2 * Copyright © 2014-2017 Broadcom
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:
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
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
24 #include "util/u_blitter.h"
25 #include "util/u_prim.h"
26 #include "util/format/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"
32 #include "v3d_context.h"
33 #include "v3d_resource.h"
35 #include "broadcom/compiler/v3d_compiler.h"
36 #include "broadcom/common/v3d_macros.h"
37 #include "broadcom/cle/v3dx_pack.h"
40 * Does the initial bining command list setup for drawing to a given FBO.
43 v3d_start_draw(struct v3d_context
*v3d
)
45 struct v3d_job
*job
= v3d
->job
;
50 /* Get space to emit our BCL state, using a branch to jump to a new BO
53 v3d_cl_ensure_space_with_branch(&job
->bcl
, 256 /* XXX */);
55 job
->submit
.bcl_start
= job
->bcl
.bo
->offset
;
56 v3d_job_add_bo(job
, job
->bcl
.bo
);
58 /* The PTB will request the tile alloc initial size per tile at start
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);
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).
70 tile_alloc_size
+= 8192;
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.
76 tile_alloc_size
+= 512 * 1024;
78 job
->tile_alloc
= v3d_bo_alloc(v3d
->screen
, tile_alloc_size
,
80 uint32_t tsda_per_tile_size
= v3d
->screen
->devinfo
.ver
>= 40 ? 256 : 64;
81 job
->tile_state
= v3d_bo_alloc(v3d
->screen
,
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);
94 config
.multisample_mode_4x
= job
->msaa
;
96 config
.maximum_bpp_of_all_render_targets
= job
->internal_bpp
;
98 #else /* V3D_VERSION < 40 */
99 /* "Binning mode lists start with a Tile Binning Mode Configuration
102 * Part1 signals the end of binning config setup.
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
;
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);
114 config
.width_in_tiles
= job
->draw_tiles_x
;
115 config
.height_in_tiles
= job
->draw_tiles_y
;
117 config
.number_of_render_targets
=
118 MAX2(v3d
->framebuffer
.nr_cbufs
, 1);
120 config
.multisample_mode_4x
= job
->msaa
;
122 config
.maximum_bpp_of_all_render_targets
= job
->internal_bpp
;
124 #endif /* V3D_VERSION < 40 */
126 /* There's definitely nothing in the VCD cache we want. */
127 cl_emit(&job
->bcl
, FLUSH_VCD_CACHE
, bin
);
129 /* Disable any leftover OQ state from another job. */
130 cl_emit(&job
->bcl
, OCCLUSION_QUERY_COUNTER
, counter
);
132 /* "Binning mode lists must have a Start Tile Binning item (6) after
133 * any prefix state data before the binning list proper starts."
135 cl_emit(&job
->bcl
, START_TILE_BINNING
, bin
);
137 job
->needs_flush
= true;
138 job
->draw_width
= v3d
->framebuffer
.width
;
139 job
->draw_height
= v3d
->framebuffer
.height
;
143 v3d_predraw_check_stage_inputs(struct pipe_context
*pctx
,
144 enum pipe_shader_type s
)
146 struct v3d_context
*v3d
= v3d_context(pctx
);
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
];
153 struct v3d_sampler_view
*view
= v3d_sampler_view(pview
);
155 if (view
->texture
!= view
->base
.texture
&&
156 view
->base
.format
!= PIPE_FORMAT_X32_S8X24_UINT
)
157 v3d_update_shadow_texture(pctx
, &view
->base
);
159 v3d_flush_jobs_writing_resource(v3d
, view
->texture
,
163 /* Flush writes to UBOs. */
164 foreach_bit(i
, v3d
->constbuf
[s
].enabled_mask
) {
165 struct pipe_constant_buffer
*cb
= &v3d
->constbuf
[s
].cb
[i
];
167 v3d_flush_jobs_writing_resource(v3d
, cb
->buffer
,
172 /* Flush reads/writes to our SSBOs */
173 foreach_bit(i
, v3d
->ssbo
[s
].enabled_mask
) {
174 struct pipe_shader_buffer
*sb
= &v3d
->ssbo
[s
].sb
[i
];
176 v3d_flush_jobs_reading_resource(v3d
, sb
->buffer
,
177 V3D_FLUSH_NOT_CURRENT_JOB
);
181 /* Flush reads/writes to our image views */
182 foreach_bit(i
, v3d
->shaderimg
[s
].enabled_mask
) {
183 struct v3d_image_view
*view
= &v3d
->shaderimg
[s
].si
[i
];
185 v3d_flush_jobs_reading_resource(v3d
, view
->base
.resource
,
186 V3D_FLUSH_NOT_CURRENT_JOB
);
189 /* Flush writes to our vertex buffers (i.e. from transform feedback) */
190 if (s
== PIPE_SHADER_VERTEX
) {
191 foreach_bit(i
, v3d
->vertexbuf
.enabled_mask
) {
192 struct pipe_vertex_buffer
*vb
= &v3d
->vertexbuf
.vb
[i
];
194 v3d_flush_jobs_writing_resource(v3d
, vb
->buffer
.resource
,
201 v3d_predraw_check_outputs(struct pipe_context
*pctx
)
203 struct v3d_context
*v3d
= v3d_context(pctx
);
205 /* Flush jobs reading from TF buffers that we are about to write. */
206 if (v3d_transform_feedback_enabled(v3d
)) {
207 struct v3d_streamout_stateobj
*so
= &v3d
->streamout
;
209 for (int i
= 0; i
< so
->num_targets
; i
++) {
213 const struct pipe_stream_output_target
*target
=
215 v3d_flush_jobs_reading_resource(v3d
, target
->buffer
,
222 * Checks if the state for the current draw reads a particular resource in
223 * in the given shader stage.
226 v3d_state_reads_resource(struct v3d_context
*v3d
,
227 struct pipe_resource
*prsc
,
228 enum pipe_shader_type s
)
230 struct v3d_resource
*rsc
= v3d_resource(prsc
);
233 if (s
== PIPE_SHADER_VERTEX
) {
234 foreach_bit(i
, v3d
->vertexbuf
.enabled_mask
) {
235 struct pipe_vertex_buffer
*vb
= &v3d
->vertexbuf
.vb
[i
];
236 if (!vb
->buffer
.resource
)
239 struct v3d_resource
*vb_rsc
=
240 v3d_resource(vb
->buffer
.resource
);
241 if (rsc
->bo
== vb_rsc
->bo
)
246 /* Constant buffers */
247 foreach_bit(i
, v3d
->constbuf
[s
].enabled_mask
) {
248 struct pipe_constant_buffer
*cb
= &v3d
->constbuf
[s
].cb
[i
];
252 struct v3d_resource
*cb_rsc
= v3d_resource(cb
->buffer
);
253 if (rsc
->bo
== cb_rsc
->bo
)
257 /* Shader storage buffers */
258 foreach_bit(i
, v3d
->ssbo
[s
].enabled_mask
) {
259 struct pipe_shader_buffer
*sb
= &v3d
->ssbo
[s
].sb
[i
];
263 struct v3d_resource
*sb_rsc
= v3d_resource(sb
->buffer
);
264 if (rsc
->bo
== sb_rsc
->bo
)
269 for (int i
= 0; i
< v3d
->tex
[s
].num_textures
; i
++) {
270 struct pipe_sampler_view
*pview
= v3d
->tex
[s
].textures
[i
];
274 struct v3d_sampler_view
*view
= v3d_sampler_view(pview
);
275 struct v3d_resource
*v_rsc
= v3d_resource(view
->texture
);
276 if (rsc
->bo
== v_rsc
->bo
)
284 v3d_emit_wait_for_tf(struct v3d_job
*job
)
286 /* XXX: we might be able to skip this in some cases, for now we
289 cl_emit(&job
->bcl
, FLUSH_TRANSFORM_FEEDBACK_DATA
, flush
);
291 cl_emit(&job
->bcl
, WAIT_FOR_TRANSFORM_FEEDBACK
, wait
) {
292 /* XXX: Wait for all outstanding writes... maybe we can do
293 * better in some cases.
