src/gallium/drivers/v3d/v3dx_draw.c

   1 /*
   2  * Copyright © 2014-2017 Broadcom
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "util/u_blitter.h"
  25 #include "util/u_prim.h"
  26 #include "util/u_format.h"
  27 #include "util/u_pack_color.h"
  28 #include "util/u_prim_restart.h"
  29 #include "util/u_upload_mgr.h"
  30 #include "indices/u_primconvert.h"
  31
  32 #include "v3d_context.h"
  33 #include "v3d_resource.h"
  34 #include "v3d_cl.h"
  35 #include "broadcom/compiler/v3d_compiler.h"
  36 #include "broadcom/common/v3d_macros.h"
  37 #include "broadcom/cle/v3dx_pack.h"
  38
  39 /**
  40  * Does the initial bining command list setup for drawing to a given FBO.
  41  */
  42 static void
  43 v3d_start_draw(struct v3d_context *v3d)
  44 {
  45         struct v3d_job *job = v3d->job;
  46
  47         if (job->needs_flush)
  48                 return;
  49
  50         /* Get space to emit our BCL state, using a branch to jump to a new BO
  51          * if necessary.
  52          */
  53         v3d_cl_ensure_space_with_branch(&job->bcl, 256 /* XXX */);
  54
  55         job->submit.bcl_start = job->bcl.bo->offset;
  56         v3d_job_add_bo(job, job->bcl.bo);
  57
  58         job->tile_alloc = v3d_bo_alloc(v3d->screen, 1024 * 1024, "tile_alloc");
  59         uint32_t tsda_per_tile_size = v3d->screen->devinfo.ver >= 40 ? 256 : 64;
  60         job->tile_state = v3d_bo_alloc(v3d->screen,
  61                                        job->draw_tiles_y *
  62                                        job->draw_tiles_x *
  63                                        tsda_per_tile_size,
  64                                        "TSDA");
  65
  66 #if V3D_VERSION >= 40
  67         cl_emit(&job->bcl, TILE_BINNING_MODE_CFG, config) {
  68                 config.width_in_pixels = v3d->framebuffer.width;
  69                 config.height_in_pixels = v3d->framebuffer.height;
  70                 config.number_of_render_targets =
  71                         MAX2(v3d->framebuffer.nr_cbufs, 1);
  72
  73                 config.multisample_mode_4x = job->msaa;
  74
  75                 config.maximum_bpp_of_all_render_targets = job->internal_bpp;
  76         }
  77 #else /* V3D_VERSION < 40 */
  78         /* "Binning mode lists start with a Tile Binning Mode Configuration
  79          * item (120)"
  80          *
  81          * Part1 signals the end of binning config setup.
  82          */
  83         cl_emit(&job->bcl, TILE_BINNING_MODE_CFG_PART2, config) {
  84                 config.tile_allocation_memory_address =
  85                         cl_address(job->tile_alloc, 0);
  86                 config.tile_allocation_memory_size = job->tile_alloc->size;
  87         }
  88
  89         cl_emit(&job->bcl, TILE_BINNING_MODE_CFG_PART1, config) {
  90                 config.tile_state_data_array_base_address =
  91                         cl_address(job->tile_state, 0);
  92
  93                 config.width_in_tiles = job->draw_tiles_x;
  94                 config.height_in_tiles = job->draw_tiles_y;
  95                 /* Must be >= 1 */
  96                 config.number_of_render_targets =
  97                         MAX2(v3d->framebuffer.nr_cbufs, 1);
  98
  99                 config.multisample_mode_4x = job->msaa;
 100
 101                 config.maximum_bpp_of_all_render_targets = job->internal_bpp;
 102         }
 103 #endif /* V3D_VERSION < 40 */
 104
 105         /* There's definitely nothing in the VCD cache we want. */
 106         cl_emit(&job->bcl, FLUSH_VCD_CACHE, bin);
 107
 108         /* Disable any leftover OQ state from another job. */
 109         cl_emit(&job->bcl, OCCLUSION_QUERY_COUNTER, counter);
 110
 111         /* "Binning mode lists must have a Start Tile Binning item (6) after
 112          *  any prefix state data before the binning list proper starts."
