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         /* XXX perf: If we're reading from the output of TF in this job, we
 128          * should instead be using the wait for transform feedback
 129          * functionality.
 130          */
 131
 132         /* Flush writes to textures we're sampling. */
 133         for (int i = 0; i < v3d->tex[s].num_textures; i++) {
 134                 struct pipe_sampler_view *pview = v3d->tex[s].textures[i];
 135                 if (!pview)
 136                         continue;
 137                 struct v3d_sampler_view *view = v3d_sampler_view(pview);
 138
 139                 if (view->texture != view->base.texture)
 140                         v3d_update_shadow_texture(pctx, &view->base);
 141
 142                 v3d_flush_jobs_writing_resource(v3d, view->texture);
 143         }
 144
 145         /* Flush writes to UBOs. */
 146         foreach_bit(i, v3d->constbuf[s].enabled_mask) {
 147                 struct pipe_constant_buffer *cb = &v3d->constbuf[s].cb[i];
 148                 if (cb->buffer)
 149                         v3d_flush_jobs_writing_resource(v3d, cb->buffer);
 150         }
 151
 152         /* Flush writes to our image views */
 153         foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
 154                 struct v3d_image_view *view = &v3d->shaderimg[s].si[i];
 155
 156                 v3d_flush_jobs_writing_resource(v3d, view->base.resource);
 157         }
 158 }
 159
 160 static void
 161 v3d_emit_gl_shader_state(struct v3d_context *v3d,
 162                          const struct pipe_draw_info *info)
 163 {
 164         struct v3d_job *job = v3d->job;
 165         /* VC5_DIRTY_VTXSTATE */
 166         struct v3d_vertex_stateobj *vtx = v3d->vtx;
 167         /* VC5_DIRTY_VTXBUF */
 168         struct v3d_vertexbuf_stateobj *vertexbuf = &v3d->vertexbuf;
 169
 170         /* Upload the uniforms to the indirect CL first */
 171         struct v3d_cl_reloc fs_uniforms =
 172                 v3d_write_uniforms(v3d, v3d->prog.fs,
 173                                    PIPE_SHADER_FRAGMENT);
 174         struct v3d_cl_reloc vs_uniforms =
 175                 v3d_write_uniforms(v3d, v3d->prog.vs,
 176                                    PIPE_SHADER_VERTEX);
 177         struct v3d_cl_reloc cs_uniforms =
 178                 v3d_write_uniforms(v3d, v3d->prog.cs,
 179                                    PIPE_SHADER_VERTEX);
 180
 181         /* See GFXH-930 workaround below */
 182         uint32_t num_elements_to_emit = MAX2(vtx->num_elements, 1);
 183         uint32_t shader_rec_offset =
 184                 v3d_cl_ensure_space(&job->indirect,
 185                                     cl_packet_length(GL_SHADER_STATE_RECORD) +
 186                                     num_elements_to_emit *
 187                                     cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD),
 188                                     32);
 189
 190         /* XXX perf: We should move most of the SHADER_STATE_RECORD setup to
 191          * compile time, so that we mostly just have to OR the VS and FS
 192          * records together at draw time.
 193          */
 194         cl_emit(&job->indirect, GL_SHADER_STATE_RECORD, shader) {
 195                 shader.enable_clipping = true;
 196                 /* VC5_DIRTY_PRIM_MODE | VC5_DIRTY_RASTERIZER */
 197                 shader.point_size_in_shaded_vertex_data =
 198                         (info->mode == PIPE_PRIM_POINTS &&
 199                          v3d->rasterizer->base.point_size_per_vertex);
 200
 201                 /* Must be set if the shader modifies Z, discards, or modifies
 202                  * the sample mask.  For any of these cases, the fragment
 203                  * shader needs to write the Z value (even just discards).
