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
= 0x958020;
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 meta
->bifrost2
.uniform_count
= MIN2(ss
->uniform_count
,
346 meta
->midgard1
.uniform_count
= MIN2(ss
->uniform_count
,
348 meta
->midgard1
.work_count
= ss
->work_reg_count
;
350 /* TODO: This is not conformant on ES3 */
351 meta
->midgard1
.flags_hi
= MALI_SUPPRESS_INF_NAN
;
353 meta
->midgard1
.flags_lo
= 0x220;
354 meta
->midgard1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
359 panfrost_translate_compare_func(enum pipe_compare_func in
)
362 case PIPE_FUNC_NEVER
:
363 return MALI_FUNC_NEVER
;
366 return MALI_FUNC_LESS
;
368 case PIPE_FUNC_EQUAL
:
369 return MALI_FUNC_EQUAL
;
371 case PIPE_FUNC_LEQUAL
:
372 return MALI_FUNC_LEQUAL
;
374 case PIPE_FUNC_GREATER
:
375 return MALI_FUNC_GREATER
;
377 case PIPE_FUNC_NOTEQUAL
:
378 return MALI_FUNC_NOTEQUAL
;
380 case PIPE_FUNC_GEQUAL
:
381 return MALI_FUNC_GEQUAL
;
383 case PIPE_FUNC_ALWAYS
:
384 return MALI_FUNC_ALWAYS
;
387 unreachable("Invalid func");
392 panfrost_translate_stencil_op(enum pipe_stencil_op in
)
395 case PIPE_STENCIL_OP_KEEP
:
396 return MALI_STENCIL_KEEP
;
398 case PIPE_STENCIL_OP_ZERO
:
399 return MALI_STENCIL_ZERO
;
401 case PIPE_STENCIL_OP_REPLACE
:
402 return MALI_STENCIL_REPLACE
;
404 case PIPE_STENCIL_OP_INCR
:
405 return MALI_STENCIL_INCR
;
407 case PIPE_STENCIL_OP_DECR
:
408 return MALI_STENCIL_DECR
;
410 case PIPE_STENCIL_OP_INCR_WRAP
:
411 return MALI_STENCIL_INCR_WRAP
;
413 case PIPE_STENCIL_OP_DECR_WRAP
:
414 return MALI_STENCIL_DECR_WRAP
;
416 case PIPE_STENCIL_OP_INVERT
:
417 return MALI_STENCIL_INVERT
;
420 unreachable("Invalid stencil op");
425 translate_tex_wrap(enum pipe_tex_wrap w
)
428 case PIPE_TEX_WRAP_REPEAT
:
429 return MALI_WRAP_REPEAT
;
431 case PIPE_TEX_WRAP_CLAMP
:
432 return MALI_WRAP_CLAMP
;
434 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
435 return MALI_WRAP_CLAMP_TO_EDGE
;
437 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
438 return MALI_WRAP_CLAMP_TO_BORDER
;
440 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
441 return MALI_WRAP_MIRRORED_REPEAT
;
443 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
444 return MALI_WRAP_MIRRORED_CLAMP
;
446 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
447 return MALI_WRAP_MIRRORED_CLAMP_TO_EDGE
;
449 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
450 return MALI_WRAP_MIRRORED_CLAMP_TO_BORDER
;
453 unreachable("Invalid wrap");
457 void panfrost_sampler_desc_init(const struct pipe_sampler_state
*cso
,
458 struct mali_sampler_descriptor
*hw
)
460 unsigned func
= panfrost_translate_compare_func(cso
->compare_func
);
461 bool min_nearest
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
;
462 bool mag_nearest
= cso
->mag_img_filter
== PIPE_TEX_FILTER_NEAREST
;
463 bool mip_linear
= cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
;
464 unsigned min_filter
= min_nearest
? MALI_SAMP_MIN_NEAREST
: 0;
465 unsigned mag_filter
= mag_nearest
? MALI_SAMP_MAG_NEAREST
: 0;
466 unsigned mip_filter
= mip_linear
?
