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                 /* Set if the EZ test must be disabled (due to shader side
 208                  * effects and the early_z flag not being present in the
 209                  * shader).
 210                  */
 211                 shader.turn_off_early_z_test =
 212                         v3d->prog.fs->prog_data.fs->disable_ez;
 213
 214                 shader.fragment_shader_uses_real_pixel_centre_w_in_addition_to_centroid_w2 =
 215                         v3d->prog.fs->prog_data.fs->uses_center_w;
 216
 217                 shader.number_of_varyings_in_fragment_shader =
 218                         v3d->prog.fs->prog_data.base->num_inputs;
 219
 220                 shader.coordinate_shader_propagate_nans = true;
 221                 shader.vertex_shader_propagate_nans = true;
 222                 shader.fragment_shader_propagate_nans = true;
 223
 224                 shader.coordinate_shader_code_address =
 225                         cl_address(v3d_resource(v3d->prog.cs->resource)->bo,
 226                                    v3d->prog.cs->offset);
 227                 shader.vertex_shader_code_address =
 228                         cl_address(v3d_resource(v3d->prog.vs->resource)->bo,
 229                                    v3d->prog.vs->offset);
 230                 shader.fragment_shader_code_address =
 231                         cl_address(v3d_resource(v3d->prog.fs->resource)->bo,
 232                                    v3d->prog.fs->offset);
 233
 234                 /* XXX: Use combined input/output size flag in the common
 235                  * case.
 236                  */
 237                 shader.coordinate_shader_has_separate_input_and_output_vpm_blocks =
 238                         v3d->prog.cs->prog_data.vs->separate_segments;
 239                 shader.vertex_shader_has_separate_input_and_output_vpm_blocks =
 240                         v3d->prog.vs->prog_data.vs->separate_segments;
 241
 242                 shader.coordinate_shader_input_vpm_segment_size =
 243                         v3d->prog.cs->prog_data.vs->vpm_input_size;
 244                 shader.vertex_shader_input_vpm_segment_size =
 245                         v3d->prog.vs->prog_data.vs->vpm_input_size;
 246
 247                 shader.coordinate_shader_output_vpm_segment_size =
 248                         v3d->prog.cs->prog_data.vs->vpm_output_size;
 249                 shader.vertex_shader_output_vpm_segment_size =
 250                         v3d->prog.vs->prog_data.vs->vpm_output_size;
 251
 252                 shader.coordinate_shader_uniforms_address = cs_uniforms;
 253                 shader.vertex_shader_uniforms_address = vs_uniforms;
 254                 shader.fragment_shader_uniforms_address = fs_uniforms;
 255
 256 #if V3D_VERSION >= 41
 257                 shader.min_coord_shader_input_segments_required_in_play = 1;
 258                 shader.min_vertex_shader_input_segments_required_in_play = 1;
 259
 260                 shader.coordinate_shader_4_way_threadable =
 261                         v3d->prog.cs->prog_data.vs->base.threads == 4;
 262                 shader.vertex_shader_4_way_threadable =
 263                         v3d->prog.vs->prog_data.vs->base.threads == 4;
 264                 shader.fragment_shader_4_way_threadable =
 265                         v3d->prog.fs->prog_data.fs->base.threads == 4;
 266
 267                 shader.coordinate_shader_start_in_final_thread_section =
 268                         v3d->prog.cs->prog_data.vs->base.single_seg;
 269                 shader.vertex_shader_start_in_final_thread_section =
 270                         v3d->prog.vs->prog_data.vs->base.single_seg;
 271                 shader.fragment_shader_start_in_final_thread_section =
 272                         v3d->prog.fs->prog_data.fs->base.single_seg;
 273 #else
 274                 shader.coordinate_shader_4_way_threadable =
 275                         v3d->prog.cs->prog_data.vs->base.threads == 4;
 276                 shader.coordinate_shader_2_way_threadable =
 277                         v3d->prog.cs->prog_data.vs->base.threads == 2;
 278                 shader.vertex_shader_4_way_threadable =
 279                         v3d->prog.vs->prog_data.vs->base.threads == 4;
 280                 shader.vertex_shader_2_way_threadable =
 281                         v3d->prog.vs->prog_data.vs->base.threads == 2;
