2 * Copyright (C) 2018 Alyssa Rosenzweig
3 * Copyright (C) 2020 Collabora Ltd.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
25 #include "util/macros.h"
26 #include "util/u_prim.h"
27 #include "util/u_vbuf.h"
29 #include "panfrost-quirks.h"
31 #include "pan_allocate.h"
33 #include "pan_cmdstream.h"
34 #include "pan_context.h"
37 /* If a BO is accessed for a particular shader stage, will it be in the primary
38 * batch (vertex/tiler) or the secondary batch (fragment)? Anything but
39 * fragment will be primary, e.g. compute jobs will be considered
40 * "vertex/tiler" by analogy */
42 static inline uint32_t
43 panfrost_bo_access_for_stage(enum pipe_shader_type stage
)
45 assert(stage
== PIPE_SHADER_FRAGMENT
||
46 stage
== PIPE_SHADER_VERTEX
||
47 stage
== PIPE_SHADER_COMPUTE
);
49 return stage
== PIPE_SHADER_FRAGMENT
?
50 PAN_BO_ACCESS_FRAGMENT
:
51 PAN_BO_ACCESS_VERTEX_TILER
;
55 panfrost_vt_emit_shared_memory(struct panfrost_context
*ctx
,
56 struct mali_vertex_tiler_postfix
*postfix
)
58 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
59 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
61 unsigned shift
= panfrost_get_stack_shift(batch
->stack_size
);
62 struct mali_shared_memory shared
= {
64 .scratchpad
= panfrost_batch_get_scratchpad(batch
, shift
, dev
->thread_tls_alloc
, dev
->core_count
)->gpu
,
65 .shared_workgroup_count
= ~0,
67 postfix
->shared_memory
= panfrost_upload_transient(batch
, &shared
, sizeof(shared
));
71 panfrost_vt_attach_framebuffer(struct panfrost_context
*ctx
,
72 struct mali_vertex_tiler_postfix
*postfix
)
74 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
75 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
77 /* If we haven't, reserve space for the framebuffer */
79 if (!batch
->framebuffer
.gpu
) {
80 unsigned size
= (dev
->quirks
& MIDGARD_SFBD
) ?
81 sizeof(struct mali_single_framebuffer
) :
82 sizeof(struct mali_framebuffer
);
84 batch
->framebuffer
= panfrost_allocate_transient(batch
, size
);
87 if (!(dev
->quirks
& MIDGARD_SFBD
))
88 batch
->framebuffer
.gpu
|= MALI_MFBD
;
91 postfix
->shared_memory
= batch
->framebuffer
.gpu
;
95 panfrost_vt_update_rasterizer(struct panfrost_context
*ctx
,
96 struct mali_vertex_tiler_prefix
*prefix
,
97 struct mali_vertex_tiler_postfix
*postfix
)
99 struct panfrost_rasterizer
*rasterizer
= ctx
->rasterizer
;
101 postfix
->gl_enables
|= 0x7;
102 SET_BIT(postfix
->gl_enables
, MALI_FRONT_CCW_TOP
,
103 rasterizer
&& rasterizer
->base
.front_ccw
);
104 SET_BIT(postfix
->gl_enables
, MALI_CULL_FACE_FRONT
,
105 rasterizer
&& (rasterizer
->base
.cull_face
& PIPE_FACE_FRONT
));
106 SET_BIT(postfix
->gl_enables
, MALI_CULL_FACE_BACK
,
107 rasterizer
&& (rasterizer
->base
.cull_face
& PIPE_FACE_BACK
));
108 SET_BIT(prefix
->unknown_draw
, MALI_DRAW_FLATSHADE_FIRST
,
109 rasterizer
&& rasterizer
->base
.flatshade_first
);
113 panfrost_vt_update_primitive_size(struct panfrost_context
*ctx
,
114 struct mali_vertex_tiler_prefix
*prefix
,
115 union midgard_primitive_size
*primitive_size
)
117 struct panfrost_rasterizer
*rasterizer
= ctx
->rasterizer
;
119 if (!panfrost_writes_point_size(ctx
)) {
120 bool points
= prefix
->draw_mode
== MALI_POINTS
;
125 rasterizer
->base
.point_size
:
126 rasterizer
->base
.line_width
;
128 primitive_size
->constant
= val
;
133 panfrost_vt_update_occlusion_query(struct panfrost_context
*ctx
,
134 struct mali_vertex_tiler_postfix
*postfix
)
136 SET_BIT(postfix
->gl_enables
, MALI_OCCLUSION_QUERY
, ctx
->occlusion_query
);
137 if (ctx
->occlusion_query
)
138 postfix
->occlusion_counter
= ctx
->occlusion_query
->bo
->gpu
;
140 postfix
->occlusion_counter
= 0;
144 panfrost_vt_init(struct panfrost_context
*ctx
,
145 enum pipe_shader_type stage
,
146 struct mali_vertex_tiler_prefix
*prefix
,
147 struct mali_vertex_tiler_postfix
*postfix
)
149 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
151 if (!ctx
->shader
[stage
])
154 memset(prefix
, 0, sizeof(*prefix
));
155 memset(postfix
, 0, sizeof(*postfix
));
157 if (device
->quirks
& IS_BIFROST
) {
158 postfix
->gl_enables
= 0x2;
159 panfrost_vt_emit_shared_memory(ctx
, postfix
);
161 postfix
->gl_enables
= 0x6;
162 panfrost_vt_attach_framebuffer(ctx
, postfix
);
165 if (stage
== PIPE_SHADER_FRAGMENT
) {
166 panfrost_vt_update_occlusion_query(ctx
, postfix
);
167 panfrost_vt_update_rasterizer(ctx
, prefix
, postfix
);
172 panfrost_translate_index_size(unsigned size
)
176 return MALI_DRAW_INDEXED_UINT8
;
179 return MALI_DRAW_INDEXED_UINT16
;
182 return MALI_DRAW_INDEXED_UINT32
;
185 unreachable("Invalid index size");
189 /* Gets a GPU address for the associated index buffer. Only gauranteed to be
190 * good for the duration of the draw (transient), could last longer. Also get
191 * the bounds on the index buffer for the range accessed by the draw. We do
192 * these operations together because there are natural optimizations which
193 * require them to be together. */
196 panfrost_get_index_buffer_bounded(struct panfrost_context
*ctx
,
197 const struct pipe_draw_info
*info
,
198 unsigned *min_index
, unsigned *max_index
)
200 struct panfrost_resource
*rsrc
= pan_resource(info
->index
.resource
);
201 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
202 off_t offset
= info
->start
* info
->index_size
;
203 bool needs_indices
= true;
206 if (info
->max_index
!= ~0u) {
207 *min_index
= info
->min_index
;
208 *max_index
= info
->max_index
;
209 needs_indices
= false;
212 if (!info
->has_user_indices
) {
213 /* Only resources can be directly mapped */
214 panfrost_batch_add_bo(batch
, rsrc
->bo
,
215 PAN_BO_ACCESS_SHARED
|
217 PAN_BO_ACCESS_VERTEX_TILER
);
218 out
= rsrc
->bo
->gpu
+ offset
;
220 /* Check the cache */
221 needs_indices
= !panfrost_minmax_cache_get(rsrc
->index_cache
,
227 /* Otherwise, we need to upload to transient memory */
228 const uint8_t *ibuf8
= (const uint8_t *) info
->index
.user
;
229 out
= panfrost_upload_transient(batch
, ibuf8
+ offset
,
236 u_vbuf_get_minmax_index(&ctx
->base
, info
, min_index
, max_index
);
238 if (!info
->has_user_indices
)
239 panfrost_minmax_cache_add(rsrc
->index_cache
,
240 info
->start
, info
->count
,
241 *min_index
, *max_index
);
248 panfrost_vt_set_draw_info(struct panfrost_context
*ctx
,
249 const struct pipe_draw_info
*info
,
250 enum mali_draw_mode draw_mode
,
251 struct mali_vertex_tiler_postfix
*vertex_postfix
,
252 struct mali_vertex_tiler_prefix
*tiler_prefix
,
253 struct mali_vertex_tiler_postfix
*tiler_postfix
,
254 unsigned *vertex_count
,
255 unsigned *padded_count
)
257 tiler_prefix
->draw_mode
= draw_mode
;
259 unsigned draw_flags
= 0;
261 if (panfrost_writes_point_size(ctx
))
262 draw_flags
|= MALI_DRAW_VARYING_SIZE
;
264 if (info
->primitive_restart
)
265 draw_flags
|= MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX
;
267 /* These doesn't make much sense */
269 draw_flags
|= 0x3000;
271 if (info
->index_size
) {
272 unsigned min_index
= 0, max_index
= 0;
274 tiler_prefix
->indices
= panfrost_get_index_buffer_bounded(ctx
,
279 /* Use the corresponding values */
280 *vertex_count
= max_index
- min_index
+ 1;
281 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= min_index
+ info
->index_bias
;
282 tiler_prefix
->offset_bias_correction
= -min_index
;
283 tiler_prefix
->index_count
= MALI_POSITIVE(info
->count
);
284 draw_flags
|= panfrost_translate_index_size(info
->index_size
);
286 tiler_prefix
->indices
= 0;
287 *vertex_count
= ctx
->vertex_count
;
288 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= info
->start
;
289 tiler_prefix
->offset_bias_correction
= 0;
290 tiler_prefix
->index_count
= MALI_POSITIVE(ctx
->vertex_count
);
293 tiler_prefix
->unknown_draw
= draw_flags
;
295 /* Encode the padded vertex count */
297 if (info
->instance_count
> 1) {
298 *padded_count
= panfrost_padded_vertex_count(*vertex_count
);
300 unsigned shift
= __builtin_ctz(ctx
->padded_count
);
301 unsigned k
= ctx
->padded_count
>> (shift
+ 1);
303 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= shift
;
304 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= k
;
306 *padded_count
= *vertex_count
;
308 /* Reset instancing state */
309 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= 0;
310 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= 0;
315 panfrost_shader_meta_init(struct panfrost_context
*ctx
,
316 enum pipe_shader_type st
,
317 struct mali_shader_meta
*meta
)
319 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
320 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
322 memset(meta
, 0, sizeof(*meta
));
323 meta
->shader
= (ss
->bo
? ss
->bo
->gpu
: 0) | ss
->first_tag
;
324 meta
->attribute_count
= ss
->attribute_count
;
325 meta
->varying_count
= ss
->varying_count
;
326 meta
->texture_count
= ctx
->sampler_view_count
[st
];
327 meta
->sampler_count
= ctx
->sampler_count
[st
];
329 if (dev
->quirks
& IS_BIFROST
) {
330 if (st
== PIPE_SHADER_VERTEX
)
331 meta
->bifrost1
.unk1
= 0x800000;
333 /* First clause ATEST |= 0x4000000.