295 wait
.block_count
= 255;
298 /* We have just flushed all our outstanding TF work in this job so make
299 * sure we don't emit TF flushes again for any of it again.
301 _mesa_set_clear(job
->tf_write_prscs
, NULL
);
305 v3d_emit_wait_for_tf_if_needed(struct v3d_context
*v3d
, struct v3d_job
*job
)
307 if (!job
->tf_enabled
)
310 set_foreach(job
->tf_write_prscs
, entry
) {
311 struct pipe_resource
*prsc
= (struct pipe_resource
*)entry
->key
;
312 for (int s
= 0; s
< PIPE_SHADER_COMPUTE
; s
++) {
313 /* Fragment shaders can only start executing after all
314 * binning (and thus TF) is complete.
316 * XXX: For VS/GS/TES, if the binning shader does not
317 * read the resource then we could also avoid emitting
320 if (s
== PIPE_SHADER_FRAGMENT
)
323 if (v3d_state_reads_resource(v3d
, prsc
, s
)) {
324 v3d_emit_wait_for_tf(job
);
331 #if V3D_VERSION >= 41
333 v3d_emit_gs_state_record(struct v3d_job
*job
,
334 struct v3d_compiled_shader
*gs_bin
,
335 struct v3d_cl_reloc gs_bin_uniforms
,
336 struct v3d_compiled_shader
*gs
,
337 struct v3d_cl_reloc gs_render_uniforms
)
339 cl_emit(&job
->indirect
, GEOMETRY_SHADER_STATE_RECORD
, shader
) {
340 shader
.geometry_bin_mode_shader_code_address
=
341 cl_address(v3d_resource(gs_bin
->resource
)->bo
,
343 shader
.geometry_bin_mode_shader_4_way_threadable
=
344 gs_bin
->prog_data
.gs
->base
.threads
== 4;
345 shader
.geometry_bin_mode_shader_start_in_final_thread_section
=
346 gs_bin
->prog_data
.gs
->base
.single_seg
;
347 shader
.geometry_bin_mode_shader_propagate_nans
= true;
348 shader
.geometry_bin_mode_shader_uniforms_address
=
351 shader
.geometry_render_mode_shader_code_address
=
352 cl_address(v3d_resource(gs
->resource
)->bo
, gs
->offset
);
353 shader
.geometry_render_mode_shader_4_way_threadable
=
354 gs
->prog_data
.gs
->base
.threads
== 4;
355 shader
.geometry_render_mode_shader_start_in_final_thread_section
=
356 gs
->prog_data
.gs
->base
.single_seg
;
357 shader
.geometry_render_mode_shader_propagate_nans
= true;
358 shader
.geometry_render_mode_shader_uniforms_address
=
364 v3d_gs_output_primitive(uint32_t prim_type
)
368 return GEOMETRY_SHADER_POINTS
;
370 return GEOMETRY_SHADER_LINE_STRIP
;
371 case GL_TRIANGLE_STRIP
:
372 return GEOMETRY_SHADER_TRI_STRIP
;
374 unreachable("Unsupported primitive type");
379 v3d_emit_tes_gs_common_params(struct v3d_job
*job
,
380 uint8_t gs_out_prim_type
)
382 /* This, and v3d_emit_tes_gs_shader_params below, fill in default
383 * values for tessellation fields even though we don't support
384 * tessellation yet because our packing functions (and the simulator)
385 * complain if we don't.
387 cl_emit(&job
->indirect
, TESSELLATION_GEOMETRY_COMMON_PARAMS
, shader
) {
388 shader
.tessellation_type
= TESSELLATION_TYPE_TRIANGLE
;
389 shader
.tessellation_point_mode
= false;
390 shader
.tessellation_edge_spacing
= TESSELLATION_EDGE_SPACING_EVEN
;
391 shader
.tessellation_clockwise
= true;
392 shader
.tessellation_invocations
= 1;
394 shader
.geometry_shader_output_format
=
395 v3d_gs_output_primitive(gs_out_prim_type
);
396 shader
.geometry_shader_instances
= 1; /* FIXME */
401 v3d_emit_tes_gs_shader_params(struct v3d_job
*job
,
402 struct v3d_gs_prog_data
*gs
)
404 cl_emit(&job
->indirect
, TESSELLATION_GEOMETRY_SHADER_PARAMS
, shader
) {
405 shader
.tcs_batch_flush_mode
= V3D_TCS_FLUSH_MODE_FULLY_PACKED
;
406 shader
.per_patch_data_column_depth
= 1;
407 shader
.tcs_output_segment_size_in_sectors
= 1;
408 shader
.tcs_output_segment_pack_mode
= V3D_PACK_MODE_16_WAY
;
409 shader
.tes_output_segment_size_in_sectors
= 1;
410 shader
.tes_output_segment_pack_mode
= V3D_PACK_MODE_16_WAY
;
411 shader
.gs_output_segment_size_in_sectors
=
413 shader
.gs_output_segment_pack_mode
= V3D_PACK_MODE_16_WAY
; /* FIXME*/
414 shader
.tbg_max_patches_per_tcs_batch
= 1;
415 shader
.tbg_max_extra_vertex_segs_for_patches_after_first
= 0;
416 shader
.tbg_min_tcs_output_segments_required_in_play
= 1;
417 shader
.tbg_min_per_patch_data_segments_required_in_play
= 1;
418 shader
.tpg_max_patches_per_tes_batch
= 1;
419 shader
.tpg_max_vertex_segments_per_tes_batch
= 0;
420 shader
.tpg_max_tcs_output_segments_per_tes_batch
= 1;
421 shader
.tpg_min_tes_output_segments_required_in_play
= 1;
422 shader
.gbg_max_tes_output_vertex_segments_per_gs_batch
= 0;
423 shader
.gbg_min_gs_output_segments_required_in_play
= 1;
430 v3d_emit_gl_shader_state(struct v3d_context
*v3d
,
431 const struct pipe_draw_info
*info
)
433 struct v3d_job
*job
= v3d
->job
;
434 /* VC5_DIRTY_VTXSTATE */
435 struct v3d_vertex_stateobj
*vtx
= v3d
->vtx
;
436 /* VC5_DIRTY_VTXBUF */
437 struct v3d_vertexbuf_stateobj
*vertexbuf
= &v3d
->vertexbuf
;
439 /* Upload the uniforms to the indirect CL first */
440 struct v3d_cl_reloc fs_uniforms
=
441 v3d_write_uniforms(v3d
, job
, v3d
->prog
.fs
,
442 PIPE_SHADER_FRAGMENT
);
444 struct v3d_cl_reloc gs_uniforms
= { NULL
, 0 };
445 struct v3d_cl_reloc gs_bin_uniforms
= { NULL
, 0 };
447 gs_uniforms
= v3d_write_uniforms(v3d
, job
, v3d
->prog
.gs
,
448 PIPE_SHADER_GEOMETRY
);
450 if (v3d
->prog
.gs_bin
) {
451 gs_bin_uniforms
= v3d_write_uniforms(v3d
, job
, v3d
->prog
.gs_bin
,
452 PIPE_SHADER_GEOMETRY
);
455 struct v3d_cl_reloc vs_uniforms
=
456 v3d_write_uniforms(v3d
, job
, v3d
->prog
.vs
,
458 struct v3d_cl_reloc cs_uniforms
=
459 v3d_write_uniforms(v3d
, job
, v3d
->prog
.cs
,
462 /* Update the cache dirty flag based on the shader progs data */
463 job
->tmu_dirty_rcl
|= v3d
->prog
.cs
->prog_data
.vs
->base
.tmu_dirty_rcl
;
464 job
->tmu_dirty_rcl
|= v3d
->prog
.vs
->prog_data
.vs
->base
.tmu_dirty_rcl
;
465 if (v3d
->prog
.gs_bin
) {
466 job
->tmu_dirty_rcl
|=
467 v3d
->prog
.gs_bin
->prog_data
.gs
->base
.tmu_dirty_rcl
;
470 job
->tmu_dirty_rcl
|=
471 v3d
->prog
.gs
->prog_data
.gs
->base
.tmu_dirty_rcl
;
473 job
->tmu_dirty_rcl
|= v3d
->prog
.fs
->prog_data
.fs
->base
.tmu_dirty_rcl
;
475 /* See GFXH-930 workaround below */
476 uint32_t num_elements_to_emit
= MAX2(vtx
->num_elements
, 1);
478 uint32_t shader_state_record_length
=
479 cl_packet_length(GL_SHADER_STATE_RECORD
);
480 #if V3D_VERSION >= 41
482 shader_state_record_length
+=
483 cl_packet_length(GEOMETRY_SHADER_STATE_RECORD
) +
484 cl_packet_length(TESSELLATION_GEOMETRY_COMMON_PARAMS
) +
485 2 * cl_packet_length(TESSELLATION_GEOMETRY_SHADER_PARAMS
);
489 uint32_t shader_rec_offset
=
490 v3d_cl_ensure_space(&job
->indirect
,
491 shader_state_record_length
+
492 num_elements_to_emit
*
493 cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD
),
496 /* XXX perf: We should move most of the SHADER_STATE_RECORD setup to
497 * compile time, so that we mostly just have to OR the VS and FS
498 * records together at draw time.
