i965_drv_video/i965_avc_hw_scoreboard.c

   1 /*
   2  * Copyright © 2010 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  *    Xiang Haihao <haihao.xiang@intel.com>
  26  *
  27  */
  28
  29 #include <stdio.h>
  30 #include <string.h>
  31 #include <assert.h>
  32
  33 #include "va_backend.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_avc_hw_scoreboard.h"
  41 #include "i965_media_h264.h"
  42 #include "i965_media.h"
  43
  44 extern struct media_kernel *h264_avc_kernels;
  45
  46 /* On Ironlake */
  47 #include "shaders/h264/mc/export.inc.gen5"
  48
  49 enum {
  50     AVC_HW_SCOREBOARD = 0,
  51     AVC_HW_SCOREBOARD_MBAFF
  52 };
  53
  54 static unsigned long avc_hw_scoreboard_kernel_offset[] = {
  55     SETHWSCOREBOARD_IP_GEN5,
  56     SETHWSCOREBOARD_MBAFF_IP_GEN5
  57 };
  58
  59 static unsigned int avc_hw_scoreboard_constants[] = {
  60     0x08040201,
  61     0x00000010,
  62     0x08000210,
  63     0x00000000,
  64     0x08040201,
  65     0x08040210,
  66     0x01000010,
  67     0x08040200
  68 };
  69
  70 static void
  71 i965_avc_hw_scoreboard_surface_state(struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
  72 {
  73     struct i965_surface_state *ss;
  74     dri_bo *bo;
  75
  76     bo = avc_hw_scoreboard_context->surface.ss_bo;
  77     dri_bo_map(bo, 1);
  78     assert(bo->virtual);
  79     ss = bo->virtual;
  80     memset(ss, 0, sizeof(*ss));
  81     ss->ss0.surface_type = I965_SURFACE_BUFFER;
  82     ss->ss1.base_addr = avc_hw_scoreboard_context->surface.s_bo->offset;
  83     ss->ss2.width = ((avc_hw_scoreboard_context->surface.total_mbs * MB_CMD_IN_OWS - 1) & 0x7f);
  84     ss->ss2.height = (((avc_hw_scoreboard_context->surface.total_mbs * MB_CMD_IN_OWS - 1) >> 7) & 0x1fff);
  85     ss->ss3.depth = (((avc_hw_scoreboard_context->surface.total_mbs * MB_CMD_IN_OWS - 1) >> 20) & 0x7f);
  86     dri_bo_emit_reloc(bo,
  87                       I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER,
  88                       0,
  89                       offsetof(struct i965_surface_state, ss1),
  90                       avc_hw_scoreboard_context->surface.s_bo);
  91     dri_bo_unmap(bo);
  92 }
  93
  94 static void
  95 i965_avc_hw_scoreboard_interface_descriptor_table(struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
  96 {
  97     struct i965_interface_descriptor *desc;
  98     dri_bo *bo;
  99
 100     bo = avc_hw_scoreboard_context->idrt.bo;
 101     dri_bo_map(bo, 1);
 102     assert(bo->virtual);
 103     desc = bo->virtual;
 104     memset(desc, 0, sizeof(*desc));
 105     desc->desc0.grf_reg_blocks = 7;
 106     desc->desc0.kernel_start_pointer = (avc_hw_scoreboard_context->hw_kernel.bo->offset +
 107                                         avc_hw_scoreboard_context->hw_kernel.offset) >> 6; /* reloc */
 108     desc->desc1.const_urb_entry_read_offset = 0;
 109     desc->desc1.const_urb_entry_read_len = 1;
 110     desc->desc3.binding_table_entry_count = 0;
 111     desc->desc3.binding_table_pointer =
 112         avc_hw_scoreboard_context->binding_table.bo->offset >> 5; /*reloc */
 113
 114     dri_bo_emit_reloc(bo,
 115                       I915_GEM_DOMAIN_INSTRUCTION, 0,
 116                       desc->desc0.grf_reg_blocks + avc_hw_scoreboard_context->hw_kernel.offset,
 117                       offsetof(struct i965_interface_descriptor, desc0),
 118                       avc_hw_scoreboard_context->hw_kernel.bo);
 119
 120     dri_bo_emit_reloc(bo,