334 * Less than 32 regs |= 0x200 */
335 meta
->bifrost1
.unk1
= 0x950020;
338 meta
->bifrost1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
339 if (st
== PIPE_SHADER_VERTEX
)
340 meta
->bifrost2
.preload_regs
= 0xC0;
342 meta
->bifrost2
.preload_regs
= 0x1;
343 SET_BIT(meta
->bifrost2
.preload_regs
, 0x10, ss
->reads_frag_coord
);
346 meta
->bifrost2
.uniform_count
= MIN2(ss
->uniform_count
,
349 meta
->midgard1
.uniform_count
= MIN2(ss
->uniform_count
,
351 meta
->midgard1
.work_count
= ss
->work_reg_count
;
353 /* TODO: This is not conformant on ES3 */
354 meta
->midgard1
.flags_hi
= MALI_SUPPRESS_INF_NAN
;
356 meta
->midgard1
.flags_lo
= 0x20;
357 meta
->midgard1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
359 SET_BIT(meta
->midgard1
.flags_hi
, MALI_WRITES_GLOBAL
, ss
->writes_global
);
364 panfrost_translate_compare_func(enum pipe_compare_func in
)
367 case PIPE_FUNC_NEVER
:
368 return MALI_FUNC_NEVER
;
371 return MALI_FUNC_LESS
;
373 case PIPE_FUNC_EQUAL
:
374 return MALI_FUNC_EQUAL
;
376 case PIPE_FUNC_LEQUAL
:
377 return MALI_FUNC_LEQUAL
;
379 case PIPE_FUNC_GREATER
:
380 return MALI_FUNC_GREATER
;
382 case PIPE_FUNC_NOTEQUAL
:
383 return MALI_FUNC_NOTEQUAL
;
385 case PIPE_FUNC_GEQUAL
:
386 return MALI_FUNC_GEQUAL
;
388 case PIPE_FUNC_ALWAYS
:
389 return MALI_FUNC_ALWAYS
;
392 unreachable("Invalid func");
397 panfrost_translate_stencil_op(enum pipe_stencil_op in
)
400 case PIPE_STENCIL_OP_KEEP
:
401 return MALI_STENCIL_KEEP
;
403 case PIPE_STENCIL_OP_ZERO
:
404 return MALI_STENCIL_ZERO
;
406 case PIPE_STENCIL_OP_REPLACE
:
407 return MALI_STENCIL_REPLACE
;
409 case PIPE_STENCIL_OP_INCR
:
410 return MALI_STENCIL_INCR
;
412 case PIPE_STENCIL_OP_DECR
:
413 return MALI_STENCIL_DECR
;
415 case PIPE_STENCIL_OP_INCR_WRAP
:
416 return MALI_STENCIL_INCR_WRAP
;
418 case PIPE_STENCIL_OP_DECR_WRAP
:
419 return MALI_STENCIL_DECR_WRAP
;
421 case PIPE_STENCIL_OP_INVERT
:
422 return MALI_STENCIL_INVERT
;
425 unreachable("Invalid stencil op");
430 translate_tex_wrap(enum pipe_tex_wrap w
)
433 case PIPE_TEX_WRAP_REPEAT
:
434 return MALI_WRAP_REPEAT
;
436 case PIPE_TEX_WRAP_CLAMP
:
437 return MALI_WRAP_CLAMP
;
439 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
440 return MALI_WRAP_CLAMP_TO_EDGE
;
442 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
443 return MALI_WRAP_CLAMP_TO_BORDER
;
445 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
446 return MALI_WRAP_MIRRORED_REPEAT
;
448 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
449 return MALI_WRAP_MIRRORED_CLAMP
;
451 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
452 return MALI_WRAP_MIRRORED_CLAMP_TO_EDGE
;
454 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
455 return MALI_WRAP_MIRRORED_CLAMP_TO_BORDER
;
458 unreachable("Invalid wrap");
462 void panfrost_sampler_desc_init(const struct pipe_sampler_state
*cso
,
463 struct mali_sampler_descriptor
*hw
)
465 unsigned func
= panfrost_translate_compare_func(cso
->compare_func
);
466 bool min_nearest
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
;
467 bool mag_nearest
= cso
->mag_img_filter
== PIPE_TEX_FILTER_NEAREST
;
468 bool mip_linear
= cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
;
469 unsigned min_filter
= min_nearest
? MALI_SAMP_MIN_NEAREST
: 0;
470 unsigned mag_filter
= mag_nearest
? MALI_SAMP_MAG_NEAREST
: 0;
471 unsigned mip_filter
= mip_linear
?
472 (MALI_SAMP_MIP_LINEAR_1
| MALI_SAMP_MIP_LINEAR_2
) : 0;
473 unsigned normalized
= cso
->normalized_coords
? MALI_SAMP_NORM_COORDS
: 0;
475 *hw
= (struct mali_sampler_descriptor
) {
476 .filter_mode
= min_filter
| mag_filter
| mip_filter
|
478 .wrap_s
= translate_tex_wrap(cso
->wrap_s
),
479 .wrap_t
= translate_tex_wrap(cso
->wrap_t
),
480 .wrap_r
= translate_tex_wrap(cso
->wrap_r
),
481 .compare_func
= panfrost_flip_compare_func(func
),
483 cso
->border_color
.f
[0],
484 cso
->border_color
.f
[1],
485 cso
->border_color
.f
[2],
486 cso
->border_color
.f
[3]
488 .min_lod
= FIXED_16(cso
->min_lod
, false), /* clamp at 0 */
489 .max_lod
= FIXED_16(cso
->max_lod
, false),
490 .lod_bias
= FIXED_16(cso
->lod_bias
, true), /* can be negative */
491 .seamless_cube_map
= cso
->seamless_cube_map
,
494 /* If necessary, we disable mipmapping in the sampler descriptor by
495 * clamping the LOD as tight as possible (from 0 to epsilon,
496 * essentially -- remember these are fixed point numbers, so
499 if (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
500 hw
->max_lod
= hw
->min_lod
+ 1;
503 void panfrost_sampler_desc_init_bifrost(const struct pipe_sampler_state
*cso
,
504 struct bifrost_sampler_descriptor
*hw
)
506 *hw
= (struct bifrost_sampler_descriptor
) {
508 .wrap_s
= translate_tex_wrap(cso
->wrap_s
),
509 .wrap_t
= translate_tex_wrap(cso
->wrap_t
),
510 .wrap_r
= translate_tex_wrap(cso
->wrap_r
),
512 .min_filter
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
,
513 .norm_coords
= cso
->normalized_coords
,
514 .mip_filter
= cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
,
515 .mag_filter
= cso
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
,
516 .min_lod
= FIXED_16(cso
->min_lod
, false), /* clamp at 0 */
517 .max_lod
= FIXED_16(cso
->max_lod
, false),
520 /* If necessary, we disable mipmapping in the sampler descriptor by
521 * clamping the LOD as tight as possible (from 0 to epsilon,
522 * essentially -- remember these are fixed point numbers, so
525 if (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
526 hw
->max_lod
= hw
->min_lod
+ 1;
530 panfrost_make_stencil_state(const struct pipe_stencil_state
*in
,
531 struct mali_stencil_test
*out
)
533 out
->ref
= 0; /* Gallium gets it from elsewhere */
535 out
->mask
= in
->valuemask
;
536 out
->func
= panfrost_translate_compare_func(in
->func
);
537 out
->sfail
= panfrost_translate_stencil_op(in
->fail_op
);
538 out
->dpfail
= panfrost_translate_stencil_op(in
->zfail_op
);
539 out
->dppass
= panfrost_translate_stencil_op(in
->zpass_op
);
543 panfrost_frag_meta_rasterizer_update(struct panfrost_context
*ctx
,
544 struct mali_shader_meta
*fragmeta
)
546 if (!ctx
->rasterizer
) {
547 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, true);
548 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, false);
549 fragmeta
->depth_units
= 0.0f
;
550 fragmeta
->depth_factor
= 0.0f
;
551 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
, false);
552 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
, false);
556 bool msaa
= ctx
->rasterizer
->base
.multisample
;
558 /* TODO: Sample size */
559 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, msaa
);
560 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, !msaa
);
561 fragmeta
->depth_units
= ctx
->rasterizer
->base
.offset_units
* 2.0f
;
562 fragmeta