 204                  */
 205                 shader.fragment_shader_does_z_writes =
 206                         (v3d->prog.fs->prog_data.fs->writes_z ||
 207                          v3d->prog.fs->prog_data.fs->discard);
 208
 209                 shader.fragment_shader_uses_real_pixel_centre_w_in_addition_to_centroid_w2 =
 210                         v3d->prog.fs->prog_data.fs->uses_center_w;
 211
 212                 shader.number_of_varyings_in_fragment_shader =
 213                         v3d->prog.fs->prog_data.base->num_inputs;
 214
 215                 shader.coordinate_shader_propagate_nans = true;
 216                 shader.vertex_shader_propagate_nans = true;
 217                 shader.fragment_shader_propagate_nans = true;
 218
 219                 shader.coordinate_shader_code_address =
 220                         cl_address(v3d_resource(v3d->prog.cs->resource)->bo,
 221                                    v3d->prog.cs->offset);
 222                 shader.vertex_shader_code_address =
 223                         cl_address(v3d_resource(v3d->prog.vs->resource)->bo,
 224                                    v3d->prog.vs->offset);
 225                 shader.fragment_shader_code_address =
 226                         cl_address(v3d_resource(v3d->prog.fs->resource)->bo,
 227                                    v3d->prog.fs->offset);
 228
 229                 /* XXX: Use combined input/output size flag in the common
 230                  * case.
 231                  */
 232                 shader.coordinate_shader_has_separate_input_and_output_vpm_blocks =
 233                         v3d->prog.cs->prog_data.vs->separate_segments;
 234                 shader.vertex_shader_has_separate_input_and_output_vpm_blocks =
 235                         v3d->prog.vs->prog_data.vs->separate_segments;
 236
 237                 shader.coordinate_shader_input_vpm_segment_size =
 238                         v3d->prog.cs->prog_data.vs->vpm_input_size;
 239                 shader.vertex_shader_input_vpm_segment_size =
 240                         v3d->prog.vs->prog_data.vs->vpm_input_size;
 241
 242                 shader.coordinate_shader_output_vpm_segment_size =
 243                         v3d->prog.cs->prog_data.vs->vpm_output_size;
 244                 shader.vertex_shader_output_vpm_segment_size =
 245                         v3d->prog.vs->prog_data.vs->vpm_output_size;
 246
 247                 shader.coordinate_shader_uniforms_address = cs_uniforms;
 248                 shader.vertex_shader_uniforms_address = vs_uniforms;
 249                 shader.fragment_shader_uniforms_address = fs_uniforms;
 250
 251 #if V3D_VERSION >= 41
 252                 shader.min_coord_shader_input_segments_required_in_play = 1;
 253                 shader.min_vertex_shader_input_segments_required_in_play = 1;
 254
 255                 shader.coordinate_shader_4_way_threadable =
 256                         v3d->prog.cs->prog_data.vs->base.threads == 4;
 257                 shader.vertex_shader_4_way_threadable =
 258                         v3d->prog.vs->prog_data.vs->base.threads == 4;
 259                 shader.fragment_shader_4_way_threadable =
 260                         v3d->prog.fs->prog_data.fs->base.threads == 4;
 261
 262                 shader.coordinate_shader_start_in_final_thread_section =
 263                         v3d->prog.cs->prog_data.vs->base.single_seg;
 264                 shader.vertex_shader_start_in_final_thread_section =
 265                         v3d->prog.vs->prog_data.vs->base.single_seg;
 266                 shader.fragment_shader_start_in_final_thread_section =
 267                         v3d->prog.fs->prog_data.fs->base.single_seg;
 268 #else
 269                 shader.coordinate_shader_4_way_threadable =
 270                         v3d->prog.cs->prog_data.vs->base.threads == 4;
 271                 shader.coordinate_shader_2_way_threadable =
 272                         v3d->prog.cs->prog_data.vs->base.threads == 2;
 273                 shader.vertex_shader_4_way_threadable =
 274                         v3d->prog.vs->prog_data.vs->base.threads == 4;
 275                 shader.vertex_shader_2_way_threadable =
 276                         v3d->prog.vs->prog_data.vs->base.threads == 2;
 277                 shader.fragment_shader_4_way_threadable =
 278                         v3d->prog.fs->prog_data.fs->base.threads == 4;
 279                 shader.fragment_shader_2_way_threadable =
 280                         v3d->prog.fs->prog_data.fs->base.threads == 2;
 281 #endif
 282
 283                 shader.vertex_id_read_by_coordinate_shader =
 284                         v3d->prog.cs->prog_data.vs->uses_vid;
 285                 shader.instance_id_read_by_coordinate_shader =
 286                         v3d->prog.cs->prog_data.vs->uses_iid;
