v3d: use the GPU to record primitives written to transform feedback
[mesa.git] / 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 /* The PTB will request the tile alloc initial size per tile at start
59 * of tile binning.
60 */
61 uint32_t tile_alloc_size = (job->draw_tiles_x *
62 job->draw_tiles_y) * 64;
63 /* The PTB allocates in aligned 4k chunks after the initial setup. */
64 tile_alloc_size = align(tile_alloc_size, 4096);
65
66 /* Include the first two chunk allocations that the PTB does so that
67 * we definitely clear the OOM condition before triggering one (the HW
68 * won't trigger OOM during the first allocations).
69 */
70 tile_alloc_size += 8192;
71
72 /* For performance, allocate some extra initial memory after the PTB's
73 * minimal allocations, so that we hopefully don't have to block the
74 * GPU on the kernel handling an OOM signal.
75 */
76 tile_alloc_size += 512 * 1024;
77
78 job->tile_alloc = v3d_bo_alloc(v3d->screen, tile_alloc_size,
79 "tile_alloc");
80 uint32_t tsda_per_tile_size = v3d->screen->devinfo.ver >= 40 ? 256 : 64;
81 job->tile_state = v3d_bo_alloc(v3d->screen,
82 job->draw_tiles_y *
83 job->draw_tiles_x *
84 tsda_per_tile_size,
85 "TSDA");
86
87 #if V3D_VERSION >= 40
88 cl_emit(&job->bcl, TILE_BINNING_MODE_CFG, config) {
89 config.width_in_pixels = v3d->framebuffer.width;
90 config.height_in_pixels = v3d->framebuffer.height;
91 config.number_of_render_targets =
92 MAX2(v3d->framebuffer.nr_cbufs, 1);
93
94 config.multisample_mode_4x = job->msaa;
95
96 config.maximum_bpp_of_all_render_targets = job->internal_bpp;
97 }
98 #else /* V3D_VERSION < 40 */
99 /* "Binning mode lists start with a Tile Binning Mode Configuration
100 * item (120)"
101 *
102 * Part1 signals the end of binning config setup.
103 */
104 cl_emit(&job->bcl, TILE_BINNING_MODE_CFG_PART2, config) {
105 config.tile_allocation_memory_address =
106 cl_address(job->tile_alloc, 0);
107 config.tile_allocation_memory_size = job->tile_alloc->size;
108 }
109
110 cl_emit(&job->bcl, TILE_BINNING_MODE_CFG_PART1, config) {
111 config.tile_state_data_array_base_address =
112 cl_address(job->tile_state, 0);
113
114 config.width_in_tiles = job->draw_tiles_x;
115 config.height_in_tiles = job->draw_tiles_y;
116 /* Must be >= 1 */
117 config.number_of_render_targets =
118 MAX2(v3d->framebuffer.nr_cbufs, 1);
119
120 config.multisample_mode_4x = job->msaa;
121
122 config.maximum_bpp_of_all_render_targets = job->internal_bpp;
123 }
124 #endif /* V3D_VERSION < 40 */
125
126 /* There's definitely nothing in the VCD cache we want. */
127 cl_emit(&job->bcl, FLUSH_VCD_CACHE, bin);
128
129 /* Disable any leftover OQ state from another job. */
130 cl_emit(&job->bcl, OCCLUSION_QUERY_COUNTER, counter);
131
132 /* "Binning mode lists must have a Start Tile Binning item (6) after
133 * any prefix state data before the binning list proper starts."
134 */
135 cl_emit(&job->bcl, START_TILE_BINNING, bin);
136
137 job->needs_flush = true;
138 job->draw_width = v3d->framebuffer.width;
139 job->draw_height = v3d->framebuffer.height;
140 }
141
142 static void
143 v3d_predraw_check_stage_inputs(struct pipe_context *pctx,
144 enum pipe_shader_type s)
145 {
146 struct v3d_context *v3d = v3d_context(pctx);
147
148 /* Flush writes to textures we're sampling. */
149 for (int i = 0; i < v3d->tex[s].num_textures; i++) {
150 struct pipe_sampler_view *pview = v3d->tex[s].textures[i];
151 if (!pview)
152 continue;
153 struct v3d_sampler_view *view = v3d_sampler_view(pview);
154
155 if (view->texture != view->base.texture &&
156 view->base.format != PIPE_FORMAT_X32_S8X24_UINT)
157 v3d_update_shadow_texture(pctx, &view->base);
158
159 v3d_flush_jobs_writing_resource(v3d, view->texture, false);
160 }
161
162 /* Flush writes to UBOs. */
163 foreach_bit(i, v3d->constbuf[s].enabled_mask) {
164 struct pipe_constant_buffer *cb = &v3d->constbuf[s].cb[i];
165 if (cb->buffer)
166 v3d_flush_jobs_writing_resource(v3d, cb->buffer, false);
167 }
168
169 /* Flush writes to our image views */
170 foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
171 struct v3d_image_view *view = &v3d->shaderimg[s].si[i];
172
173 v3d_flush_jobs_writing_resource(v3d, view->base.resource,
174 false);
175 }
176
177 /* Flush writes to our vertex buffers (i.e. from transform feedback) */
178 if (s == PIPE_SHADER_VERTEX) {
179 foreach_bit(i, v3d->vertexbuf.enabled_mask) {
180 struct pipe_vertex_buffer *vb = &v3d->vertexbuf.vb[i];
181
182 v3d_flush_jobs_writing_resource(v3d, vb->buffer.resource,
183 false);
184 }
185 }
186 }
187
188 static void
189 v3d_predraw_check_outputs(struct pipe_context *pctx)
190 {
191 struct v3d_context *v3d = v3d_context(pctx);
192
193 /* Flush jobs reading from TF buffers that we are about to write. */
194 if (v3d_transform_feedback_enabled(v3d)) {
195 struct v3d_streamout_stateobj *so = &v3d->streamout;
196
197 for (int i = 0; i < so->num_targets; i++) {
198 if (!so->targets[i])
199 continue;
200
201 const struct pipe_stream_output_target *target =
202 so->targets[i];
203 v3d_flush_jobs_reading_resource(v3d, target->buffer);
204 }
205 }
206 }
207
208 /**
209 * Checks if the state for the current draw reads a particular resource in
210 * in the given shader stage.