500 #if V3D_VERSION >= 41
502 v3d_emit_gs_state_record(v3d
->job
,
503 v3d
->prog
.gs_bin
, gs_bin_uniforms
,
504 v3d
->prog
.gs
, gs_uniforms
);
506 struct v3d_gs_prog_data
*gs
= v3d
->prog
.gs
->prog_data
.gs
;
507 struct v3d_gs_prog_data
*gs_bin
= v3d
->prog
.gs_bin
->prog_data
.gs
;
509 v3d_emit_tes_gs_common_params(v3d
->job
, gs
->out_prim_type
);
510 v3d_emit_tes_gs_shader_params(v3d
->job
, gs_bin
);
511 v3d_emit_tes_gs_shader_params(v3d
->job
, gs
);
515 cl_emit(&job
->indirect
, GL_SHADER_STATE_RECORD
, shader
) {
516 shader
.enable_clipping
= true;
517 /* VC5_DIRTY_PRIM_MODE | VC5_DIRTY_RASTERIZER */
518 shader
.point_size_in_shaded_vertex_data
=
519 (info
->mode
== PIPE_PRIM_POINTS
&&
520 v3d
->rasterizer
->base
.point_size_per_vertex
);
522 /* Must be set if the shader modifies Z, discards, or modifies
523 * the sample mask. For any of these cases, the fragment
524 * shader needs to write the Z value (even just discards).
526 shader
.fragment_shader_does_z_writes
=
527 v3d
->prog
.fs
->prog_data
.fs
->writes_z
;
528 /* Set if the EZ test must be disabled (due to shader side
529 * effects and the early_z flag not being present in the
532 shader
.turn_off_early_z_test
=
533 v3d
->prog
.fs
->prog_data
.fs
->disable_ez
;
535 shader
.fragment_shader_uses_real_pixel_centre_w_in_addition_to_centroid_w2
=
536 v3d
->prog
.fs
->prog_data
.fs
->uses_center_w
;
538 #if V3D_VERSION >= 41
539 shader
.any_shader_reads_hardware_written_primitive_id
=
540 v3d
->prog
.gs
? v3d
->prog
.gs
->prog_data
.gs
->uses_pid
:
544 #if V3D_VERSION >= 40
545 shader
.do_scoreboard_wait_on_first_thread_switch
=
546 v3d
->prog
.fs
->prog_data
.fs
->lock_scoreboard_on_first_thrsw
;
547 shader
.disable_implicit_point_line_varyings
=
548 !v3d
->prog
.fs
->prog_data
.fs
->uses_implicit_point_line_varyings
;
551 shader
.number_of_varyings_in_fragment_shader
=
552 v3d
->prog
.fs
->prog_data
.fs
->num_inputs
;
554 shader
.coordinate_shader_propagate_nans
= true;
555 shader
.vertex_shader_propagate_nans
= true;
556 shader
.fragment_shader_propagate_nans
= true;
558 shader
.coordinate_shader_code_address
=
559 cl_address(v3d_resource(v3d
->prog
.cs
->resource
)->bo
,
560 v3d
->prog
.cs
->offset
);
561 shader
.vertex_shader_code_address
=
562 cl_address(v3d_resource(v3d
->prog
.vs
->resource
)->bo
,
563 v3d
->prog
.vs
->offset
);
564 shader
.fragment_shader_code_address
=
565 cl_address(v3d_resource(v3d
->prog
.fs
->resource
)->bo
,
566 v3d
->prog
.fs
->offset
);
568 /* XXX: Use combined input/output size flag in the common
571 shader
.coordinate_shader_has_separate_input_and_output_vpm_blocks
=
572 v3d
->prog
.cs
->prog_data
.vs
->separate_segments
;
573 shader
.vertex_shader_has_separate_input_and_output_vpm_blocks
=
574 v3d
->prog
.vs
->prog_data
.vs
->separate_segments
;
576 shader
.coordinate_shader_input_vpm_segment_size
=
577 v3d
->prog
.cs
->prog_data
.vs
->separate_segments
?
578 v3d
->prog
.cs
->prog_data
.vs
->vpm_input_size
: 1;
579 shader
.vertex_shader_input_vpm_segment_size
=
580 v3d
->prog
.vs
->prog_data
.vs
->separate_segments
?