 121                       I915_GEM_DOMAIN_INSTRUCTION, 0,
 122                       desc->desc3.binding_table_entry_count,
 123                       offsetof(struct i965_interface_descriptor, desc3),
 124                       avc_hw_scoreboard_context->binding_table.bo);
 125
 126     dri_bo_unmap(bo);
 127 }
 128
 129 static void
 130 i965_avc_hw_scoreboard_binding_table(struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 131 {
 132     unsigned int *binding_table;
 133     dri_bo *bo = avc_hw_scoreboard_context->binding_table.bo;
 134
 135     dri_bo_map(bo, 1);
 136     assert(bo->virtual);
 137     binding_table = bo->virtual;
 138     memset(binding_table, 0, bo->size);
 139     binding_table[0] = avc_hw_scoreboard_context->surface.ss_bo->offset;
 140     dri_bo_emit_reloc(bo,
 141                       I915_GEM_DOMAIN_INSTRUCTION, 0,
 142                       0,
 143                       0,
 144                       avc_hw_scoreboard_context->surface.ss_bo);
 145     dri_bo_unmap(bo);
 146 }
 147
 148 static void
 149 i965_avc_hw_scoreboard_vfe_state(struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 150 {
 151     struct i965_vfe_state *vfe_state;
 152     dri_bo *bo;
 153
 154     bo = avc_hw_scoreboard_context->vfe_state.bo;
 155     dri_bo_map(bo, 1);
 156     assert(bo->virtual);
 157     vfe_state = bo->virtual;
 158     memset(vfe_state, 0, sizeof(*vfe_state));
 159     vfe_state->vfe1.max_threads = avc_hw_scoreboard_context->urb.num_vfe_entries - 1;
 160     vfe_state->vfe1.urb_entry_alloc_size = avc_hw_scoreboard_context->urb.size_vfe_entry - 1;
 161     vfe_state->vfe1.num_urb_entries = avc_hw_scoreboard_context->urb.num_vfe_entries;
 162     vfe_state->vfe1.vfe_mode = VFE_GENERIC_MODE;
 163     vfe_state->vfe1.children_present = 0;
 164     vfe_state->vfe2.interface_descriptor_base =
 165         avc_hw_scoreboard_context->idrt.bo->offset >> 4; /* reloc */
 166     dri_bo_emit_reloc(bo,
 167                       I915_GEM_DOMAIN_INSTRUCTION, 0,
 168                       0,
 169                       offsetof(struct i965_vfe_state, vfe2),
 170                       avc_hw_scoreboard_context->idrt.bo);
 171     dri_bo_unmap(bo);
 172 }
 173
 174 static void
 175 i965_avc_hw_scoreboard_upload_constants(struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 176 {
 177     unsigned char *constant_buffer;
 178
 179     dri_bo_map(avc_hw_scoreboard_context->curbe.bo, 1);
 180     assert(avc_hw_scoreboard_context->curbe.bo->virtual);
 181     constant_buffer = avc_hw_scoreboard_context->curbe.bo->virtual;
 182     memcpy(constant_buffer, avc_hw_scoreboard_constants, sizeof(avc_hw_scoreboard_constants));
 183     dri_bo_unmap(avc_hw_scoreboard_context->curbe.bo);
 184 }
 185
 186 static void
 187 i965_avc_hw_scoreboard_states_setup(struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 188 {
 189     i965_avc_hw_scoreboard_surface_state(avc_hw_scoreboard_context);
 190     i965_avc_hw_scoreboard_binding_table(avc_hw_scoreboard_context);
 191     i965_avc_hw_scoreboard_interface_descriptor_table(avc_hw_scoreboard_context);
 192     i965_avc_hw_scoreboard_vfe_state(avc_hw_scoreboard_context);
 193     i965_avc_hw_scoreboard_upload_constants(avc_hw_scoreboard_context);
 194 }
 195
 196 static void
 197 i965_avc_hw_scoreboard_pipeline_select(VADriverContextP ctx)
 198 {
 199     BEGIN_BATCH(ctx, 1);
 200     OUT_BATCH(ctx, CMD_PIPELINE_SELECT | PIPELINE_SELECT_MEDIA);
 201     ADVANCE_BATCH(ctx);