211 */
212 static bool
213 v3d_state_reads_resource(struct v3d_context *v3d,
214 struct pipe_resource *prsc,
215 enum pipe_shader_type s)
216 {
217 struct v3d_resource *rsc = v3d_resource(prsc);
218
219 /* Vertex buffers */
220 if (s == PIPE_SHADER_VERTEX) {
221 foreach_bit(i, v3d->vertexbuf.enabled_mask) {
222 struct pipe_vertex_buffer *vb = &v3d->vertexbuf.vb[i];
223 if (!vb->buffer.resource)
224 continue;
225
226 struct v3d_resource *vb_rsc =
227 v3d_resource(vb->buffer.resource);
228 if (rsc->bo == vb_rsc->bo)
229 return true;
230 }
231 }
232
233 /* Constant buffers */
234 foreach_bit(i, v3d->constbuf[s].enabled_mask) {
235 struct pipe_constant_buffer *cb = &v3d->constbuf[s].cb[i];
236 if (!cb->buffer)
237 continue;
238
239 struct v3d_resource *cb_rsc = v3d_resource(cb->buffer);
240 if (rsc->bo == cb_rsc->bo)
241 return true;
242 }
243
244 /* Shader storage buffers */
245 foreach_bit(i, v3d->ssbo[s].enabled_mask) {
246 struct pipe_shader_buffer *sb = &v3d->ssbo[s].sb[i];
247 if (!sb->buffer)
248 continue;
249
250 struct v3d_resource *sb_rsc = v3d_resource(sb->buffer);
251 if (rsc->bo == sb_rsc->bo)
252 return true;
253 }
254
255 /* Textures */
256 for (int i = 0; i < v3d->tex[s].num_textures; i++) {
257 struct pipe_sampler_view *pview = v3d->tex[s].textures[i];
258 if (!pview)
259 continue;
260
261 struct v3d_sampler_view *view = v3d_sampler_view(pview);
262 struct v3d_resource *v_rsc = v3d_resource(view->texture);
263 if (rsc->bo == v_rsc->bo)
264 return true;
265 }
266
267 return false;
268 }
269
270 static void
271 v3d_emit_wait_for_tf(struct v3d_job *job)
272 {
273 /* XXX: we might be able to skip this in some cases, for now we
274 * always emit it.
275 */
276 cl_emit(&job->bcl, FLUSH_TRANSFORM_FEEDBACK_DATA, flush);
277
278 cl_emit(&job->bcl, WAIT_FOR_TRANSFORM_FEEDBACK, wait) {
279 /* XXX: Wait for all outstanding writes... maybe we can do
280 * better in some cases.
281 */
282 wait.block_count = 255;
283 }
284
285 /* We have just flushed all our outstanding TF work in this job so make
286 * sure we don't emit TF flushes again for any of it again.
287 */
288 _mesa_set_clear(job->tf_write_prscs, NULL);
289 }
290
291 static void
292 v3d_emit_wait_for_tf_if_needed(struct v3d_context *v3d, struct v3d_job *job)
293 {
294 if (!job->tf_enabled)
295 return;
296
297 set_foreach(job->tf_write_prscs, entry) {
298 struct pipe_resource *prsc = (struct pipe_resource *)entry->key;
299 for (int s = 0; s < PIPE_SHADER_COMPUTE; s++) {
300 /* Fragment shaders can only start executing after all
301 * binning (and thus TF) is complete.
302 *
303 * XXX: For VS/GS/TES, if the binning shader does not
304 * read the resource then we could also avoid emitting
305 * the wait.
306 */
307 if (s == PIPE_SHADER_FRAGMENT)
308 continue;
309
310 if (v3d_state_reads_resource(v3d, prsc, s)) {
311 v3d_emit_wait_for_tf(job);
312 return;
313 }
314 }
315 }
316 }
317
318 static void
319 v3d_emit_gl_shader_state(struct v3d_context *v3d,
320 const struct pipe_draw_info *info)
321 {
322 struct v3d_job *job = v3d->job;
323 /* VC5_DIRTY_VTXSTATE */
324 struct v3d_vertex_stateobj *vtx = v3d->vtx;
325 /* VC5_DIRTY_VTXBUF */
326 struct v3d_vertexbuf_stateobj *vertexbuf = &v3d->vertexbuf;
327
328 /* Upload the uniforms to the indirect CL first */
329 struct v3d_cl_reloc fs_uniforms =
330 v3d_write_uniforms(v3d, v3d->prog.fs,
331 PIPE_SHADER_FRAGMENT);
332 struct v3d_cl_reloc vs_uniforms =
333 v3d_write_uniforms(v3d, v3d->prog.vs,
334 PIPE_SHADER_VERTEX);
335 struct v3d_cl_reloc cs_uniforms =
336 v3d_write_uniforms(v3d, v3d->prog.cs,
337 PIPE_SHADER_VERTEX);
338
339 /* See GFXH-930 workaround below */
340 uint32_t num_elements_to_emit = MAX2(vtx->num_elements, 1);
341 uint32_t shader_rec_offset =
342 v3d_cl_ensure_space(&job->indirect,
343 cl_packet_length(GL_SHADER_STATE_RECORD) +
344 num_elements_to_emit *
345 cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD),
346 32);
347
348 /* XXX perf: We should move most of the SHADER_STATE_RECORD setup to
349 * compile time, so that we mostly just have to OR the VS and FS
350 * records together at draw time.