581 v3d
->prog
.vs
->prog_data
.vs
->vpm_input_size
: 1;
583 shader
.coordinate_shader_output_vpm_segment_size
=
584 v3d
->prog
.cs
->prog_data
.vs
->vpm_output_size
;
585 shader
.vertex_shader_output_vpm_segment_size
=
586 v3d
->prog
.vs
->prog_data
.vs
->vpm_output_size
;
588 shader
.coordinate_shader_uniforms_address
= cs_uniforms
;
589 shader
.vertex_shader_uniforms_address
= vs_uniforms
;
590 shader
.fragment_shader_uniforms_address
= fs_uniforms
;
592 #if V3D_VERSION >= 41
593 shader
.min_coord_shader_input_segments_required_in_play
= 1;
594 shader
.min_vertex_shader_input_segments_required_in_play
= 1;
596 shader
.coordinate_shader_4_way_threadable
=
597 v3d
->prog
.cs
->prog_data
.vs
->base
.threads
== 4;
598 shader
.vertex_shader_4_way_threadable
=
599 v3d
->prog
.vs
->prog_data
.vs
->base
.threads
== 4;
600 shader
.fragment_shader_4_way_threadable
=
601 v3d
->prog
.fs
->prog_data
.fs
->base
.threads
== 4;
603 shader
.coordinate_shader_start_in_final_thread_section
=
604 v3d
->prog
.cs
->prog_data
.vs
->base
.single_seg
;
605 shader
.vertex_shader_start_in_final_thread_section
=
606 v3d
->prog
.vs
->prog_data
.vs
->base
.single_seg
;
607 shader
.fragment_shader_start_in_final_thread_section
=
608 v3d
->prog
.fs
->prog_data
.fs
->base
.single_seg
;
610 shader
.coordinate_shader_4_way_threadable
=
611 v3d
->prog
.cs
->prog_data
.vs
->base
.threads
== 4;
612 shader
.coordinate_shader_2_way_threadable
=
613 v3d
->prog
.cs
->prog_data
.vs
->base
.threads
== 2;
614 shader
.vertex_shader_4_way_threadable
=
615 v3d
->prog
.vs
->prog_data
.vs
->base
.threads
== 4;
616 shader
.vertex_shader_2_way_threadable
=
617 v3d
->prog
.vs
->prog_data
.vs
->base
.threads
== 2;
618 shader
.fragment_shader_4_way_threadable
=
619 v3d
->prog
.fs
->prog_data
.fs
->base
.threads
== 4;
620 shader
.fragment_shader_2_way_threadable
=
621 v3d
->prog
.fs
->prog_data
.fs
->base
.threads
== 2;
624 shader
.vertex_id_read_by_coordinate_shader
=
625 v3d
->prog
.cs
->prog_data
.vs
->uses_vid
;
626 shader
.instance_id_read_by_coordinate_shader
=
627 v3d
->prog
.cs
->prog_data
.vs
->uses_iid
;
628 shader
.vertex_id_read_by_vertex_shader
=
629 v3d
->prog
.vs
->prog_data
.vs
->uses_vid
;
630 shader
.instance_id_read_by_vertex_shader
=
631 v3d
->prog
.vs
->prog_data
.vs
->uses_iid
;
633 shader
.address_of_default_attribute_values
=
634 cl_address(v3d_resource(vtx
->defaults
)->bo
,
635 vtx
->defaults_offset
);
638 bool cs_loaded_any
= false;
639 for (int i
= 0; i
< vtx
->num_elements
; i
++) {
640 struct pipe_vertex_element
*elem
= &vtx
->pipe
[i
];
641 struct pipe_vertex_buffer
*vb
=
642 &vertexbuf
->vb
[elem
->vertex_buffer_index
];
643 struct v3d_resource
*rsc
= v3d_resource(vb
->buffer
.resource
);
645 const uint32_t size
=
646 cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD
);
647 cl_emit_with_prepacked(&job
->indirect
,
648 GL_SHADER_STATE_ATTRIBUTE_RECORD
,
649 &vtx
->attrs
[i
* size
], attr
) {
650 attr
.stride
= vb
->stride
;
651 attr
.address
= cl_address(rsc
->bo
,
654 attr
.number_of_values_read_by_coordinate_shader
=
655 v3d
->prog
.cs
->prog_data
.vs
->vattr_sizes
[i
];
656 attr
.number_of_values_read_by_vertex_shader
=
657 v3d
->prog
.vs
->prog_data
.vs
->vattr_sizes
[i
];
659 /* GFXH-930: At least one attribute must be enabled
660 * and read by CS and VS. If we have attributes being
661 * consumed by the VS but not the CS, then set up a
662 * dummy load of the last attribute into the CS's VPM
663 * inputs. (Since CS is just dead-code-elimination
664 * compared to VS, we can't have CS loading but not
667 if (v3d
->prog
.cs
->prog_data
.vs
->vattr_sizes
[i
])
668 cs_loaded_any
= true;
669 if (i
== vtx
->num_elements
- 1 && !cs_loaded_any
) {
670 attr
.number_of_values_read_by_coordinate_shader
= 1;
672 #if V3D_VERSION >= 41
673 attr
.maximum_index
= 0xffffff;
676 STATIC_ASSERT(sizeof(vtx
->attrs
) >= V3D_MAX_VS_INPUTS
/ 4 * size
);
679 if (vtx
->num_elements
== 0) {
680 /* GFXH-930: At least one attribute must be enabled and read
681 * by CS and VS. If we have no attributes being consumed by
682 * the shader, set up a dummy to be loaded into the VPM.
684 cl_emit(&job
->indirect
, GL_SHADER_STATE_ATTRIBUTE_RECORD
, attr
) {
685 /* Valid address of data whose value will be unused. */
686 attr
.address
= cl_address(job
->indirect
.bo
, 0);
688 attr
.type
= ATTRIBUTE_FLOAT
;
692 attr
.number_of_values_read_by_coordinate_shader
= 1;
693 attr
.number_of_values_read_by_vertex_shader
= 1;
697 cl_emit(&job
->bcl
, VCM_CACHE_SIZE
, vcm
) {
698 vcm
.number_of_16_vertex_batches_for_binning
=
699 v3d
->prog
.cs
->prog_data
.vs
->vcm_cache_size
;
700 vcm
.number_of_16_vertex_batches_for_rendering
=
701 v3d
->prog
.vs
->prog_data
.vs
->vcm_cache_size
;
704 #if V3D_VERSION >= 41
706 cl_emit(&job
->bcl
, GL_SHADER_STATE_INCLUDING_GS
, state
) {
707 state
.address
= cl_address(job
->indirect
.bo
,
709 state
.number_of_attribute_arrays
= num_elements_to_emit
;
712 cl_emit(&job
->bcl
, GL_SHADER_STATE
, state
) {
713 state
.address
= cl_address(job
->indirect
.bo
,
715 state
.number_of_attribute_arrays
= num_elements_to_emit
;
719 assert(!v3d
->prog
.gs
);
720 cl_emit(&job
->bcl
, GL_SHADER_STATE
, state
) {
721 state
.address
= cl_address(job
->indirect
.bo
, shader_rec_offset
);
722 state
.number_of_attribute_arrays
= num_elements_to_emit
;
726 v3d_bo_unreference(&cs_uniforms
.bo
);
727 v3d_bo_unreference(&vs_uniforms
.bo
);
729 v3d_bo_unreference(&gs_uniforms
.bo
);
730 if (gs_bin_uniforms
.bo
)
731 v3d_bo_unreference(&gs_bin_uniforms
.bo
);
732 v3d_bo_unreference(&fs_uniforms
.bo
);
736 * Updates the number of primitvies generated from the number of vertices
737 * to draw. We do this here instead of using PRIMITIVE_COUNTS_FEEDBACK because
738 * using the GPU packet for this might require sync waits and this is trivial
739 * to handle in the CPU instead.
742 v3d_update_primitives_generated_counter(struct v3d_context
*v3d
,
743 const struct pipe_draw_info
*info
)
745 if (!v3d
->active_queries
)
748 uint32_t prims
= u_prims_for_vertices(info
->mode
, info
->count
);
749 v3d
->prims_generated
+= prims
;
753 v3d_update_job_ez(struct v3d_context
*v3d
, struct v3d_job
*job
)
755 switch (v3d
->zsa
->ez_state
) {
756 case VC5_EZ_UNDECIDED
:
757 /* If the Z/S state didn't pick a direction but didn't
758 * disable, then go along with the current EZ state. This
759 * allows EZ optimization for Z func == EQUAL or NEVER.
765 /* If the Z/S state picked a direction, then it needs to match
766 * the current direction if we've decided on one.
