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 */
42static void
43v3d_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
142static void
143v3d_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        /* XXX perf: If we're reading from the output of TF in this job, we
149         * should instead be using the wait for transform feedback
150         * functionality.
151         */
152
153        /* Flush writes to textures we're sampling. */
154        for (int i = 0; i < v3d->tex[s].num_textures; i++) {
155                struct pipe_sampler_view *pview = v3d->tex[s].textures[i];
156                if (!pview)
157                        continue;
158                struct v3d_sampler_view *view = v3d_sampler_view(pview);
159
160                if (view->texture != view->base.texture &&
161                    view->base.format != PIPE_FORMAT_X32_S8X24_UINT)
162                        v3d_update_shadow_texture(pctx, &view->base);
163
164                v3d_flush_jobs_writing_resource(v3d, view->texture);
165        }
166
167        /* Flush writes to UBOs. */
168        foreach_bit(i, v3d->constbuf[s].enabled_mask) {
169                struct pipe_constant_buffer *cb = &v3d->constbuf[s].cb[i];
170                if (cb->buffer)
171                        v3d_flush_jobs_writing_resource(v3d, cb->buffer);
172        }
173
174        /* Flush writes to our image views */
175        foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
176                struct v3d_image_view *view = &v3d->shaderimg[s].si[i];
177
178                v3d_flush_jobs_writing_resource(v3d, view->base.resource);
179        }
180}
181
182static void
183v3d_emit_gl_shader_state(struct v3d_context *v3d,
184                         const struct pipe_draw_info *info)
185{
186        struct v3d_job *job = v3d->job;
187        /* VC5_DIRTY_VTXSTATE */
188        struct v3d_vertex_stateobj *vtx = v3d->vtx;
189        /* VC5_DIRTY_VTXBUF */
190        struct v3d_vertexbuf_stateobj *vertexbuf = &v3d->vertexbuf;
191
192        /* Upload the uniforms to the indirect CL first */
193        struct v3d_cl_reloc fs_uniforms =
194                v3d_write_uniforms(v3d, v3d->prog.fs,
195                                   PIPE_SHADER_FRAGMENT);
196        struct v3d_cl_reloc vs_uniforms =
197                v3d_write_uniforms(v3d, v3d->prog.vs,
198                                   PIPE_SHADER_VERTEX);
199        struct v3d_cl_reloc cs_uniforms =
200                v3d_write_uniforms(v3d, v3d->prog.cs,
201                                   PIPE_SHADER_VERTEX);
202
203        /* See GFXH-930 workaround below */
204        uint32_t num_elements_to_emit = MAX2(vtx->num_elements, 1);
205        uint32_t shader_rec_offset =
206                v3d_cl_ensure_space(&job->indirect,
207                                    cl_packet_length(GL_SHADER_STATE_RECORD) +
208                                    num_elements_to_emit *
209                                    cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD),
210                                    32);
211
212        /* XXX perf: We should move most of the SHADER_STATE_RECORD setup to
213         * compile time, so that we mostly just have to OR the VS and FS
214         * records together at draw time.
215         */
216        cl_emit(&job->indirect, GL_SHADER_STATE_RECORD, shader) {
217                shader.enable_clipping = true;
218                /* VC5_DIRTY_PRIM_MODE | VC5_DIRTY_RASTERIZER */
219                shader.point_size_in_shaded_vertex_data =
220                        (info->mode == PIPE_PRIM_POINTS &&
221                         v3d->rasterizer->base.point_size_per_vertex);
222
223                /* Must be set if the shader modifies Z, discards, or modifies
224                 * the sample mask.  For any of these cases, the fragment
225                 * shader needs to write the Z value (even just discards).
226                 */
227                shader.fragment_shader_does_z_writes =
228                        v3d->prog.fs->prog_data.fs->writes_z;
229                /* Set if the EZ test must be disabled (due to shader side
230                 * effects and the early_z flag not being present in the
231                 * shader).