 282                 shader.fragment_shader_4_way_threadable =
 283                         v3d->prog.fs->prog_data.fs->base.threads == 4;
 284                 shader.fragment_shader_2_way_threadable =
 285                         v3d->prog.fs->prog_data.fs->base.threads == 2;
 286 #endif
 287
 288                 shader.vertex_id_read_by_coordinate_shader =
 289                         v3d->prog.cs->prog_data.vs->uses_vid;
 290                 shader.instance_id_read_by_coordinate_shader =
 291                         v3d->prog.cs->prog_data.vs->uses_iid;
 292                 shader.vertex_id_read_by_vertex_shader =
 293                         v3d->prog.vs->prog_data.vs->uses_vid;
 294                 shader.instance_id_read_by_vertex_shader =
 295                         v3d->prog.vs->prog_data.vs->uses_iid;
 296
 297                 shader.address_of_default_attribute_values =
 298                         cl_address(v3d_resource(vtx->defaults)->bo,
 299                                    vtx->defaults_offset);
 300         }
 301
 302         for (int i = 0; i < vtx->num_elements; i++) {
 303                 struct pipe_vertex_element *elem = &vtx->pipe[i];
 304                 struct pipe_vertex_buffer *vb =
 305                         &vertexbuf->vb[elem->vertex_buffer_index];
 306                 struct v3d_resource *rsc = v3d_resource(vb->buffer.resource);
 307
 308                 const uint32_t size =
 309                         cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD);
 310                 cl_emit_with_prepacked(&job->indirect,
 311                                        GL_SHADER_STATE_ATTRIBUTE_RECORD,
 312                                        &vtx->attrs[i * size], attr) {
 313                         attr.stride = vb->stride;
 314                         attr.address = cl_address(rsc->bo,
 315                                                   vb->buffer_offset +
 316                                                   elem->src_offset);
 317                         attr.number_of_values_read_by_coordinate_shader =
 318                                 v3d->prog.cs->prog_data.vs->vattr_sizes[i];
 319                         attr.number_of_values_read_by_vertex_shader =
 320                                 v3d->prog.vs->prog_data.vs->vattr_sizes[i];
 321 #if V3D_VERSION >= 41
 322                         attr.maximum_index = 0xffffff;
 323 #endif
 324                 }
 325                 STATIC_ASSERT(sizeof(vtx->attrs) >= V3D_MAX_VS_INPUTS / 4 * size);
 326         }
 327
 328         if (vtx->num_elements == 0) {
 329                 /* GFXH-930: At least one attribute must be enabled and read
 330                  * by CS and VS.  If we have no attributes being consumed by
 331                  * the shader, set up a dummy to be loaded into the VPM.
 332                  */
 333                 cl_emit(&job->indirect, GL_SHADER_STATE_ATTRIBUTE_RECORD, attr) {
 334                         /* Valid address of data whose value will be unused. */
 335                         attr.address = cl_address(job->indirect.bo, 0);
 336
 337                         attr.type = ATTRIBUTE_FLOAT;
 338                         attr.stride = 0;
 339                         attr.vec_size = 1;
 340
 341                         attr.number_of_values_read_by_coordinate_shader = 1;
 342                         attr.number_of_values_read_by_vertex_shader = 1;
 343                 }
 344         }
 345
 346         cl_emit(&job->bcl, VCM_CACHE_SIZE, vcm) {
 347                 vcm.number_of_16_vertex_batches_for_binning =
 348                         v3d->prog.cs->prog_data.vs->vcm_cache_size;
 349                 vcm.number_of_16_vertex_batches_for_rendering =
 350                         v3d->prog.vs->prog_data.vs->vcm_cache_size;
 351         }
 352
 353         cl_emit(&job->bcl, GL_SHADER_STATE, state) {
 354                 state.address = cl_address(job->indirect.bo, shader_rec_offset);
 355                 state.number_of_attribute_arrays = num_elements_to_emit;
 356         }
 357
 358         v3d_bo_unreference(&cs_uniforms.bo);
 359         v3d_bo_unreference(&vs_uniforms.bo);
 360         v3d_bo_unreference(&fs_uniforms.bo);
 361
 362         job->shader_rec_count++;
 363 }
 364
 365 /**
 366  * Computes the various transform feedback statistics, since they can't be
 367  * recorded by CL packets.