768 if (job
->ez_state
== VC5_EZ_UNDECIDED
)
769 job
->ez_state
= v3d
->zsa
->ez_state
;
770 else if (job
->ez_state
!= v3d
->zsa
->ez_state
)
771 job
->ez_state
= VC5_EZ_DISABLED
;
774 case VC5_EZ_DISABLED
:
775 /* If the current Z/S state disables EZ because of a bad Z
776 * func or stencil operation, then we can't do any more EZ in
779 job
->ez_state
= VC5_EZ_DISABLED
;
783 /* If the FS affects the Z of the pixels, then it may update against
784 * the chosen EZ direction (though we could use
785 * ARB_conservative_depth's hints to avoid this)
787 if (v3d
->prog
.fs
->prog_data
.fs
->writes_z
) {
788 job
->ez_state
= VC5_EZ_DISABLED
;
791 if (job
->first_ez_state
== VC5_EZ_UNDECIDED
&&
792 (job
->ez_state
!= VC5_EZ_DISABLED
|| job
->draw_calls_queued
== 0))
793 job
->first_ez_state
= job
->ez_state
;
797 v3d_draw_vbo(struct pipe_context
*pctx
, const struct pipe_draw_info
*info
)
799 struct v3d_context
*v3d
= v3d_context(pctx
);
801 if (!info
->count_from_stream_output
&& !info
->indirect
&&
802 !info
->primitive_restart
&&
803 !u_trim_pipe_prim(info
->mode
, (unsigned*)&info
->count
))
806 /* Fall back for weird desktop GL primitive restart values. */
807 if (info
->primitive_restart
&&
811 switch (info
->index_size
) {
820 if (info
->restart_index
!= mask
) {
821 util_draw_vbo_without_prim_restart(pctx
, info
);
826 if (info
->mode
>= PIPE_PRIM_QUADS
) {
827 util_primconvert_save_rasterizer_state(v3d
->primconvert
, &v3d
->rasterizer
->base
);
828 util_primconvert_draw_vbo(v3d
->primconvert
, info
);
829 perf_debug("Fallback conversion for %d %s vertices\n",
830 info
->count
, u_prim_name(info
->mode
));
834 /* Before setting up the draw, flush anything writing to the resources
835 * that we read from or reading from resources we write to.
837 for (int s
= 0; s
< PIPE_SHADER_COMPUTE
; s
++)
838 v3d_predraw_check_stage_inputs(pctx
, s
);
840 if (info
->indirect
) {
841 v3d_flush_jobs_writing_resource(v3d
, info
->indirect
->buffer
,
845 v3d_predraw_check_outputs(pctx
);
847 /* If transform feedback is active and we are switching primitive type
848 * we need to submit the job before drawing and update the vertex count
849 * written to TF based on the primitive type since we will need to
850 * know the exact vertex count if the application decides to call
851 * glDrawTransformFeedback() later.
853 if (v3d
->streamout
.num_targets
> 0 &&
854 u_base_prim_type(info
->mode
) != u_base_prim_type(v3d
->prim_mode
)) {
855 v3d_tf_update_counters(v3d
);
858 struct v3d_job
*job
= v3d_get_job_for_fbo(v3d
);
860 /* If vertex texturing depends on the output of rendering, we need to
861 * ensure that that rendering is complete before we run a coordinate
862 * shader that depends on it.
864 * Given that doing that is unusual, for now we just block the binner
865 * on the last submitted render, rather than tracking the last
866 * rendering to each texture's BO.
868 if (v3d
->tex
[PIPE_SHADER_VERTEX
].num_textures
|| info
->indirect
) {
869 perf_debug("Blocking binner on last render "
870 "due to vertex texturing or indirect drawing.\n");
871 job
->submit
.in_sync_bcl
= v3d
->out_sync
;
874 /* Mark SSBOs and images as being written. We don't actually know
875 * which ones are read vs written, so just assume the worst.
877 for (int s
= 0; s
< PIPE_SHADER_COMPUTE
; s
++) {
878 foreach_bit(i
, v3d
->ssbo
[s
].enabled_mask
) {
879 v3d_job_add_write_resource(job
,
880 v3d
->ssbo
[s
].sb
[i
].buffer
);
881 job
->tmu_dirty_rcl
= true;
884 foreach_bit(i
, v3d
->shaderimg
[s
].enabled_mask
) {
885 v3d_job_add_write_resource(job
,
886 v3d
->shaderimg
[s
].si
[i
].base
.resource
);
887 job
->tmu_dirty_rcl
= true;
891 /* Get space to emit our draw call into the BCL, using a branch to
892 * jump to a new BO if necessary.
894 v3d_cl_ensure_space_with_branch(&job
->bcl
, 256 /* XXX */);
896 if (v3d
->prim_mode
!= info
->mode
) {
897 v3d
->prim_mode
= info
->mode
;
898 v3d
->dirty
|= VC5_DIRTY_PRIM_MODE
;
902 v3d_update_compiled_shaders(v3d
, info
->mode
);
903 v3d_update_job_ez(v3d
, job
);
905 /* If this job was writing to transform feedback buffers before this
906 * draw and we are reading from them here, then we need to wait for TF
907 * to complete before we emit this draw.
909 * Notice this check needs to happen before we emit state for the
910 * current draw call, where we update job->tf_enabled, so we can ensure
911 * that we only check TF writes for prior draws.
913 v3d_emit_wait_for_tf_if_needed(v3d
, job
);
915 #if V3D_VERSION >= 41
916 v3d41_emit_state(pctx
);
918 v3d33_emit_state(pctx
);
921 if (v3d
->dirty
& (VC5_DIRTY_VTXBUF
|
923 VC5_DIRTY_PRIM_MODE
|
924 VC5_DIRTY_RASTERIZER
|
925 VC5_DIRTY_COMPILED_CS
|
926 VC5_DIRTY_COMPILED_VS
|
927 VC5_DIRTY_COMPILED_GS_BIN
|
928 VC5_DIRTY_COMPILED_GS
|
929 VC5_DIRTY_COMPILED_FS
|
930 v3d
->prog
.cs
->uniform_dirty_bits
|
931 v3d
->prog
.vs
->uniform_dirty_bits
|
933 v3d
->prog
.gs_bin
->uniform_dirty_bits
: 0) |
935 v3d
->prog
.gs
->uniform_dirty_bits
: 0) |
936 v3d
->prog
.fs
->uniform_dirty_bits
)) {
937 v3d_emit_gl_shader_state(v3d
, info
);
942 /* The Base Vertex/Base Instance packet sets those values to nonzero
943 * for the next draw call only.
945 if (info
->index_bias
|| info
->start_instance
) {
946 cl_emit(&job
->bcl
, BASE_VERTEX_BASE_INSTANCE
, base
) {
947 base
.base_instance
= info
->start_instance
;
948 base
.base_vertex
= info
->index_bias
;
952 uint32_t prim_tf_enable
= 0;
954 /* V3D 3.x: The HW only processes transform feedback on primitives
957 if (v3d
->streamout
.num_targets
)
958 prim_tf_enable
= (V3D_PRIM_POINTS_TF
- V3D_PRIM_POINTS
);
961 v3d_update_primitives_generated_counter(v3d
, info
);
963 /* Note that the primitive type fields match with OpenGL/gallium
964 * definitions, up to but not including QUADS.