 113          */
 114         cl_emit(&job->bcl, START_TILE_BINNING, bin);
 115
 116         job->needs_flush = true;
 117         job->draw_width = v3d->framebuffer.width;
 118         job->draw_height = v3d->framebuffer.height;
 119 }
 120
 121 static void
 122 v3d_predraw_check_stage_inputs(struct pipe_context *pctx,
 123                                enum pipe_shader_type s)
 124 {
 125         struct v3d_context *v3d = v3d_context(pctx);
 126
 127         /* Flush writes to textures we're sampling. */
 128         for (int i = 0; i < v3d->tex[s].num_textures; i++) {
 129                 struct pipe_sampler_view *pview = v3d->tex[s].textures[i];
 130                 if (!pview)
 131                         continue;
 132                 struct v3d_sampler_view *view = v3d_sampler_view(pview);
 133
 134                 if (view->texture != view->base.texture)
 135                         v3d_update_shadow_texture(pctx, &view->base);
 136
 137                 v3d_flush_jobs_writing_resource(v3d, view->texture);
 138         }
 139
 140         /* Flush writes to UBOs. */
 141         foreach_bit(i, v3d->constbuf[s].enabled_mask) {
 142                 struct pipe_constant_buffer *cb = &v3d->constbuf[s].cb[i];
 143                 if (cb->buffer)
 144                         v3d_flush_jobs_writing_resource(v3d, cb->buffer);
 145         }
 146 }
 147
 148 static void
 149 v3d_emit_gl_shader_state(struct v3d_context *v3d,
 150                          const struct pipe_draw_info *info)
 151 {
 152         struct v3d_job *job = v3d->job;
 153         /* VC5_DIRTY_VTXSTATE */
 154         struct v3d_vertex_stateobj *vtx = v3d->vtx;
 155         /* VC5_DIRTY_VTXBUF */
 156         struct v3d_vertexbuf_stateobj *vertexbuf = &v3d->vertexbuf;
 157
 158         /* Upload the uniforms to the indirect CL first */
 159         struct v3d_cl_reloc fs_uniforms =
 160                 v3d_write_uniforms(v3d, v3d->prog.fs,
 161                                    PIPE_SHADER_FRAGMENT);
 162         struct v3d_cl_reloc vs_uniforms =
 163                 v3d_write_uniforms(v3d, v3d->prog.vs,
 164                                    PIPE_SHADER_VERTEX);
 165         struct v3d_cl_reloc cs_uniforms =
 166                 v3d_write_uniforms(v3d, v3d->prog.cs,
 167                                    PIPE_SHADER_VERTEX);
 168
 169         /* See GFXH-930 workaround below */
 170         uint32_t num_elements_to_emit = MAX2(vtx->num_elements, 1);
 171         uint32_t shader_rec_offset =
 172                 v3d_cl_ensure_space(&job->indirect,
 173                                     cl_packet_length(GL_SHADER_STATE_RECORD) +
 174                                     num_elements_to_emit *
 175                                     cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD),
 176                                     32);
 177
 178         cl_emit(&job->indirect, GL_SHADER_STATE_RECORD, shader) {
 179                 shader.enable_clipping = true;
 180                 /* VC5_DIRTY_PRIM_MODE | VC5_DIRTY_RASTERIZER */
 181                 shader.point_size_in_shaded_vertex_data =
 182                         (info->mode == PIPE_PRIM_POINTS &&
 183                          v3d->rasterizer->base.point_size_per_vertex);
 184
 185                 /* Must be set if the shader modifies Z, discards, or modifies
 186                  * the sample mask.  For any of these cases, the fragment
 187                  * shader needs to write the Z value (even just discards).
 188                  */
 189                 shader.fragment_shader_does_z_writes =
 190                         (v3d->prog.fs->prog_data.fs->writes_z ||
 191                          v3d->prog.fs->prog_data.fs->discard);
 192
 193                 shader.fragment_shader_uses_real_pixel_centre_w_in_addition_to_centroid_w2 =
 194                         v3d->prog.fs->prog_data.fs->uses_center_w;
 195
 196                 shader.number_of_varyings_in_fragment_shader =
 197                         v3d->prog.fs->prog_data.base->num_inputs;
 198
 199                 shader.coordinate_shader_propagate_nans = true;
 200                 shader.vertex_shader_propagate_nans = true;
 201                 shader.fragment_shader_propagate_nans = true;
 202
 203                 shader.coordinate_shader_code_address =
 204                         cl_address(v3d_resource(v3d->prog.cs->resource)->bo,
 205                                    v3d->prog.cs->offset);
 206                 shader.vertex_shader_code_address =
 207                         cl_address(v3d_resource(v3d->prog.vs->resource)->bo,
 208                                    v3d->prog.vs->offset);
 209                 shader.fragment_shader_code_address =
 210                         cl_address(v3d_resource(v3d->prog.fs->resource)->bo,
 211                                    v3d->prog.fs->offset);
 212
 213                 /* XXX: Use combined input/output size flag in the common
 214                  * case.