 368  */
 369 static void
 370 v3d_tf_statistics_record(struct v3d_context *v3d,
 371                          const struct pipe_draw_info *info,
 372                          bool prim_tf)
 373 {
 374         if (!v3d->active_queries)
 375                 return;
 376
 377         uint32_t prims = u_prims_for_vertices(info->mode, info->count);
 378         v3d->prims_generated += prims;
 379
 380         if (prim_tf) {
 381                 /* XXX: Only count if we didn't overflow. */
 382                 v3d->tf_prims_generated += prims;
 383         }
 384 }
 385
 386 static void
 387 v3d_update_job_ez(struct v3d_context *v3d, struct v3d_job *job)
 388 {
 389         switch (v3d->zsa->ez_state) {
 390         case VC5_EZ_UNDECIDED:
 391                 /* If the Z/S state didn't pick a direction but didn't
 392                  * disable, then go along with the current EZ state.  This
 393                  * allows EZ optimization for Z func == EQUAL or NEVER.
 394                  */
 395                 break;
 396
 397         case VC5_EZ_LT_LE:
 398         case VC5_EZ_GT_GE:
 399                 /* If the Z/S state picked a direction, then it needs to match
 400                  * the current direction if we've decided on one.
 401                  */
 402                 if (job->ez_state == VC5_EZ_UNDECIDED)
 403                         job->ez_state = v3d->zsa->ez_state;
 404                 else if (job->ez_state != v3d->zsa->ez_state)
 405                         job->ez_state = VC5_EZ_DISABLED;
 406                 break;
 407
 408         case VC5_EZ_DISABLED:
 409                 /* If the current Z/S state disables EZ because of a bad Z
 410                  * func or stencil operation, then we can't do any more EZ in
 411                  * this frame.
 412                  */
 413                 job->ez_state = VC5_EZ_DISABLED;
 414                 break;
 415         }
 416
 417         /* If the FS affects the Z of the pixels, then it may update against
 418          * the chosen EZ direction (though we could use
 419          * ARB_conservative_depth's hints to avoid this)
 420          */
 421         if (v3d->prog.fs->prog_data.fs->writes_z) {
 422                 job->ez_state = VC5_EZ_DISABLED;
 423         }
 424
 425         if (job->first_ez_state == VC5_EZ_UNDECIDED &&
 426             (job->ez_state != VC5_EZ_DISABLED || job->draw_calls_queued == 0))
 427                 job->first_ez_state = job->ez_state;
 428 }
 429
 430 static void
 431 v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 432 {
 433         struct v3d_context *v3d = v3d_context(pctx);
 434
 435         if (!info->count_from_stream_output && !info->indirect &&
 436             !info->primitive_restart &&
 437             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
 438                 return;
 439
 440         /* Fall back for weird desktop GL primitive restart values. */
 441         if (info->primitive_restart &&
 442             info->index_size) {
 443                 uint32_t mask = ~0;
 444
 445                 switch (info->index_size) {
 446                 case 2:
 447                         mask = 0xffff;
 448                         break;
 449                 case 1:
 450                         mask = 0xff;
 451                         break;
 452                 }
 453
 454                 if (info->restart_index != mask) {
 455                         util_draw_vbo_without_prim_restart(pctx, info);
 456                         return;
 457                 }
 458         }
 459
 460         if (info->mode >= PIPE_PRIM_QUADS) {
 461                 util_primconvert_save_rasterizer_state(v3d->primconvert, &v3d->rasterizer->base);
 462                 util_primconvert_draw_vbo(v3d->primconvert, info);
 463                 perf_debug("Fallback conversion for %d %s vertices\n",
 464                            info->count, u_prim_name(info->mode));
 465                 return;
 466         }
 467
 468         /* Before setting up the draw, flush anything writing to the textures
 469          * that we read from.