232                 */
233                shader.turn_off_early_z_test =
234                        v3d->prog.fs->prog_data.fs->disable_ez;
235
236                shader.fragment_shader_uses_real_pixel_centre_w_in_addition_to_centroid_w2 =
237                        v3d->prog.fs->prog_data.fs->uses_center_w;
238
239                shader.number_of_varyings_in_fragment_shader =
240                        v3d->prog.fs->prog_data.fs->num_inputs;
241
242                shader.coordinate_shader_propagate_nans = true;
243                shader.vertex_shader_propagate_nans = true;
244                shader.fragment_shader_propagate_nans = true;
245
246                shader.coordinate_shader_code_address =
247                        cl_address(v3d_resource(v3d->prog.cs->resource)->bo,
248                                   v3d->prog.cs->offset);
249                shader.vertex_shader_code_address =
250                        cl_address(v3d_resource(v3d->prog.vs->resource)->bo,
251                                   v3d->prog.vs->offset);
252                shader.fragment_shader_code_address =
253                        cl_address(v3d_resource(v3d->prog.fs->resource)->bo,
254                                   v3d->prog.fs->offset);
255
256                /* XXX: Use combined input/output size flag in the common
257                 * case.
258                 */
259                shader.coordinate_shader_has_separate_input_and_output_vpm_blocks =
260                        v3d->prog.cs->prog_data.vs->separate_segments;
261                shader.vertex_shader_has_separate_input_and_output_vpm_blocks =
262                        v3d->prog.vs->prog_data.vs->separate_segments;
263
264                shader.coordinate_shader_input_vpm_segment_size =
265                        v3d->prog.cs->prog_data.vs->separate_segments ?
266                        v3d->prog.cs->prog_data.vs->vpm_input_size : 1;
267                shader.vertex_shader_input_vpm_segment_size =
268                        v3d->prog.vs->prog_data.vs->separate_segments ?
269                        v3d->prog.vs->prog_data.vs->vpm_input_size : 1;
270
271                shader.coordinate_shader_output_vpm_segment_size =
272                        v3d->prog.cs->prog_data.vs->vpm_output_size;
273                shader.vertex_shader_output_vpm_segment_size =
274                        v3d->prog.vs->prog_data.vs->vpm_output_size;
275
276                shader.coordinate_shader_uniforms_address = cs_uniforms;
277                shader.vertex_shader_uniforms_address = vs_uniforms;
278                shader.fragment_shader_uniforms_address = fs_uniforms;
279
280#if V3D_VERSION >= 41
281                shader.min_coord_shader_input_segments_required_in_play = 1;
282                shader.min_vertex_shader_input_segments_required_in_play = 1;
283
284                shader.coordinate_shader_4_way_threadable =
285                        v3d->prog.cs->prog_data.vs->base.threads == 4;
286                shader.vertex_shader_4_way_threadable =
287                        v3d->prog.vs->prog_data.vs->base.threads == 4;
288                shader.fragment_shader_4_way_threadable =
289                        v3d->prog.fs->prog_data.fs->base.threads == 4;
290
291                shader.coordinate_shader_start_in_final_thread_section =
292                        v3d->prog.cs->prog_data.vs->base.single_seg;
293                shader.vertex_shader_start_in_final_thread_section =
294                        v3d->prog.vs->prog_data.vs->base.single_seg;
295                shader.fragment_shader_start_in_final_thread_section =
296                        v3d->prog.fs->prog_data.fs->base.single_seg;
297#else
298                shader.coordinate_shader_4_way_threadable =
299                        v3d->prog.cs->prog_data.vs->base.threads == 4;
300                shader.coordinate_shader_2_way_threadable =
301                        v3d->prog.cs->prog_data.vs->base.threads == 2;
302                shader.vertex_shader_4_way_threadable =
303                        v3d->prog.vs->prog_data.vs->base.threads == 4;
304                shader.vertex_shader_2_way_threadable =
305                        v3d->prog.vs->prog_data.vs->base.threads == 2;
306                shader.fragment_shader_4_way_threadable =
307                        v3d->prog.fs->prog_data.fs->base.threads == 4;
308                shader.fragment_shader_2_way_threadable =
309                        v3d->prog.fs->prog_data.fs->base.threads == 2;
310#endif
311
312                shader.vertex_id_read_by_coordinate_shader =
313                        v3d->prog.cs->prog_data.vs->uses_vid;
314                shader.instance_id_read_by_coordinate_shader =