 215                  */
 216                 shader.coordinate_shader_has_separate_input_and_output_vpm_blocks =
 217                         v3d->prog.cs->prog_data.vs->separate_segments;
 218                 shader.vertex_shader_has_separate_input_and_output_vpm_blocks =
 219                         v3d->prog.vs->prog_data.vs->separate_segments;
 220
 221                 shader.coordinate_shader_input_vpm_segment_size =
 222                         v3d->prog.cs->prog_data.vs->vpm_input_size;
 223                 shader.vertex_shader_input_vpm_segment_size =
 224                         v3d->prog.vs->prog_data.vs->vpm_input_size;
 225
 226                 shader.coordinate_shader_output_vpm_segment_size =
 227                         v3d->prog.cs->prog_data.vs->vpm_output_size;
 228                 shader.vertex_shader_output_vpm_segment_size =
 229                         v3d->prog.vs->prog_data.vs->vpm_output_size;
 230
 231                 shader.coordinate_shader_uniforms_address = cs_uniforms;
 232                 shader.vertex_shader_uniforms_address = vs_uniforms;
 233                 shader.fragment_shader_uniforms_address = fs_uniforms;
 234
 235 #if V3D_VERSION >= 41
 236                 shader.min_coord_shader_input_segments_required_in_play = 1;
 237                 shader.min_vertex_shader_input_segments_required_in_play = 1;
 238
 239                 shader.coordinate_shader_4_way_threadable =
 240                         v3d->prog.cs->prog_data.vs->base.threads == 4;
 241                 shader.vertex_shader_4_way_threadable =
 242                         v3d->prog.vs->prog_data.vs->base.threads == 4;
 243                 shader.fragment_shader_4_way_threadable =
 244                         v3d->prog.fs->prog_data.fs->base.threads == 4;
 245
 246                 shader.coordinate_shader_start_in_final_thread_section =
 247                         v3d->prog.cs->prog_data.vs->base.single_seg;
 248                 shader.vertex_shader_start_in_final_thread_section =
 249                         v3d->prog.vs->prog_data.vs->base.single_seg;
 250                 shader.fragment_shader_start_in_final_thread_section =
 251                         v3d->prog.fs->prog_data.fs->base.single_seg;
 252 #else
 253                 shader.coordinate_shader_4_way_threadable =
 254                         v3d->prog.cs->prog_data.vs->base.threads == 4;
 255                 shader.coordinate_shader_2_way_threadable =
 256                         v3d->prog.cs->prog_data.vs->base.threads == 2;
 257                 shader.vertex_shader_4_way_threadable =
 258                         v3d->prog.vs->prog_data.vs->base.threads == 4;
 259                 shader.vertex_shader_2_way_threadable =
 260                         v3d->prog.vs->prog_data.vs->base.threads == 2;
 261                 shader.fragment_shader_4_way_threadable =
 262                         v3d->prog.fs->prog_data.fs->base.threads == 4;
 263                 shader.fragment_shader_2_way_threadable =
 264                         v3d->prog.fs->prog_data.fs->base.threads == 2;
 265 #endif
 266
 267                 shader.vertex_id_read_by_coordinate_shader =
 268                         v3d->prog.cs->prog_data.vs->uses_vid;
 269                 shader.instance_id_read_by_coordinate_shader =
 270                         v3d->prog.cs->prog_data.vs->uses_iid;
 271                 shader.vertex_id_read_by_vertex_shader =
 272                         v3d->prog.vs->prog_data.vs->uses_vid;
 273                 shader.instance_id_read_by_vertex_shader =
 274                         v3d->prog.vs->prog_data.vs->uses_iid;
 275
 276                 shader.address_of_default_attribute_values =
 277                         cl_address(v3d_resource(vtx->defaults)->bo,
 278                                    vtx->defaults_offset);
 279         }
 280
 281         for (int i = 0; i < vtx->num_elements; i++) {
 282                 struct pipe_vertex_element *elem = &vtx->pipe[i];
 283                 struct pipe_vertex_buffer *vb =
 284                         &vertexbuf->vb[elem->vertex_buffer_index];
 285                 struct v3d_resource *rsc = v3d_resource(vb->buffer.resource);
 286
 287                 const uint32_t size =
 288                         cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD);
 289                 cl_emit_with_prepacked(&job->indirect,
 290                                        GL_SHADER_STATE_ATTRIBUTE_RECORD,
 291                                        &vtx->attrs[i * size], attr) {
 292                         attr.stride = vb->stride;
 293                         attr.address = cl_address(rsc->bo,
 294                                                   vb->buffer_offset +
 295                                                   elem->src_offset);
 296                         attr.number_of_values_read_by_coordinate_shader =
 297                                 v3d->prog.cs->prog_data.vs->vattr_sizes[i];
 298                         attr.number_of_values_read_by_vertex_shader =
 299                                 v3d->prog.vs->prog_data.vs->vattr_sizes[i];
 300 #if V3D_VERSION >= 41
 301                         attr.maximum_index = 0xffffff;
 302 #endif
 303                 }
 304                 STATIC_ASSERT(sizeof(vtx->attrs) >= VC5_MAX_ATTRIBUTES * size);
 305         }
 306
 307         if (vtx->num_elements == 0) {
 308                 /* GFXH-930: At least one attribute must be enabled and read
 309                  * by CS and VS.  If we have no attributes being consumed by
 310                  * the shader, set up a dummy to be loaded into the VPM.