 470          */
 471         for (int s = 0; s < PIPE_SHADER_TYPES; s++)
 472                 v3d_predraw_check_stage_inputs(pctx, s);
 473
 474         if (info->indirect)
 475                 v3d_flush_jobs_writing_resource(v3d, info->indirect->buffer);
 476
 477         struct v3d_job *job = v3d_get_job_for_fbo(v3d);
 478
 479         /* If vertex texturing depends on the output of rendering, we need to
 480          * ensure that that rendering is complete before we run a coordinate
 481          * shader that depends on it.
 482          *
 483          * Given that doing that is unusual, for now we just block the binner
 484          * on the last submitted render, rather than tracking the last
 485          * rendering to each texture's BO.
 486          */
 487         if (v3d->tex[PIPE_SHADER_VERTEX].num_textures || info->indirect) {
 488                 perf_debug("Blocking binner on last render "
 489                            "due to vertex texturing or indirect drawing.\n");
 490                 job->submit.in_sync_bcl = v3d->out_sync;
 491         }
 492
 493         /* Mark SSBOs as being written.  We don't actually know which ones are
 494          * read vs written, so just assume the worst
 495          */
 496         for (int s = 0; s < PIPE_SHADER_TYPES; s++) {
 497                 foreach_bit(i, v3d->ssbo[s].enabled_mask) {
 498                         v3d_job_add_write_resource(job,
 499                                                    v3d->ssbo[s].sb[i].buffer);
 500                         job->tmu_dirty_rcl = true;
 501                 }
 502
 503                 foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
 504                         v3d_job_add_write_resource(job,
 505                                                    v3d->shaderimg[s].si[i].base.resource);
 506                         job->tmu_dirty_rcl = true;
 507                 }
 508         }
 509
 510         /* Get space to emit our draw call into the BCL, using a branch to
 511          * jump to a new BO if necessary.
 512          */
 513         v3d_cl_ensure_space_with_branch(&job->bcl, 256 /* XXX */);
 514
 515         if (v3d->prim_mode != info->mode) {
 516                 v3d->prim_mode = info->mode;
 517                 v3d->dirty |= VC5_DIRTY_PRIM_MODE;
 518         }
 519
 520         v3d_start_draw(v3d);
 521         v3d_update_compiled_shaders(v3d, info->mode);
 522         v3d_update_job_ez(v3d, job);
 523
 524 #if V3D_VERSION >= 41
 525         v3d41_emit_state(pctx);
 526 #else
 527         v3d33_emit_state(pctx);
 528 #endif
 529
 530         if (v3d->dirty & (VC5_DIRTY_VTXBUF |
 531                           VC5_DIRTY_VTXSTATE |
 532                           VC5_DIRTY_PRIM_MODE |
 533                           VC5_DIRTY_RASTERIZER |
 534                           VC5_DIRTY_COMPILED_CS |
 535                           VC5_DIRTY_COMPILED_VS |
 536                           VC5_DIRTY_COMPILED_FS |
 537                           v3d->prog.cs->uniform_dirty_bits |
 538                           v3d->prog.vs->uniform_dirty_bits |
 539                           v3d->prog.fs->uniform_dirty_bits)) {
 540                 v3d_emit_gl_shader_state(v3d, info);
 541         }
 542
 543         v3d->dirty = 0;
 544
 545         /* The Base Vertex/Base Instance packet sets those values to nonzero
 546          * for the next draw call only.