315                        v3d->prog.cs->prog_data.vs->uses_iid;
316                shader.vertex_id_read_by_vertex_shader =
317                        v3d->prog.vs->prog_data.vs->uses_vid;
318                shader.instance_id_read_by_vertex_shader =
319                        v3d->prog.vs->prog_data.vs->uses_iid;
320
321                shader.address_of_default_attribute_values =
322                        cl_address(v3d_resource(vtx->defaults)->bo,
323                                   vtx->defaults_offset);
324        }
325
326        bool cs_loaded_any = false;
327        for (int i = 0; i < vtx->num_elements; i++) {
328                struct pipe_vertex_element *elem = &vtx->pipe[i];
329                struct pipe_vertex_buffer *vb =
330                        &vertexbuf->vb[elem->vertex_buffer_index];
331                struct v3d_resource *rsc = v3d_resource(vb->buffer.resource);
332
333                const uint32_t size =
334                        cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD);
335                cl_emit_with_prepacked(&job->indirect,
336                                       GL_SHADER_STATE_ATTRIBUTE_RECORD,
337                                       &vtx->attrs[i * size], attr) {
338                        attr.stride = vb->stride;
339                        attr.address = cl_address(rsc->bo,
340                                                  vb->buffer_offset +
341                                                  elem->src_offset);
342                        attr.number_of_values_read_by_coordinate_shader =
343                                v3d->prog.cs->prog_data.vs->vattr_sizes[i];
344                        attr.number_of_values_read_by_vertex_shader =
345                                v3d->prog.vs->prog_data.vs->vattr_sizes[i];
346
347                        /* GFXH-930: At least one attribute must be enabled
348                         * and read by CS and VS.  If we have attributes being
349                         * consumed by the VS but not the CS, then set up a
350                         * dummy load of the last attribute into the CS's VPM
351                         * inputs.  (Since CS is just dead-code-elimination
352                         * compared to VS, we can't have CS loading but not
353                         * VS).
354                         */
355                        if (v3d->prog.cs->prog_data.vs->vattr_sizes[i])
356                                cs_loaded_any = true;
357                        if (i == vtx->num_elements - 1 && !cs_loaded_any) {
358                                attr.number_of_values_read_by_coordinate_shader = 1;
359                        }
360#if V3D_VERSION >= 41
361                        attr.maximum_index = 0xffffff;
362#endif
363                }
364                STATIC_ASSERT(sizeof(vtx->attrs) >= V3D_MAX_VS_INPUTS / 4 * size);
365        }
366
367        if (vtx->num_elements == 0) {
368                /* GFXH-930: At least one attribute must be enabled and read
369                 * by CS and VS.  If we have no attributes being consumed by
370                 * the shader, set up a dummy to be loaded into the VPM.
371                 */
372                cl_emit(&job->indirect, GL_SHADER_STATE_ATTRIBUTE_RECORD, attr) {
373                        /* Valid address of data whose value will be unused. */
374                        attr.address = cl_address(job->indirect.bo, 0);
375
376                        attr.type = ATTRIBUTE_FLOAT;
377                        attr.stride = 0;
378                        attr.vec_size = 1;
379
380                        attr.number_of_values_read_by_coordinate_shader = 1;
381                        attr.number_of_values_read_by_vertex_shader = 1;
382                }
383        }
384
385        cl_emit(&job->bcl, VCM_CACHE_SIZE, vcm) {
386                vcm.number_of_16_vertex_batches_for_binning =
387                        v3d->prog.cs->prog_data.vs->vcm_cache_size;
388                vcm.number_of_16_vertex_batches_for_rendering =
389                        v3d->prog.vs->prog_data.vs->vcm_cache_size;
390        }
391
392        cl_emit(&job->bcl, GL_SHADER_STATE, state) {
393                state.address = cl_address(job->indirect.bo, shader_rec_offset);
394                state.number_of_attribute_arrays = num_elements_to_emit;
395        }
396
397        v3d_bo_unreference(&cs_uniforms.bo);
398        v3d_bo_unreference(&vs_uniforms.bo);
399        v3d_bo_unreference(&fs_uniforms.bo);
400
401        job->shader_rec_count++;
402}
403
404/**
405 * Computes the various transform feedback statistics, since they can't be
406 * recorded by CL packets.