->depth_factor
= ctx
->rasterizer
->base
.offset_scale
;
564 /* XXX: Which bit is which? Does this maybe allow offseting not-tri? */
566 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
,
567 ctx
->rasterizer
->base
.offset_tri
);
568 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
,
569 ctx
->rasterizer
->base
.offset_tri
);
573 panfrost_frag_meta_zsa_update(struct panfrost_context
*ctx
,
574 struct mali_shader_meta
*fragmeta
)
576 const struct pipe_depth_stencil_alpha_state
*zsa
= ctx
->depth_stencil
;
577 int zfunc
= PIPE_FUNC_ALWAYS
;
580 struct pipe_stencil_state default_stencil
= {
582 .func
= PIPE_FUNC_ALWAYS
,
583 .fail_op
= MALI_STENCIL_KEEP
,
584 .zfail_op
= MALI_STENCIL_KEEP
,
585 .zpass_op
= MALI_STENCIL_KEEP
,
590 panfrost_make_stencil_state(&default_stencil
,
591 &fragmeta
->stencil_front
);
592 fragmeta
->stencil_mask_front
= default_stencil
.writemask
;
593 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
594 fragmeta
->stencil_mask_back
= default_stencil
.writemask
;
595 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
, false);
596 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
, false);
598 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
,
599 zsa
->stencil
[0].enabled
);
600 panfrost_make_stencil_state(&zsa
->stencil
[0],
601 &fragmeta
->stencil_front
);
602 fragmeta
->stencil_mask_front
= zsa
->stencil
[0].writemask
;
603 fragmeta
->stencil_front
.ref
= ctx
->stencil_ref
.ref_value
[0];
605 /* If back-stencil is not enabled, use the front values */
607 if (zsa
->stencil
[1].enabled
) {
608 panfrost_make_stencil_state(&zsa
->stencil
[1],
609 &fragmeta
->stencil_back
);
610 fragmeta
->stencil_mask_back
= zsa
->stencil
[1].writemask
;
611 fragmeta
->stencil_back
.ref
= ctx
->stencil_ref
.ref_value
[1];
613 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
614 fragmeta
->stencil_mask_back
= fragmeta
->stencil_mask_front
;
615 fragmeta
->stencil_back
.ref
= fragmeta
->stencil_front
.ref
;
618 if (zsa
->depth
.enabled
)
619 zfunc
= zsa
->depth
.func
;
621 /* Depth state (TODO: Refactor) */
623 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
,
624 zsa
->depth
.writemask
);
627 fragmeta
->unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
628 fragmeta
->unknown2_3
|= MALI_DEPTH_FUNC(panfrost_translate_compare_func(zfunc
));
632 panfrost_fs_required(
633 struct panfrost_shader_state
*fs
,
634 struct panfrost_blend_final
*blend
,
637 /* If we generally have side effects */
641 /* If colour is written we need to execute */
642 for (unsigned i
= 0; i
< rt_count
; ++i
) {
643 if (!blend
[i
].no_colour
)
647 /* If depth is written and not implied we need to execute.
648 * TODO: Predicate on Z/S writes being enabled */
649 return (fs
->writes_depth
|| fs
->writes_stencil
);
653 panfrost_frag_meta_blend_update(struct panfrost_context
*ctx
,
654 struct mali_shader_meta
*fragmeta
,
657 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
658 struct panfrost_shader_state
*fs
;
659 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
661 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_DITHER
,
662 (dev
->quirks
& MIDGARD_SFBD
) && ctx
->blend
&&
663 !ctx
->blend
->base
.dither
);
665 /* Get blending setup */
666 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
668 struct panfrost_blend_final blend
[PIPE_MAX_COLOR_BUFS
];
669 unsigned shader_offset
= 0;
670 struct panfrost_bo
*shader_bo
= NULL
;
672 for (unsigned c
= 0; c
< rt_count
; ++c
)
673 blend
[c
] = panfrost_get_blend_for_context(ctx
, c
, &shader_bo
,
676 /* Disable shader execution if we can */
677 if (dev
->quirks
& MIDGARD_SHADERLESS
678 && !panfrost_fs_required(fs
, blend
, rt_count
)) {
679 fragmeta
->shader
= 0;
680 fragmeta
->attribute_count
= 0;
681 fragmeta
->varying_count
= 0;
682 fragmeta
->texture_count
= 0;
683 fragmeta
->sampler_count
= 0;
685 /* This feature is not known to work on Bifrost */
686 fragmeta
->midgard1
.work_count
= 1;
687 fragmeta
->midgard1
.uniform_count
= 0;
688 fragmeta
->midgard1
.uniform_buffer_count
= 0;
691 /* If there is a blend shader, work registers are shared. We impose 8
692 * work registers as a limit for blend shaders. Should be lower XXX */
694 if (!(dev
->quirks
& IS_BIFROST
)) {
695 for (unsigned c
= 0; c
< rt_count
; ++c
) {
696 if (blend
[c
].is_shader
) {
697 fragmeta
->midgard1
.work_count
=
698 MAX2(fragmeta
->midgard1
.work_count
, 8);
703 /* Even on MFBD, the shader descriptor gets blend shaders. It's *also*
704 * copied to the blend_meta appended (by convention), but this is the
705 * field actually read by the hardware. (Or maybe both are read...?).
706 * Specify the last RTi with a blend shader. */
708 fragmeta
->blend
.shader
= 0;
710 for (signed rt
= (rt_count
- 1); rt
>= 0; --rt
) {
711 if (!blend
[rt
].is_shader
)
714 fragmeta
->blend
.shader
= blend
[rt
].shader
.gpu
|
715 blend
[rt
].shader
.first_tag
;
719 if (dev
->quirks
& MIDGARD_SFBD
) {
720 /* When only a single render target platform is used, the blend
721 * information is inside the shader meta itself. We additionally
722 * need to signal CAN_DISCARD for nontrivial blend modes (so
723 * we're able to read back the destination buffer) */
725 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_BLEND_SHADER
,
728 if (!blend
[0].is_shader
) {
729 fragmeta
->blend
.equation
= *blend
[0].equation
.equation
;
730 fragmeta
->blend
.constant
= blend
[0].equation
.constant
;
733 SET_BIT(fragmeta
->unknown2_3
, MALI_CAN_DISCARD
,
734 !blend
[0].no_blending
|| fs
->can_discard
);
738 if (dev
->quirks
& IS_BIFROST
) {
739 bool no_blend
= true;
741 for (unsigned i
= 0; i
< rt_count
; ++i
)
742 no_blend
&= (blend
[i
].no_blending
| blend
[i
].no_colour
);
744 SET_BIT(fragmeta
->bifrost1
.unk1
, MALI_BIFROST_EARLY_Z
,
745 !fs
->can_discard
&& !fs
->writes_depth
&& no_blend
);
748 /* Additional blend descriptor tacked on for jobs using MFBD */
750 for (unsigned i
= 0; i
< rt_count
; ++i
) {
753 if (ctx
->pipe_framebuffer
.nr_cbufs
> i
&& !blend
[i
].no_colour
) {
756 bool is_srgb
= (ctx
->pipe_framebuffer
.nr_cbufs
> i
) &&
757 (ctx
->pipe_framebuffer
.cbufs
[i
]) &&
758 util_format_is_srgb(ctx
->pipe_framebuffer
.cbufs
[i
]->format
);
760 SET_BIT(flags
, MALI_BLEND_MRT_SHADER
, blend
[i
].is_shader
);
761 SET_BIT(flags
, MALI_BLEND_LOAD_TIB
, !blend
[i
].no_blending
);
762 SET_BIT(flags
, MALI_BLEND_SRGB
, is_srgb
);
763 SET_BIT(flags
, MALI_BLEND_NO_DITHER
, !ctx
->blend
->base
.dither
);
766 if (dev
->quirks
& IS_BIFROST
) {
767 struct bifrost_blend_rt
*brts
= rts
;
769 brts
[i
].flags
= flags
;
771 if (blend
[i
].is_shader
) {
772 /* The blend shader's address needs to be at
773 * the same top 32 bit as the fragment shader.
774 * TODO: Ensure that's always the case.