 547          */
 548         if (info->index_bias || info->start_instance) {
 549                 cl_emit(&job->bcl, BASE_VERTEX_BASE_INSTANCE, base) {
 550                         base.base_instance = info->start_instance;
 551                         base.base_vertex = info->index_bias;
 552                 }
 553         }
 554
 555         uint32_t prim_tf_enable = 0;
 556 #if V3D_VERSION < 40
 557         /* V3D 3.x: The HW only processes transform feedback on primitives
 558          * with the flag set.
 559          */
 560         if (v3d->streamout.num_targets)
 561                 prim_tf_enable = (V3D_PRIM_POINTS_TF - V3D_PRIM_POINTS);
 562 #endif
 563
 564         v3d_tf_statistics_record(v3d, info, v3d->streamout.num_targets);
 565
 566         /* Note that the primitive type fields match with OpenGL/gallium
 567          * definitions, up to but not including QUADS.
 568          */
 569         if (info->index_size) {
 570                 uint32_t index_size = info->index_size;
 571                 uint32_t offset = info->start * index_size;
 572                 struct pipe_resource *prsc;
 573                 if (info->has_user_indices) {
 574                         prsc = NULL;
 575                         u_upload_data(v3d->uploader, 0,
 576                                       info->count * info->index_size, 4,
 577                                       info->index.user,
 578                                       &offset, &prsc);
 579                 } else {
 580                         prsc = info->index.resource;
 581                 }
 582                 struct v3d_resource *rsc = v3d_resource(prsc);
 583
 584 #if V3D_VERSION >= 40
 585                 cl_emit(&job->bcl, INDEX_BUFFER_SETUP, ib) {
 586                         ib.address = cl_address(rsc->bo, 0);
 587                         ib.size = rsc->bo->size;
 588                 }
 589 #endif
 590
 591                 if (info->indirect) {
 592                         cl_emit(&job->bcl, INDIRECT_INDEXED_INSTANCED_PRIM_LIST, prim) {
 593                                 prim.index_type = ffs(info->index_size) - 1;
 594 #if V3D_VERSION < 40
 595                                 prim.address_of_indices_list =
 596                                         cl_address(rsc->bo, offset);
 597 #endif /* V3D_VERSION < 40 */
 598                                 prim.mode = info->mode | prim_tf_enable;
 599                                 prim.enable_primitive_restarts = info->primitive_restart;
 600
 601                                 prim.number_of_draw_indirect_indexed_records = info->indirect->draw_count;
 602
 603                                 prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
 604                                 prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
 605                                                           info->indirect->offset);
 606                         }
 607                 } else if (info->instance_count > 1) {
 608                         cl_emit(&job->bcl, INDEXED_INSTANCED_PRIM_LIST, prim) {
 609                                 prim.index_type = ffs(info->index_size) - 1;
 610 #if V3D_VERSION >= 40
 611                                 prim.index_offset = offset;
 612 #else /* V3D_VERSION < 40 */
 613                                 prim.maximum_index = (1u << 31) - 1; /* XXX */
 614                                 prim.address_of_indices_list =
 615                                         cl_address(rsc->bo, offset);
 616 #endif /* V3D_VERSION < 40 */
 617                                 prim.mode = info->mode | prim_tf_enable;
 618                                 prim.enable_primitive_restarts = info->primitive_restart;
 619
 620                                 prim.number_of_instances = info->instance_count;
 621                                 prim.instance_length = info->count;
 622                         }
 623                 } else {
 624                         cl_emit(&job->bcl, INDEXED_PRIM_LIST, prim) {
 625                                 prim.index_type = ffs(info->index_size) - 1;
 626                                 prim.length = info->count;
 627 #if V3D_VERSION >= 40
 628                                 prim.index_offset = offset;
 629 #else /* V3D_VERSION < 40 */
 630                                 prim.maximum_index = (1u << 31) - 1; /* XXX */
 631                                 prim.address_of_indices_list =
 632                                         cl_address(rsc->bo, offset);
 633 #endif /* V3D_VERSION < 40 */
 634                                 prim.mode = info->mode | prim_tf_enable;
 635                                 prim.enable_primitive_restarts = info->primitive_restart;