467 (MALI_SAMP_MIP_LINEAR_1
| MALI_SAMP_MIP_LINEAR_2
) : 0;
468 unsigned normalized
= cso
->normalized_coords
? MALI_SAMP_NORM_COORDS
: 0;
470 *hw
= (struct mali_sampler_descriptor
) {
471 .filter_mode
= min_filter
| mag_filter
| mip_filter
|
473 .wrap_s
= translate_tex_wrap(cso
->wrap_s
),
474 .wrap_t
= translate_tex_wrap(cso
->wrap_t
),
475 .wrap_r
= translate_tex_wrap(cso
->wrap_r
),
476 .compare_func
= panfrost_flip_compare_func(func
),
478 cso
->border_color
.f
[0],
479 cso
->border_color
.f
[1],
480 cso
->border_color
.f
[2],
481 cso
->border_color
.f
[3]
483 .min_lod
= FIXED_16(cso
->min_lod
, false), /* clamp at 0 */
484 .max_lod
= FIXED_16(cso
->max_lod
, false),
485 .lod_bias
= FIXED_16(cso
->lod_bias
, true), /* can be negative */
486 .seamless_cube_map
= cso
->seamless_cube_map
,
489 /* If necessary, we disable mipmapping in the sampler descriptor by
490 * clamping the LOD as tight as possible (from 0 to epsilon,
491 * essentially -- remember these are fixed point numbers, so
494 if (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
495 hw
->max_lod
= hw
->min_lod
+ 1;
498 void panfrost_sampler_desc_init_bifrost(const struct pipe_sampler_state
*cso
,
499 struct bifrost_sampler_descriptor
*hw
)
501 *hw
= (struct bifrost_sampler_descriptor
) {
503 .wrap_s
= translate_tex_wrap(cso
->wrap_s
),
504 .wrap_t
= translate_tex_wrap(cso
->wrap_t
),
505 .wrap_r
= translate_tex_wrap(cso
->wrap_r
),
507 .min_filter
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
,
508 .norm_coords
= cso
->normalized_coords
,
509 .mip_filter
= cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
,
510 .mag_filter
= cso
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
,
511 .min_lod
= FIXED_16(cso
->min_lod
, false), /* clamp at 0 */
512 .max_lod
= FIXED_16(cso
->max_lod
, false),
515 /* If necessary, we disable mipmapping in the sampler descriptor by
516 * clamping the LOD as tight as possible (from 0 to epsilon,
517 * essentially -- remember these are fixed point numbers, so
520 if (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
521 hw
->max_lod
= hw
->min_lod
+ 1;
525 panfrost_make_stencil_state(const struct pipe_stencil_state
*in
,
526 struct mali_stencil_test
*out
)
528 out
->ref
= 0; /* Gallium gets it from elsewhere */
530 out
->mask
= in
->valuemask
;
531 out
->func
= panfrost_translate_compare_func(in
->func
);
532 out
->sfail
= panfrost_translate_stencil_op(in
->fail_op
);
533 out
->dpfail
= panfrost_translate_stencil_op(in
->zfail_op
);
534 out
->dppass
= panfrost_translate_stencil_op(in
->zpass_op
);
538 panfrost_frag_meta_rasterizer_update(struct panfrost_context
*ctx
,
539 struct mali_shader_meta
*fragmeta
)
541 if (!ctx
->rasterizer
) {
542 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, true);
543 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, false);
544 fragmeta
->depth_units
= 0.0f
;
545 fragmeta
->depth_factor
= 0.0f
;
546 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
, false);
547 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
, false);
551 bool msaa
= ctx
->rasterizer
->base
.multisample
;
553 /* TODO: Sample size */
554 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, msaa
);
555 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, !msaa
);
556 fragmeta
->depth_units
= ctx
->rasterizer
->base
.offset_units
* 2.0f
;
557 fragmeta
->depth_factor
= ctx
->rasterizer
->base
.offset_scale
;
559 /* XXX: Which bit is which? Does this maybe allow offseting not-tri? */
561 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
,
562 ctx
->rasterizer
->base
.offset_tri
);
563 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
,
564 ctx
->rasterizer
->base
.offset_tri
);
568 panfrost_frag_meta_zsa_update(struct panfrost_context
*ctx
,
569 struct mali_shader_meta
*fragmeta
)
571 const struct pipe_depth_stencil_alpha_state
*zsa
= ctx
->depth_stencil
;
572 int zfunc
= PIPE_FUNC_ALWAYS
;
575 struct pipe_stencil_state default_stencil
= {
577 .func
= PIPE_FUNC_ALWAYS
,
578 .fail_op
= MALI_STENCIL_KEEP
,
579 .zfail_op
= MALI_STENCIL_KEEP
,
580 .zpass_op
= MALI_STENCIL_KEEP
,
585 panfrost_make_stencil_state(&default_stencil
,
586 &fragmeta
->stencil_front
);
587 fragmeta
->stencil_mask_front
= default_stencil
.writemask
;
588 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
589 fragmeta
->stencil_mask_back
= default_stencil
.writemask
;
590 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
, false);
591 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
, false);
593 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
,
594 zsa
->stencil
[0].enabled
);
595 panfrost_make_stencil_state(&zsa
->stencil
[0],
596 &fragmeta
->stencil_front
);
597 fragmeta
->stencil_mask_front
= zsa
->stencil
[0].writemask
;
598 fragmeta
->stencil_front
.ref
= ctx
->stencil_ref
.ref_value
[0];
600 /* If back-stencil is not enabled, use the front values */
602 if (zsa
->stencil
[1].enabled
) {
603 panfrost_make_stencil_state(&zsa
->stencil
[1],
604 &fragmeta
->stencil_back
);
605 fragmeta
->stencil_mask_back
= zsa
->stencil
[1].writemask
;
606 fragmeta
->stencil_back
.ref
= ctx
->stencil_ref
.ref_value
[1];
608 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
609 fragmeta
->stencil_mask_back
= fragmeta
->stencil_mask_front
;
610 fragmeta
->stencil_back
.ref
= fragmeta
->stencil_front
.ref
;
613 if (zsa
->depth
.enabled
)
614 zfunc
= zsa
->depth
.func
;
616 /* Depth state (TODO: Refactor) */
618 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
,
619 zsa
->depth
.writemask
);
622 fragmeta
->unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
623 fragmeta
->unknown2_3
|= MALI_DEPTH_FUNC(panfrost_translate_compare_func(zfunc
));
627 panfrost_fs_required(
628 struct panfrost_shader_state
*fs
,
629 struct panfrost_blend_final
*blend
,
632 /* If we generally have side effects */
636 /* If colour is written we need to execute */
637 for (unsigned i
= 0; i
< rt_count
; ++i
) {
638 if (!blend
[i
].no_colour
)
642 /* If depth is written and not implied we need to execute.