 202 }
 203
 204 static void
 205 i965_avc_hw_scoreboard_urb_layout(VADriverContextP ctx, struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 206 {
 207     struct i965_driver_data *i965 = i965_driver_data(ctx);
 208     unsigned int vfe_fence, cs_fence;
 209
 210     vfe_fence = avc_hw_scoreboard_context->urb.cs_start;
 211     cs_fence = URB_SIZE((&i965->intel));
 212
 213     BEGIN_BATCH(ctx, 3);
 214     OUT_BATCH(ctx, CMD_URB_FENCE | UF0_VFE_REALLOC | UF0_CS_REALLOC | 1);
 215     OUT_BATCH(ctx, 0);
 216     OUT_BATCH(ctx,
 217               (vfe_fence << UF2_VFE_FENCE_SHIFT) |      /* VFE_SIZE */
 218               (cs_fence << UF2_CS_FENCE_SHIFT));        /* CS_SIZE */
 219     ADVANCE_BATCH(ctx);
 220 }
 221
 222 static void
 223 i965_avc_hw_scoreboard_state_base_address(VADriverContextP ctx)
 224 {
 225     BEGIN_BATCH(ctx, 8);
 226     OUT_BATCH(ctx, CMD_STATE_BASE_ADDRESS | 6);
 227     OUT_BATCH(ctx, 0 | BASE_ADDRESS_MODIFY);
 228     OUT_BATCH(ctx, 0 | BASE_ADDRESS_MODIFY);
 229     OUT_BATCH(ctx, 0 | BASE_ADDRESS_MODIFY);
 230     OUT_BATCH(ctx, 0 | BASE_ADDRESS_MODIFY);
 231     OUT_BATCH(ctx, 0 | BASE_ADDRESS_MODIFY);
 232     OUT_BATCH(ctx, 0 | BASE_ADDRESS_MODIFY);
 233     OUT_BATCH(ctx, 0 | BASE_ADDRESS_MODIFY);
 234     ADVANCE_BATCH(ctx);
 235 }
 236
 237 static void
 238 i965_avc_hw_scoreboard_state_pointers(VADriverContextP ctx, struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 239 {
 240     BEGIN_BATCH(ctx, 3);
 241     OUT_BATCH(ctx, CMD_MEDIA_STATE_POINTERS | 1);
 242     OUT_BATCH(ctx, 0);
 243     OUT_RELOC(ctx, avc_hw_scoreboard_context->vfe_state.bo, I915_GEM_DOMAIN_INSTRUCTION, 0, 0);
 244     ADVANCE_BATCH(ctx);
 245 }
 246
 247 static void
 248 i965_avc_hw_scoreboard_cs_urb_layout(VADriverContextP ctx, struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 249 {
 250     BEGIN_BATCH(ctx, 2);
 251     OUT_BATCH(ctx, CMD_CS_URB_STATE | 0);
 252     OUT_BATCH(ctx,
 253               ((avc_hw_scoreboard_context->urb.size_cs_entry - 1) << 4) |     /* URB Entry Allocation Size */
 254               (avc_hw_scoreboard_context->urb.num_cs_entries << 0));          /* Number of URB Entries */
 255     ADVANCE_BATCH(ctx);
 256 }
 257
 258 static void
 259 i965_avc_hw_scoreboard_constant_buffer(VADriverContextP ctx, struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 260 {
 261     BEGIN_BATCH(ctx, 2);
 262     OUT_BATCH(ctx, CMD_CONSTANT_BUFFER | (1 << 8) | (2 - 2));
 263     OUT_RELOC(ctx, avc_hw_scoreboard_context->curbe.bo,
 264               I915_GEM_DOMAIN_INSTRUCTION, 0,
 265               avc_hw_scoreboard_context->urb.size_cs_entry - 1);
 266     ADVANCE_BATCH(ctx);
 267 }
 268
 269 static void
 270 i965_avc_hw_scoreboard_objects(VADriverContextP ctx, struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 271 {
 272     BEGIN_BATCH(ctx, 6);
 273     OUT_BATCH(ctx, CMD_MEDIA_OBJECT | 4);
 274     OUT_BATCH(ctx, 0); /* interface descriptor offset: 0 */
 275     OUT_BATCH(ctx, 0); /* no indirect data */
 276     OUT_BATCH(ctx, 0);
 277     OUT_BATCH(ctx, ((avc_hw_scoreboard_context->inline_data.num_mb_cmds << 16) |
 278                     (avc_hw_scoreboard_context->inline_data.starting_mb_number << 0)));
 279     OUT_BATCH(ctx, avc_hw_scoreboard_context->inline_data.pic_width_in_mbs);
 280     ADVANCE_BATCH(ctx);
 281 }
 282
 283 static void
 284 i965_avc_hw_scoreboard_pipeline_setup(VADriverContextP ctx, struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 285 {