 287                 shader.vertex_id_read_by_vertex_shader =
 288                         v3d->prog.vs->prog_data.vs->uses_vid;
 289                 shader.instance_id_read_by_vertex_shader =
 290                         v3d->prog.vs->prog_data.vs->uses_iid;
 291
 292                 shader.address_of_default_attribute_values =
 293                         cl_address(v3d_resource(vtx->defaults)->bo,
 294                                    vtx->defaults_offset);
 295         }
 296
 297         for (int i = 0; i < vtx->num_elements; i++) {
 298                 struct pipe_vertex_element *elem = &vtx->pipe[i];
 299                 struct pipe_vertex_buffer *vb =
 300                         &vertexbuf->vb[elem->vertex_buffer_index];
 301                 struct v3d_resource *rsc = v3d_resource(vb->buffer.resource);
 302
 303                 const uint32_t size =
 304                         cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD);
 305                 cl_emit_with_prepacked(&job->indirect,
 306                                        GL_SHADER_STATE_ATTRIBUTE_RECORD,
 307                                        &vtx->attrs[i * size], attr) {
 308                         attr.stride = vb->stride;
 309                         attr.address = cl_address(rsc->bo,
 310                                                   vb->buffer_offset +
 311                                                   elem->src_offset);
 312                         attr.number_of_values_read_by_coordinate_shader =
 313                                 v3d->prog.cs->prog_data.vs->vattr_sizes[i];
 314                         attr.number_of_values_read_by_vertex_shader =
 315                                 v3d->prog.vs->prog_data.vs->vattr_sizes[i];
 316 #if V3D_VERSION >= 41
 317                         attr.maximum_index = 0xffffff;
 318 #endif
 319                 }
 320                 STATIC_ASSERT(sizeof(vtx->attrs) >= VC5_MAX_ATTRIBUTES * size);
 321         }
 322
 323         if (vtx->num_elements == 0) {
 324                 /* GFXH-930: At least one attribute must be enabled and read
 325                  * by CS and VS.  If we have no attributes being consumed by
 326                  * the shader, set up a dummy to be loaded into the VPM.
 327                  */
 328                 cl_emit(&job->indirect, GL_SHADER_STATE_ATTRIBUTE_RECORD, attr) {
 329                         /* Valid address of data whose value will be unused. */
 330                         attr.address = cl_address(job->indirect.bo, 0);
 331
 332                         attr.type = ATTRIBUTE_FLOAT;
 333                         attr.stride = 0;
 334                         attr.vec_size = 1;
 335
 336                         attr.number_of_values_read_by_coordinate_shader = 1;
 337                         attr.number_of_values_read_by_vertex_shader = 1;
 338                 }
 339         }
 340
 341         cl_emit(&job->bcl, VCM_CACHE_SIZE, vcm) {
 342                 vcm.number_of_16_vertex_batches_for_binning =
 343                         v3d->prog.cs->prog_data.vs->vcm_cache_size;
 344                 vcm.number_of_16_vertex_batches_for_rendering =
 345                         v3d->prog.vs->prog_data.vs->vcm_cache_size;
 346         }
 347
 348         cl_emit(&job->bcl, GL_SHADER_STATE, state) {
 349                 state.address = cl_address(job->indirect.bo, shader_rec_offset);
 350                 state.number_of_attribute_arrays = num_elements_to_emit;
 351         }
 352
 353         v3d_bo_unreference(&cs_uniforms.bo);
 354         v3d_bo_unreference(&vs_uniforms.bo);
 355         v3d_bo_unreference(&fs_uniforms.bo);
 356
 357         job->shader_rec_count++;
 358 }
 359
 360 /**
 361  * Computes the various transform feedback statistics, since they can't be
 362  * recorded by CL packets.
 363  */
 364 static void
 365 v3d_tf_statistics_record(struct v3d_context *v3d,
 366                          const struct pipe_draw_info *info,
 367                          bool prim_tf)
 368 {
 369         if (!v3d->active_queries)
 370                 return;
 371
 372         uint32_t prims = u_prims_for_vertices(info->mode, info->count);
 373         v3d->prims_generated += prims;
 374
 375         if (prim_tf) {
 376                 /* XXX: Only count if we didn't overflow. */
 377                 v3d->tf_prims_generated += prims;
 378         }
 379 }
 380
 381 static void
 382 v3d_update_job_ez(struct v3d_context *v3d, struct v3d_job *job)
 383 {
 384         switch (v3d->zsa->ez_state) {
 385         case VC5_EZ_UNDECIDED:
 386                 /* If the Z/S state didn't pick a direction but didn't
 387                  * disable, then go along with the current EZ state.  This
 388                  * allows EZ optimization for Z func == EQUAL or NEVER.