351 */
352 cl_emit(&job->indirect, GL_SHADER_STATE_RECORD, shader) {
353 shader.enable_clipping = true;
354 /* VC5_DIRTY_PRIM_MODE | VC5_DIRTY_RASTERIZER */
355 shader.point_size_in_shaded_vertex_data =
356 (info->mode == PIPE_PRIM_POINTS &&
357 v3d->rasterizer->base.point_size_per_vertex);
358
359 /* Must be set if the shader modifies Z, discards, or modifies
360 * the sample mask. For any of these cases, the fragment
361 * shader needs to write the Z value (even just discards).
362 */
363 shader.fragment_shader_does_z_writes =
364 v3d->prog.fs->prog_data.fs->writes_z;
365 /* Set if the EZ test must be disabled (due to shader side
366 * effects and the early_z flag not being present in the
367 * shader).
368 */
369 shader.turn_off_early_z_test =
370 v3d->prog.fs->prog_data.fs->disable_ez;
371
372 shader.fragment_shader_uses_real_pixel_centre_w_in_addition_to_centroid_w2 =
373 v3d->prog.fs->prog_data.fs->uses_center_w;
374
375 #if V3D_VERSION >= 40
376 shader.do_scoreboard_wait_on_first_thread_switch =
377 v3d->prog.fs->prog_data.fs->lock_scoreboard_on_first_thrsw;
378 shader.disable_implicit_point_line_varyings =
379 !v3d->prog.fs->prog_data.fs->uses_implicit_point_line_varyings;
380 #endif
381
382 shader.number_of_varyings_in_fragment_shader =
383 v3d->prog.fs->prog_data.fs->num_inputs;
384
385 shader.coordinate_shader_propagate_nans = true;
386 shader.vertex_shader_propagate_nans = true;
387 shader.fragment_shader_propagate_nans = true;
388
389 shader.coordinate_shader_code_address =
390 cl_address(v3d_resource(v3d->prog.cs->resource)->bo,
391 v3d->prog.cs->offset);
392 shader.vertex_shader_code_address =
393 cl_address(v3d_resource(v3d->prog.vs->resource)->bo,
394 v3d->prog.vs->offset);
395 shader.fragment_shader_code_address =
396 cl_address(v3d_resource(v3d->prog.fs->resource)->bo,
397 v3d->prog.fs->offset);
398
399 /* XXX: Use combined input/output size flag in the common
400 * case.
401 */
402 shader.coordinate_shader_has_separate_input_and_output_vpm_blocks =
403 v3d->prog.cs->prog_data.vs->separate_segments;
404 shader.vertex_shader_has_separate_input_and_output_vpm_blocks =
405 v3d->prog.vs->prog_data.vs->separate_segments;
406
407 shader.coordinate_shader_input_vpm_segment_size =
408 v3d->prog.cs->prog_data.vs->separate_segments ?
409 v3d->prog.cs->prog_data.vs->vpm_input_size : 1;
410 shader.vertex_shader_input_vpm_segment_size =
411 v3d->prog.vs->prog_data.vs->separate_segments ?
