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