 389                  */
 390                 break;
 391
 392         case VC5_EZ_LT_LE:
 393         case VC5_EZ_GT_GE:
 394                 /* If the Z/S state picked a direction, then it needs to match
 395                  * the current direction if we've decided on one.
 396                  */
 397                 if (job->ez_state == VC5_EZ_UNDECIDED)
 398                         job->ez_state = v3d->zsa->ez_state;
 399                 else if (job->ez_state != v3d->zsa->ez_state)
 400                         job->ez_state = VC5_EZ_DISABLED;
 401                 break;
 402
 403         case VC5_EZ_DISABLED:
 404                 /* If the current Z/S state disables EZ because of a bad Z
 405                  * func or stencil operation, then we can't do any more EZ in
 406                  * this frame.
 407                  */
 408                 job->ez_state = VC5_EZ_DISABLED;
 409                 break;
 410         }
 411
 412         /* If the FS affects the Z of the pixels, then it may update against
 413          * the chosen EZ direction (though we could use
 414          * ARB_conservative_depth's hints to avoid this)
 415          */
 416         if (v3d->prog.fs->prog_data.fs->writes_z) {
 417                 job->ez_state = VC5_EZ_DISABLED;
 418         }
 419
 420         if (job->first_ez_state == VC5_EZ_UNDECIDED &&
 421             (job->ez_state != VC5_EZ_DISABLED || job->draw_calls_queued == 0))
 422                 job->first_ez_state = job->ez_state;
 423 }
 424
 425 static void
 426 v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 427 {
 428         struct v3d_context *v3d = v3d_context(pctx);
 429
 430         if (!info->count_from_stream_output && !info->indirect &&
 431             !info->primitive_restart &&
 432             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
 433                 return;
 434
 435         /* Fall back for weird desktop GL primitive restart values. */
 436         if (info->primitive_restart &&
 437             info->index_size) {
 438                 uint32_t mask = ~0;
 439
 440                 switch (info->index_size) {
 441                 case 2:
 442                         mask = 0xffff;
 443                         break;
 444                 case 1:
 445                         mask = 0xff;
 446                         break;
 447                 }
 448
 449                 if (info->restart_index != mask) {
 450                         util_draw_vbo_without_prim_restart(pctx, info);
 451                         return;
 452                 }
 453         }
 454
 455         if (info->mode >= PIPE_PRIM_QUADS) {
 456                 util_primconvert_save_rasterizer_state(v3d->primconvert, &v3d->rasterizer->base);
 457                 util_primconvert_draw_vbo(v3d->primconvert, info);
 458                 perf_debug("Fallback conversion for %d %s vertices\n",
 459                            info->count, u_prim_name(info->mode));
 460                 return;
 461         }
 462
 463         /* Before setting up the draw, flush anything writing to the textures
 464          * that we read from.
 465          */
 466         for (int s = 0; s < PIPE_SHADER_TYPES; s++)
 467                 v3d_predraw_check_stage_inputs(pctx, s);
 468
 469         if (info->indirect)
 470                 v3d_flush_jobs_writing_resource(v3d, info->indirect->buffer);
 471
 472         struct v3d_job *job = v3d_get_job_for_fbo(v3d);
 473
 474         /* If vertex texturing depends on the output of rendering, we need to
 475          * ensure that that rendering is complete before we run a coordinate
 476          * shader that depends on it.
 477          *
 478          * Given that doing that is unusual, for now we just block the binner
 479          * on the last submitted render, rather than tracking the last
 480          * rendering to each texture's BO.
 481          */
 482         if (v3d->tex[PIPE_SHADER_VERTEX].num_textures) {
 483                 perf_debug("Blocking binner on last render "
 484                            "due to vertex texturing.\n");
 485                 job->submit.in_sync_bcl = v3d->out_sync;
 486         }
 487
 488         /* Mark SSBOs as being written.  We don't actually know which ones are
 489          * read vs written, so just assume the worst
 490          */
 491         for (int s = 0; s < PIPE_SHADER_TYPES; s++) {
 492                 foreach_bit(i, v3d->ssbo[s].enabled_mask) {
 493                         v3d_job_add_write_resource(job,
 494                                                    v3d->ssbo[s].sb[i].buffer);
 495                         job->tmu_dirty_rcl = true;
 496                 }
 497
 498                 foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
 499                         v3d_job_add_write_resource(job,
 500                                                    v3d->shaderimg[s].si[i].base.resource);
 501                         job->tmu_dirty_rcl = true;
 502                 }
 503         }
 504
 505         /* Get space to emit our draw call into the BCL, using a branch to
 506          * jump to a new BO if necessary.