 311                  */
 312                 cl_emit(&job->indirect, GL_SHADER_STATE_ATTRIBUTE_RECORD, attr) {
 313                         /* Valid address of data whose value will be unused. */
 314                         attr.address = cl_address(job->indirect.bo, 0);
 315
 316                         attr.type = ATTRIBUTE_FLOAT;
 317                         attr.stride = 0;
 318                         attr.vec_size = 1;
 319
 320                         attr.number_of_values_read_by_coordinate_shader = 1;
 321                         attr.number_of_values_read_by_vertex_shader = 1;
 322                 }
 323         }
 324
 325         cl_emit(&job->bcl, VCM_CACHE_SIZE, vcm) {
 326                 vcm.number_of_16_vertex_batches_for_binning =
 327                         v3d->prog.cs->prog_data.vs->vcm_cache_size;
 328                 vcm.number_of_16_vertex_batches_for_rendering =
 329                         v3d->prog.vs->prog_data.vs->vcm_cache_size;
 330         }
 331
 332         cl_emit(&job->bcl, GL_SHADER_STATE, state) {
 333                 state.address = cl_address(job->indirect.bo, shader_rec_offset);
 334                 state.number_of_attribute_arrays = num_elements_to_emit;
 335         }
 336
 337         v3d_bo_unreference(&cs_uniforms.bo);
 338         v3d_bo_unreference(&vs_uniforms.bo);
 339         v3d_bo_unreference(&fs_uniforms.bo);
 340
 341         job->shader_rec_count++;
 342 }
 343
 344 /**
 345  * Computes the various transform feedback statistics, since they can't be
 346  * recorded by CL packets.
 347  */
 348 static void
 349 v3d_tf_statistics_record(struct v3d_context *v3d,
 350                          const struct pipe_draw_info *info,
 351                          bool prim_tf)
 352 {
 353         if (!v3d->active_queries)
 354                 return;
 355
 356         uint32_t prims = u_prims_for_vertices(info->mode, info->count);
 357         v3d->prims_generated += prims;
 358
 359         if (prim_tf) {
 360                 /* XXX: Only count if we didn't overflow. */
 361                 v3d->tf_prims_generated += prims;
 362         }
 363 }
 364
 365 static void
 366 v3d_update_job_ez(struct v3d_context *v3d, struct v3d_job *job)
 367 {
 368         switch (v3d->zsa->ez_state) {
 369         case VC5_EZ_UNDECIDED:
 370                 /* If the Z/S state didn't pick a direction but didn't
 371                  * disable, then go along with the current EZ state.  This
 372                  * allows EZ optimization for Z func == EQUAL or NEVER.
 373                  */
 374                 break;
 375
 376         case VC5_EZ_LT_LE:
 377         case VC5_EZ_GT_GE:
 378                 /* If the Z/S state picked a direction, then it needs to match
 379                  * the current direction if we've decided on one.
 380                  */
 381                 if (job->ez_state == VC5_EZ_UNDECIDED)
 382                         job->ez_state = v3d->zsa->ez_state;
 383                 else if (job->ez_state != v3d->zsa->ez_state)
 384                         job->ez_state = VC5_EZ_DISABLED;
 385                 break;
 386
 387         case VC5_EZ_DISABLED:
 388                 /* If the current Z/S state disables EZ because of a bad Z
 389                  * func or stencil operation, then we can't do any more EZ in
 390                  * this frame.