776 assert((blend
[i
].shader
.gpu
& (0xffffffffull
<< 32)) ==
777 (fs
->bo
->gpu
& (0xffffffffull
<< 32)));
778 brts
[i
].shader
= blend
[i
].shader
.gpu
;
780 } else if (ctx
->pipe_framebuffer
.nr_cbufs
> i
) {
781 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
782 const struct util_format_description
*format_desc
;
783 format_desc
= util_format_description(format
);
785 brts
[i
].equation
= *blend
[i
].equation
.equation
;
787 /* TODO: this is a bit more complicated */
788 brts
[i
].constant
= blend
[i
].equation
.constant
;
790 brts
[i
].format
= panfrost_format_to_bifrost_blend(format_desc
);
792 /* 0x19 disables blending and forces REPLACE
793 * mode (equivalent to rgb_mode = alpha_mode =
794 * x122, colour mask = 0xF). 0x1a allows
796 brts
[i
].unk2
= blend
[i
].no_blending
? 0x19 : 0x1a;
798 brts
[i
].shader_type
= fs
->blend_types
[i
];
800 /* Dummy attachment for depth-only */
802 brts
[i
].shader_type
= fs
->blend_types
[i
];
805 struct midgard_blend_rt
*mrts
= rts
;
806 mrts
[i
].flags
= flags
;
808 if (blend
[i
].is_shader
) {
809 mrts
[i
].blend
.shader
= blend
[i
].shader
.gpu
| blend
[i
].shader
.first_tag
;
811 mrts
[i
].blend
.equation
= *blend
[i
].equation
.equation
;
812 mrts
[i
].blend
.constant
= blend
[i
].equation
.constant
;
819 panfrost_frag_shader_meta_init(struct panfrost_context
*ctx
,
820 struct mali_shader_meta
*fragmeta
,
823 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
824 struct panfrost_shader_state
*fs
;
826 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
828 fragmeta
->alpha_coverage
= ~MALI_ALPHA_COVERAGE(0.000000);
829 fragmeta
->unknown2_3
= MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS
) | 0x3010;
830 fragmeta
->unknown2_4
= 0x4e0;
832 /* unknown2_4 has 0x10 bit set on T6XX and T720. We don't know why this
833 * is required (independent of 32-bit/64-bit descriptors), or why it's
834 * not used on later GPU revisions. Otherwise, all shader jobs fault on
835 * these earlier chips (perhaps this is a chicken bit of some kind).
836 * More investigation is needed. */
838 SET_BIT(fragmeta
->unknown2_4
, 0x10, dev
->quirks
& MIDGARD_SFBD
);
840 if (dev
->quirks
& IS_BIFROST
) {
843 /* Depending on whether it's legal to in the given shader, we try to
844 * enable early-z testing. TODO: respect e-z force */
846 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_EARLY_Z
,
847 !fs
->can_discard
&& !fs
->writes_global
&&
848 !fs
->writes_depth
&& !fs
->writes_stencil
);
850 /* Add the writes Z/S flags if needed. */
851 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_WRITES_Z
, fs
->writes_depth
);
852 SET_BIT(fragmeta
->midgard1
.flags_hi
, MALI_WRITES_S
, fs
->writes_stencil
);
854 /* Any time texturing is used, derivatives are implicitly calculated,
855 * so we need to enable helper invocations */
857 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_HELPER_INVOCATIONS
,
858 fs
->helper_invocations
);
860 const struct pipe_depth_stencil_alpha_state
*zsa
= ctx
->depth_stencil
;
862 bool depth_enabled
= fs
->writes_depth
||
863 (zsa
&& zsa
->depth
.enabled
&& zsa
->depth
.func
!= PIPE_FUNC_ALWAYS
);
865 SET_BIT(fragmeta
->midgard1
.flags_lo
, 0x400, !depth_enabled
&& fs
->can_discard
);
866 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_READS_ZS
, depth_enabled
&& fs
->can_discard
);
869 panfrost_frag_meta_rasterizer_update(ctx
, fragmeta
);
870 panfrost_frag_meta_zsa_update(ctx
, fragmeta
);
871 panfrost_frag_meta_blend_update(ctx
, fragmeta
, rts
);
875 panfrost_emit_shader_meta(struct panfrost_batch
*batch
,
876 enum pipe_shader_type st
,
877 struct mali_vertex_tiler_postfix
*postfix
)
879 struct panfrost_context
*ctx
= batch
->ctx
;
880 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
887 struct mali_shader_meta meta
;
889 panfrost_shader_meta_init(ctx
, st
, &meta
);
891 /* Add the shader BO to the batch. */
892 panfrost_batch_add_bo(batch
, ss
->bo
,
893 PAN_BO_ACCESS_PRIVATE
|
895 panfrost_bo_access_for_stage(st
));
899 if (st
== PIPE_SHADER_FRAGMENT
) {
900 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
901 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
902 size_t desc_size
= sizeof(meta
);
904 struct panfrost_transfer xfer
;
907 if (dev
->quirks
& MIDGARD_SFBD
)
909 else if (dev
->quirks
& IS_BIFROST
)
910 rt_size
= sizeof(struct bifrost_blend_rt
);
912 rt_size
= sizeof(struct midgard_blend_rt
);
914 desc_size
+= rt_size
* rt_count
;
917 rts
= rzalloc_size(ctx
, rt_size
* rt_count
);
919 panfrost_frag_shader_meta_init(ctx
, &meta
, rts
);
921 xfer
= panfrost_allocate_transient(batch
, desc_size
);
923 memcpy(xfer
.cpu
, &meta
, sizeof(meta
));
924 memcpy(xfer
.cpu
+ sizeof(meta
), rts
, rt_size
* rt_count
);
929 shader_ptr
= xfer
.gpu
;
931 shader_ptr
= panfrost_upload_transient(batch
, &meta
,
935 postfix
->shader
= shader_ptr
;
939 panfrost_mali_viewport_init(struct panfrost_context
*ctx
,
940 struct mali_viewport
*mvp
)
942 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
944 /* Clip bounds are encoded as floats. The viewport itself is encoded as
945 * (somewhat) asymmetric ints. */
947 const struct pipe_scissor_state
*ss
= &ctx
->scissor
;
949 memset(mvp
, 0, sizeof(*mvp
));
951 /* By default, do no viewport clipping, i.e. clip to (-inf, inf) in
952 * each direction. Clipping to the viewport in theory should work, but
953 * in practice causes issues when we're not explicitly trying to
956 *mvp
= (struct mali_viewport
) {
957 .clip_minx
= -INFINITY
,
958 .clip_miny
= -INFINITY
,
959 .clip_maxx
= INFINITY
,
960 .clip_maxy
= INFINITY
,
963 /* Always scissor to the viewport by default. */
964 float vp_minx
= (int) (vp
->translate
[0] - fabsf(vp
->scale
[0]));
965 float vp_maxx
= (int) (vp
->translate
[0] + fabsf(vp
->scale
[0]));
967 float vp_miny
= (int) (vp
->translate
[1] - fabsf(vp
->scale
[1]));
968 float vp_maxy
= (int) (vp
->translate
[1] + fabsf(vp
->scale
[1]));
970 float minz
= (vp
->translate
[2] - fabsf(vp
->scale
[2]));
971 float maxz
= (vp
->translate
[2] + fabsf(vp
->scale
[2]));
973 /* Apply the scissor test */
975 unsigned minx
, miny
, maxx
, maxy
;
977 if (ss
&& ctx
->rasterizer
&& ctx
->rasterizer
->base
.scissor
) {
978 minx
= MAX2(ss
->minx
, vp_minx
);
979 miny
= MAX2(ss
->miny
, vp_miny
);
980 maxx
= MIN2(ss
->maxx
, vp_maxx
);
981 maxy
= MIN2(ss
->maxy
, vp_maxy
);
989 /* Hardware needs the min/max to be strictly ordered, so flip if we
990 * need to. The viewport transformation in the vertex shader will
991 * handle the negatives if we don't */
994 unsigned temp
= miny
;
1000 unsigned temp
= minx
;
1011 /* Clamp to the framebuffer size as a last check */
1013 minx
= MIN2(ctx
->pipe_framebuffer
.width
, minx
);
1014 maxx
= MIN2(ctx
->pipe_framebuffer
.width
, maxx
);
1016 miny
= MIN2(ctx
->pipe_framebuffer
.height
, miny
);
1017 maxy
= MIN2(ctx
->pipe_framebuffer
.height
, maxy
);
1021 mvp
->viewport0
[0] = minx
;
1022 mvp
->viewport1
[0] = MALI_POSITIVE(maxx
);
1024 mvp
->viewport0
[1] = miny
;
1025 mvp
->viewport1
[1] = MALI_POSITIVE(maxy
);
1027 mvp
->clip_minz
= minz
;
1028 mvp
->clip_maxz
= maxz
;
1032 panfrost_emit_viewport(struct panfrost_batch
*batch
,
1033 struct mali_vertex_tiler_postfix
*tiler_postfix
)
1035 struct panfrost_context
*ctx
= batch
->ctx
;
1036 struct mali_viewport mvp
;
1038 panfrost_mali_viewport_init(batch
->ctx
, &mvp
);
1040 /* Update the job, unless we're doing wallpapering (whose lack of
1041 * scissor we can ignore, since if we "miss" a tile of wallpaper, it'll
1042 * just... be faster :) */
1044 if (!ctx
->wallpaper_batch
)
1045 panfrost_batch_union_scissor(batch
, mvp
.viewport0
[0],
1047 mvp
.viewport1
[0] + 1,
1048 mvp
.viewport1
[1] + 1);
1050 tiler_postfix
->viewport
= panfrost_upload_transient(batch
, &mvp
,
1055 panfrost_map_constant_buffer_gpu(struct panfrost_batch
*batch
,
1056 enum pipe_shader_type st
,
1057 struct panfrost_constant_buffer
*buf
,
1060 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1061 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1064 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1065 PAN_BO_ACCESS_SHARED
|
1066 PAN_BO_ACCESS_READ
|
1067 panfrost_bo_access_for_stage(st
));
1069 /* Alignment gauranteed by
1070 * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
1071 return rsrc
->bo
->gpu
+ cb
->buffer_offset
;
1072 } else if (cb
->user_buffer
) {
1073 return panfrost_upload_transient(batch
,