 507          */
 508         v3d_cl_ensure_space_with_branch(&job->bcl, 256 /* XXX */);
 509
 510         if (v3d->prim_mode != info->mode) {
 511                 v3d->prim_mode = info->mode;
 512                 v3d->dirty |= VC5_DIRTY_PRIM_MODE;
 513         }
 514
 515         v3d_start_draw(v3d);
 516         v3d_update_compiled_shaders(v3d, info->mode);
 517         v3d_update_job_ez(v3d, job);
 518
 519 #if V3D_VERSION >= 41
 520         v3d41_emit_state(pctx);
 521 #else
 522         v3d33_emit_state(pctx);
 523 #endif
 524
 525         if (v3d->dirty & (VC5_DIRTY_VTXBUF |
 526                           VC5_DIRTY_VTXSTATE |
 527                           VC5_DIRTY_PRIM_MODE |
 528                           VC5_DIRTY_RASTERIZER |
 529                           VC5_DIRTY_COMPILED_CS |
 530                           VC5_DIRTY_COMPILED_VS |
 531                           VC5_DIRTY_COMPILED_FS |
 532                           v3d->prog.cs->uniform_dirty_bits |
 533                           v3d->prog.vs->uniform_dirty_bits |
 534                           v3d->prog.fs->uniform_dirty_bits)) {
 535                 v3d_emit_gl_shader_state(v3d, info);
 536         }
 537
 538         v3d->dirty = 0;
 539
 540         /* The Base Vertex/Base Instance packet sets those values to nonzero
 541          * for the next draw call only.
 542          */
 543         if (info->index_bias || info->start_instance) {
 544                 cl_emit(&job->bcl, BASE_VERTEX_BASE_INSTANCE, base) {
 545                         base.base_instance = info->start_instance;
 546                         base.base_vertex = info->index_bias;
 547                 }
 548         }
 549
 550         uint32_t prim_tf_enable = 0;
 551 #if V3D_VERSION < 40
 552         /* V3D 3.x: The HW only processes transform feedback on primitives
 553          * with the flag set.
 554          */
 555         if (v3d->streamout.num_targets)
 556                 prim_tf_enable = (V3D_PRIM_POINTS_TF - V3D_PRIM_POINTS);
 557 #endif
 558
 559         v3d_tf_statistics_record(v3d, info, v3d->streamout.num_targets);
 560
 561         /* Note that the primitive type fields match with OpenGL/gallium
 562          * definitions, up to but not including QUADS.
 563          */
 564         if (info->index_size) {
 565                 uint32_t index_size = info->index_size;
 566                 uint32_t offset = info->start * index_size;
 567                 struct pipe_resource *prsc;
 568                 if (info->has_user_indices) {
 569                         prsc = NULL;
 570                         u_upload_data(v3d->uploader, 0,
 571                                       info->count * info->index_size, 4,
 572                                       info->index.user,
 573                                       &offset, &prsc);
 574                 } else {
 575                         prsc = info->index.resource;
 576                 }
 577                 struct v3d_resource *rsc = v3d_resource(prsc);
 578
 579 #if V3D_VERSION >= 40
 580                 cl_emit(&job->bcl, INDEX_BUFFER_SETUP, ib) {
 581                         ib.address = cl_address(rsc->bo, 0);
 582                         ib.size = rsc->bo->size;
 583                 }
 584 #endif
 585
 586                 if (info->indirect) {
 587                         cl_emit(&job->bcl, INDIRECT_INDEXED_INSTANCED_PRIM_LIST, prim) {
 588                                 prim.index_type = ffs(info->index_size) - 1;
 589 #if V3D_VERSION < 40
 590                                 prim.address_of_indices_list =
 591                                         cl_address(rsc->bo, offset);
 592 #endif /* V3D_VERSION < 40 */
 593                                 prim.mode = info->mode | prim_tf_enable;
 594                                 prim.enable_primitive_restarts = info->primitive_restart;
 595
 596                                 prim.number_of_draw_indirect_indexed_records = info->indirect->draw_count;
 597
 598                                 prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
 599                                 prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
