliceParameter->num_macroblocks;
 506         unsigned int mb_intra_ub;
 507         int slice_mb_x = pSliceParameter->macroblock_address % mb_width;
 508         for (i = 0; i < slice_mb_number;  ) {
 509             int mb_count = i + slice_mb_begin;
 510             mb_x = mb_count % mb_width;
 511             mb_y = mb_count / mb_width;
 512             mb_intra_ub = 0;
 513             if (mb_x != 0) {
 514                 mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 515             }
 516             if (mb_y != 0) {
 517                 mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 518                 if (mb_x != 0)
 519                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 520                 if (mb_x != (mb_width -1))
 521                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 522             }
 523             if (i < mb_width) {
 524                 if (i == 0)
 525                     mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_AE);
 526                 mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_BCD_MASK);
 527                 if ((i == (mb_width - 1)) && slice_mb_x) {
 528                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 529                 }
 530             }
 531
 532             if ((i == mb_width) && slice_mb_x) {
 533                 mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_D);
 534             }
 535             *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 536             *command_ptr++ = kernel;
 537             *command_ptr++ = 0;
 538             *command_ptr++ = 0;
 539             *command_ptr++ = 0;
 540             *command_ptr++ = 0;
 541
 542             /*inline data */
 543             *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 544             *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 545
 546             i += 1;
 547         }
 548     }
 549
 550     *command_ptr++ = 0;
 551     *command_ptr++ = MI_BATCH_BUFFER_END;
 552
 553     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 554 }
 555
 556 static void gen8_vme_media_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 557 {
 558     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 559
 560     i965_gpe_context_init(ctx, &vme_context->gpe_context);
 561
 562     /* VME output buffer */
 563     dri_bo_unreference(vme_context->vme_output.bo);
 564     vme_context->vme_output.bo = NULL;
 565
 566     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 567     vme_context->vme_batchbuffer.bo = NULL;
 568
 569     /* VME state */
 570     dri_bo_unreference(vme_context->vme_state.bo);
 571     vme_context->vme_state.bo = NULL;
 572 }
 573
 574 static void gen8_vme_pipeline_programing(VADriverContextP ctx,
 575                                           struct encode_state *encode_state,
 576                                           struct intel_encoder_context *encoder_context)
 577 {
 578     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 579     struct intel_batchbuffer *batch = encoder_context->base.batch;
 580     VAEncPictureParameterBufferH264 *pPicParameter = (VAEncPictureParameterBufferH264 *)encode_state->pic_param_ext->buffer;
 581     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 582     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 583     int width_in_mbs = pSequenceParameter->picture_width_in_mbs;
 584     int height_in_mbs = pSequenceParameter->picture_height_in_mbs;
 585     int kernel_shader;
 586     bool allow_hwscore = true;
 587     int s;
 588
 589     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 590         pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[s]->buffer;
 591         if ((pSliceParameter->macroblock_address % width_in_mbs)) {
 592                 allow_hwscore = false;
 593                 break;
 594         }
 595     }
 596     if ((pSliceParameter->slice_type == SLICE_TYPE_I) ||
 597         (pSliceParameter->slice_type == SLICE_TYPE_I)) {
 598         kernel_shader = VME_INTRA_SHADER;
 599    } else if ((pSliceParameter->slice_type == SLICE_TYPE_P) ||
 600         (pSliceParameter->slice_type == SLICE_TYPE_SP)) {
 601         kernel_shader = VME_INTER_SHADER;
 602    } else {
 603         kernel_shader = VME_BINTER_SHADER;
 604         if (!allow_hwscore)
 605              kernel_shader = VME_INTER_SHADER;
 606    }
 607     if (allow_hwscore)
 608         gen7_vme_walker_fill_vme_batchbuffer(ctx,
 609                                   encode_state,
 610                                   width_in_mbs, height_in_mbs,
 611                                   kernel_shader,
 612                                   pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 613                                   encoder_context);