966 if (info
->index_size
) {
967 uint32_t index_size
= info
->index_size
;
968 uint32_t offset
= info
->start
* index_size
;
969 struct pipe_resource
*prsc
;
970 if (info
->has_user_indices
) {
972 u_upload_data(v3d
->uploader
, 0,
973 info
->count
* info
->index_size
, 4,
977 prsc
= info
->index
.resource
;
979 struct v3d_resource
*rsc
= v3d_resource(prsc
);
981 #if V3D_VERSION >= 40
982 cl_emit(&job
->bcl
, INDEX_BUFFER_SETUP
, ib
) {
983 ib
.address
= cl_address(rsc
->bo
, 0);
984 ib
.size
= rsc
->bo
->size
;
988 if (info
->indirect
) {
989 cl_emit(&job
->bcl
, INDIRECT_INDEXED_INSTANCED_PRIM_LIST
, prim
) {
990 prim
.index_type
= ffs(info
->index_size
) - 1;
992 prim
.address_of_indices_list
=
993 cl_address(rsc
->bo
, offset
);
994 #endif /* V3D_VERSION < 40 */
995 prim
.mode
= info
->mode
| prim_tf_enable
;
996 prim
.enable_primitive_restarts
= info
->primitive_restart
;
998 prim
.number_of_draw_indirect_indexed_records
= info
->indirect
->draw_count
;
1000 prim
.stride_in_multiples_of_4_bytes
= info
->indirect
->stride
>> 2;
1001 prim
.address
= cl_address(v3d_resource(info
->indirect
->buffer
)->bo
,
1002 info
->indirect
->offset
);
1004 } else if (info
->instance_count
> 1) {
1005 cl_emit(&job
->bcl
, INDEXED_INSTANCED_PRIM_LIST
, prim
) {
1006 prim
.index_type
= ffs(info
->index_size
) - 1;
1007 #if V3D_VERSION >= 40
1008 prim
.index_offset
= offset
;
1009 #else /* V3D_VERSION < 40 */
1010 prim
.maximum_index
= (1u << 31) - 1; /* XXX */
1011 prim
.address_of_indices_list
=
1012 cl_address(rsc
->bo
, offset
);
1013 #endif /* V3D_VERSION < 40 */
1014 prim
.mode
= info
->mode
| prim_tf_enable
;
1015 prim
.enable_primitive_restarts
= info
->primitive_restart
;
1017 prim
.number_of_instances
= info
->instance_count
;
1018 prim
.instance_length
= info
->count
;
1021 cl_emit(&job
->bcl
, INDEXED_PRIM_LIST
, prim
) {
1022 prim
.index_type
= ffs(info
->index_size
) - 1;
1023 prim
.length
= info
->count
;
1024 #if V3D_VERSION >= 40
1025 prim
.index_offset
= offset
;
1026 #else /* V3D_VERSION < 40 */
1027 prim
.maximum_index
= (1u << 31) - 1; /* XXX */
1028 prim
.address_of_indices_list
=
1029 cl_address(rsc
->bo
, offset
);
1030 #endif /* V3D_VERSION < 40 */
1031 prim
.mode
= info
->mode
| prim_tf_enable
;
1032 prim
.enable_primitive_restarts
= info
->primitive_restart
;
1036 if (info
->has_user_indices
)
1037 pipe_resource_reference(&prsc
, NULL
);
1039 if (info
->indirect
) {
1040 cl_emit(&job
->bcl
, INDIRECT_VERTEX_ARRAY_INSTANCED_PRIMS
, prim
) {
1041 prim
.mode
= info
->mode
| prim_tf_enable
;
1042 prim
.number_of_draw_indirect_array_records
= info
->indirect
->draw_count
;
1044 prim
.stride_in_multiples_of_4_bytes
= info
->indirect
->stride
>> 2;
1045 prim
.address
= cl_address(v3d_resource(info
->indirect
->buffer
)->bo
,
1046 info
->indirect
->offset
);
1048 } else if (info
->instance_count
> 1) {
1049 struct pipe_stream_output_target
*so
=
1050 info
->count_from_stream_output
;
1051 uint32_t vert_count
= so
?
1052 v3d_stream_output_target_get_vertex_count(so
) :
1054 cl_emit(&job
->bcl
, VERTEX_ARRAY_INSTANCED_PRIMS
, prim
) {
1055 prim
.mode
= info
->mode
| prim_tf_enable
;
1056 prim
.index_of_first_vertex
= info
->start
;
1057 prim
.number_of_instances
= info
->instance_count
;
1058 prim
.instance_length
= vert_count
;
1061 struct pipe_stream_output_target
*so
=
1062 info
->count_from_stream_output
;
1063 uint32_t vert_count
= so
?
1064 v3d_stream_output_target_get_vertex_count(so
) :
1066 cl_emit(&job
->bcl
, VERTEX_ARRAY_PRIMS
, prim
) {
1067 prim
.mode
= info
->mode
| prim_tf_enable
;
1068 prim
.length
= vert_count
;
1069 prim
.index_of_first_vertex
= info
->start
;
1074 /* A flush is required in between a TF draw and any following TF specs
1075 * packet, or the GPU may hang. Just flush each time for now.
1077 if (v3d
->streamout
.num_targets
)
1078 cl_emit(&job
->bcl
, TRANSFORM_FEEDBACK_FLUSH_AND_COUNT
, flush
);
1080 job
->draw_calls_queued
++;
1081 if (v3d
->streamout
.num_targets
)
1082 job
->tf_draw_calls_queued
++;
1084 /* Increment the TF offsets by how many verts we wrote. XXX: This
1085 * needs some clamping to the buffer size.
1087 for (int i
= 0; i
< v3d
->streamout
.num_targets
; i
++)
1088 v3d
->streamout
.offsets
[i
] += info
->count
;
1090 if (v3d
->zsa
&& job
->zsbuf
&& v3d
->zsa
->base
.depth
.enabled
) {
1091 struct v3d_resource
*rsc
= v3d_resource(job
->zsbuf
->texture
);
1092 v3d_job_add_bo(job
, rsc
->bo
);
1094 job
->load
|= PIPE_CLEAR_DEPTH
& ~job
->clear
;
1095 if (v3d
->zsa
->base
.depth
.writemask
)
1096 job
->store
|= PIPE_CLEAR_DEPTH
;
1097 rsc
->initialized_buffers
= PIPE_CLEAR_DEPTH
;
1100 if (v3d
->zsa
&& job
->zsbuf
&& v3d
->zsa
->base
.stencil
[0].enabled
) {
1101 struct v3d_resource
*rsc
= v3d_resource(job
->zsbuf
->texture
);
1102 if (rsc
->separate_stencil
)
1103 rsc
= rsc
->separate_stencil
;
1105 v3d_job_add_bo(job
, rsc
->bo
);
1107 job
->load
|= PIPE_CLEAR_STENCIL
& ~job
->clear
;
1108 if (v3d
->zsa
->base
.stencil
[0].writemask
||
1109 v3d
->zsa
->base
.stencil
[1].writemask
) {
1110 job
->store
|= PIPE_CLEAR_STENCIL
;
1112 rsc
->initialized_buffers
|= PIPE_CLEAR_STENCIL
;
1115 for (int i
= 0; i
< V3D_MAX_DRAW_BUFFERS
; i
++) {
1116 uint32_t bit
= PIPE_CLEAR_COLOR0
<< i
;
1117 int blend_rt
= v3d
->blend
->base
.independent_blend_enable
? i
: 0;
1119 if (job
->store
& bit
|| !job
->cbufs
[i
])
1121 struct v3d_resource
*rsc
= v3d_resource(job
->cbufs
[i
]->texture
);
1123 job
->load
|= bit
& ~job
->clear
;
1124 if (v3d
->blend
->base
.rt
[blend_rt
].colormask
)
1126 v3d_job_add_bo(job
, rsc
->bo
);
1129 if (job
->referenced_size
> 768 * 1024 * 1024) {
1130 perf_debug("Flushing job with %dkb to try to free up memory\n",
1131 job
->referenced_size
/ 1024);
1135 if (V3D_DEBUG
& V3D_DEBUG_ALWAYS_FLUSH
)
1139 #if V3D_VERSION >= 41
1140 #define V3D_CSD_CFG012_WG_COUNT_SHIFT 16
1141 #define V3D_CSD_CFG012_WG_OFFSET_SHIFT 0
1142 /* Allow this dispatch to start while the last one is still running. */
1143 #define V3D_CSD_CFG3_OVERLAP_WITH_PREV (1 << 26)
1144 /* Maximum supergroup ID. 6 bits. */
1145 #define V3D_CSD_CFG3_MAX_SG_ID_SHIFT 20
1146 /* Batches per supergroup minus 1. 8 bits. */
1147 #define V3D_CSD_CFG3_BATCHES_PER_SG_M1_SHIFT 12
1148 /* Workgroups per supergroup, 0 means 16 */
1149 #define V3D_CSD_CFG3_WGS_PER_SG_SHIFT 8
1150 #define V3D_CSD_CFG3_WG_SIZE_SHIFT 0
1152 #define V3D_CSD_CFG5_PROPAGATE_NANS (1 << 2)
1153 #define V3D_CSD_CFG5_SINGLE_SEG (1 << 1)
1154 #define V3D_CSD_CFG5_THREADING (1 << 0)
1157 v3d_launch_grid(struct pipe_context
*pctx
, const struct pipe_grid_info
*info
)
1159 struct v3d_context
*v3d
= v3d_context(pctx
);
1160 struct v3d_screen
*screen
= v3d
->screen
;
1162 v3d_predraw_check_stage_inputs(pctx
, PIPE_SHADER_COMPUTE
);
1164 v3d_update_compiled_cs(v3d
);
1166 if (!v3d
->prog
.compute
->resource
) {
1167 static bool warned
= false;
1170 "Compute shader failed to compile. "
1171 "Expect corruption.\n");
1177 /* Some of the units of scale:
1179 * - Batches of 16 work items (shader invocations) that will be queued
1180 * to the run on a QPU at once.