 636                         }
 637                 }
 638
 639                 job->draw_calls_queued++;
 640
 641                 if (info->has_user_indices)
 642                         pipe_resource_reference(&prsc, NULL);
 643         } else {
 644                 if (info->indirect) {
 645                         cl_emit(&job->bcl, INDIRECT_VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
 646                                 prim.mode = info->mode | prim_tf_enable;
 647                                 prim.number_of_draw_indirect_array_records = info->indirect->draw_count;
 648
 649                                 prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
 650                                 prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
 651                                                           info->indirect->offset);
 652                         }
 653                 } else if (info->instance_count > 1) {
 654                         cl_emit(&job->bcl, VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
 655                                 prim.mode = info->mode | prim_tf_enable;
 656                                 prim.index_of_first_vertex = info->start;
 657                                 prim.number_of_instances = info->instance_count;
 658                                 prim.instance_length = info->count;
 659                         }
 660                 } else {
 661                         cl_emit(&job->bcl, VERTEX_ARRAY_PRIMS, prim) {
 662                                 prim.mode = info->mode | prim_tf_enable;
 663                                 prim.length = info->count;
 664                                 prim.index_of_first_vertex = info->start;
 665                         }
 666                 }
 667         }
 668
 669         /* A flush is required in between a TF draw and any following TF specs
 670          * packet, or the GPU may hang.  Just flush each time for now.
 671          */
 672         if (v3d->streamout.num_targets)
 673                 cl_emit(&job->bcl, TRANSFORM_FEEDBACK_FLUSH_AND_COUNT, flush);
 674
 675         job->draw_calls_queued++;
 676
 677         /* Increment the TF offsets by how many verts we wrote.  XXX: This
 678          * needs some clamping to the buffer size.
 679          */
 680         for (int i = 0; i < v3d->streamout.num_targets; i++)
 681                 v3d->streamout.offsets[i] += info->count;
 682
 683         if (v3d->zsa && job->zsbuf && v3d->zsa->base.depth.enabled) {
 684                 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
 685                 v3d_job_add_bo(job, rsc->bo);
 686
 687                 job->load |= PIPE_CLEAR_DEPTH & ~job->clear;
 688                 if (v3d->zsa->base.depth.writemask)
 689                         job->store |= PIPE_CLEAR_DEPTH;
 690                 rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
 691         }
 692
 693         if (v3d->zsa && job->zsbuf && v3d->zsa->base.stencil[0].enabled) {
 694                 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
 695                 if (rsc->separate_stencil)
 696                         rsc = rsc->separate_stencil;
 697
 698                 v3d_job_add_bo(job, rsc->bo);
 699
 700                 job->load |= PIPE_CLEAR_STENCIL & ~job->clear;
 701                 if (v3d->zsa->base.stencil[0].writemask ||
 702                     v3d->zsa->base.stencil[1].writemask) {
 703                         job->store |= PIPE_CLEAR_STENCIL;
 704                 }
 705                 rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
 706         }
 707
 708         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
 709                 uint32_t bit = PIPE_CLEAR_COLOR0 << i;
 710                 int blend_rt = v3d->blend->base.independent_blend_enable ? i : 0;
 711
 712                 if (job->store & bit || !job->cbufs[i])
 713                         continue;
 714                 struct v3d_resource *rsc = v3d_resource(job->cbufs[i]->texture);
 715
 716                 job->load |= bit & ~job->clear;
 717                 if (v3d->blend->base.rt[blend_rt].colormask)
 718                         job->store |= bit;
 719                 v3d_job_add_bo(job, rsc->bo);
 720         }
 721
 722         if (job->referenced_size > 768 * 1024 * 1024) {
 723                 perf_debug("Flushing job with %dkb to try to free up memory\n",
 724                         job->referenced_size / 1024);
 725                 v3d_flush(pctx);
 726         }
 727
 728         if (V3D_DEBUG & V3D_DEBUG_ALWAYS_FLUSH)
 729                 v3d_flush(pctx);
 730 }
 731
 732 /**
 733  * Implements gallium's clear() hook (glClear()) by drawing a pair of triangles.