 614     else
 615         gen8_vme_fill_vme_batchbuffer(ctx,
 616                                    encode_state,
 617                                    width_in_mbs, height_in_mbs,
 618                                    kernel_shader,
 619                                    pPicParameter->pic_fields.bits.transform_8x8_mode_flag,
 620                                    encoder_context);
 621
 622     intel_batchbuffer_start_atomic(batch, 0x1000);
 623     gen6_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 624     BEGIN_BATCH(batch, 2);
 625     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (2 << 6));
 626     OUT_RELOC(batch,
 627               vme_context->vme_batchbuffer.bo,
 628               I915_GEM_DOMAIN_COMMAND, 0,
 629               0);
 630     ADVANCE_BATCH(batch);
 631
 632     intel_batchbuffer_end_atomic(batch);
 633 }
 634
 635 static VAStatus gen8_vme_prepare(VADriverContextP ctx,
 636                                   struct encode_state *encode_state,
 637                                   struct intel_encoder_context *encoder_context)
 638 {
 639     VAStatus vaStatus = VA_STATUS_SUCCESS;
 640     VAEncSliceParameterBufferH264 *pSliceParameter = (VAEncSliceParameterBufferH264 *)encode_state->slice_params_ext[0]->buffer;
 641     int is_intra = pSliceParameter->slice_type == SLICE_TYPE_I;
 642     VAEncSequenceParameterBufferH264 *pSequenceParameter = (VAEncSequenceParameterBufferH264 *)encode_state->seq_param_ext->buffer;
 643     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 644
 645     if (!vme_context->h264_level ||
 646         (vme_context->h264_level != pSequenceParameter->level_idc)) {
 647         vme_context->h264_level = pSequenceParameter->level_idc;
 648     }
 649
 650     intel_vme_update_mbmv_cost(ctx, encode_state, encoder_context);
 651
 652     /*Setup all the memory object*/
 653     gen8_vme_surface_setup(ctx, encode_state, is_intra, encoder_context);
 654     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
 655     //gen8_vme_vme_state_setup(ctx, encode_state, is_intra, encoder_context);
 656     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 657
 658     /*Programing media pipeline*/
 659     gen8_vme_pipeline_programing(ctx, encode_state, encoder_context);
 660
 661     return vaStatus;
 662 }
 663
 664 static VAStatus gen8_vme_run(VADriverContextP ctx,
 665                               struct encode_state *encode_state,
 666                               struct intel_encoder_context *encoder_context)
 667 {
 668     struct intel_batchbuffer *batch = encoder_context->base.batch;
 669
 670     intel_batchbuffer_flush(batch);
 671
 672     return VA_STATUS_SUCCESS;
 673 }
 674
 675 static VAStatus gen8_vme_stop(VADriverContextP ctx,
 676                                struct encode_state *encode_state,
 677                                struct intel_encoder_context *encoder_context)
 678 {
 679     return VA_STATUS_SUCCESS;
 680 }
 681
 682 static VAStatus
 683 gen8_vme_pipeline(VADriverContextP ctx,
 684                    VAProfile profile,
 685                    struct encode_state *encode_state,
 686                    struct intel_encoder_context *encoder_context)
 687 {
 688     gen8_vme_media_init(ctx, encoder_context);
 689     gen8_vme_prepare(ctx, encode_state, encoder_context);
 690     gen8_vme_run(ctx, encode_state, encoder_context);
 691     gen8_vme_stop(ctx, encode_state, encoder_context);
 692
 693     return VA_STATUS_SUCCESS;
 694 }
 695
 696 static void
 697 gen8_vme_mpeg2_output_buffer_setup(VADriverContextP ctx,
 698                                     struct encode_state *encode_state,
 699                                     int index,
 700                                     int is_intra,
 701                                     struct intel_encoder_context *encoder_context)
 702
 703 {
 704     struct i965_driver_data *i965 = i965_driver_data(ctx);
 705     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 706     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 707     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 708     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 709
 710     vme_context->vme_output.num_blocks = width_in_mbs * height_in_mbs;
 711     vme_context->vme_output.pitch = 16; /* in bytes, always 16 */
 712
 713     if (is_intra)
 714         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 2;
 715     else
 716         vme_context->vme_output.size_block = INTRA_VME_OUTPUT_IN_BYTES * 24;
 717     /*
 718      * Inter MV . 32-byte Intra search + 16 IME info + 128 IME MV + 32 IME Ref
 719      * + 16 FBR Info + 128 FBR MV + 32 FBR Ref.
 720      * 16 * (2 + 2 * (1 + 8 + 2))= 16 * 24.