412 v3d->prog.vs->prog_data.vs->vpm_input_size : 1;
413
414 shader.coordinate_shader_output_vpm_segment_size =
415 v3d->prog.cs->prog_data.vs->vpm_output_size;
416 shader.vertex_shader_output_vpm_segment_size =
417 v3d->prog.vs->prog_data.vs->vpm_output_size;
418
419 shader.coordinate_shader_uniforms_address = cs_uniforms;
420 shader.vertex_shader_uniforms_address = vs_uniforms;
421 shader.fragment_shader_uniforms_address = fs_uniforms;
422
423 #if V3D_VERSION >= 41
424 shader.min_coord_shader_input_segments_required_in_play = 1;
425 shader.min_vertex_shader_input_segments_required_in_play = 1;
426
427 shader.coordinate_shader_4_way_threadable =
428 v3d->prog.cs->prog_data.vs->base.threads == 4;
429 shader.vertex_shader_4_way_threadable =
430 v3d->prog.vs->prog_data.vs->base.threads == 4;
431 shader.fragment_shader_4_way_threadable =
432 v3d->prog.fs->prog_data.fs->base.threads == 4;
433
434 shader.coordinate_shader_start_in_final_thread_section =
435 v3d->prog.cs->prog_data.vs->base.single_seg;
436 shader.vertex_shader_start_in_final_thread_section =
437 v3d->prog.vs->prog_data.vs->base.single_seg;
438 shader.fragment_shader_start_in_final_thread_section =
439 v3d->prog.fs->prog_data.fs->base.single_seg;
440 #else
441 shader.coordinate_shader_4_way_threadable =
442 v3d->prog.cs->prog_data.vs->base.threads == 4;
443 shader.coordinate_shader_2_way_threadable =
444 v3d->prog.cs->prog_data.vs->base.threads == 2;
445 shader.vertex_shader_4_way_threadable =
446 v3d->prog.vs->prog_data.vs->base.threads == 4;
447 shader.vertex_shader_2_way_threadable =
448 v3d->prog.vs->prog_data.vs->base.threads == 2;
449 shader.fragment_shader_4_way_threadable =
450 v3d->prog.fs->prog_data.fs->base.threads == 4;
451 shader.fragment_shader_2_way_threadable =
452 v3d->prog.fs->prog_data.fs->base.threads == 2;
453 #endif
454
455 shader.vertex_id_read_by_coordinate_shader =
456 v3d->prog.cs->prog_data.vs->uses_vid;
457 shader.instance_id_read_by_coordinate_shader =
458 v3d->prog.cs->prog_data.vs->uses_iid;
459 shader.vertex_id_read_by_vertex_shader =
460 v3d->prog.vs->prog_data.vs->uses_vid;
461 shader.instance_id_read_by_vertex_shader =
462 v3d->prog.vs->prog_data.vs->uses_iid;
463
464 shader.address_of_default_attribute_values =
465 cl_address(v3d_resource(vtx->defaults)->bo,
466 vtx->defaults_offset);
467 }
468
469 bool cs_loaded_any = false;
470 for (int i = 0; i < vtx->num_elements; i++) {
471 struct pipe_vertex_element *elem = &vtx->pipe[i];
472 struct pipe_vertex_buffer *vb =
473 &vertexbuf->vb[elem->vertex_buffer_index];
474 struct v3d_resource *rsc = v3d_resource(vb->buffer.resource);
475
476 const uint32_t size =
477 cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD);
478 cl_emit_with_prepacked(&job->indirect,
479 GL_SHADER_STATE_ATTRIBUTE_RECORD,
480 &vtx->attrs[i * size], attr) {
481 attr.stride = vb->stride;
482 attr.address = cl_address(rsc->bo,
483 vb->buffer_offset +
484 elem->src_offset);
485 attr.number_of_values_read_by_coordinate_shader =
486 v3d->prog.cs->prog_data.vs->vattr_sizes[i];
487 attr.number_of_values_read_by_vertex_shader =
488 v3d->prog.vs->prog_data.vs->vattr_sizes[i];
489
490 /* GFXH-930: At least one attribute must be enabled
491 * and read by CS and VS. If we have attributes being
492 * consumed by the VS but not the CS, then set up a
493 * dummy load of the last attribute into the CS's VPM
494 * inputs. (Since CS is just dead-code-elimination
495 * compared to VS, we can't have CS loading but not
496 * VS).
497 */
498 if (v3d->prog.cs->prog_data.vs->vattr_sizes[i])
499 cs_loaded_any = true;
500 if (i == vtx->num_elements - 1 && !cs_loaded_any) {
501 attr.number_of_values_read_by_coordinate_shader = 1;
502 }
503 #if V3D_VERSION >= 41
504 attr.maximum_index = 0xffffff;
505 #endif
506 }
507 STATIC_ASSERT(sizeof(vtx->attrs) >= V3D_MAX_VS_INPUTS / 4 * size);
508 }
509
510 if (vtx->num_elements == 0) {
511 /* GFXH-930: At least one attribute must be enabled and read
512 * by CS and VS. If we have no attributes being consumed by
513 * the shader, set up a dummy to be loaded into the VPM.
514 */
515 cl_emit(&job->indirect, GL_SHADER_STATE_ATTRIBUTE_RECORD, attr) {
516 /* Valid address of data whose value will be unused. */
517 attr.address = cl_address(job->indirect.bo, 0);
518
519 attr.type = ATTRIBUTE_FLOAT;
520 attr.stride = 0;
521 attr.vec_size = 1;
522
523 attr.number_of_values_read_by_coordinate_shader = 1;
524 attr.number_of_values_read_by_vertex_shader = 1;
525 }
526 }
527
528 cl_emit(&job->bcl, VCM_CACHE_SIZE, vcm) {
529 vcm.number_of_16_vertex_batches_for_binning =
530 v3d->prog.cs->prog_data.vs->vcm_cache_size;
531 vcm.number_of_16_vertex_batches_for_rendering =
532 v3d->prog.vs->prog_data.vs->vcm_cache_size;
533 }
534
535 cl_emit(&job->bcl, GL_SHADER_STATE, state) {
536 state.address = cl_address(job->indirect.bo, shader_rec_offset);
537 state.number_of_attribute_arrays = num_elements_to_emit;
538 }
539
540 v3d_bo_unreference(&cs_uniforms.bo);
541 v3d_bo_unreference(&vs_uniforms.bo);
542 v3d_bo_unreference(&fs_uniforms.bo);
543
544 job->shader_rec_count++;
545 }
546
547 /**
548 * Updates the number of primitvies generated from the number of vertices
549 * to draw. We do this here instead of using PRIMITIVE_COUNTS_FEEDBACK because
550 * using the GPU packet for this might require sync waits and this is trivial
551 * to handle in the CPU instead.