 600                                                           info->indirect->offset);
 601                         }
 602                 } else if (info->instance_count > 1) {
 603                         cl_emit(&job->bcl, INDEXED_INSTANCED_PRIM_LIST, prim) {
 604                                 prim.index_type = ffs(info->index_size) - 1;
 605 #if V3D_VERSION >= 40
 606                                 prim.index_offset = offset;
 607 #else /* V3D_VERSION < 40 */
 608                                 prim.maximum_index = (1u << 31) - 1; /* XXX */
 609                                 prim.address_of_indices_list =
 610                                         cl_address(rsc->bo, offset);
 611 #endif /* V3D_VERSION < 40 */
 612                                 prim.mode = info->mode | prim_tf_enable;
 613                                 prim.enable_primitive_restarts = info->primitive_restart;
 614
 615                                 prim.number_of_instances = info->instance_count;
 616                                 prim.instance_length = info->count;
 617                         }
 618                 } else {
 619                         cl_emit(&job->bcl, INDEXED_PRIM_LIST, prim) {
 620                                 prim.index_type = ffs(info->index_size) - 1;
 621                                 prim.length = info->count;
 622 #if V3D_VERSION >= 40
 623                                 prim.index_offset = offset;
 624 #else /* V3D_VERSION < 40 */
 625                                 prim.maximum_index = (1u << 31) - 1; /* XXX */
 626                                 prim.address_of_indices_list =
 627                                         cl_address(rsc->bo, offset);
 628 #endif /* V3D_VERSION < 40 */
 629                                 prim.mode = info->mode | prim_tf_enable;
 630                                 prim.enable_primitive_restarts = info->primitive_restart;
 631                         }
 632                 }
 633
 634                 job->draw_calls_queued++;
 635
 636                 if (info->has_user_indices)
 637                         pipe_resource_reference(&prsc, NULL);
 638         } else {
 639                 if (info->indirect) {
 640                         cl_emit(&job->bcl, INDIRECT_VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
 641                                 prim.mode = info->mode | prim_tf_enable;
 642                                 prim.number_of_draw_indirect_array_records = info->indirect->draw_count;
 643
 644                                 prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
 645                                 prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
 646                                                           info->indirect->offset);
 647                         }
 648                 } else if (info->instance_count > 1) {
 649                         cl_emit(&job->bcl, VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
 650                                 prim.mode = info->mode | prim_tf_enable;
 651                                 prim.index_of_first_vertex = info->start;
 652                                 prim.number_of_instances = info->instance_count;
 653                                 prim.instance_length = info->count;
 654                         }
 655                 } else {
 656                         cl_emit(&job->bcl, VERTEX_ARRAY_PRIMS, prim) {
 657                                 prim.mode = info->mode | prim_tf_enable;
 658                                 prim.length = info->count;
 659                                 prim.index_of_first_vertex = info->start;
 660                         }
 661                 }
 662         }
 663
 664         /* A flush is required in between a TF draw and any following TF specs
 665          * packet, or the GPU may hang.  Just flush each time for now.
 666          */
 667         if (v3d->streamout.num_targets)
 668                 cl_emit(&job->bcl, TRANSFORM_FEEDBACK_FLUSH_AND_COUNT, flush);
 669
 670         job->draw_calls_queued++;
 671
 672         /* Increment the TF offsets by how many verts we wrote.  XXX: This
 673          * needs some clamping to the buffer size.