 286     intel_batchbuffer_start_atomic(ctx, 0x1000);
 287     intel_batchbuffer_emit_mi_flush(ctx);
 288     i965_avc_hw_scoreboard_pipeline_select(ctx);
 289     i965_avc_hw_scoreboard_state_base_address(ctx);
 290     i965_avc_hw_scoreboard_state_pointers(ctx, avc_hw_scoreboard_context);
 291     i965_avc_hw_scoreboard_urb_layout(ctx, avc_hw_scoreboard_context);
 292     i965_avc_hw_scoreboard_cs_urb_layout(ctx, avc_hw_scoreboard_context);
 293     i965_avc_hw_scoreboard_constant_buffer(ctx, avc_hw_scoreboard_context);
 294     i965_avc_hw_scoreboard_objects(ctx, avc_hw_scoreboard_context);
 295     intel_batchbuffer_end_atomic(ctx);
 296 }
 297
 298 void
 299 i965_avc_hw_scoreboard(VADriverContextP ctx, struct decode_state *decode_state)
 300 {
 301     struct i965_driver_data *i965 = i965_driver_data(ctx);
 302     struct i965_media_state *media_state = &i965->media_state;
 303     struct i965_h264_context *i965_h264_context = (struct i965_h264_context *)media_state->private_context;
 304     struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context;
 305
 306     if (i965_h264_context->use_avc_hw_scoreboard) {
 307         assert(i965_h264_context != NULL);
 308         avc_hw_scoreboard_context = &i965_h264_context->avc_hw_scoreboard_context;
 309
 310         avc_hw_scoreboard_context->inline_data.num_mb_cmds = i965_h264_context->avc_it_command_mb_info.mbs;
 311         avc_hw_scoreboard_context->inline_data.starting_mb_number = i965_h264_context->avc_it_command_mb_info.mbs;
 312         avc_hw_scoreboard_context->inline_data.pic_width_in_mbs = i965_h264_context->picture.width_in_mbs;
 313         avc_hw_scoreboard_context->surface.total_mbs = i965_h264_context->avc_it_command_mb_info.mbs * 2;
 314
 315         dri_bo_unreference(avc_hw_scoreboard_context->hw_kernel.bo);
 316         avc_hw_scoreboard_context->hw_kernel.bo = h264_avc_kernels[H264_AVC_COMBINED].bo;
 317         assert(avc_hw_scoreboard_context->hw_kernel.bo != NULL);
 318         dri_bo_reference(avc_hw_scoreboard_context->hw_kernel.bo);
 319
 320         if (i965_h264_context->picture.mbaff_frame_flag)
 321             avc_hw_scoreboard_context->hw_kernel.offset = avc_hw_scoreboard_kernel_offset[AVC_HW_SCOREBOARD_MBAFF];
 322         else
 323             avc_hw_scoreboard_context->hw_kernel.offset = avc_hw_scoreboard_kernel_offset[AVC_HW_SCOREBOARD];
 324
 325         i965_avc_hw_scoreboard_states_setup(avc_hw_scoreboard_context);
 326         i965_avc_hw_scoreboard_pipeline_setup(ctx, avc_hw_scoreboard_context);
 327     }
 328 }
 329
 330 void
 331 i965_avc_hw_scoreboard_decode_init(VADriverContextP ctx)
 332 {
 333     struct i965_driver_data *i965 = i965_driver_data(ctx);
 334     struct i965_media_state *media_state = &i965->media_state;
 335     struct i965_h264_context *i965_h264_context = (struct i965_h264_context *)media_state->private_context;
 336     struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context;
 337     dri_bo *bo;
 338
 339     if (i965_h264_context->use_avc_hw_scoreboard) {
 340         assert(i965_h264_context != NULL);
 341         avc_hw_scoreboard_context = &i965_h264_context->avc_hw_scoreboard_context;
 342
 343         dri_bo_unreference(avc_hw_scoreboard_context->curbe.bo);
 344         bo = dri_bo_alloc(i965->intel.bufmgr,
 345                           "constant buffer",
 346                           4096, 64);
 347         assert(bo);
 348         avc_hw_scoreboard_context->curbe.bo = bo;
 349