552 */
553 static void
554 v3d_update_primitives_generated_counter(struct v3d_context *v3d,
555 const struct pipe_draw_info *info)
556 {
557 if (!v3d->active_queries)
558 return;
559
560 uint32_t prims = u_prims_for_vertices(info->mode, info->count);
561 v3d->prims_generated += prims;
562 }
563
564 static void
565 v3d_update_job_ez(struct v3d_context *v3d, struct v3d_job *job)
566 {
567 switch (v3d->zsa->ez_state) {
568 case VC5_EZ_UNDECIDED:
569 /* If the Z/S state didn't pick a direction but didn't
570 * disable, then go along with the current EZ state. This
571 * allows EZ optimization for Z func == EQUAL or NEVER.
572 */
573 break;
574
575 case VC5_EZ_LT_LE:
576 case VC5_EZ_GT_GE:
577 /* If the Z/S state picked a direction, then it needs to match
578 * the current direction if we've decided on one.
579 */
580 if (job->ez_state == VC5_EZ_UNDECIDED)
581 job->ez_state = v3d->zsa->ez_state;
582 else if (job->ez_state != v3d->zsa->ez_state)
583 job->ez_state = VC5_EZ_DISABLED;
584 break;
585
586 case VC5_EZ_DISABLED:
587 /* If the current Z/S state disables EZ because of a bad Z
588 * func or stencil operation, then we can't do any more EZ in
589 * this frame.
590 */
591 job->ez_state = VC5_EZ_DISABLED;
592 break;
593 }
594
595 /* If the FS affects the Z of the pixels, then it may update against
596 * the chosen EZ direction (though we could use
597 * ARB_conservative_depth's hints to avoid this)
598 */
599 if (v3d->prog.fs->prog_data.fs->writes_z) {
600 job->ez_state = VC5_EZ_DISABLED;
601 }
602
603 if (job->first_ez_state == VC5_EZ_UNDECIDED &&
604 (job->ez_state != VC5_EZ_DISABLED || job->draw_calls_queued == 0))
605 job->first_ez_state = job->ez_state;
606 }
607
608 static void
609 v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
610 {
611 struct v3d_context *v3d = v3d_context(pctx);
612
613 if (!info->count_from_stream_output && !info->indirect &&
614 !info->primitive_restart &&
615 !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
616 return;
617
618 /* Fall back for weird desktop GL primitive restart values. */
619 if (info->primitive_restart &&
620 info->index_size) {
621 uint32_t mask = ~0;
622
623 switch (info->index_size) {
624 case 2:
625 mask = 0xffff;
626 break;
627 case 1:
628 mask = 0xff;
629 break;
630 }
631
632 if (info->restart_index != mask) {
633 util_draw_vbo_without_prim_restart(pctx, info);
634 return;
635 }
636 }
637
638 if (info->mode >= PIPE_PRIM_QUADS) {
639 util_primconvert_save_rasterizer_state(v3d->primconvert, &v3d->rasterizer->base);
640 util_primconvert_draw_vbo(v3d->primconvert, info);
641 perf_debug("Fallback conversion for %d %s vertices\n",
642 info->count, u_prim_name(info->mode));
643 return;
644 }
645
646 /* Before setting up the draw, flush anything writing to the resources
647 * that we read from or reading from resources we write to.
648 */
649 for (int s = 0; s < PIPE_SHADER_COMPUTE; s++)
650 v3d_predraw_check_stage_inputs(pctx, s);
651
652 if (info->indirect) {
653 v3d_flush_jobs_writing_resource(v3d, info->indirect->buffer,
654 false);
655 }
656
657 v3d_predraw_check_outputs(pctx);
658
659 /* If transform feedback is active and we are switching primitive type
660 * we need to submit the job before drawing and update the vertex count
661 * written to TF based on the primitive type since we will need to
662 * know the exact vertex count if the application decides to call
663 * glDrawTransformFeedback() later.
664 */
665 if (v3d->streamout.num_targets > 0 &&
666 u_base_prim_type(info->mode) != u_base_prim_type(v3d->prim_mode)) {
667 v3d_tf_update_counters(v3d);
668 }
669
670 struct v3d_job *job = v3d_get_job_for_fbo(v3d);
671
672 /* If vertex texturing depends on the output of rendering, we need to
673 * ensure that that rendering is complete before we run a coordinate
674 * shader that depends on it.
675 *
676 * Given that doing that is unusual, for now we just block the binner
677 * on the last submitted render, rather than tracking the last
678 * rendering to each texture's BO.
679 */
680 if (v3d->tex[PIPE_SHADER_VERTEX].num_textures || info->indirect) {
681 perf_debug("Blocking binner on last render "
682 "due to vertex texturing or indirect drawing.\n");
683 job->submit.in_sync_bcl = v3d->out_sync;
684 }
685
686 /* Mark SSBOs as being written. We don't actually know which ones are
687 * read vs written, so just assume the worst
688 */
689 for (int s = 0; s < PIPE_SHADER_COMPUTE; s++) {
690 foreach_bit(i, v3d->ssbo[s].enabled_mask) {
691 v3d_job_add_write_resource(job,
692 v3d->ssbo[s].sb[i].buffer);
693 job->tmu_dirty_rcl = true;
694 }
695
696 foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
697 v3d_job_add_write_resource(job,
698 v3d->shaderimg[s].si[i].base.resource);
699 job->tmu_dirty_rcl = true;
700 }
701 }
702
703 /* Get space to emit our draw call into the BCL, using a branch to
704 * jump to a new BO if necessary.