 391                  */
 392                 job->ez_state = VC5_EZ_DISABLED;
 393                 break;
 394         }
 395
 396         /* If the FS affects the Z of the pixels, then it may update against
 397          * the chosen EZ direction (though we could use
 398          * ARB_conservative_depth's hints to avoid this)
 399          */
 400         if (v3d->prog.fs->prog_data.fs->writes_z) {
 401                 job->ez_state = VC5_EZ_DISABLED;
 402         }
 403
 404         if (job->first_ez_state == VC5_EZ_UNDECIDED &&
 405             (job->ez_state != VC5_EZ_DISABLED || job->draw_calls_queued == 0))
 406                 job->first_ez_state = job->ez_state;
 407 }
 408
 409 static void
 410 v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 411 {
 412         struct v3d_context *v3d = v3d_context(pctx);
 413
 414         if (!info->count_from_stream_output && !info->indirect &&
 415             !info->primitive_restart &&
 416             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
 417                 return;
 418
 419         /* Fall back for weird desktop GL primitive restart values. */
 420         if (info->primitive_restart &&
 421             info->index_size) {
 422                 uint32_t mask = ~0;
 423
 424                 switch (info->index_size) {
 425                 case 2:
 426                         mask = 0xffff;
 427                         break;
 428                 case 1:
 429                         mask = 0xff;
 430                         break;
 431                 }
 432
 433                 if (info->restart_index != mask) {
 434                         util_draw_vbo_without_prim_restart(pctx, info);
 435                         return;
 436                 }
 437         }
 438
 439         if (info->mode >= PIPE_PRIM_QUADS) {
 440                 util_primconvert_save_rasterizer_state(v3d->primconvert, &v3d->rasterizer->base);
 441                 util_primconvert_draw_vbo(v3d->primconvert, info);
 442                 perf_debug("Fallback conversion for %d %s vertices\n",
 443                            info->count, u_prim_name(info->mode));
 444                 return;
 445         }
 446
 447         /* Before setting up the draw, flush anything writing to the textures
 448          * that we read from.
 449          */
 450         for (int s = 0; s < PIPE_SHADER_TYPES; s++)
 451                 v3d_predraw_check_stage_inputs(pctx, s);
 452
 453         struct v3d_job *job = v3d_get_job_for_fbo(v3d);
 454
 455         /* If vertex texturing depends on the output of rendering, we need to
 456          * ensure that that rendering is complete before we run a coordinate
 457          * shader that depends on it.
 458          *
 459          * Given that doing that is unusual, for now we just block the binner
 460          * on the last submitted render, rather than tracking the last
 461          * rendering to each texture's BO.
 462          */
 463         if (v3d->tex[PIPE_SHADER_VERTEX].num_textures) {
 464                 perf_debug("Blocking binner on last render "
 465                            "due to vertex texturing.\n");
 466                 job->submit.in_sync_bcl = v3d->out_sync;
 467         }
 468
 469         /* Get space to emit our draw call into the BCL, using a branch to
 470          * jump to a new BO if necessary.
 471          */
 472         v3d_cl_ensure_space_with_branch(&job->bcl, 256 /* XXX */);
 473
 474         if (v3d->prim_mode != info->mode) {
 475                 v3d->prim_mode = info->mode;
 476                 v3d->dirty |= VC5_DIRTY_PRIM_MODE;
 477         }
 478
 479         v3d_start_draw(v3d);
 480         v3d_update_compiled_shaders(v3d, info->mode);
 481         v3d_update_job_ez(v3d, job);
 482
 483 #if V3D_VERSION >= 41
 484         v3d41_emit_state(pctx);
 485 #else
 486         v3d33_emit_state(pctx);
 487 #endif
 488
 489         if (v3d->dirty & (VC5_DIRTY_VTXBUF |
 490                           VC5_DIRTY_VTXSTATE |
 491                           VC5_DIRTY_PRIM_MODE |
 492                           VC5_DIRTY_RASTERIZER |
 493                           VC5_DIRTY_COMPILED_CS |
 494                           VC5_DIRTY_COMPILED_VS |
 495                           VC5_DIRTY_COMPILED_FS |
 496                           v3d->prog.cs->uniform_dirty_bits |
 497                           v3d->prog.vs->uniform_dirty_bits |
 498                           v3d->prog.fs->uniform_dirty_bits)) {
 499                 v3d_emit_gl_shader_state(v3d, info);
 500         }
 501
 502         v3d->dirty = 0;
 503
 504         /* The Base Vertex/Base Instance packet sets those values to nonzero
 505          * for the next draw call only.
 506          */
 507         if (info->index_bias || info->start_instance) {
 508                 cl_emit(&job->bcl, BASE_VERTEX_BASE_INSTANCE, base) {
 509                         base.base_instance = info->start_instance;
 510                         base.base_vertex = info->index_bias;
 511                 }
 512         }
 513
 514         uint32_t prim_tf_enable = 0;
 515 #if V3D_VERSION < 40
 516         /* V3D 3.x: The HW only processes transform feedback on primitives
 517          * with the flag set.