 734  */
 735 static void
 736 v3d_draw_clear(struct v3d_context *v3d,
 737                unsigned buffers,
 738                const union pipe_color_union *color,
 739                double depth, unsigned stencil)
 740 {
 741         static const union pipe_color_union dummy_color = {};
 742
 743         /* The blitter util dereferences the color regardless, even though the
 744          * gallium clear API may not pass one in when only Z/S are cleared.
 745          */
 746         if (!color)
 747                 color = &dummy_color;
 748
 749         v3d_blitter_save(v3d);
 750         util_blitter_clear(v3d->blitter,
 751                            v3d->framebuffer.width,
 752                            v3d->framebuffer.height,
 753                            util_framebuffer_get_num_layers(&v3d->framebuffer),
 754                            buffers, color, depth, stencil);
 755 }
 756
 757 /**
 758  * Attempts to perform the GL clear by using the TLB's fast clear at the start
 759  * of the frame.
 760  */
 761 static unsigned
 762 v3d_tlb_clear(struct v3d_job *job, unsigned buffers,
 763               const union pipe_color_union *color,
 764               double depth, unsigned stencil)
 765 {
 766         struct v3d_context *v3d = job->v3d;
 767
 768         if (job->draw_calls_queued) {
 769                 /* If anything in the CL has drawn using the buffer, then the
 770                  * TLB clear we're trying to add now would happen before that
 771                  * drawing.
 772                  */
 773                 buffers &= ~(job->load | job->store);
 774         }
 775
 776         /* GFXH-1461: If we were to emit a load of just depth or just stencil,
 777          * then the clear for the other may get lost.  We need to decide now
 778          * if it would be possible to need to emit a load of just one after
 779          * we've set up our TLB clears.
 780          */
 781         if (buffers & PIPE_CLEAR_DEPTHSTENCIL &&
 782             (buffers & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL &&
 783             job->zsbuf &&
 784             util_format_is_depth_and_stencil(job->zsbuf->texture->format)) {
 785                 buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
 786         }
 787
 788         for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
 789                 uint32_t bit = PIPE_CLEAR_COLOR0 << i;
 790                 if (!(buffers & bit))
 791                         continue;
 792
 793                 struct pipe_surface *psurf = v3d->framebuffer.cbufs[i];
 794                 struct v3d_surface *surf = v3d_surface(psurf);
 795                 struct v3d_resource *rsc = v3d_resource(psurf->texture);
 796
 797                 union util_color uc;
 798                 uint32_t internal_size = 4 << surf->internal_bpp;
 799
 800                 static union pipe_color_union swapped_color;
 801                 if (v3d->swap_color_rb & (1 << i)) {
 802                         swapped_color.f[0] = color->f[2];
 803                         swapped_color.f[1] = color->f[1];
 804                         swapped_color.f[2] = color->f[0];
 805                         swapped_color.f[3] = color->f[3];
 806                         color = &swapped_color;
 807                 }
 808
 809                 switch (surf->internal_type) {
 810                 case V3D_INTERNAL_TYPE_8:
 811                         util_pack_color(color->f, PIPE_FORMAT_R8G8B8A8_UNORM,
 812                                         &uc);
 813                         memcpy(job->clear_color[i], uc.ui, internal_size);
 814                         break;
 815                 case V3D_INTERNAL_TYPE_8I:
 816                 case V3D_INTERNAL_TYPE_8UI:
 817                         job->clear_color[i][0] = ((color->ui[0] & 0xff) |