705 */
706 v3d_cl_ensure_space_with_branch(&job->bcl, 256 /* XXX */);
707
708 if (v3d->prim_mode != info->mode) {
709 v3d->prim_mode = info->mode;
710 v3d->dirty |= VC5_DIRTY_PRIM_MODE;
711 }
712
713 v3d_start_draw(v3d);
714 v3d_update_compiled_shaders(v3d, info->mode);
715 v3d_update_job_ez(v3d, job);
716
717 /* If this job was writing to transform feedback buffers before this
718 * draw and we are reading from them here, then we need to wait for TF
719 * to complete before we emit this draw.
720 *
721 * Notice this check needs to happen before we emit state for the
722 * current draw call, where we update job->tf_enabled, so we can ensure
723 * that we only check TF writes for prior draws.
724 */
725 v3d_emit_wait_for_tf_if_needed(v3d, job);
726
727 #if V3D_VERSION >= 41
728 v3d41_emit_state(pctx);
729 #else
730 v3d33_emit_state(pctx);
731 #endif
732
733 if (v3d->dirty & (VC5_DIRTY_VTXBUF |
734 VC5_DIRTY_VTXSTATE |
735 VC5_DIRTY_PRIM_MODE |
736 VC5_DIRTY_RASTERIZER |
737 VC5_DIRTY_COMPILED_CS |
738 VC5_DIRTY_COMPILED_VS |
739 VC5_DIRTY_COMPILED_FS |
740 v3d->prog.cs->uniform_dirty_bits |
741 v3d->prog.vs->uniform_dirty_bits |
742 v3d->prog.fs->uniform_dirty_bits)) {
743 v3d_emit_gl_shader_state(v3d, info);
744 }
745
746 v3d->dirty = 0;
747
748 /* The Base Vertex/Base Instance packet sets those values to nonzero
749 * for the next draw call only.
750 */
751 if (info->index_bias || info->start_instance) {
752 cl_emit(&job->bcl, BASE_VERTEX_BASE_INSTANCE, base) {
753 base.base_instance = info->start_instance;
754 base.base_vertex = info->index_bias;
755 }
756 }
757
758 uint32_t prim_tf_enable = 0;
759 #if V3D_VERSION < 40
760 /* V3D 3.x: The HW only processes transform feedback on primitives
761 * with the flag set.
762 */
763 if (v3d->streamout.num_targets)
764 prim_tf_enable = (V3D_PRIM_POINTS_TF - V3D_PRIM_POINTS);
765 #endif
766
767 v3d_update_primitives_generated_counter(v3d, info);
768
769 /* Note that the primitive type fields match with OpenGL/gallium
770 * definitions, up to but not including QUADS.
771 */
772 if (info->index_size) {
773 uint32_t index_size = info->index_size;
774 uint32_t offset = info->start * index_size;
775 struct pipe_resource *prsc;
776 if (info->has_user_indices) {
777 prsc = NULL;
778 u_upload_data(v3d->uploader, 0,
779 info->count * info->index_size, 4,
780 info->index.user,
781 &offset, &prsc);
782 } else {
783 prsc = info->index.resource;
784 }
785 struct v3d_resource *rsc = v3d_resource(prsc);
786
787 #if V3D_VERSION >= 40
788 cl_emit(&job->bcl, INDEX_BUFFER_SETUP, ib) {
789 ib.address = cl_address(rsc->bo, 0);
790 ib.size = rsc->bo->size;
791 }
792 #endif
793
794 if (info->indirect) {
795 cl_emit(&job->bcl, INDIRECT_INDEXED_INSTANCED_PRIM_LIST, prim) {
796 prim.index_type = ffs(info->index_size) - 1;
797 #if V3D_VERSION < 40
798 prim.address_of_indices_list =
799 cl_address(rsc->bo, offset);
800 #endif /* V3D_VERSION < 40 */
801 prim.mode = info->mode | prim_tf_enable;
802 prim.enable_primitive_restarts = info->primitive_restart;
803
804 prim.number_of_draw_indirect_indexed_records = info->indirect->draw_count;
805
806 prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
807 prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
808 info->indirect->offset);
809 }
810 } else if (info->instance_count > 1) {
811 cl_emit(&job->bcl, INDEXED_INSTANCED_PRIM_LIST, prim) {
812 prim.index_type = ffs(info->index_size) - 1;
813 #if V3D_VERSION >= 40
814 prim.index_offset = offset;
815 #else /* V3D_VERSION < 40 */
816 prim.maximum_index = (1u << 31) - 1; /* XXX */
817 prim.address_of_indices_list =
818 cl_address(rsc->bo, offset);
819 #endif /* V3D_VERSION < 40 */
820 prim.mode = info->mode | prim_tf_enable;
821 prim.enable_primitive_restarts = info->primitive_restart;
822
823 prim.number_of_instances = info->instance_count;
824 prim.instance_length = info->count;
825 }
826 } else {
827 cl_emit(&job->bcl, INDEXED_PRIM_LIST, prim) {
828 prim.index_type = ffs(info->index_size) - 1;
829 prim.length = info->count;
830 #if V3D_VERSION >= 40
831 prim.index_offset = offset;
832 #else /* V3D_VERSION < 40 */
833 prim.maximum_index = (1u << 31) - 1; /* XXX */
834 prim.address_of_indices_list =
835 cl_address(rsc->bo, offset);
836 #endif /* V3D_VERSION < 40 */
837 prim.mode = info->mode | prim_tf_enable;
838 prim.enable_primitive_restarts = info->primitive_restart;
839 }
840 }
841
842 job->draw_calls_queued++;
843
844 if (info->has_user_indices)
845 pipe_resource_reference(&prsc, NULL);
846 } else {
847 if (info->indirect) {
848 cl_emit(&job->bcl, INDIRECT_VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
849 prim.mode = info->mode | prim_tf_enable;
850 prim.number_of_draw_indirect_array_records = info->indirect->draw_count;
851
852 prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
853 prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
854 info->indirect->offset);
855 }
856 } else if (info->instance_count > 1) {
857 struct pipe_stream_output_target *so =
858 info->count_from_stream_output;
859 uint32_t vert_count = so ?