 518          */
 519         if (v3d->streamout.num_targets)
 520                 prim_tf_enable = (V3D_PRIM_POINTS_TF - V3D_PRIM_POINTS);
 521 #endif
 522
 523         v3d_tf_statistics_record(v3d, info, v3d->streamout.num_targets);
 524
 525         /* Note that the primitive type fields match with OpenGL/gallium
 526          * definitions, up to but not including QUADS.
 527          */
 528         if (info->index_size) {
 529                 uint32_t index_size = info->index_size;
 530                 uint32_t offset = info->start * index_size;
 531                 struct pipe_resource *prsc;
 532                 if (info->has_user_indices) {
 533                         prsc = NULL;
 534                         u_upload_data(v3d->uploader, 0,
 535                                       info->count * info->index_size, 4,
 536                                       info->index.user,
 537                                       &offset, &prsc);
 538                 } else {
 539                         prsc = info->index.resource;
 540                 }
 541                 struct v3d_resource *rsc = v3d_resource(prsc);
 542
 543 #if V3D_VERSION >= 40
 544                 cl_emit(&job->bcl, INDEX_BUFFER_SETUP, ib) {
 545                         ib.address = cl_address(rsc->bo, 0);
 546                         ib.size = rsc->bo->size;
 547                 }
 548 #endif
 549
 550                 if (info->instance_count > 1) {
 551                         cl_emit(&job->bcl, INDEXED_INSTANCED_PRIM_LIST, prim) {
 552                                 prim.index_type = ffs(info->index_size) - 1;
 553 #if V3D_VERSION >= 40
 554                                 prim.index_offset = offset;
 555 #else /* V3D_VERSION < 40 */
 556                                 prim.maximum_index = (1u << 31) - 1; /* XXX */
 557                                 prim.address_of_indices_list =
 558                                         cl_address(rsc->bo, offset);
 559 #endif /* V3D_VERSION < 40 */
 560                                 prim.mode = info->mode | prim_tf_enable;
 561                                 prim.enable_primitive_restarts = info->primitive_restart;
 562
 563                                 prim.number_of_instances = info->instance_count;
 564                                 prim.instance_length = info->count;
 565                         }
 566                 } else {
 567                         cl_emit(&job->bcl, INDEXED_PRIM_LIST, prim) {
 568                                 prim.index_type = ffs(info->index_size) - 1;
 569                                 prim.length = info->count;
 570 #if V3D_VERSION >= 40
 571                                 prim.index_offset = offset;
 572 #else /* V3D_VERSION < 40 */
 573                                 prim.maximum_index = (1u << 31) - 1; /* XXX */
 574                                 prim.address_of_indices_list =
 575                                         cl_address(rsc->bo, offset);
 576 #endif /* V3D_VERSION < 40 */
 577                                 prim.mode = info->mode | prim_tf_enable;
 578                                 prim.enable_primitive_restarts = info->primitive_restart;
 579                         }
 580                 }
 581
 582                 job->draw_calls_queued++;
 583
 584                 if (info->has_user_indices)
 585                         pipe_resource_reference(&prsc, NULL);
 586         } else {
 587                 if (info->instance_count > 1) {
 588                         cl_emit(&job->bcl, VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
 589                                 prim.mode = info->mode | prim_tf_enable;
 590                                 prim.index_of_first_vertex = info->start;
 591                                 prim.number_of_instances = info->instance_count;
 592                                 prim.instance_length = info->count;
 593                         }
 594                 } else {
 595                         cl_emit(&job->bcl, VERTEX_ARRAY_PRIMS, prim) {
 596                                 prim.mode = info->mode | prim_tf_enable;
 597                                 prim.length = info->count;
 598                                 prim.index_of_first_vertex = info->start;
 599                         }
 600                 }
 601         }
 602
 603         /* A flush is required in between a TF draw and any following TF specs
 604          * packet, or the GPU may hang.  Just flush each time for now.
 605          */
 606         if (v3d->streamout.num_targets)
 607                 cl_emit(&job->bcl, TRANSFORM_FEEDBACK_FLUSH_AND_COUNT, flush);
 608
 609         job->draw_calls_queued++;
 610
 611         /* Increment the TF offsets by how many verts we wrote.  XXX: This
 612          * needs some clamping to the buffer size.