1182 * - Workgroups composed of work items based on the shader's layout
1185 * - Supergroups of 1-16 workgroups. There can only be 16 supergroups
1186 * running at a time on the core, so we want to keep them large to
1187 * keep the QPUs busy, but a whole supergroup will sync at a barrier
1188 * so we want to keep them small if one is present.
1190 struct drm_v3d_submit_csd submit
= { 0 };
1191 struct v3d_job
*job
= v3d_job_create(v3d
);
1193 /* Set up the actual number of workgroups, synchronously mapping the
1194 * indirect buffer if necessary to get the dimensions.
1196 if (info
->indirect
) {
1197 struct pipe_transfer
*transfer
;
1198 uint32_t *map
= pipe_buffer_map_range(pctx
, info
->indirect
,
1199 info
->indirect_offset
,
1200 3 * sizeof(uint32_t),
1203 memcpy(v3d
->compute_num_workgroups
, map
, 3 * sizeof(uint32_t));
1204 pipe_buffer_unmap(pctx
, transfer
);
1206 if (v3d
->compute_num_workgroups
[0] == 0 ||
1207 v3d
->compute_num_workgroups
[1] == 0 ||
1208 v3d
->compute_num_workgroups
[2] == 0) {
1209 /* Nothing to dispatch, so skip the draw (CSD can't
1210 * handle 0 workgroups).
1215 v3d
->compute_num_workgroups
[0] = info
->grid
[0];
1216 v3d
->compute_num_workgroups
[1] = info
->grid
[1];
1217 v3d
->compute_num_workgroups
[2] = info
->grid
[2];
1220 for (int i
= 0; i
< 3; i
++) {
1221 submit
.cfg
[i
] |= (v3d
->compute_num_workgroups
[i
] <<
1222 V3D_CSD_CFG012_WG_COUNT_SHIFT
);
1225 perf_debug("CSD only using single WG per SG currently, "
1226 "should increase that when possible.");
1228 int wg_size
= info
->block
[0] * info
->block
[1] * info
->block
[2];
1229 submit
.cfg
[3] |= wgs_per_sg
<< V3D_CSD_CFG3_WGS_PER_SG_SHIFT
;
1230 submit
.cfg
[3] |= ((DIV_ROUND_UP(wgs_per_sg
* wg_size
, 16) - 1) <<
1231 V3D_CSD_CFG3_BATCHES_PER_SG_M1_SHIFT
);
1232 submit
.cfg
[3] |= (wg_size
& 0xff) << V3D_CSD_CFG3_WG_SIZE_SHIFT
;
1234 int batches_per_wg
= DIV_ROUND_UP(wg_size
, 16);
1235 /* Number of batches the dispatch will invoke (minus 1). */
1236 submit
.cfg
[4] = batches_per_wg
* (v3d
->compute_num_workgroups
[0] *
1237 v3d
->compute_num_workgroups
[1] *
1238 v3d
->compute_num_workgroups
[2]) - 1;
1240 /* Make sure we didn't accidentally underflow. */
1241 assert(submit
.cfg
[4] != ~0);
1243 v3d_job_add_bo(job
, v3d_resource(v3d
->prog
.compute
->resource
)->bo
);
1244 submit
.cfg
[5] = (v3d_resource(v3d
->prog
.compute
->resource
)->bo
->offset
+
1245 v3d
->prog
.compute
->offset
);
1246 submit
.cfg
[5] |= V3D_CSD_CFG5_PROPAGATE_NANS
;
1247 if (v3d
->prog
.compute
->prog_data
.base
->single_seg
)
1248 submit
.cfg
[5] |= V3D_CSD_CFG5_SINGLE_SEG
;
1249 if (v3d
->prog
.compute
->prog_data
.base
->threads
== 4)
1250 submit
.cfg
[5] |= V3D_CSD_CFG5_THREADING
;
1252 if (v3d
->prog
.compute
->prog_data
.compute
->shared_size
) {
1253 v3d
->compute_shared_memory
=
1254 v3d_bo_alloc(v3d
->screen
,
1255 v3d
->prog
.compute
->prog_data
.compute
->shared_size
*
1260 struct v3d_cl_reloc uniforms
= v3d_write_uniforms(v3d
, job
,
1262 PIPE_SHADER_COMPUTE
);
1263 v3d_job_add_bo(job
, uniforms
.bo
);
1264 submit
.cfg
[6] = uniforms
.bo
->offset
+ uniforms
.offset
;
1266 /* Pull some job state that was stored in a SUBMIT_CL struct out to
1267 * our SUBMIT_CSD struct
1269 submit
.bo_handles
= job
->submit
.bo_handles
;
1270 submit
.bo_handle_count
= job
->submit
.bo_handle_count
;
1272 /* Serialize this in the rest of our command stream. */
1273 submit
.in_sync
= v3d
->out_sync
;
1274 submit
.out_sync
= v3d
->out_sync
;
1276 if (!(V3D_DEBUG
& V3D_DEBUG_NORAST
)) {
1277 int ret
= v3d_ioctl(screen
->fd
, DRM_IOCTL_V3D_SUBMIT_CSD
,
1279 static bool warned
= false;
1280 if (ret
&& !warned
) {
1281 fprintf(stderr
, "CSD submit call returned %s. "
1282 "Expect corruption.\n", strerror(errno
));
1287 v3d_job_free(v3d
, job
);
1289 /* Mark SSBOs as being written.. we don't actually know which ones are
1290 * read vs written, so just assume the worst
1292 foreach_bit(i
, v3d
->ssbo
[PIPE_SHADER_COMPUTE
].enabled_mask
) {
1293 struct v3d_resource
*rsc
= v3d_resource(
1294 v3d
->ssbo
[PIPE_SHADER_COMPUTE
].sb
[i
].buffer
);
1295 rsc
->writes
++; /* XXX */
1298 foreach_bit(i
, v3d
->shaderimg
[PIPE_SHADER_COMPUTE
].enabled_mask
) {
1299 struct v3d_resource
*rsc
= v3d_resource(
1300 v3d
->shaderimg
[PIPE_SHADER_COMPUTE
].si
[i
].base
.resource
);
1304 v3d_bo_unreference(&uniforms
.bo
);
1305 v3d_bo_unreference(&v3d
->compute_shared_memory
);
1310 * Implements gallium's clear() hook (glClear()) by drawing a pair of triangles.