 674          */
 675         for (int i = 0; i < v3d->streamout.num_targets; i++)
 676                 v3d->streamout.offsets[i] += info->count;
 677
 678         if (v3d->zsa && job->zsbuf && v3d->zsa->base.depth.enabled) {
 679                 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
 680                 v3d_job_add_bo(job, rsc->bo);
 681
 682                 job->load |= PIPE_CLEAR_DEPTH & ~job->clear;
 683                 if (v3d->zsa->base.depth.writemask)
 684                         job->store |= PIPE_CLEAR_DEPTH;
 685                 rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
 686         }
 687
 688         if (v3d->zsa && job->zsbuf && v3d->zsa->base.stencil[0].enabled) {
 689                 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
 690                 if (rsc->separate_stencil)
 691                         rsc = rsc->separate_stencil;
 692
 693                 v3d_job_add_bo(job, rsc->bo);
 694
 695                 job->load |= PIPE_CLEAR_STENCIL & ~job->clear;
 696                 if (v3d->zsa->base.stencil[0].writemask ||
 697                     v3d->zsa->base.stencil[1].writemask) {
 698                         job->store |= PIPE_CLEAR_STENCIL;
 699                 }
 700                 rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
 701         }
 702
 703         for (int i = 0; i < VC5_MAX_DRAW_BUFFERS; i++) {
 704                 uint32_t bit = PIPE_CLEAR_COLOR0 << i;
 705                 int blend_rt = v3d->blend->base.independent_blend_enable ? i : 0;
 706
 707                 if (job->store & bit || !job->cbufs[i])
 708                         continue;
 709                 struct v3d_resource *rsc = v3d_resource(job->cbufs[i]->texture);
 710
 711                 job->load |= bit & ~job->clear;
 712                 if (v3d->blend->base.rt[blend_rt].colormask)
 713                         job->store |= bit;
 714                 v3d_job_add_bo(job, rsc->bo);
 715         }
 716
 717         if (job->referenced_size > 768 * 1024 * 1024) {
 718                 perf_debug("Flushing job with %dkb to try to free up memory\n",
 719                         job->referenced_size / 1024);
 720                 v3d_flush(pctx);
 721         }
 722
 723         if (V3D_DEBUG & V3D_DEBUG_ALWAYS_FLUSH)
 724                 v3d_flush(pctx);
 725 }
 726
 727 /**
 728  * Implements gallium's clear() hook (glClear()) by drawing a pair of triangles.
 729  */
 730 static void
 731 v3d_draw_clear(struct v3d_context *v3d,
 732                unsigned buffers,
 733                const union pipe_color_union *color,
 734                double depth, unsigned stencil)
 735 {
 736         static const union pipe_color_union dummy_color = {};
 737
 738         /* The blitter util dereferences the color regardless, even though the
 739          * gallium clear API may not pass one in when only Z/S are cleared.
 740          */
 741         if (!color)
 742                 color = &dummy_color;
 743
 744         v3d_blitter_save(v3d);
 745         util_blitter_clear(v3d->blitter,
 746                            v3d->framebuffer.width,
 747                            v3d->framebuffer.height,
 748                            util_framebuffer_get_num_layers(&v3d->framebuffer),
 749                            buffers, color, depth, stencil);
 750 }
 751
 752 /**
 753  * Attempts to perform the GL clear by using the TLB's fast clear at the start
 754  * of the frame.
 755  */
 756 static unsigned
 757 v3d_tlb_clear(struct v3d_job *job, unsigned buffers,
 758               const union pipe_color_union *color,
 759               double depth, unsigned stencil)
 760 {
 761         struct v3d_context *v3d = job->v3d;
 762
 763         if (job->draw_calls_queued) {
 764                 /* If anything in the CL has drawn using the buffer, then the
 765                  * TLB clear we're trying to add now would happen before that
 766                  * drawing.
 767                  */
 768                 buffers &= ~(job->load | job->store);
 769         }
 770
 771         /* GFXH-1461: If we were to emit a load of just depth or just stencil,
 772          * then the clear for the other may get lost.  We need to decide now
 773          * if it would be possible to need to emit a load of just one after
 774          * we've set up our TLB clears.