 613          */
 614         for (int i = 0; i < v3d->streamout.num_targets; i++)
 615                 v3d->streamout.offsets[i] += info->count;
 616
 617         if (v3d->zsa && job->zsbuf && v3d->zsa->base.depth.enabled) {
 618                 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
 619                 v3d_job_add_bo(job, rsc->bo);
 620
 621                 job->load |= PIPE_CLEAR_DEPTH & ~job->clear;
 622                 if (v3d->zsa->base.depth.writemask)
 623                         job->store |= PIPE_CLEAR_DEPTH;
 624                 rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
 625         }
 626
 627         if (v3d->zsa && job->zsbuf && v3d->zsa->base.stencil[0].enabled) {
 628                 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
 629                 if (rsc->separate_stencil)
 630                         rsc = rsc->separate_stencil;
 631
 632                 v3d_job_add_bo(job, rsc->bo);
 633
 634                 job->load |= PIPE_CLEAR_STENCIL & ~job->clear;
 635                 if (v3d->zsa->base.stencil[0].writemask ||
 636                     v3d->zsa->base.stencil[1].writemask) {
 637                         job->store |= PIPE_CLEAR_STENCIL;
 638                 }
 639                 rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
 640         }
 641
 642         for (int i = 0; i < VC5_MAX_DRAW_BUFFERS; i++) {
 643                 uint32_t bit = PIPE_CLEAR_COLOR0 << i;
 644                 int blend_rt = v3d->blend->base.independent_blend_enable ? i : 0;
 645
 646                 if (job->store & bit || !job->cbufs[i])
 647                         continue;
 648                 struct v3d_resource *rsc = v3d_resource(job->cbufs[i]->texture);
 649
 650                 job->load |= bit & ~job->clear;
 651                 if (v3d->blend->base.rt[blend_rt].colormask)
 652                         job->store |= bit;
 653                 v3d_job_add_bo(job, rsc->bo);
 654         }
 655
 656         if (job->referenced_size > 768 * 1024 * 1024) {
 657                 perf_debug("Flushing job with %dkb to try to free up memory\n",
 658                         job->referenced_size / 1024);
 659                 v3d_flush(pctx);
 660         }
 661
 662         if (V3D_DEBUG & V3D_DEBUG_ALWAYS_FLUSH)
 663                 v3d_flush(pctx);
 664 }
 665
 666 /**
 667  * Implements gallium's clear() hook (glClear()) by drawing a pair of triangles.
 668  */
 669 static void
 670 v3d_draw_clear(struct v3d_context *v3d,
 671                unsigned buffers,
 672                const union pipe_color_union *color,
 673                double depth, unsigned stencil)
 674 {
 675         static const union pipe_color_union dummy_color = {};
 676
 677         /* The blitter util dereferences the color regardless, even though the
 678          * gallium clear API may not pass one in when only Z/S are cleared.
 679          */
 680         if (!color)
 681                 color = &dummy_color;
 682
 683         v3d_blitter_save(v3d);
 684         util_blitter_clear(v3d->blitter,
 685                            v3d->framebuffer.width,
 686                            v3d->framebuffer.height,
 687                            util_framebuffer_get_num_layers(&v3d->framebuffer),
 688                            buffers, color, depth, stencil);
 689 }
 690
 691 /**
 692  * Attempts to perform the GL clear by using the TLB's fast clear at the start
 693  * of the frame.
 694  */
 695 static unsigned
 696 v3d_tlb_clear(struct v3d_job *job, unsigned buffers,
 697               const union pipe_color_union *color,
 698               double depth, unsigned stencil)
 699 {
 700         struct v3d_context *v3d = job->v3d;
 701
 702         if (job->draw_calls_queued) {
 703                 /* If anything in the CL has drawn using the buffer, then the
 704                  * TLB clear we're trying to add now would happen before that
 705                  * drawing.
 706                  */
 707                 buffers &= ~(job->load | job->store);
 708         }
 709
 710         /* GFXH-1461: If we were to emit a load of just depth or just stencil,
 711          * then the clear for the other may get lost.  We need to decide now
 712          * if it would be possible to need to emit a load of just one after
 713          * we've set up our TLB clears.