407 */
408static void
409v3d_tf_statistics_record(struct v3d_context *v3d,
410                         const struct pipe_draw_info *info,
411                         bool prim_tf)
412{
413        if (!v3d->active_queries)
414                return;
415
416        uint32_t prims = u_prims_for_vertices(info->mode, info->count);
417        v3d->prims_generated += prims;
418
419        if (prim_tf) {
420                /* XXX: Only count if we didn't overflow. */
421                v3d->tf_prims_generated += prims;
422        }
423}
424
425static void
426v3d_update_job_ez(struct v3d_context *v3d, struct v3d_job *job)
427{
428        switch (v3d->zsa->ez_state) {
429        case VC5_EZ_UNDECIDED:
430                /* If the Z/S state didn't pick a direction but didn't
431                 * disable, then go along with the current EZ state.  This
432                 * allows EZ optimization for Z func == EQUAL or NEVER.
433                 */
434                break;
435
436        case VC5_EZ_LT_LE:
437        case VC5_EZ_GT_GE:
438                /* If the Z/S state picked a direction, then it needs to match
439                 * the current direction if we've decided on one.
440                 */
441                if (job->ez_state == VC5_EZ_UNDECIDED)
442                        job->ez_state = v3d->zsa->ez_state;
443                else if (job->ez_state != v3d->zsa->ez_state)
444                        job->ez_state = VC5_EZ_DISABLED;
445                break;
446
447        case VC5_EZ_DISABLED:
448                /* If the current Z/S state disables EZ because of a bad Z
449                 * func or stencil operation, then we can't do any more EZ in
450                 * this frame.
451                 */
452                job->ez_state = VC5_EZ_DISABLED;
453                break;
454        }
455
456        /* If the FS affects the Z of the pixels, then it may update against
457         * the chosen EZ direction (though we could use
458         * ARB_conservative_depth's hints to avoid this)
459         */
460        if (v3d->prog.fs->prog_data.fs->writes_z) {
461                job->ez_state = VC5_EZ_DISABLED;
462        }
463
464        if (job->first_ez_state == VC5_EZ_UNDECIDED &&
465            (job->ez_state != VC5_EZ_DISABLED || job->draw_calls_queued == 0))
466                job->first_ez_state = job->ez_state;
467}
468
469static void
470v3d_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
471{
472        struct v3d_context *v3d = v3d_context(pctx);
473
474        if (!info->count_from_stream_output && !info->indirect &&
475            !info->primitive_restart &&
476            !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
477                return;
478
479        /* Fall back for weird desktop GL primitive restart values. */
480        if (info->primitive_restart &&
481            info->index_size) {
482                uint32_t mask = ~0;
483
484                switch (info->index_size) {
485                case 2:
486                        mask = 0xffff;
487                        break;
488                case 1:
489                        mask = 0xff;
490                        break;
491                }
492
493                if (info->restart_index != mask) {
494                        util_draw_vbo_without_prim_restart(pctx, info);
495                        return;
496                }
497        }
498
499        if (info->mode >= PIPE_PRIM_QUADS) {
500                util_primconvert_save_rasterizer_state(v3d->primconvert, &v3d->rasterizer->base);
501                util_primconvert_draw_vbo(v3d->primconvert, info);
502                perf_debug("Fallback conversion for %d %s vertices\n",
503                           info->count, u_prim_name(info->mode));
504                return;
505        }
506
507        /* Before setting up the draw, flush anything writing to the textures
508         * that we read from.
509         */
510        for (int s = 0; s < PIPE_SHADER_COMPUTE; s++)
511                v3d_predraw_check_stage_inputs(pctx, s);
512
513        if (info->indirect)
514                v3d_flush_jobs_writing_resource(v3d, info->indirect->buffer);
515
516        struct v3d_job *job = v3d_get_job_for_fbo(v3d);
517
518        /* If vertex texturing depends on the output of rendering, we need to
519         * ensure that that rendering is complete before we run a coordinate
520         * shader that depends on it.
521         *
522         * Given that doing that is unusual, for now we just block the binner
523         * on the last submitted render, rather than tracking the last
524         * rendering to each texture's BO.