1078 unreachable("No constant buffer");
1082 struct sysval_uniform
{
1092 panfrost_upload_viewport_scale_sysval(struct panfrost_batch
*batch
,
1093 struct sysval_uniform
*uniform
)
1095 struct panfrost_context
*ctx
= batch
->ctx
;
1096 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
1098 uniform
->f
[0] = vp
->scale
[0];
1099 uniform
->f
[1] = vp
->scale
[1];
1100 uniform
->f
[2] = vp
->scale
[2];
1104 panfrost_upload_viewport_offset_sysval(struct panfrost_batch
*batch
,
1105 struct sysval_uniform
*uniform
)
1107 struct panfrost_context
*ctx
= batch
->ctx
;
1108 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
1110 uniform
->f
[0] = vp
->translate
[0];
1111 uniform
->f
[1] = vp
->translate
[1];
1112 uniform
->f
[2] = vp
->translate
[2];
1115 static void panfrost_upload_txs_sysval(struct panfrost_batch
*batch
,
1116 enum pipe_shader_type st
,
1117 unsigned int sysvalid
,
1118 struct sysval_uniform
*uniform
)
1120 struct panfrost_context
*ctx
= batch
->ctx
;
1121 unsigned texidx
= PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid
);
1122 unsigned dim
= PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid
);
1123 bool is_array
= PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid
);
1124 struct pipe_sampler_view
*tex
= &ctx
->sampler_views
[st
][texidx
]->base
;
1127 uniform
->i
[0] = u_minify(tex
->texture
->width0
, tex
->u
.tex
.first_level
);
1130 uniform
->i
[1] = u_minify(tex
->texture
->height0
,
1131 tex
->u
.tex
.first_level
);
1134 uniform
->i
[2] = u_minify(tex
->texture
->depth0
,
1135 tex
->u
.tex
.first_level
);
1138 uniform
->i
[dim
] = tex
->texture
->array_size
;
1142 panfrost_upload_ssbo_sysval(struct panfrost_batch
*batch
,
1143 enum pipe_shader_type st
,
1145 struct sysval_uniform
*uniform
)
1147 struct panfrost_context
*ctx
= batch
->ctx
;
1149 assert(ctx
->ssbo_mask
[st
] & (1 << ssbo_id
));
1150 struct pipe_shader_buffer sb
= ctx
->ssbo
[st
][ssbo_id
];
1152 /* Compute address */
1153 struct panfrost_bo
*bo
= pan_resource(sb
.buffer
)->bo
;
1155 panfrost_batch_add_bo(batch
, bo
,
1156 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_RW
|
1157 panfrost_bo_access_for_stage(st
));
1159 /* Upload address and size as sysval */
1160 uniform
->du
[0] = bo
->gpu
+ sb
.buffer_offset
;
1161 uniform
->u
[2] = sb
.buffer_size
;
1165 panfrost_upload_sampler_sysval(struct panfrost_batch
*batch
,
1166 enum pipe_shader_type st
,
1168 struct sysval_uniform
*uniform
)
1170 struct panfrost_context
*ctx
= batch
->ctx
;
1171 struct pipe_sampler_state
*sampl
= &ctx
->samplers
[st
][samp_idx
]->base
;
1173 uniform
->f
[0] = sampl
->min_lod
;
1174 uniform
->f
[1] = sampl
->max_lod
;
1175 uniform
->f
[2] = sampl
->lod_bias
;
1177 /* Even without any errata, Midgard represents "no mipmapping" as
1178 * fixing the LOD with the clamps; keep behaviour consistent. c.f.
1179 * panfrost_create_sampler_state which also explains our choice of
1180 * epsilon value (again to keep behaviour consistent) */
1182 if (sampl
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
1183 uniform
->f
[1] = uniform
->f
[0] + (1.0/256.0);
1187 panfrost_upload_num_work_groups_sysval(struct panfrost_batch
*batch
,
1188 struct sysval_uniform
*uniform
)
1190 struct panfrost_context
*ctx
= batch
->ctx
;
1192 uniform
->u
[0] = ctx
->compute_grid
->grid
[0];
1193 uniform
->u
[1] = ctx
->compute_grid
->grid
[1];
1194 uniform
->u
[2] = ctx
->compute_grid
->grid
[2];
1198 panfrost_upload_sysvals(struct panfrost_batch
*batch
, void *buf
,
1199 struct panfrost_shader_state
*ss
,
1200 enum pipe_shader_type st
)
1202 struct sysval_uniform
*uniforms
= (void *)buf
;
1204 for (unsigned i
= 0; i
< ss
->sysval_count
; ++i
) {
1205 int sysval
= ss
->sysval
[i
];
1207 switch (PAN_SYSVAL_TYPE(sysval
)) {
1208 case PAN_SYSVAL_VIEWPORT_SCALE
:
1209 panfrost_upload_viewport_scale_sysval(batch
,
1212 case PAN_SYSVAL_VIEWPORT_OFFSET
:
1213 panfrost_upload_viewport_offset_sysval(batch
,
1216 case PAN_SYSVAL_TEXTURE_SIZE
:
1217 panfrost_upload_txs_sysval(batch
, st
,
1218 PAN_SYSVAL_ID(sysval
),
1221 case PAN_SYSVAL_SSBO
:
1222 panfrost_upload_ssbo_sysval(batch
, st
,
1223 PAN_SYSVAL_ID(sysval
),
1226 case PAN_SYSVAL_NUM_WORK_GROUPS
:
1227 panfrost_upload_num_work_groups_sysval(batch
,
1230 case PAN_SYSVAL_SAMPLER
:
1231 panfrost_upload_sampler_sysval(batch
, st
,
1232 PAN_SYSVAL_ID(sysval
),
1242 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer
*buf
,
1245 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1246 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1249 return rsrc
->bo
->cpu
;
1250 else if (cb
->user_buffer
)
1251 return cb
->user_buffer
;
1253 unreachable("No constant buffer");
1257 panfrost_emit_const_buf(struct panfrost_batch
*batch
,
1258 enum pipe_shader_type stage
,
1259 struct mali_vertex_tiler_postfix
*postfix
)
1261 struct panfrost_context
*ctx
= batch
->ctx
;
1262 struct panfrost_shader_variants
*all
= ctx
->shader
[stage
];
1267 struct panfrost_constant_buffer
*buf
= &ctx
->constant_buffer
[stage
];
1269 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1271 /* Uniforms are implicitly UBO #0 */
1272 bool has_uniforms
= buf
->enabled_mask
& (1 << 0);
1274 /* Allocate room for the sysval and the uniforms */
1275 size_t sys_size
= sizeof(float) * 4 * ss
->sysval_count
;
1276 size_t uniform_size
= has_uniforms
? (buf
->cb
[0].buffer_size
) : 0;
1277 size_t size
= sys_size
+ uniform_size
;
1278 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1281 /* Upload sysvals requested by the shader */
1282 panfrost_upload_sysvals(batch
, transfer
.cpu
, ss
, stage
);
1284 /* Upload uniforms */
1285 if (has_uniforms
&& uniform_size
) {
1286 const void *cpu
= panfrost_map_constant_buffer_cpu(buf
, 0);
1287 memcpy(transfer
.cpu
+ sys_size
, cpu
, uniform_size
);
1290 /* Next up, attach UBOs. UBO #0 is the uniforms we just
1293 unsigned ubo_count
= panfrost_ubo_count(ctx
, stage
);
1294 assert(ubo_count
>= 1);
1296 size_t sz
= sizeof(uint64_t) * ubo_count
;
1297 uint64_t ubos
[PAN_MAX_CONST_BUFFERS
];
1298 int uniform_count
= ss
->uniform_count
;
1300 /* Upload uniforms as a UBO */
1301 ubos
[0] = MALI_MAKE_UBO(2 + uniform_count
, transfer
.gpu
);
1303 /* The rest are honest-to-goodness UBOs */
1305 for (unsigned ubo
= 1; ubo
< ubo_count
; ++ubo
) {
1306 size_t usz
= buf
->cb
[ubo
].buffer_size
;
1307 bool enabled
= buf
->enabled_mask
& (1 << ubo
);
1308 bool empty
= usz
== 0;
1310 if (!enabled
|| empty
) {
1311 /* Stub out disabled UBOs to catch accesses */
1312 ubos
[ubo
] = MALI_MAKE_UBO(0, 0xDEAD0000);
1316 mali_ptr gpu
= panfrost_map_constant_buffer_gpu(batch
, stage
,
1319 unsigned bytes_per_field
= 16;
1320 unsigned aligned
= ALIGN_POT(usz
, bytes_per_field
);
1321 ubos
[ubo
] = MALI_MAKE_UBO(aligned
/ bytes_per_field
, gpu
);
1324 mali_ptr ubufs
= panfrost_upload_transient(batch
, ubos
, sz
);
1325 postfix
->uniforms
= transfer
.gpu
;
1326 postfix
->uniform_buffers
= ubufs
;
1328 buf
->dirty_mask
= 0;
1332 panfrost_emit_shared_memory(struct panfrost_batch
*batch
,
1333 const struct pipe_grid_info
*info
,
1334 struct midgard_payload_vertex_tiler
*vtp
)
1336 struct panfrost_context
*ctx
= batch
->ctx
;
1337 struct panfrost_shader_variants
*all
= ctx
->shader
[PIPE_SHADER_COMPUTE
];
1338 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1339 unsigned single_size
= util_next_power_of_two(MAX2(ss
->shared_size
,
1341 unsigned shared_size
= single_size
* info
->grid
[0] * info
->grid
[1] *
1343 struct panfrost_bo
*bo
= panfrost_batch_get_shared_memory(batch
,
1347 struct mali_shared_memory shared
= {
1348 .shared_memory
= bo
->gpu
,
1349 .shared_workgroup_count
=
1350 util_logbase2_ceil(info
->grid
[0]) +
1351 util_logbase2_ceil(info
->grid
[1]) +
1352 util_logbase2_ceil(info
->grid
[2]),
1354 .shared_shift
= util_logbase2(single_size
) - 1
1357 vtp
->postfix
.shared_memory
= panfrost_upload_transient(batch
, &shared
,
1362 panfrost_get_tex_desc(struct panfrost_batch
*batch
,
1363 enum pipe_shader_type st
,
1364 struct panfrost_sampler_view
*view
)
1367 return (mali_ptr
) 0;
1369 struct pipe_sampler_view
*pview
= &view
->base
;
1370 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1372 /* Add the BO to the job so it's retained until the job is done. */
1374 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1375 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1376 panfrost_bo_access_for_stage(st
));
1378 panfrost_batch_add_bo(batch