 714          */
 715         if (buffers & PIPE_CLEAR_DEPTHSTENCIL &&
 716             (buffers & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL &&
 717             job->zsbuf &&
 718             util_format_is_depth_and_stencil(job->zsbuf->texture->format)) {
 719                 buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
 720         }
 721
 722         for (int i = 0; i < VC5_MAX_DRAW_BUFFERS; i++) {
 723                 uint32_t bit = PIPE_CLEAR_COLOR0 << i;
 724                 if (!(buffers & bit))
 725                         continue;
 726
 727                 struct pipe_surface *psurf = v3d->framebuffer.cbufs[i];
 728                 struct v3d_surface *surf = v3d_surface(psurf);
 729                 struct v3d_resource *rsc = v3d_resource(psurf->texture);
 730
 731                 union util_color uc;
 732                 uint32_t internal_size = 4 << surf->internal_bpp;
 733
 734                 static union pipe_color_union swapped_color;
 735                 if (v3d->swap_color_rb & (1 << i)) {
 736                         swapped_color.f[0] = color->f[2];
 737                         swapped_color.f[1] = color->f[1];
 738                         swapped_color.f[2] = color->f[0];
 739                         swapped_color.f[3] = color->f[3];
 740                         color = &swapped_color;
 741                 }
 742
 743                 switch (surf->internal_type) {
 744                 case V3D_INTERNAL_TYPE_8:
 745                         util_pack_color(color->f, PIPE_FORMAT_R8G8B8A8_UNORM,
 746                                         &uc);
 747                         memcpy(job->clear_color[i], uc.ui, internal_size);
 748                         break;
 749                 case V3D_INTERNAL_TYPE_8I:
 750                 case V3D_INTERNAL_TYPE_8UI:
 751                         job->clear_color[i][0] = ((color->ui[0] & 0xff) |
 752                                                   (color->ui[1] & 0xff) << 8 |
 753                                                   (color->ui[2] & 0xff) << 16 |
 754                                                   (color->ui[3] & 0xff) << 24);
 755                         break;
 756                 case V3D_INTERNAL_TYPE_16F:
 757                         util_pack_color(color->f, PIPE_FORMAT_R16G16B16A16_FLOAT,
 758                                         &uc);
 759                         memcpy(job->clear_color[i], uc.ui, internal_size);
 760                         break;
 761                 case V3D_INTERNAL_TYPE_16I:
 762                 case V3D_INTERNAL_TYPE_16UI:
 763                         job->clear_color[i][0] = ((color->ui[0] & 0xffff) |
 764                                                   color->ui[1] << 16);
 765                         job->clear_color[i][1] = ((color->ui[2] & 0xffff) |
 766                                                   color->ui[3] << 16);
 767                         break;
 768                 case V3D_INTERNAL_TYPE_32F:
 769                 case V3D_INTERNAL_TYPE_32I:
 770                 case V3D_INTERNAL_TYPE_32UI:
 771                         memcpy(job->clear_color[i], color->ui, internal_size);
 772                         break;
 773                 }
 774
 775                 rsc->initialized_buffers |= bit;
 776         }
 777
 778         unsigned zsclear = buffers & PIPE_CLEAR_DEPTHSTENCIL;
 779         if (zsclear) {
 780                 struct v3d_resource *rsc =
 781                         v3d_resource(v3d->framebuffer.zsbuf->texture);
 782
 783                 if (zsclear & PIPE_CLEAR_DEPTH)
 784                         job->clear_z = depth;
 785                 if (zsclear & PIPE_CLEAR_STENCIL)
 786                         job->clear_s = stencil;
 787
 788                 rsc->initialized_buffers |= zsclear;
 789         }
 790
 791         job->draw_min_x = 0;
 792         job->draw_min_y = 0;
 793         job->draw_max_x = v3d->framebuffer.width;
 794         job->draw_max_y = v3d->framebuffer.height;
 795         job->clear |= buffers;
 796         job->store |= buffers;
 797
 798         v3d_start_draw(v3d);
 799
 800         return buffers;
 801 }
 802
 803 static void
 804 v3d_clear(struct pipe_context *pctx, unsigned buffers,
 805           const union pipe_color_union *color, double depth, unsigned stencil)
 806 {
 807         struct v3d_context *v3d = v3d_context(pctx);
 808         struct v3d_job *job = v3d_get_job_for_fbo(v3d);
 809
 810         buffers &= ~v3d_tlb_clear(job, buffers, color, depth, stencil);
 811
 812         if (buffers)
 813                 v3d_draw_clear(v3d, buffers, color, depth, stencil);
 814 }
 815
 816 static void
 817 v3d_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 818                         const union pipe_color_union *color,
 819                         unsigned x, unsigned y, unsigned w, unsigned h,
 820                         bool render_condition_enabled)
 821 {
 822         fprintf(stderr, "unimpl: clear RT\n");
 823 }
 824
 825 static void
 826 v3d_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 827                         unsigned buffers, double depth, unsigned stencil,
 828                         unsigned x, unsigned y, unsigned w, unsigned h,
 829                         bool render_condition_enabled)
 830 {
 831         fprintf(stderr, "unimpl: clear DS\n");
 832 }
 833
 834 void
 835 v3dX(draw_init)(struct pipe_context *pctx)
 836 {
 837         pctx->draw_vbo = v3d_draw_vbo;
 838         pctx->clear = v3d_clear;
 839         pctx->clear_render_target = v3d_clear_render_target;
 840         pctx->clear_depth_stencil = v3d_clear_depth_stencil;
 841 }