525         */
526        if (v3d->tex[PIPE_SHADER_VERTEX].num_textures || info->indirect) {
527                perf_debug("Blocking binner on last render "
528                           "due to vertex texturing or indirect drawing.\n");
529                job->submit.in_sync_bcl = v3d->out_sync;
530        }
531
532        /* Mark SSBOs as being written.  We don't actually know which ones are
533         * read vs written, so just assume the worst
534         */
535        for (int s = 0; s < PIPE_SHADER_COMPUTE; s++) {
536                foreach_bit(i, v3d->ssbo[s].enabled_mask) {
537                        v3d_job_add_write_resource(job,
538                                                   v3d->ssbo[s].sb[i].buffer);
539                        job->tmu_dirty_rcl = true;
540                }
541
542                foreach_bit(i, v3d->shaderimg[s].enabled_mask) {
543                        v3d_job_add_write_resource(job,
544                                                   v3d->shaderimg[s].si[i].base.resource);
545                        job->tmu_dirty_rcl = true;
546                }
547        }
548
549        /* Get space to emit our draw call into the BCL, using a branch to
550         * jump to a new BO if necessary.
551         */
552        v3d_cl_ensure_space_with_branch(&job->bcl, 256 /* XXX */);
553
554        if (v3d->prim_mode != info->mode) {
555                v3d->prim_mode = info->mode;
556                v3d->dirty |= VC5_DIRTY_PRIM_MODE;
557        }
558
559        v3d_start_draw(v3d);
560        v3d_update_compiled_shaders(v3d, info->mode);
561        v3d_update_job_ez(v3d, job);
562
563#if V3D_VERSION >= 41
564        v3d41_emit_state(pctx);
565#else
566        v3d33_emit_state(pctx);
567#endif
568
569        if (v3d->dirty & (VC5_DIRTY_VTXBUF |
570                          VC5_DIRTY_VTXSTATE |
571                          VC5_DIRTY_PRIM_MODE |
572                          VC5_DIRTY_RASTERIZER |
573                          VC5_DIRTY_COMPILED_CS |
574                          VC5_DIRTY_COMPILED_VS |
575                          VC5_DIRTY_COMPILED_FS |
576                          v3d->prog.cs->uniform_dirty_bits |
577                          v3d->prog.vs->uniform_dirty_bits |
578                          v3d->prog.fs->uniform_dirty_bits)) {
579                v3d_emit_gl_shader_state(v3d, info);
580        }
581
582        v3d->dirty = 0;
583
584        /* The Base Vertex/Base Instance packet sets those values to nonzero
585         * for the next draw call only.
586         */
587        if (info->index_bias || info->start_instance) {
588                cl_emit(&job->bcl, BASE_VERTEX_BASE_INSTANCE, base) {
589                        base.base_instance = info->start_instance;
590                        base.base_vertex = info->index_bias;
591                }
592        }
593
594        uint32_t prim_tf_enable = 0;
595#if V3D_VERSION < 40
596        /* V3D 3.x: The HW only processes transform feedback on primitives
597         * with the flag set.
598         */
599        if (v3d->streamout.num_targets)
600                prim_tf_enable = (V3D_PRIM_POINTS_TF - V3D_PRIM_POINTS);
601#endif
602
603        v3d_tf_statistics_record(v3d, info, v3d->streamout.num_targets);
604
605        /* Note that the primitive type fields match with OpenGL/gallium
606         * definitions, up to but not including QUADS.