, view
->midgard_bo
,
1379 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1380 panfrost_bo_access_for_stage(st
));
1382 return view
->midgard_bo
->gpu
;
1386 panfrost_emit_texture_descriptors(struct panfrost_batch
*batch
,
1387 enum pipe_shader_type stage
,
1388 struct mali_vertex_tiler_postfix
*postfix
)
1390 struct panfrost_context
*ctx
= batch
->ctx
;
1391 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1393 if (!ctx
->sampler_view_count
[stage
])
1396 if (device
->quirks
& IS_BIFROST
) {
1397 struct bifrost_texture_descriptor
*descriptors
;
1399 descriptors
= malloc(sizeof(struct bifrost_texture_descriptor
) *
1400 ctx
->sampler_view_count
[stage
]);
1402 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1403 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1404 struct pipe_sampler_view
*pview
= &view
->base
;
1405 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1407 /* Add the BOs to the job so they are retained until the job is done. */
1409 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1410 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1411 panfrost_bo_access_for_stage(stage
));
1413 panfrost_batch_add_bo(batch
, view
->bifrost_bo
,
1414 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1415 panfrost_bo_access_for_stage(stage
));
1417 memcpy(&descriptors
[i
], view
->bifrost_descriptor
, sizeof(*view
->bifrost_descriptor
));
1420 postfix
->textures
= panfrost_upload_transient(batch
,
1422 sizeof(struct bifrost_texture_descriptor
) *
1423 ctx
->sampler_view_count
[stage
]);
1427 uint64_t trampolines
[PIPE_MAX_SHADER_SAMPLER_VIEWS
];
1429 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
)
1430 trampolines
[i
] = panfrost_get_tex_desc(batch
, stage
,
1431 ctx
->sampler_views
[stage
][i
]);
1433 postfix
->textures
= panfrost_upload_transient(batch
,
1436 ctx
->sampler_view_count
[stage
]);
1441 panfrost_emit_sampler_descriptors(struct panfrost_batch
*batch
,
1442 enum pipe_shader_type stage
,
1443 struct mali_vertex_tiler_postfix
*postfix
)
1445 struct panfrost_context
*ctx
= batch
->ctx
;
1446 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1448 if (!ctx
->sampler_count
[stage
])
1451 if (device
->quirks
& IS_BIFROST
) {
1452 size_t desc_size
= sizeof(struct bifrost_sampler_descriptor
);
1453 size_t transfer_size
= desc_size
* ctx
->sampler_count
[stage
];
1454 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1456 struct bifrost_sampler_descriptor
*desc
= (struct bifrost_sampler_descriptor
*)transfer
.cpu
;
1458 for (int i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1459 desc
[i
] = ctx
->samplers
[stage
][i
]->bifrost_hw
;
1461 postfix
->sampler_descriptor
= transfer
.gpu
;
1463 size_t desc_size
= sizeof(struct mali_sampler_descriptor
);
1464 size_t transfer_size
= desc_size
* ctx
->sampler_count
[stage
];
1465 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1467 struct mali_sampler_descriptor
*desc
= (struct mali_sampler_descriptor
*)transfer
.cpu
;
1469 for (int i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1470 desc
[i
] = ctx
->samplers
[stage
][i
]->midgard_hw
;
1472 postfix
->sampler_descriptor
= transfer
.gpu
;
1477 panfrost_emit_vertex_attr_meta(struct panfrost_batch
*batch
,
1478 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1480 struct panfrost_context
*ctx
= batch
->ctx
;
1485 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1487 panfrost_vertex_state_upd_attr_offs(ctx
, vertex_postfix
);
1488 vertex_postfix
->attribute_meta
= panfrost_upload_transient(batch
, so
->hw
,
1494 panfrost_emit_vertex_data(struct panfrost_batch
*batch
,
1495 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1497 struct panfrost_context
*ctx
= batch
->ctx
;
1498 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1500 /* Staged mali_attr, and index into them. i =/= k, depending on the
1501 * vertex buffer mask and instancing. Twice as much room is allocated,
1502 * for a worst case of NPOT_DIVIDEs which take up extra slot */
1503 union mali_attr attrs
[PIPE_MAX_ATTRIBS
* 2];
1506 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1507 /* We map a mali_attr to be 1:1 with the mali_attr_meta, which
1508 * means duplicating some vertex buffers (who cares? aside from
1509 * maybe some caching implications but I somehow doubt that
1512 struct pipe_vertex_element
*elem
= &so
->pipe
[i
];
1513 unsigned vbi
= elem
->vertex_buffer_index
;
1515 /* The exception to 1:1 mapping is that we can have multiple
1516 * entries (NPOT divisors), so we fixup anyways */
1518 so
->hw
[i
].index
= k
;
1520 if (!(ctx
->vb_mask
& (1 << vbi
)))
1523 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1524 struct panfrost_resource
*rsrc
;
1526 rsrc
= pan_resource(buf
->buffer
.resource
);
1530 /* Align to 64 bytes by masking off the lower bits. This
1531 * will be adjusted back when we fixup the src_offset in
1534 mali_ptr raw_addr
= rsrc
->bo
->gpu
+ buf
->buffer_offset
;
1535 mali_ptr addr
= raw_addr
& ~63;
1536 unsigned chopped_addr
= raw_addr
- addr
;
1538 /* Add a dependency of the batch on the vertex buffer */
1539 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1540 PAN_BO_ACCESS_SHARED
|
1541 PAN_BO_ACCESS_READ
|
1542 PAN_BO_ACCESS_VERTEX_TILER
);
1544 /* Set common fields */
1545 attrs
[k
].elements
= addr
;
1546 attrs
[k
].stride
= buf
->stride
;
1548 /* Since we advanced the base pointer, we shrink the buffer
1550 attrs
[k
].size
= rsrc
->base
.width0
- buf
->buffer_offset
;
1552 /* We need to add the extra size we masked off (for
1553 * correctness) so the data doesn't get clamped away */
1554 attrs
[k
].size
+= chopped_addr
;
1556 /* For non-instancing make sure we initialize */
1557 attrs
[k
].shift
= attrs
[k
].extra_flags
= 0;
1559 /* Instancing uses a dramatically different code path than
1560 * linear, so dispatch for the actual emission now that the
1561 * common code is finished */
1563 unsigned divisor
= elem
->instance_divisor
;
1565 if (divisor
&& ctx
->instance_count
== 1) {
1566 /* Silly corner case where there's a divisor(=1) but
1567 * there's no legitimate instancing. So we want *every*
1568 * attribute to be the same. So set stride to zero so
1569 * we don't go anywhere. */
1571 attrs
[k
].size
= attrs
[k
].stride
+ chopped_addr
;
1572 attrs
[k
].stride
= 0;
1573 attrs
[k
++].elements
|= MALI_ATTR_LINEAR
;
1574 } else if (ctx
->instance_count
<= 1) {
1575 /* Normal, non-instanced attributes */
1576 attrs
[k
++].elements
|= MALI_ATTR_LINEAR
;
1578 unsigned instance_shift
= vertex_postfix
->instance_shift
;
1579 unsigned instance_odd
= vertex_postfix
->instance_odd
;
1581 k
+= panfrost_vertex_instanced(ctx
->padded_count
,
1584 divisor
, &attrs
[k
]);
1588 /* Add special gl_VertexID/gl_InstanceID buffers */
1590 panfrost_vertex_id(ctx
->padded_count
, &attrs
[k
]);
1591 so
->hw
[PAN_VERTEX_ID
].index
= k
++;
1592 panfrost_instance_id(ctx
->padded_count
, &attrs
[k
]);
1593 so
->hw
[PAN_INSTANCE_ID
].index
= k
++;
1595 /* Upload whatever we emitted and go */
1597 vertex_postfix
->attributes
= panfrost_upload_transient(batch
, attrs
,
1598 k
* sizeof(*attrs
));
1602 panfrost_emit_varyings(struct panfrost_batch
*batch
, union mali_attr
*slot
,
1603 unsigned stride
, unsigned count
)
1605 /* Fill out the descriptor */
1606 slot
->stride
= stride
;
1607 slot
->size
= stride
* count
;
1608 slot
->shift
= slot
->extra_flags
= 0;
1610 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1613 slot
->elements
= transfer
.gpu
| MALI_ATTR_LINEAR
;
1615 return transfer
.gpu
;
1619 panfrost_streamout_offset(unsigned stride
, unsigned offset
,
1620 struct pipe_stream_output_target
*target
)
1622 return (target
->buffer_offset
+ (offset
* stride
* 4)) & 63;
1626 panfrost_emit_streamout(struct panfrost_batch
*batch
, union mali_attr
*slot
,
1627 unsigned stride
, unsigned offset
, unsigned count
,
1628 struct pipe_stream_output_target
*target
)
1630 /* Fill out the descriptor */
1631 slot
->stride
= stride
* 4;
1632 slot
->shift
= slot
->extra_flags
= 0;
1634 unsigned max_size
= target
->buffer_size
;
1635 unsigned expected_size
= slot
->stride
* count
;
1637 slot
->size
= MIN2(max_size
, expected_size
);
1639 /* Grab the BO and bind it to the batch */
1640 struct panfrost_bo
*bo
= pan_resource(target
->buffer
)->bo
;
1642 /* Varyings are WRITE from the perspective of the VERTEX but READ from
1643 * the perspective of the TILER and FRAGMENT.