 350         dri_bo_unreference(avc_hw_scoreboard_context->surface.s_bo);
 351         avc_hw_scoreboard_context->surface.s_bo = i965_h264_context->avc_it_command_mb_info.bo;
 352         assert(avc_hw_scoreboard_context->surface.s_bo != NULL);
 353         dri_bo_reference(avc_hw_scoreboard_context->surface.s_bo);
 354
 355         dri_bo_unreference(avc_hw_scoreboard_context->surface.ss_bo);
 356         bo = dri_bo_alloc(i965->intel.bufmgr,
 357                           "surface state",
 358                           sizeof(struct i965_surface_state), 32);
 359         assert(bo);
 360         avc_hw_scoreboard_context->surface.ss_bo = bo;
 361
 362         dri_bo_unreference(avc_hw_scoreboard_context->binding_table.bo);
 363         bo = dri_bo_alloc(i965->intel.bufmgr,
 364                           "binding table",
 365                           MAX_MEDIA_SURFACES * sizeof(unsigned int), 32);
 366         assert(bo);
 367         avc_hw_scoreboard_context->binding_table.bo = bo;
 368
 369         dri_bo_unreference(avc_hw_scoreboard_context->idrt.bo);
 370         bo = dri_bo_alloc(i965->intel.bufmgr,
 371                           "interface discriptor",
 372                           MAX_INTERFACE_DESC * sizeof(struct i965_interface_descriptor), 16);
 373         assert(bo);
 374         avc_hw_scoreboard_context->idrt.bo = bo;
 375
 376         dri_bo_unreference(avc_hw_scoreboard_context->vfe_state.bo);
 377         bo = dri_bo_alloc(i965->intel.bufmgr,
 378                           "vfe state",
 379                           sizeof(struct i965_vfe_state), 32);
 380         assert(bo);
 381         avc_hw_scoreboard_context->vfe_state.bo = bo;
 382
 383         avc_hw_scoreboard_context->urb.num_vfe_entries = 1;
 384         avc_hw_scoreboard_context->urb.size_vfe_entry = 2;
 385         avc_hw_scoreboard_context->urb.num_cs_entries = 1;
 386         avc_hw_scoreboard_context->urb.size_cs_entry = 1;
 387         avc_hw_scoreboard_context->urb.vfe_start = 0;
 388         avc_hw_scoreboard_context->urb.cs_start = avc_hw_scoreboard_context->urb.vfe_start +
 389             avc_hw_scoreboard_context->urb.num_vfe_entries * avc_hw_scoreboard_context->urb.size_vfe_entry;
 390         assert(avc_hw_scoreboard_context->urb.cs_start +
 391                avc_hw_scoreboard_context->urb.num_cs_entries * avc_hw_scoreboard_context->urb.size_cs_entry <= URB_SIZE((&i965->intel)));
 392     }
 393 }
 394
 395 Bool
 396 i965_avc_hw_scoreboard_ternimate(struct i965_avc_hw_scoreboard_context *avc_hw_scoreboard_context)
 397 {
 398     dri_bo_unreference(avc_hw_scoreboard_context->curbe.bo);
 399     avc_hw_scoreboard_context->curbe.bo = NULL;
 400
 401     dri_bo_unreference(avc_hw_scoreboard_context->surface.ss_bo);
 402     avc_hw_scoreboard_context->surface.ss_bo = NULL;
 403
 404     dri_bo_unreference(avc_hw_scoreboard_context->surface.s_bo);
 405     avc_hw_scoreboard_context->surface.s_bo = NULL;
 406
 407     dri_bo_unreference(avc_hw_scoreboard_context->binding_table.bo);
 408     avc_hw_scoreboard_context->binding_table.bo = NULL;
 409
 410     dri_bo_unreference(avc_hw_scoreboard_context->idrt.bo);
 411     avc_hw_scoreboard_context->idrt.bo = NULL;
 412
 413     dri_bo_unreference(avc_hw_scoreboard_context->vfe_state.bo);
 414     avc_hw_scoreboard_context->vfe_state.bo = NULL;
 415
 416     dri_bo_unreference(avc_hw_scoreboard_context->hw_kernel.bo);
 417     avc_hw_scoreboard_context->hw_kernel.bo = NULL;
 418
 419     return True;
 420 }