607         */
608        if (info->index_size) {
609                uint32_t index_size = info->index_size;
610                uint32_t offset = info->start * index_size;
611                struct pipe_resource *prsc;
612                if (info->has_user_indices) {
613                        prsc = NULL;
614                        u_upload_data(v3d->uploader, 0,
615                                      info->count * info->index_size, 4,
616                                      info->index.user,
617                                      &offset, &prsc);
618                } else {
619                        prsc = info->index.resource;
620                }
621                struct v3d_resource *rsc = v3d_resource(prsc);
622
623#if V3D_VERSION >= 40
624                cl_emit(&job->bcl, INDEX_BUFFER_SETUP, ib) {
625                        ib.address = cl_address(rsc->bo, 0);
626                        ib.size = rsc->bo->size;
627                }
628#endif
629
630                if (info->indirect) {
631                        cl_emit(&job->bcl, INDIRECT_INDEXED_INSTANCED_PRIM_LIST, prim) {
632                                prim.index_type = ffs(info->index_size) - 1;
633#if V3D_VERSION < 40
634                                prim.address_of_indices_list =
635                                        cl_address(rsc->bo, offset);
636#endif /* V3D_VERSION < 40 */
637                                prim.mode = info->mode | prim_tf_enable;
638                                prim.enable_primitive_restarts = info->primitive_restart;
639
640                                prim.number_of_draw_indirect_indexed_records = info->indirect->draw_count;
641
642                                prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
643                                prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
644                                                          info->indirect->offset);
645                        }
646                } else if (info->instance_count > 1) {
647                        cl_emit(&job->bcl, INDEXED_INSTANCED_PRIM_LIST, prim) {
648                                prim.index_type = ffs(info->index_size) - 1;
649#if V3D_VERSION >= 40
650                                prim.index_offset = offset;
651#else /* V3D_VERSION < 40 */
652                                prim.maximum_index = (1u << 31) - 1; /* XXX */
653                                prim.address_of_indices_list =
654                                        cl_address(rsc->bo, offset);
655#endif /* V3D_VERSION < 40 */
656                                prim.mode = info->mode | prim_tf_enable;
657                                prim.enable_primitive_restarts = info->primitive_restart;
658
659                                prim.number_of_instances = info->instance_count;
660                                prim.instance_length = info->count;
661                        }
662                } else {
663                        cl_emit(&job->bcl, INDEXED_PRIM_LIST, prim) {
664                                prim.index_type = ffs(info->index_size) - 1;
665                                prim.length = info->count;
666#if V3D_VERSION >= 40
667                                prim.index_offset = offset;
668#else /* V3D_VERSION < 40 */
669                                prim.maximum_index = (1u << 31) - 1; /* XXX */
670                                prim.address_of_indices_list =
671                                        cl_address(rsc->bo, offset);
672#endif /* V3D_VERSION < 40 */
673                                prim.mode = info->mode | prim_tf_enable;
674                                prim.enable_primitive_restarts = info->primitive_restart;
675                        }
676                }
677
678                job->draw_calls_queued++;
679
680                if (info->has_user_indices)
681                        pipe_resource_reference(&prsc, NULL);
682        } else {
683                if (info->indirect) {
684                        cl_emit(&job->bcl, INDIRECT_VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
685                                prim.mode = info->mode | prim_tf_enable;
686                                prim.number_of_draw_indirect_array_records = info->indirect->draw_count;
687
688                                prim.stride_in_multiples_of_4_bytes = info->indirect->stride >> 2;
689                                prim.address = cl_address(v3d_resource(info->indirect->buffer)->bo,
690                                                          info->indirect->offset);
691                        }
692                } else if (info->instance_count > 1) {
693                        cl_emit(&job->bcl, VERTEX_ARRAY_INSTANCED_PRIMS, prim) {
694                                prim.mode = info->mode | prim_tf_enable;
695                                prim.index_of_first_vertex = info->start;
696                                prim.number_of_instances = info->instance_count;
697                                prim.instance_length = info->count;
698                        }
699                } else {
700                        cl_emit(&job->bcl, VERTEX_ARRAY_PRIMS, prim) {
701                                prim.mode = info->mode | prim_tf_enable;
702                                prim.length = info->count;
703                                prim.index_of_first_vertex = info->start;
704                        }
705                }
706        }
707
708        /* A flush is required in between a TF draw and any following TF specs
709         * packet, or the GPU may hang.  Just flush each time for now.
710         */
711        if (v3d->streamout.num_targets)
712                cl_emit(&job->bcl, TRANSFORM_FEEDBACK_FLUSH_AND_COUNT, flush);
713
714        job->draw_calls_queued++;
715
716        /* Increment the TF offsets by how many verts we wrote.  XXX: This
717         * needs some clamping to the buffer size.