 721      */
 722
 723     vme_context->vme_output.bo = dri_bo_alloc(i965->intel.bufmgr,
 724                                               "VME output buffer",
 725                                               vme_context->vme_output.num_blocks * vme_context->vme_output.size_block,
 726                                               0x1000);
 727     assert(vme_context->vme_output.bo);
 728     vme_context->vme_buffer_suface_setup(ctx,
 729                                          &vme_context->gpe_context,
 730                                          &vme_context->vme_output,
 731                                          BINDING_TABLE_OFFSET(index),
 732                                          SURFACE_STATE_OFFSET(index));
 733 }
 734
 735 static void
 736 gen8_vme_mpeg2_output_vme_batchbuffer_setup(VADriverContextP ctx,
 737                                              struct encode_state *encode_state,
 738                                              int index,
 739                                              struct intel_encoder_context *encoder_context)
 740
 741 {
 742     struct i965_driver_data *i965 = i965_driver_data(ctx);
 743     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 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
 748     vme_context->vme_batchbuffer.num_blocks = width_in_mbs * height_in_mbs + 1;
 749     vme_context->vme_batchbuffer.size_block = 64; /* 4 OWORDs */
 750     vme_context->vme_batchbuffer.pitch = 16;
 751     vme_context->vme_batchbuffer.bo = dri_bo_alloc(i965->intel.bufmgr,
 752                                                    "VME batchbuffer",
 753                                                    vme_context->vme_batchbuffer.num_blocks * vme_context->vme_batchbuffer.size_block,
 754                                                    0x1000);
 755     vme_context->vme_buffer_suface_setup(ctx,
 756                                          &vme_context->gpe_context,
 757                                          &vme_context->vme_batchbuffer,
 758                                          BINDING_TABLE_OFFSET(index),
 759                                          SURFACE_STATE_OFFSET(index));
 760 }
 761
 762 static VAStatus
 763 gen8_vme_mpeg2_surface_setup(VADriverContextP ctx,
 764                               struct encode_state *encode_state,
 765                               int is_intra,
 766                               struct intel_encoder_context *encoder_context)
 767 {
 768     struct object_surface *obj_surface;
 769
 770     /*Setup surfaces state*/
 771     /* current picture for encoding */
 772     obj_surface = encode_state->input_yuv_object;
 773     gen8_vme_source_surface_state(ctx, 0, obj_surface, encoder_context);
 774     gen8_vme_media_source_surface_state(ctx, 4, obj_surface, encoder_context);
 775     gen8_vme_media_chroma_source_surface_state(ctx, 6, obj_surface, encoder_context);
 776
 777     if (!is_intra) {
 778         /* reference 0 */
 779         obj_surface = encode_state->reference_objects[0];
 780
 781         if (obj_surface->bo != NULL)
 782             gen8_vme_source_surface_state(ctx, 1, obj_surface, encoder_context);
 783
 784         /* reference 1 */
 785         obj_surface = encode_state->reference_objects[1];
 786
 787         if (obj_surface && obj_surface->bo != NULL)
 788             gen8_vme_source_surface_state(ctx, 2, obj_surface, encoder_context);
 789     }
 790
 791     /* VME output */
 792     gen8_vme_mpeg2_output_buffer_setup(ctx, encode_state, 3, is_intra, encoder_context);
 793     gen8_vme_mpeg2_output_vme_batchbuffer_setup(ctx, encode_state, 5, encoder_context);
 794
 795     return VA_STATUS_SUCCESS;
 796 }
 797
 798 static void
 799 gen8_vme_mpeg2_fill_vme_batchbuffer(VADriverContextP ctx,
 800                                      struct encode_state *encode_state,
 801                                      int mb_width, int mb_height,
 802                                      int kernel,
 803                                      int transform_8x8_mode_flag,
 804                                      struct intel_encoder_context *encoder_context)
 805 {
 806     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 807     int mb_x = 0, mb_y = 0;
 808     int i, s, j;
 809     unsigned int *command_ptr;
 810
 811
 812     dri_bo_map(vme_context->vme_batchbuffer.bo, 1);
 813     command_ptr = vme_context->vme_batchbuffer.bo->virtual;
 814
 815     for (s = 0; s < encode_state->num_slice_params_ext; s++) {