860 v3d_stream_output_target_get_vertex_count(so) :
861 info->count;
862 cl_emit(&job->bcl, VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
863 prim.mode = info->mode | prim_tf_enable;
864 prim.index_of_first_vertex = info->start;
865 prim.number_of_instances = info->instance_count;
866 prim.instance_length = vert_count;
867 }
868 } else {
869 struct pipe_stream_output_target *so =
870 info->count_from_stream_output;
871 uint32_t vert_count = so ?
872 v3d_stream_output_target_get_vertex_count(so) :
873 info->count;
874 cl_emit(&job->bcl, VERTEX_ARRAY_PRIMS, prim) {
875 prim.mode = info->mode | prim_tf_enable;
876 prim.length = vert_count;
877 prim.index_of_first_vertex = info->start;
878 }
879 }
880 }
881
882 /* A flush is required in between a TF draw and any following TF specs
883 * packet, or the GPU may hang. Just flush each time for now.
884 */
885 if (v3d->streamout.num_targets)
886 cl_emit(&job->bcl, TRANSFORM_FEEDBACK_FLUSH_AND_COUNT, flush);
887
888 job->draw_calls_queued++;
889
890 /* Increment the TF offsets by how many verts we wrote. XXX: This
891 * needs some clamping to the buffer size.
892 */
893 for (int i = 0; i < v3d->streamout.num_targets; i++)
894 v3d->streamout.offsets[i] += info->count;
895
896 if (v3d->zsa && job->zsbuf && v3d->zsa->base.depth.enabled) {
897 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
898 v3d_job_add_bo(job, rsc->bo);
899
900 job->load |= PIPE_CLEAR_DEPTH & ~job->clear;
901 if (v3d->zsa->base.depth.writemask)
902 job->store |= PIPE_CLEAR_DEPTH;
903 rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
904 }
905
906 if (v3d->zsa && job->zsbuf && v3d->zsa->base.stencil[0].enabled) {
907 struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
908 if (rsc->separate_stencil)
909 rsc = rsc->separate_stencil;
910
911 v3d_job_add_bo(job, rsc->bo);
912
913 job->load |= PIPE_CLEAR_STENCIL & ~job->clear;
914 if (v3d->zsa->base.stencil[0].writemask ||
915 v3d->zsa->base.stencil[1].writemask) {
916 job->store |= PIPE_CLEAR_STENCIL;
917 }
918 rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
919 }
920
921 for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
922 uint32_t bit = PIPE_CLEAR_COLOR0 << i;
923 int blend_rt = v3d->blend->base.independent_blend_enable ? i : 0;
924
925 if (job->store & bit || !job->cbufs[i])
926 continue;
927 struct v3d_resource *rsc = v3d_resource(job->cbufs[i]->texture);
928
929 job->load |= bit & ~job->clear;
930 if (v3d->blend->base.rt[blend_rt].colormask)
931 job->store |= bit;
932 v3d_job_add_bo(job, rsc->bo);
933 }
934
935 if (job->referenced_size > 768 * 1024 * 1024) {
936 perf_debug("Flushing job with %dkb to try to free up memory\n",
937 job->referenced_size / 1024);
938 v3d_flush(pctx);
939 }
940
941 if (V3D_DEBUG & V3D_DEBUG_ALWAYS_FLUSH)
942 v3d_flush(pctx);
943 }
944
945 /**
946 * Implements gallium's clear() hook (glClear()) by drawing a pair of triangles.
947 */
948 static void
949 v3d_draw_clear(struct v3d_context *v3d,
950 unsigned buffers,
951 const union pipe_color_union *color,
952 double depth, unsigned stencil)
953 {
954 static const union pipe_color_union dummy_color = {};
955
956 /* The blitter util dereferences the color regardless, even though the
957 * gallium clear API may not pass one in when only Z/S are cleared.
958 */
959 if (!color)
960 color = &dummy_color;
961
962 v3d_blitter_save(v3d);
963 util_blitter_clear(v3d->blitter,
964 v3d->framebuffer.width,
965 v3d->framebuffer.height,
966 util_framebuffer_get_num_layers(&v3d->framebuffer),
967 buffers, color, depth, stencil,
968 util_framebuffer_get_num_samples(&v3d->framebuffer) > 1);
969 }
970
971 /**
972 * Attempts to perform the GL clear by using the TLB's fast clear at the start
973 * of the frame.
974 */
975 static unsigned
976 v3d_tlb_clear(struct v3d_job *job, unsigned buffers,
977 const union pipe_color_union *color,
978 double depth, unsigned stencil)
979 {
980 struct v3d_context *v3d = job->v3d;
981
982 if (job->draw_calls_queued) {
983 /* If anything in the CL has drawn using the buffer, then the
984 * TLB clear we're trying to add now would happen before that
985 * drawing.