718         */
719        for (int i = 0; i < v3d->streamout.num_targets; i++)
720                v3d->streamout.offsets[i] += info->count;
721
722        if (v3d->zsa && job->zsbuf && v3d->zsa->base.depth.enabled) {
723                struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
724                v3d_job_add_bo(job, rsc->bo);
725
726                job->load |= PIPE_CLEAR_DEPTH & ~job->clear;
727                if (v3d->zsa->base.depth.writemask)
728                        job->store |= PIPE_CLEAR_DEPTH;
729                rsc->initialized_buffers = PIPE_CLEAR_DEPTH;
730        }
731
732        if (v3d->zsa && job->zsbuf && v3d->zsa->base.stencil[0].enabled) {
733                struct v3d_resource *rsc = v3d_resource(job->zsbuf->texture);
734                if (rsc->separate_stencil)
735                        rsc = rsc->separate_stencil;
736
737                v3d_job_add_bo(job, rsc->bo);
738
739                job->load |= PIPE_CLEAR_STENCIL & ~job->clear;
740                if (v3d->zsa->base.stencil[0].writemask ||
741                    v3d->zsa->base.stencil[1].writemask) {
742                        job->store |= PIPE_CLEAR_STENCIL;
743                }
744                rsc->initialized_buffers |= PIPE_CLEAR_STENCIL;
745        }
746
747        for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
748                uint32_t bit = PIPE_CLEAR_COLOR0 << i;
749                int blend_rt = v3d->blend->base.independent_blend_enable ? i : 0;
750
751                if (job->store & bit || !job->cbufs[i])
752                        continue;
753                struct v3d_resource *rsc = v3d_resource(job->cbufs[i]->texture);
754
755                job->load |= bit & ~job->clear;
756                if (v3d->blend->base.rt[blend_rt].colormask)
757                        job->store |= bit;
758                v3d_job_add_bo(job, rsc->bo);
759        }
760
761        if (job->referenced_size > 768 * 1024 * 1024) {
762                perf_debug("Flushing job with %dkb to try to free up memory\n",
763                        job->referenced_size / 1024);
764                v3d_flush(pctx);
765        }
766
767        if (V3D_DEBUG & V3D_DEBUG_ALWAYS_FLUSH)
768                v3d_flush(pctx);
769}
770
771/**
772 * Implements gallium's clear() hook (glClear()) by drawing a pair of triangles.
773 */
774static void
775v3d_draw_clear(struct v3d_context *v3d,
776               unsigned buffers,
777               const union pipe_color_union *color,
778               double depth, unsigned stencil)
779{
780        static const union pipe_color_union dummy_color = {};
781
782        /* The blitter util dereferences the color regardless, even though the
783         * gallium clear API may not pass one in when only Z/S are cleared.
784         */
785        if (!color)
786                color = &dummy_color;
787
788        v3d_blitter_save(v3d);
789        util_blitter_clear(v3d->blitter,
790                           v3d->framebuffer.width,
791                           v3d->framebuffer.height,
792                           util_framebuffer_get_num_layers(&v3d->framebuffer),
793                           buffers, color, depth, stencil);
794}
795
796/**
797 * Attempts to perform the GL clear by using the TLB's fast clear at the start
798 * of the frame.
799 */
800static unsigned
801v3d_tlb_clear(struct v3d_job *job, unsigned buffers,
802              const union pipe_color_union *color,
803              double depth, unsigned stencil)
804{
805        struct v3d_context *v3d = job->v3d;
806
807        if (job->draw_calls_queued) {
808                /* If anything in the CL has drawn using the buffer, then the
809                 * TLB clear we're trying to add now would happen before that
810                 * drawing.
811                 */
812                buffers &= ~(job->load | job->store);
813        }
814
815        /* GFXH-1461: If we were to emit a load of just depth or just stencil,
816         * then the clear for the other may get lost.  We need to decide now
817         * if it would be possible to need to emit a load of just one after
818         * we've set up our TLB clears.