 816         VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[s]->buffer;
 817
 818         for (j = 0; j < encode_state->slice_params_ext[s]->num_elements; j++) {
 819             int slice_mb_begin = slice_param->macroblock_address;
 820             int slice_mb_number = slice_param->num_macroblocks;
 821             unsigned int mb_intra_ub;
 822             int slice_mb_x = slice_param->macroblock_address % mb_width;
 823
 824             for (i = 0; i < slice_mb_number;) {
 825                 int mb_count = i + slice_mb_begin;
 826
 827                 mb_x = mb_count % mb_width;
 828                 mb_y = mb_count / mb_width;
 829                 mb_intra_ub = 0;
 830
 831                 if (mb_x != 0) {
 832                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_AE;
 833                 }
 834
 835                 if (mb_y != 0) {
 836                     mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_B;
 837
 838                     if (mb_x != 0)
 839                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_D;
 840
 841                     if (mb_x != (mb_width -1))
 842                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 843                 }
 844
 845                 if (i < mb_width) {
 846                     if (i == 0)
 847                         mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_AE);
 848
 849                     mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_BCD_MASK);
 850
 851                     if ((i == (mb_width - 1)) && slice_mb_x) {
 852                         mb_intra_ub |= INTRA_PRED_AVAIL_FLAG_C;
 853                     }
 854                 }
 855
 856                 if ((i == mb_width) && slice_mb_x) {
 857                     mb_intra_ub &= ~(INTRA_PRED_AVAIL_FLAG_D);
 858                 }
 859
 860                 *command_ptr++ = (CMD_MEDIA_OBJECT | (8 - 2));
 861                 *command_ptr++ = kernel;
 862                 *command_ptr++ = 0;
 863                 *command_ptr++ = 0;
 864                 *command_ptr++ = 0;
 865                 *command_ptr++ = 0;
 866
 867                 /*inline data */
 868                 *command_ptr++ = (mb_width << 16 | mb_y << 8 | mb_x);
 869                 *command_ptr++ = ( (1 << 16) | transform_8x8_mode_flag | (mb_intra_ub << 8));
 870
 871                 i += 1;
 872             }
 873
 874             slice_param++;
 875         }
 876     }
 877
 878     *command_ptr++ = 0;
 879     *command_ptr++ = MI_BATCH_BUFFER_END;
 880
 881     dri_bo_unmap(vme_context->vme_batchbuffer.bo);
 882 }
 883
 884 static void
 885 gen8_vme_mpeg2_pipeline_programing(VADriverContextP ctx,
 886                                     struct encode_state *encode_state,
 887                                     int is_intra,
 888                                     struct intel_encoder_context *encoder_context)
 889 {
 890     struct gen6_vme_context *vme_context = encoder_context->vme_context;
 891     struct intel_batchbuffer *batch = encoder_context->base.batch;
 892     VAEncSequenceParameterBufferMPEG2 *seq_param = (VAEncSequenceParameterBufferMPEG2 *)encode_state->seq_param_ext->buffer;
 893     int width_in_mbs = ALIGN(seq_param->picture_width, 16) / 16;
 894     int height_in_mbs = ALIGN(seq_param->picture_height, 16) / 16;
 895
 896     gen8_vme_mpeg2_fill_vme_batchbuffer(ctx,
 897                                          encode_state,
 898                                          width_in_mbs, height_in_mbs,
 899                                          is_intra ? VME_INTRA_SHADER : VME_INTER_SHADER,
 900                                          0,
 901                                          encoder_context);
 902
 903     intel_batchbuffer_start_atomic(batch, 0x1000);
 904     gen6_gpe_pipeline_setup(ctx, &vme_context->gpe_context, batch);
 905     BEGIN_BATCH(batch, 2);
 906     OUT_BATCH(batch, MI_BATCH_BUFFER_START | (2 << 6));
 907     OUT_RELOC(batch,
 908               vme_context->vme_batchbuffer.bo,
 909               I915_GEM_DOMAIN_COMMAND, 0,
 910               0);
 911     ADVANCE_BATCH(batch);
 912
 913     intel_batchbuffer_end_atomic(batch);
 914 }
 915
 916 static VAStatus
 917 gen8_vme_mpeg2_prepare(VADriverContextP ctx,
 918                         struct encode_state *encode_state,
 919                         struct intel_encoder_context *encoder_context)
 920 {
 921     VAStatus vaStatus = VA_STATUS_SUCCESS;
 922     VAEncSliceParameterBufferMPEG2 *slice_param = (VAEncSliceParameterBufferMPEG2 *)encode_state->slice_params_ext[0]->buffer;