1645 panfrost_batch_add_bo(batch
, bo
,
1646 PAN_BO_ACCESS_SHARED
|
1648 PAN_BO_ACCESS_VERTEX_TILER
|
1649 PAN_BO_ACCESS_FRAGMENT
);
1651 mali_ptr addr
= bo
->gpu
+ target
->buffer_offset
+ (offset
* slot
->stride
);
1652 slot
->elements
= addr
;
1655 /* Given a shader and buffer indices, link varying metadata together */
1658 is_special_varying(gl_varying_slot loc
)
1661 case VARYING_SLOT_POS
:
1662 case VARYING_SLOT_PSIZ
:
1663 case VARYING_SLOT_PNTC
:
1664 case VARYING_SLOT_FACE
:
1672 panfrost_emit_varying_meta(void *outptr
, struct panfrost_shader_state
*ss
,
1673 signed general
, signed gl_Position
,
1674 signed gl_PointSize
, signed gl_PointCoord
,
1675 signed gl_FrontFacing
)
1677 struct mali_attr_meta
*out
= (struct mali_attr_meta
*) outptr
;
1679 for (unsigned i
= 0; i
< ss
->varying_count
; ++i
) {
1680 gl_varying_slot location
= ss
->varyings_loc
[i
];
1684 case VARYING_SLOT_POS
:
1685 index
= gl_Position
;
1687 case VARYING_SLOT_PSIZ
:
1688 index
= gl_PointSize
;
1690 case VARYING_SLOT_PNTC
:
1691 index
= gl_PointCoord
;
1693 case VARYING_SLOT_FACE
:
1694 index
= gl_FrontFacing
;
1702 out
[i
].index
= index
;
1707 has_point_coord(unsigned mask
, gl_varying_slot loc
)
1709 if ((loc
>= VARYING_SLOT_TEX0
) && (loc
<= VARYING_SLOT_TEX7
))
1710 return (mask
& (1 << (loc
- VARYING_SLOT_TEX0
)));
1711 else if (loc
== VARYING_SLOT_PNTC
)
1712 return (mask
& (1 << 8));
1717 /* Helpers for manipulating stream out information so we can pack varyings
1718 * accordingly. Compute the src_offset for a given captured varying */
1720 static struct pipe_stream_output
*
1721 pan_get_so(struct pipe_stream_output_info
*info
, gl_varying_slot loc
)
1723 for (unsigned i
= 0; i
< info
->num_outputs
; ++i
) {
1724 if (info
->output
[i
].register_index
== loc
)
1725 return &info
->output
[i
];
1728 unreachable("Varying not captured");
1732 pan_varying_size(enum mali_format fmt
)
1734 unsigned type
= MALI_EXTRACT_TYPE(fmt
);
1735 unsigned chan
= MALI_EXTRACT_CHANNELS(fmt
);
1736 unsigned bits
= MALI_EXTRACT_BITS(fmt
);
1739 if (bits
== MALI_CHANNEL_FLOAT
) {
1741 bool fp16
= (type
== MALI_FORMAT_SINT
);
1742 assert(fp16
|| (type
== MALI_FORMAT_UNORM
));
1746 assert(type
>= MALI_FORMAT_SNORM
&& type
<= MALI_FORMAT_SINT
);
1757 /* Indices for named (non-XFB) varyings that are present. These are packed
1758 * tightly so they correspond to a bitfield present (P) indexed by (1 <<
1759 * PAN_VARY_*). This has the nice property that you can lookup the buffer index
1760 * of a given special field given a shift S by:
1762 * idx = popcount(P & ((1 << S) - 1))
1764 * That is... look at all of the varyings that come earlier and count them, the
1765 * count is the new index since plus one. Likewise, the total number of special
1766 * buffers required is simply popcount(P)
1769 enum pan_special_varying
{
1770 PAN_VARY_GENERAL
= 0,
1771 PAN_VARY_POSITION
= 1,
1773 PAN_VARY_PNTCOORD
= 3,
1775 PAN_VARY_FRAGCOORD
= 5,
1781 /* Given a varying, figure out which index it correpsonds to */
1783 static inline unsigned
1784 pan_varying_index(unsigned present
, enum pan_special_varying v
)
1786 unsigned mask
= (1 << v
) - 1;
1787 return util_bitcount(present
& mask
);
1790 /* Get the base offset for XFB buffers, which by convention come after
1791 * everything else. Wrapper function for semantic reasons; by construction this
1792 * is just popcount. */
1794 static inline unsigned
1795 pan_xfb_base(unsigned present
)
1797 return util_bitcount(present
);
1801 panfrost_emit_varying_descriptor(struct panfrost_batch
*batch
,
1802 unsigned vertex_count
,
1803 struct mali_vertex_tiler_postfix
*vertex_postfix
,
1804 struct mali_vertex_tiler_postfix
*tiler_postfix
,
1805 union midgard_primitive_size
*primitive_size
)
1807 /* Load the shaders */
1808 struct panfrost_context
*ctx
= batch
->ctx
;
1809 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1810 struct panfrost_shader_state
*vs
, *fs
;
1811 unsigned int num_gen_varyings
= 0;
1812 size_t vs_size
, fs_size
;
1814 /* Allocate the varying descriptor */
1816 vs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_VERTEX
);
1817 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
1818 vs_size
= sizeof(struct mali_attr_meta
) * vs
->varying_count
;
1819 fs_size
= sizeof(struct mali_attr_meta
) * fs
->varying_count
;
1821 struct panfrost_transfer trans
= panfrost_allocate_transient(batch
,
1825 struct pipe_stream_output_info
*so
= &vs
->stream_output
;
1827 /* Check if this varying is linked by us. This is the case for
1828 * general-purpose, non-captured varyings. If it is, link it. If it's
1829 * not, use the provided stream out information to determine the
1830 * offset, since it was already linked for us. */
1832 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1833 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1835 bool special
= is_special_varying(loc
);
1836 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1839 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1841 unsigned dst_offset
= o
->dst_offset
* 4; /* dwords */
1842 vs
->varyings
[i
].src_offset
= dst_offset
;
1843 } else if (!special
) {
1844 vs
->varyings
[i
].src_offset
= 16 * (num_gen_varyings
++);
1848 /* Conversely, we need to set src_offset for the captured varyings.