1313 v3d_draw_clear(struct v3d_context
*v3d
,
1315 const union pipe_color_union
*color
,
1316 double depth
, unsigned stencil
)
1318 static const union pipe_color_union dummy_color
= {};
1320 /* The blitter util dereferences the color regardless, even though the
1321 * gallium clear API may not pass one in when only Z/S are cleared.
1324 color
= &dummy_color
;
1326 v3d_blitter_save(v3d
);
1327 util_blitter_clear(v3d
->blitter
,
1328 v3d
->framebuffer
.width
,
1329 v3d
->framebuffer
.height
,
1330 util_framebuffer_get_num_layers(&v3d
->framebuffer
),
1331 buffers
, color
, depth
, stencil
,
1332 util_framebuffer_get_num_samples(&v3d
->framebuffer
) > 1);
1336 * Attempts to perform the GL clear by using the TLB's fast clear at the start
1340 v3d_tlb_clear(struct v3d_job
*job
, unsigned buffers
,
1341 const union pipe_color_union
*color
,
1342 double depth
, unsigned stencil
)
1344 struct v3d_context
*v3d
= job
->v3d
;
1346 if (job
->draw_calls_queued
) {
1347 /* If anything in the CL has drawn using the buffer, then the
1348 * TLB clear we're trying to add now would happen before that
1351 buffers
&= ~(job
->load
| job
->store
);
1354 /* GFXH-1461: If we were to emit a load of just depth or just stencil,
1355 * then the clear for the other may get lost. We need to decide now
1356 * if it would be possible to need to emit a load of just one after
1357 * we've set up our TLB clears.
1359 if (buffers
& PIPE_CLEAR_DEPTHSTENCIL
&&
1360 (buffers
& PIPE_CLEAR_DEPTHSTENCIL
) != PIPE_CLEAR_DEPTHSTENCIL
&&
1362 util_format_is_depth_and_stencil(job
->zsbuf
->texture
->format
)) {
1363 buffers
&= ~PIPE_CLEAR_DEPTHSTENCIL
;
1366 for (int i
= 0; i
< V3D_MAX_DRAW_BUFFERS
; i
++) {
1367 uint32_t bit
= PIPE_CLEAR_COLOR0
<< i
;
1368 if (!(buffers
& bit
))
1371 struct pipe_surface
*psurf
= v3d
->framebuffer
.cbufs
[i
];
1372 struct v3d_surface
*surf
= v3d_surface(psurf
);
1373 struct v3d_resource
*rsc
= v3d_resource(psurf
->texture
);
1375 union util_color uc
;
1376 uint32_t internal_size
= 4 << surf
->internal_bpp
;
1378 static union pipe_color_union swapped_color
;
1379 if (v3d
->swap_color_rb
& (1 << i
)) {
1380 swapped_color
.f
[0] = color
->f
[2];
1381 swapped_color
.f
[1] = color
->f
[1];
1382 swapped_color
.f
[2] = color
->f
[0];
1383 swapped_color
.f
[3] = color
->f
[3];
1384 color
= &swapped_color
;
1387 switch (surf
->internal_type
) {
1388 case V3D_INTERNAL_TYPE_8
:
1389 util_pack_color(color
->f
, PIPE_FORMAT_R8G8B8A8_UNORM
,
1391 memcpy(job
->clear_color
[i
], uc
.ui
, internal_size
);
1393 case V3D_INTERNAL_TYPE_8I
:
1394 case V3D_INTERNAL_TYPE_8UI
:
1395 job
->clear_color
[i
][0] = ((color
->ui
[0] & 0xff) |
1396 (color
->ui
[1] & 0xff) << 8 |
1397 (color
->ui
[2] & 0xff) << 16 |
1398 (color
->ui
[3] & 0xff) << 24);
1400 case V3D_INTERNAL_TYPE_16F
:
1401 util_pack_color(color
->f
, PIPE_FORMAT_R16G16B16A16_FLOAT
,
1403 memcpy(job
->clear_color
[i
], uc
.ui
, internal_size
);
1405 case V3D_INTERNAL_TYPE_16I
:
1406 case V3D_INTERNAL_TYPE_16UI
:
1407 job
->clear_color
[i
][0] = ((color
->ui
[0] & 0xffff) |
1408 color
->ui
[1] << 16);
1409 job
->clear_color
[i
][1] = ((color
->ui
[2] & 0xffff) |
1410 color
->ui
[3] << 16);
1412 case V3D_INTERNAL_TYPE_32F
:
1413 case V3D_INTERNAL_TYPE_32I
:
1414 case V3D_INTERNAL_TYPE_32UI
:
1415 memcpy(job
->clear_color
[i
], color
->ui
, internal_size
);
1419 rsc
->initialized_buffers
|= bit
;
1422 unsigned zsclear
= buffers
& PIPE_CLEAR_DEPTHSTENCIL
;
1424 struct v3d_resource
*rsc
=
1425 v3d_resource(v3d
->framebuffer
.zsbuf
->texture
);
1427 if (zsclear
& PIPE_CLEAR_DEPTH
)
1428 job
->clear_z
= depth
;
1429 if (zsclear
& PIPE_CLEAR_STENCIL
)
1430 job
->clear_s
= stencil
;
1432 rsc
->initialized_buffers
|= zsclear
;
1435 job
->draw_min_x
= 0;
1436 job
->draw_min_y
= 0;
1437 job
->draw_max_x
= v3d
->framebuffer
.width
;
1438 job
->draw_max_y
= v3d
->framebuffer
.height
;
1439 job
->clear
|= buffers
;
1440 job
->store
|= buffers
;
1442 v3d_start_draw(v3d
);
1448 v3d_clear(struct pipe_context
*pctx
, unsigned buffers
,
1449 const union pipe_color_union
*color
, double depth
, unsigned stencil
)
1451 struct v3d_context
*v3d
= v3d_context(pctx
);
1452 struct v3d_job
*job
= v3d_get_job_for_fbo(v3d
);
1454 buffers
&= ~v3d_tlb_clear(job
, buffers
, color
, depth
, stencil
);
1457 v3d_draw_clear(v3d
, buffers
, color
, depth
, stencil
);
1461 v3d_clear_render_target(struct pipe_context
*pctx
, struct pipe_surface
*ps
,
1462 const union pipe_color_union
*color
,
1463 unsigned x
, unsigned y
, unsigned w
, unsigned h
,
1464 bool render_condition_enabled
)
1466 fprintf(stderr
, "unimpl: clear RT\n");
1470 v3d_clear_depth_stencil(struct pipe_context
*pctx
, struct pipe_surface
*ps
,
1471 unsigned buffers
, double depth
, unsigned stencil
,
1472 unsigned x
, unsigned y
, unsigned w
, unsigned h
,
1473 bool render_condition_enabled
)
1475 fprintf(stderr
, "unimpl: clear DS\n");
1479 v3dX(draw_init
)(struct pipe_context
*pctx
)
1481 pctx
->draw_vbo
= v3d_draw_vbo
;
1482 pctx
->clear
= v3d_clear
;
1483 pctx
->clear_render_target
= v3d_clear_render_target
;
1484 pctx
->clear_depth_stencil
= v3d_clear_depth_stencil
;
1485 #if V3D_VERSION >= 41
1486 if (v3d_context(pctx
)->screen
->has_csd
)
1487 pctx
->launch_grid
= v3d_launch_grid
;