 923
 924     /*Setup all the memory object*/
 925     gen8_vme_mpeg2_surface_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 926     gen8_vme_interface_setup(ctx, encode_state, encoder_context);
 927     gen8_vme_vme_state_setup(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 928     gen8_vme_constant_setup(ctx, encode_state, encoder_context);
 929
 930     /*Programing media pipeline*/
 931     gen8_vme_mpeg2_pipeline_programing(ctx, encode_state, slice_param->is_intra_slice, encoder_context);
 932
 933     return vaStatus;
 934 }
 935
 936 static VAStatus
 937 gen8_vme_mpeg2_pipeline(VADriverContextP ctx,
 938                          VAProfile profile,
 939                          struct encode_state *encode_state,
 940                          struct intel_encoder_context *encoder_context)
 941 {
 942     gen8_vme_media_init(ctx, encoder_context);
 943     gen8_vme_mpeg2_prepare(ctx, encode_state, encoder_context);
 944     gen8_vme_run(ctx, encode_state, encoder_context);
 945     gen8_vme_stop(ctx, encode_state, encoder_context);
 946
 947     return VA_STATUS_SUCCESS;
 948 }
 949
 950 static void
 951 gen8_vme_context_destroy(void *context)
 952 {
 953     struct gen6_vme_context *vme_context = context;
 954
 955     i965_gpe_context_destroy(&vme_context->gpe_context);
 956
 957     dri_bo_unreference(vme_context->vme_output.bo);
 958     vme_context->vme_output.bo = NULL;
 959
 960     dri_bo_unreference(vme_context->vme_state.bo);
 961     vme_context->vme_state.bo = NULL;
 962
 963     dri_bo_unreference(vme_context->vme_batchbuffer.bo);
 964     vme_context->vme_batchbuffer.bo = NULL;
 965
 966     if (vme_context->vme_state_message) {
 967         free(vme_context->vme_state_message);
 968         vme_context->vme_state_message = NULL;
 969     }
 970
 971     free(vme_context);
 972 }
 973
 974 Bool gen8_vme_context_init(VADriverContextP ctx, struct intel_encoder_context *encoder_context)
 975 {
 976     struct gen6_vme_context *vme_context = calloc(1, sizeof(struct gen6_vme_context));
 977     struct i965_kernel *vme_kernel_list = NULL;
 978         int i965_kernel_num;
 979
 980     switch (encoder_context->codec) {
 981     case CODEC_H264:
 982         vme_kernel_list = gen8_vme_kernels;
 983         encoder_context->vme_pipeline = gen8_vme_pipeline;
 984         i965_kernel_num = sizeof(gen8_vme_kernels) / sizeof(struct i965_kernel);
 985         break;
 986
 987     case CODEC_MPEG2:
 988         vme_kernel_list = gen8_vme_mpeg2_kernels;
 989         encoder_context->vme_pipeline = gen8_vme_mpeg2_pipeline;
 990         i965_kernel_num = sizeof(gen8_vme_mpeg2_kernels) / sizeof(struct i965_kernel);
 991
 992         break;
 993
 994     default:
 995         /* never get here */
 996         assert(0);
 997
 998         break;
 999     }
1000     vme_context->vme_kernel_sum = i965_kernel_num;
1001     vme_context->gpe_context.surface_state_binding_table.length = (SURFACE_STATE_PADDED_SIZE + sizeof(unsigned int)) * MAX_MEDIA_SURFACES_GEN6;
1002
1003     vme_context->gpe_context.idrt.max_entries = MAX_INTERFACE_DESC_GEN6;
1004     vme_context->gpe_context.idrt.entry_size = sizeof(struct gen6_interface_descriptor_data);
1005
1006     vme_context->gpe_context.curbe.length = CURBE_TOTAL_DATA_LENGTH;
1007
1008     vme_context->gpe_context.vfe_state.max_num_threads = 60 - 1;
1009     vme_context->gpe_context.vfe_state.num_urb_entries = 16;
1010     vme_context->gpe_context.vfe_state.gpgpu_mode = 0;
1011     vme_context->gpe_context.vfe_state.urb_entry_size = 59 - 1;
1012     vme_context->gpe_context.vfe_state.curbe_allocation_size = CURBE_ALLOCATION_SIZE - 1;
1013
1014     gen7_vme_scoreboard_init(ctx, vme_context);
1015
1016     i965_gpe_load_kernels(ctx,
1017                           &vme_context->gpe_context,
1018                           vme_kernel_list,
1019                           i965_kernel_num);
1020     vme_context->vme_surface2_setup = gen7_gpe_surface2_setup;
1021     vme_context->vme_media_rw_surface_setup = gen7_gpe_media_rw_surface_setup;
1022     vme_context->vme_buffer_suface_setup = gen7_gpe_buffer_suface_setup;
1023     vme_context->vme_media_chroma_surface_setup = gen75_gpe_media_chroma_surface_setup;
1024
1025     encoder_context->vme_context = vme_context;
1026     encoder_context->vme_context_destroy = gen8_vme_context_destroy;
1027
1028     vme_context->vme_state_message = malloc(VME_MSG_LENGTH * sizeof(int));
1029
1030     return True;
1031 }