1849 * Here, the layout is defined by the stream out info, not us */
1851 /* Link up with fragment varyings */
1852 bool reads_point_coord
= fs
->reads_point_coord
;
1854 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1855 gl_varying_slot loc
= fs
->varyings_loc
[i
];
1856 unsigned src_offset
;
1860 for (unsigned j
= 0; j
< vs
->varying_count
; ++j
) {
1861 if (vs
->varyings_loc
[j
] == loc
) {
1867 /* Either assign or reuse */
1869 src_offset
= vs
->varyings
[vs_idx
].src_offset
;
1871 src_offset
= 16 * (num_gen_varyings
++);
1873 fs
->varyings
[i
].src_offset
= src_offset
;
1875 if (has_point_coord(fs
->point_sprite_mask
, loc
))
1876 reads_point_coord
= true;
1879 memcpy(trans
.cpu
, vs
->varyings
, vs_size
);
1880 memcpy(trans
.cpu
+ vs_size
, fs
->varyings
, fs_size
);
1882 union mali_attr varyings
[PIPE_MAX_ATTRIBS
] = {0};
1884 /* Figure out how many streamout buffers could be bound */
1885 unsigned so_count
= ctx
->streamout
.num_targets
;
1886 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1887 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1889 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1890 if (!captured
) continue;
1892 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1893 so_count
= MAX2(so_count
, o
->output_buffer
+ 1);
1896 signed idx
= so_count
;
1897 signed general
= idx
++;
1898 signed gl_Position
= idx
++;
1899 signed gl_PointSize
= vs
->writes_point_size
? (idx
++) : -1;
1900 signed gl_PointCoord
= reads_point_coord
? (idx
++) : -1;
1901 signed gl_FrontFacing
= fs
->reads_face
? (idx
++) : -1;
1902 signed gl_FragCoord
= (fs
->reads_frag_coord
&&
1903 !(device
->quirks
& IS_BIFROST
))
1906 /* Emit the stream out buffers */
1908 unsigned out_count
= u_stream_outputs_for_vertices(ctx
->active_prim
,
1911 for (unsigned i
= 0; i
< so_count
; ++i
) {
1912 if (i
< ctx
->streamout
.num_targets
) {
1913 panfrost_emit_streamout(batch
, &varyings
[i
],
1915 ctx
->streamout
.offsets
[i
],
1917 ctx
->streamout
.targets
[i
]);
1919 /* Emit a dummy buffer */
1920 panfrost_emit_varyings(batch
, &varyings
[i
],
1924 /* Clear the attribute type */
1925 varyings
[i
].elements
&= ~0xF;
1929 panfrost_emit_varyings(batch
, &varyings
[general
],
1930 num_gen_varyings
* 16,
1933 mali_ptr varyings_p
;
1935 /* fp32 vec4 gl_Position */
1936 varyings_p
= panfrost_emit_varyings(batch
, &varyings
[gl_Position
],
1937 sizeof(float) * 4, vertex_count
);
1938 tiler_postfix
->position_varying
= varyings_p
;
1941 if (panfrost_writes_point_size(ctx
)) {
1942 varyings_p
= panfrost_emit_varyings(batch
,
1943 &varyings
[gl_PointSize
],
1945 primitive_size
->pointer
= varyings_p
;
1948 if (gl_PointCoord
>= 0)
1949 varyings
[gl_PointCoord
].elements
= MALI_VARYING_POINT_COORD
;
1951 if (gl_FrontFacing
>= 0)
1952 varyings
[gl_FrontFacing
].elements
= MALI_VARYING_FRONT_FACING
;
1954 if (gl_FragCoord
>= 0)
1955 varyings
[gl_FragCoord
].elements
= MALI_VARYING_FRAG_COORD
;
1957 assert(!(device
->quirks
& IS_BIFROST
) || !(reads_point_coord
));
1959 /* Let's go ahead and link varying meta to the buffer in question, now
1960 * that that information is available. VARYING_SLOT_POS is mapped to
1961 * gl_FragCoord for fragment shaders but gl_Positionf or vertex shaders
1964 panfrost_emit_varying_meta(trans
.cpu
, vs
, general
, gl_Position
,
1965 gl_PointSize
, gl_PointCoord
,
1968 panfrost_emit_varying_meta(trans
.cpu
+ vs_size
, fs
, general
,
1969 gl_FragCoord
, gl_PointSize
,
1970 gl_PointCoord
, gl_FrontFacing
);
1972 /* Replace streamout */
1974 struct mali_attr_meta
*ovs
= (struct mali_attr_meta
*)trans
.cpu
;
1975 struct mali_attr_meta
*ofs
= ovs
+ vs
->varying_count
;
1977 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1978 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1980 /* If we write gl_PointSize from the vertex shader but don't
1981 * consume it, no memory will be allocated for it, so if we
1982 * attempted to write anyway we would dereference a NULL
1983 * pointer on the GPU. Midgard seems fine with this; Bifrost
1986 if (loc
== VARYING_SLOT_PSIZ
&& !panfrost_writes_point_size(ctx
))
1987 ovs
[i
].format
= MALI_VARYING_DISCARD
;
1989 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1993 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1994 ovs
[i
].index
= o
->output_buffer
;
1996 assert(o
->stream
== 0);
1997 ovs
[i
].format
= (vs
->varyings
[i
].format
& ~MALI_NR_CHANNELS(4))
1998 | MALI_NR_CHANNELS(o
->num_components
);
2000 if (device
->quirks
& HAS_SWIZZLES
)
2001 ovs
[i
].swizzle
= panfrost_get_default_swizzle(o
->num_components
);
2003 ovs
[i
].swizzle
= panfrost_bifrost_swizzle(o
->num_components
);
2005 /* Link to the fragment */
2009 for (unsigned j
= 0; j
< fs
->varying_count
; ++j
) {
2010 if (fs
->varyings_loc
[j
] == loc
) {
2017 ofs
[fs_idx
].index
= ovs
[i
].index
;
2018 ofs
[fs_idx
].format
= ovs
[i
].format
;
2019 ofs
[fs_idx
].swizzle
= ovs
[i
].swizzle
;
2023 /* Replace point sprite */
2024 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
2025 /* If we have a point sprite replacement, handle that here. We
2026 * have to translate location first. TODO: Flip y in shader.
2027 * We're already keying ... just time crunch .. */
2029 if (has_point_coord(fs
->point_sprite_mask
,
2030 fs
->varyings_loc
[i
])) {
2031 ofs
[i
].index
= gl_PointCoord
;
2033 /* Swizzle out the z/w to 0/1 */
2034 ofs
[i
].format
= MALI_RG16F
;
2035 ofs
[i
].swizzle
= panfrost_get_default_swizzle(2);
2039 /* Fix up unaligned addresses */
2040 for (unsigned i
= 0; i
< so_count
; ++i
) {
2041 if (varyings
[i
].elements
< MALI_RECORD_SPECIAL
)
2044 unsigned align
= (varyings
[i
].elements
& 63);
2046 /* While we're at it, the SO buffers are linear */
2049 varyings
[i
].elements
|= MALI_ATTR_LINEAR
;
2053 /* We need to adjust alignment */
2054 varyings
[i
].elements
&= ~63;
2055 varyings
[i
].elements
|= MALI_ATTR_LINEAR
;
2056 varyings
[i
].size
+= align
;
2058 for (unsigned v
= 0; v
< vs
->varying_count
; ++v
) {
2059 if (ovs
[v
].index
!= i
)
2062 ovs
[v
].src_offset
= vs
->varyings
[v
].src_offset
+ align
;
2065 for (unsigned f
= 0; f
< fs
->varying_count
; ++f
) {
2066 if (ofs
[f
].index
!= i
)
2069 ofs
[f
].src_offset
= fs
->varyings
[f
].src_offset
+ align
;
2073 varyings_p
= panfrost_upload_transient(batch
, varyings
,
2074 idx
* sizeof(*varyings
));
2075 vertex_postfix
->varyings
= varyings_p
;
2076 tiler_postfix
->varyings
= varyings_p
;
2078 vertex_postfix
->varying_meta
= trans
.gpu
;
2079 tiler_postfix
->varying_meta
= trans
.gpu
+ vs_size
;
2083 panfrost_emit_vertex_tiler_jobs(struct panfrost_batch
*batch
,
2084 struct mali_vertex_tiler_prefix
*vertex_prefix
,
2085 struct mali_vertex_tiler_postfix
*vertex_postfix
,
2086 struct mali_vertex_tiler_prefix
*tiler_prefix
,
2087 struct mali_vertex_tiler_postfix
*tiler_postfix
,
2088 union midgard_primitive_size
*primitive_size
)
2090 struct panfrost_context
*ctx
= batch
->ctx
;
2091 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
2092 bool wallpapering
= ctx
->wallpaper_batch
&& batch
->tiler_dep
;
2093 struct bifrost_payload_vertex bifrost_vertex
= {0,};
2094 struct bifrost_payload_tiler bifrost_tiler
= {0,};
2095 struct midgard_payload_vertex_tiler midgard_vertex
= {0,};
2096 struct midgard_payload_vertex_tiler midgard_tiler
= {0,};
2098 size_t vp_size
, tp_size
;
2100 if (device
->quirks
& IS_BIFROST
) {
2101 bifrost_vertex
.prefix
= *vertex_prefix
;
2102 bifrost_vertex
.postfix
= *vertex_postfix
;
2103 vp
= &bifrost_vertex
;
2104 vp_size
= sizeof(bifrost_vertex
);
2106 bifrost_tiler
.prefix
= *tiler_prefix
;
2107 bifrost_tiler
.tiler
.primitive_size
= *primitive_size
;
2108 bifrost_tiler
.tiler
.tiler_meta
= panfrost_batch_get_tiler_meta(batch
, ~0);
2109 bifrost_tiler
.postfix
= *tiler_postfix
;
2110 tp
= &bifrost_tiler
;
2111 tp_size
= sizeof(bifrost_tiler
);
2113 midgard_vertex
.prefix
= *vertex_prefix
;
2114 midgard_vertex
.postfix
= *vertex_postfix
;
2115 vp
= &midgard_vertex
;
2116 vp_size
= sizeof(midgard_vertex
);
2118 midgard_tiler
.prefix
= *tiler_prefix
;
2119 midgard_tiler
.postfix
= *tiler_postfix
;
2120 midgard_tiler
.primitive_size
= *primitive_size
;
2121 tp
= &midgard_tiler
;
2122 tp_size
= sizeof(midgard_tiler
);
2126 /* Inject in reverse order, with "predicted" job indices.
2127 * THIS IS A HACK XXX */
2128 panfrost_new_job(batch
, JOB_TYPE_TILER
, false,
2129 batch
->job_index
+ 2, tp
, tp_size
, true);
2130 panfrost_new_job(batch
, JOB_TYPE_VERTEX
, false, 0,
2135 /* If rasterizer discard is enable, only submit the vertex */
2137 bool rasterizer_discard
= ctx
->rasterizer
&&
2138 ctx
->rasterizer
->base
.rasterizer_discard
;
2140 unsigned vertex
= panfrost_new_job(batch
, JOB_TYPE_VERTEX
, false, 0,
2141 vp
, vp_size
, false);
2143 if (rasterizer_discard
)
2146 panfrost_new_job(batch
, JOB_TYPE_TILER
, false, vertex
, tp
, tp_size
,
2150 /* TODO: stop hardcoding this */
2152 panfrost_emit_sample_locations(struct panfrost_batch
*batch
)
2154 uint16_t locations
[] = {
2205 return panfrost_upload_transient(batch
, locations
, 96 * sizeof(uint16_t));