 775          */
 776         if (buffers & PIPE_CLEAR_DEPTHSTENCIL &&
 777             (buffers & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL &&
 778             job->zsbuf &&
 779             util_format_is_depth_and_stencil(job->zsbuf->texture->format)) {
 780                 buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
 781         }
 782
 783         for (int i = 0; i < VC5_MAX_DRAW_BUFFERS; i++) {
 784                 uint32_t bit = PIPE_CLEAR_COLOR0 << i;
 785                 if (!(buffers & bit))
 786                         continue;
 787
 788                 struct pipe_surface *psurf = v3d->framebuffer.cbufs[i];
 789                 struct v3d_surface *surf = v3d_surface(psurf);
 790                 struct v3d_resource *rsc = v3d_resource(psurf->texture);
 791
 792                 union util_color uc;
 793                 uint32_t internal_size = 4 << surf->internal_bpp;
 794
 795                 static union pipe_color_union swapped_color;
 796                 if (v3d->swap_color_rb & (1 << i)) {
 797                         swapped_color.f[0] = color->f[2];
 798                         swapped_color.f[1] = color->f[1];
 799                         swapped_color.f[2] = color->f[0];
 800                         swapped_color.f[3] = color->f[3];
 801                         color = &swapped_color;
 802                 }
 803
 804                 switch (surf->internal_type) {
 805                 case V3D_INTERNAL_TYPE_8:
 806                         util_pack_color(color->f, PIPE_FORMAT_R8G8B8A8_UNORM,
 807                                         &uc);
 808                         memcpy(job->clear_color[i], uc.ui, internal_size);
 809                         break;
 810                 case V3D_INTERNAL_TYPE_8I:
 811                 case V3D_INTERNAL_TYPE_8UI:
 812                         job->clear_color[i][0] = ((color->ui[0] & 0xff) |
 813                                                   (color->ui[1] & 0xff) << 8 |
 814                                                   (color->ui[2] & 0xff) << 16 |
 815                                                   (color->ui[3] & 0xff) << 24);
 816                         break;
 817                 case V3D_INTERNAL_TYPE_16F:
 818                         util_pack_color(color->f, PIPE_FORMAT_R16G16B16A16_FLOAT,
 819                                         &uc);
 820                         memcpy(job->clear_color[i], uc.ui, internal_size);
 821                         break;
 822                 case V3D_INTERNAL_TYPE_16I:
 823                 case V3D_INTERNAL_TYPE_16UI:
 824                         job->clear_color[i][0] = ((color->ui[0] & 0xffff) |
 825                                                   color->ui[1] << 16);
 826                         job->clear_color[i][1] = ((color->ui[2] & 0xffff) |
 827                                                   color->ui[3] << 16);
 828                         break;
 829                 case V3D_INTERNAL_TYPE_32F:
 830                 case V3D_INTERNAL_TYPE_32I:
 831                 case V3D_INTERNAL_TYPE_32UI:
 832                         memcpy(job->clear_color[i], color->ui, internal_size);
 833                         break;
 834                 }
 835
 836                 rsc->initialized_buffers |= bit;
 837         }
 838
 839         unsigned zsclear = buffers & PIPE_CLEAR_DEPTHSTENCIL;
 840         if (zsclear) {
 841                 struct v3d_resource *rsc =
 842                         v3d_resource(v3d->framebuffer.zsbuf->texture);
 843
 844                 if (zsclear & PIPE_CLEAR_DEPTH)
 845                         job->clear_z = depth;
 846                 if (zsclear & PIPE_CLEAR_STENCIL)
 847                         job->clear_s = stencil;
 848
 849                 rsc->initialized_buffers |= zsclear;
 850         }
 851
 852         job->draw_min_x = 0;
 853         job->draw_min_y = 0;
 854         job->draw_max_x = v3d->framebuffer.width;
 855         job->draw_max_y = v3d->framebuffer.height;
 856         job->clear |= buffers;
 857         job->store |= buffers;
 858
 859         v3d_start_draw(v3d);
 860
 861         return buffers;
 862 }
 863
 864 static void
 865 v3d_clear(struct pipe_context *pctx, unsigned buffers,
 866           const union pipe_color_union *color, double depth, unsigned stencil)
 867 {
 868         struct v3d_context *v3d = v3d_context(pctx);
 869         struct v3d_job *job = v3d_get_job_for_fbo(v3d);
 870
 871         buffers &= ~v3d_tlb_clear(job, buffers, color, depth, stencil);
 872
 873         if (buffers)
 874                 v3d_draw_clear(v3d, buffers, color, depth, stencil);
 875 }
 876
 877 static void
 878 v3d_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 879                         const union pipe_color_union *color,
 880                         unsigned x, unsigned y, unsigned w, unsigned h,
 881                         bool render_condition_enabled)
 882 {
 883         fprintf(stderr, "unimpl: clear RT\n");
 884 }
 885
 886 static void
 887 v3d_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 888                         unsigned buffers, double depth, unsigned stencil,
 889                         unsigned x, unsigned y, unsigned w, unsigned h,
 890                         bool render_condition_enabled)
 891 {
 892         fprintf(stderr, "unimpl: clear DS\n");
 893 }
 894
 895 void
 896 v3dX(draw_init)(struct pipe_context *pctx)
 897 {
 898         pctx->draw_vbo = v3d_draw_vbo;
 899         pctx->clear = v3d_clear;
 900         pctx->clear_render_target = v3d_clear_render_target;
 901         pctx->clear_depth_stencil = v3d_clear_depth_stencil;
 902 }