 818                                                   (color->ui[1] & 0xff) << 8 |
 819                                                   (color->ui[2] & 0xff) << 16 |
 820                                                   (color->ui[3] & 0xff) << 24);
 821                         break;
 822                 case V3D_INTERNAL_TYPE_16F:
 823                         util_pack_color(color->f, PIPE_FORMAT_R16G16B16A16_FLOAT,
 824                                         &uc);
 825                         memcpy(job->clear_color[i], uc.ui, internal_size);
 826                         break;
 827                 case V3D_INTERNAL_TYPE_16I:
 828                 case V3D_INTERNAL_TYPE_16UI:
 829                         job->clear_color[i][0] = ((color->ui[0] & 0xffff) |
 830                                                   color->ui[1] << 16);
 831                         job->clear_color[i][1] = ((color->ui[2] & 0xffff) |
 832                                                   color->ui[3] << 16);
 833                         break;
 834                 case V3D_INTERNAL_TYPE_32F:
 835                 case V3D_INTERNAL_TYPE_32I:
 836                 case V3D_INTERNAL_TYPE_32UI:
 837                         memcpy(job->clear_color[i], color->ui, internal_size);
 838                         break;
 839                 }
 840
 841                 rsc->initialized_buffers |= bit;
 842         }
 843
 844         unsigned zsclear = buffers & PIPE_CLEAR_DEPTHSTENCIL;
 845         if (zsclear) {
 846                 struct v3d_resource *rsc =
 847                         v3d_resource(v3d->framebuffer.zsbuf->texture);
 848
 849                 if (zsclear & PIPE_CLEAR_DEPTH)
 850                         job->clear_z = depth;
 851                 if (zsclear & PIPE_CLEAR_STENCIL)
 852                         job->clear_s = stencil;
 853
 854                 rsc->initialized_buffers |= zsclear;
 855         }
 856
 857         job->draw_min_x = 0;
 858         job->draw_min_y = 0;
 859         job->draw_max_x = v3d->framebuffer.width;
 860         job->draw_max_y = v3d->framebuffer.height;
 861         job->clear |= buffers;
 862         job->store |= buffers;
 863
 864         v3d_start_draw(v3d);
 865
 866         return buffers;
 867 }
 868
 869 static void
 870 v3d_clear(struct pipe_context *pctx, unsigned buffers,
 871           const union pipe_color_union *color, double depth, unsigned stencil)
 872 {
 873         struct v3d_context *v3d = v3d_context(pctx);
 874         struct v3d_job *job = v3d_get_job_for_fbo(v3d);
 875
 876         buffers &= ~v3d_tlb_clear(job, buffers, color, depth, stencil);
 877
 878         if (buffers)
 879                 v3d_draw_clear(v3d, buffers, color, depth, stencil);
 880 }
 881
 882 static void
 883 v3d_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 884                         const union pipe_color_union *color,
 885                         unsigned x, unsigned y, unsigned w, unsigned h,
 886                         bool render_condition_enabled)
 887 {
 888         fprintf(stderr, "unimpl: clear RT\n");
 889 }
 890
 891 static void
 892 v3d_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 893                         unsigned buffers, double depth, unsigned stencil,
 894                         unsigned x, unsigned y, unsigned w, unsigned h,
 895                         bool render_condition_enabled)
 896 {
 897         fprintf(stderr, "unimpl: clear DS\n");
 898 }
 899
 900 void
 901 v3dX(draw_init)(struct pipe_context *pctx)
 902 {
 903         pctx->draw_vbo = v3d_draw_vbo;
 904         pctx->clear = v3d_clear;
 905         pctx->clear_render_target = v3d_clear_render_target;
 906         pctx->clear_depth_stencil = v3d_clear_depth_stencil;
 907 }