819         */
820        if (buffers & PIPE_CLEAR_DEPTHSTENCIL &&
821            (buffers & PIPE_CLEAR_DEPTHSTENCIL) != PIPE_CLEAR_DEPTHSTENCIL &&
822            job->zsbuf &&
823            util_format_is_depth_and_stencil(job->zsbuf->texture->format)) {
824                buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
825        }
826
827        for (int i = 0; i < V3D_MAX_DRAW_BUFFERS; i++) {
828                uint32_t bit = PIPE_CLEAR_COLOR0 << i;
829                if (!(buffers & bit))
830                        continue;
831
832                struct pipe_surface *psurf = v3d->framebuffer.cbufs[i];
833                struct v3d_surface *surf = v3d_surface(psurf);
834                struct v3d_resource *rsc = v3d_resource(psurf->texture);
835
836                union util_color uc;
837                uint32_t internal_size = 4 << surf->internal_bpp;
838
839                static union pipe_color_union swapped_color;
840                if (v3d->swap_color_rb & (1 << i)) {
841                        swapped_color.f[0] = color->f[2];
842                        swapped_color.f[1] = color->f[1];
843                        swapped_color.f[2] = color->f[0];
844                        swapped_color.f[3] = color->f[3];
845                        color = &swapped_color;
846                }
847
848                switch (surf->internal_type) {
849                case V3D_INTERNAL_TYPE_8:
850                        util_pack_color(color->f, PIPE_FORMAT_R8G8B8A8_UNORM,
851                                        &uc);
852                        memcpy(job->clear_color[i], uc.ui, internal_size);
853                        break;
854                case V3D_INTERNAL_TYPE_8I:
855                case V3D_INTERNAL_TYPE_8UI:
856                        job->clear_color[i][0] = ((color->ui[0] & 0xff) |
857                                                  (color->ui[1] & 0xff) << 8 |
858                                                  (color->ui[2] & 0xff) << 16 |
859                                                  (color->ui[3] & 0xff) << 24);
860                        break;
861                case V3D_INTERNAL_TYPE_16F:
862                        util_pack_color(color->f, PIPE_FORMAT_R16G16B16A16_FLOAT,
863                                        &uc);
864                        memcpy(job->clear_color[i], uc.ui, internal_size);
865                        break;
866                case V3D_INTERNAL_TYPE_16I:
867                case V3D_INTERNAL_TYPE_16UI:
868                        job->clear_color[i][0] = ((color->ui[0] & 0xffff) |
869                                                  color->ui[1] << 16);
870                        job->clear_color[i][1] = ((color->ui[2] & 0xffff) |
871                                                  color->ui[3] << 16);
872                        break;
873                case V3D_INTERNAL_TYPE_32F:
874                case V3D_INTERNAL_TYPE_32I:
875                case V3D_INTERNAL_TYPE_32UI:
876                        memcpy(job->clear_color[i], color->ui, internal_size);
877                        break;
878                }
879
880                rsc->initialized_buffers |= bit;
881        }
882
883        unsigned zsclear = buffers & PIPE_CLEAR_DEPTHSTENCIL;
884        if (zsclear) {
885                struct v3d_resource *rsc =
886                        v3d_resource(v3d->framebuffer.zsbuf->texture);
887
888                if (zsclear & PIPE_CLEAR_DEPTH)
889                        job->clear_z = depth;
890                if (zsclear & PIPE_CLEAR_STENCIL)
891                        job->clear_s = stencil;
892
893                rsc->initialized_buffers |= zsclear;
894        }
895
896        job->draw_min_x = 0;
897        job->draw_min_y = 0;
898        job->draw_max_x = v3d->framebuffer.width;
899        job->draw_max_y = v3d->framebuffer.height;
900        job->clear |= buffers;
901        job->store |= buffers;
902
903        v3d_start_draw(v3d);
904
905        return buffers;
906}
907
908static void
909v3d_clear(struct pipe_context *pctx, unsigned buffers,
910          const union pipe_color_union *color, double depth, unsigned stencil)
911{
912        struct v3d_context *v3d = v3d_context(pctx);
913        struct v3d_job *job = v3d_get_job_for_fbo(v3d);
914
915        buffers &= ~v3d_tlb_clear(job, buffers, color, depth, stencil);
916
917        if (buffers)
918                v3d_draw_clear(v3d, buffers, color, depth, stencil);
919}
920
921static void
922v3d_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
923                        const union pipe_color_union *color,
924                        unsigned x, unsigned y, unsigned w, unsigned h,
925                        bool render_condition_enabled)
926{
927        fprintf(stderr, "unimpl: clear RT\n");
928}
929
930static void
931v3d_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
932                        unsigned buffers, double depth, unsigned stencil,
933                        unsigned x, unsigned y, unsigned w, unsigned h,
934                        bool render_condition_enabled)
935{
936        fprintf(stderr, "unimpl: clear DS\n");
937}
938
939void
940v3dX(draw_init)(struct pipe_context *pctx)
941{
942        pctx->draw_vbo = v3d_draw_vbo;
943        pctx->clear = v3d_clear;
944        pctx->clear_render_target = v3d_clear_render_target;
945        pctx->clear_depth_stencil = v3d_clear_depth_stencil;
946}
947