986 */
987 buffers &= ~(job->load | job->store);
988 }
989
990 /* GFXH-1461: If we were to emit a load of just depth or just stencil,
991 * then the clear for the other may get lost. We need to decide now
992 * if it would be possible to need to emit a load of just one after
993 * we've set up our TLB clears.
994 */
995 if (buffers & PIPE_CLEAR_DEPTHSTENCIL &&
996 (buffers & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL &&
997 job->zsbuf &&
998 util_format_is_depth_and_stencil(job->zsbuf->texture->format)) {
999 buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
1000 }
1001
1002 for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
1003 uint32_t bit = PIPE_CLEAR_COLOR0 << i;
1004 if (!(buffers & bit))
1005 continue;
1006
1007 struct pipe_surface *psurf = v3d->framebuffer.cbufs[i];
1008 struct v3d_surface *surf = v3d_surface(psurf);
1009 struct v3d_resource *rsc = v3d_resource(psurf->texture);
1010
1011 union util_color uc;
1012 uint32_t internal_size = 4 << surf->internal_bpp;
1013
1014 static union pipe_color_union swapped_color;
1015 if (v3d->swap_color_rb & (1 << i)) {
1016 swapped_color.f[0] = color->f[2];
1017 swapped_color.f[1] = color->f[1];
1018 swapped_color.f[2] = color->f[0];
1019 swapped_color.f[3] = color->f[3];
1020 color = &swapped_color;
1021 }
1022
1023 switch (surf->internal_type) {
1024 case V3D_INTERNAL_TYPE_8:
1025 util_pack_color(color->f, PIPE_FORMAT_R8G8B8A8_UNORM,
1026 &uc);
1027 memcpy(job->clear_color[i], uc.ui, internal_size);
1028 break;
1029 case V3D_INTERNAL_TYPE_8I:
1030 case V3D_INTERNAL_TYPE_8UI:
1031 job->clear_color[i][0] = ((color->ui[0] & 0xff) |
1032 (color->ui[1] & 0xff) << 8 |
1033 (color->ui[2] & 0xff) << 16 |
1034 (color->ui[3] & 0xff) << 24);
1035 break;
1036 case V3D_INTERNAL_TYPE_16F:
1037 util_pack_color(color->f, PIPE_FORMAT_R16G16B16A16_FLOAT,
1038 &uc);
1039 memcpy(job->clear_color[i], uc.ui, internal_size);
1040 break;
1041 case V3D_INTERNAL_TYPE_16I:
1042 case V3D_INTERNAL_TYPE_16UI:
1043 job->clear_color[i][0] = ((color->ui[0] & 0xffff) |
1044 color->ui[1] << 16);
1045 job->clear_color[i][1] = ((color->ui[2] & 0xffff) |
1046 color->ui[3] << 16);
1047 break;
1048 case V3D_INTERNAL_TYPE_32F:
1049 case V3D_INTERNAL_TYPE_32I:
1050 case V3D_INTERNAL_TYPE_32UI:
1051 memcpy(job->clear_color[i], color->ui, internal_size);
1052 break;
1053 }
1054
1055 rsc->initialized_buffers |= bit;
1056 }
1057
1058 unsigned zsclear = buffers & PIPE_CLEAR_DEPTHSTENCIL;
1059 if (zsclear) {
1060 struct v3d_resource *rsc =
1061 v3d_resource(v3d->framebuffer.zsbuf->texture);
1062
1063 if (zsclear & PIPE_CLEAR_DEPTH)
1064 job->clear_z = depth;
1065 if (zsclear & PIPE_CLEAR_STENCIL)
1066 job->clear_s = stencil;
1067
1068 rsc->initialized_buffers |= zsclear;
1069 }
1070
1071 job->draw_min_x = 0;
1072 job->draw_min_y = 0;
1073 job->draw_max_x = v3d->framebuffer.width;
1074 job->draw_max_y = v3d->framebuffer.height;
1075 job->clear |= buffers;
1076 job->store |= buffers;
1077
1078 v3d_start_draw(v3d);
1079
1080 return buffers;
1081 }
1082
1083 static void
1084 v3d_clear(struct pipe_context *pctx, unsigned buffers,
1085 const union pipe_color_union *color, double depth, unsigned stencil)
1086 {
1087 struct v3d_context *v3d = v3d_context(pctx);
1088 struct v3d_job *job = v3d_get_job_for_fbo(v3d);
1089
1090 buffers &= ~v3d_tlb_clear(job, buffers, color, depth, stencil);
1091
1092 if (buffers)
1093 v3d_draw_clear(v3d, buffers, color, depth, stencil);
1094 }
1095
1096 static void
1097 v3d_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
1098 const union pipe_color_union *color,
1099 unsigned x, unsigned y, unsigned w, unsigned h,
1100 bool render_condition_enabled)
1101 {
1102 fprintf(stderr, "unimpl: clear RT\n");
1103 }
1104
1105 static void
1106 v3d_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
1107 unsigned buffers, double depth, unsigned stencil,
1108 unsigned x, unsigned y, unsigned w, unsigned h,
1109 bool render_condition_enabled)
1110 {
1111 fprintf(stderr, "unimpl: clear DS\n");
1112 }
1113
1114 void
1115 v3dX(draw_init)(struct pipe_context *pctx)
1116 {
1117 pctx->draw_vbo = v3d_draw_vbo;
1118 pctx->clear = v3d_clear;
1119 pctx->clear_render_target = v3d_clear_render_target;
1120 pctx->clear_depth_stencil = v3d_clear_depth_stencil;
1121 }