643 * TODO: Predicate on Z/S writes being enabled */
644 return (fs
->writes_depth
|| fs
->writes_stencil
);
648 panfrost_frag_meta_blend_update(struct panfrost_context
*ctx
,
649 struct mali_shader_meta
*fragmeta
,
652 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
653 struct panfrost_shader_state
*fs
;
654 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
656 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_DITHER
,
657 (dev
->quirks
& MIDGARD_SFBD
) && ctx
->blend
&&
658 !ctx
->blend
->base
.dither
);
660 /* Get blending setup */
661 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
663 struct panfrost_blend_final blend
[PIPE_MAX_COLOR_BUFS
];
664 unsigned shader_offset
= 0;
665 struct panfrost_bo
*shader_bo
= NULL
;
667 for (unsigned c
= 0; c
< rt_count
; ++c
)
668 blend
[c
] = panfrost_get_blend_for_context(ctx
, c
, &shader_bo
,
671 /* Disable shader execution if we can */
672 if (dev
->quirks
& MIDGARD_SHADERLESS
673 && !panfrost_fs_required(fs
, blend
, rt_count
)) {
674 fragmeta
->shader
= 0;
675 fragmeta
->attribute_count
= 0;
676 fragmeta
->varying_count
= 0;
677 fragmeta
->texture_count
= 0;
678 fragmeta
->sampler_count
= 0;
680 /* This feature is not known to work on Bifrost */
681 fragmeta
->midgard1
.work_count
= 1;
682 fragmeta
->midgard1
.uniform_count
= 0;
683 fragmeta
->midgard1
.uniform_buffer_count
= 0;
686 /* If there is a blend shader, work registers are shared. We impose 8
687 * work registers as a limit for blend shaders. Should be lower XXX */
689 if (!(dev
->quirks
& IS_BIFROST
)) {
690 for (unsigned c
= 0; c
< rt_count
; ++c
) {
691 if (blend
[c
].is_shader
) {
692 fragmeta
->midgard1
.work_count
=
693 MAX2(fragmeta
->midgard1
.work_count
, 8);
698 /* Even on MFBD, the shader descriptor gets blend shaders. It's *also*
699 * copied to the blend_meta appended (by convention), but this is the
700 * field actually read by the hardware. (Or maybe both are read...?).
701 * Specify the last RTi with a blend shader. */
703 fragmeta
->blend
.shader
= 0;
705 for (signed rt
= (rt_count
- 1); rt
>= 0; --rt
) {
706 if (!blend
[rt
].is_shader
)
709 fragmeta
->blend
.shader
= blend
[rt
].shader
.gpu
|
710 blend
[rt
].shader
.first_tag
;
714 if (dev
->quirks
& MIDGARD_SFBD
) {
715 /* When only a single render target platform is used, the blend
716 * information is inside the shader meta itself. We additionally
717 * need to signal CAN_DISCARD for nontrivial blend modes (so
718 * we're able to read back the destination buffer) */
720 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_BLEND_SHADER
,
723 if (!blend
[0].is_shader
) {
724 fragmeta
->blend
.equation
= *blend
[0].equation
.equation
;
725 fragmeta
->blend
.constant
= blend
[0].equation
.constant
;
728 SET_BIT(fragmeta
->unknown2_3
, MALI_CAN_DISCARD
,
729 !blend
[0].no_blending
|| fs
->can_discard
);
733 /* Additional blend descriptor tacked on for jobs using MFBD */
735 for (unsigned i
= 0; i
< rt_count
; ++i
) {
738 if (ctx
->pipe_framebuffer
.nr_cbufs
> i
&& !blend
[i
].no_colour
) {
741 bool is_srgb
= (ctx
->pipe_framebuffer
.nr_cbufs
> i
) &&
742 (ctx
->pipe_framebuffer
.cbufs
[i
]) &&
743 util_format_is_srgb(ctx
->pipe_framebuffer
.cbufs
[i
]->format
);
745 SET_BIT(flags
, MALI_BLEND_MRT_SHADER
, blend
[i
].is_shader
);
746 SET_BIT(flags
, MALI_BLEND_LOAD_TIB
, !blend
[i
].no_blending
);
747 SET_BIT(flags
, MALI_BLEND_SRGB
, is_srgb
);
748 SET_BIT(flags
, MALI_BLEND_NO_DITHER
, !ctx
->blend
->base
.dither
);
751 if (dev
->quirks
& IS_BIFROST
) {
752 struct bifrost_blend_rt
*brts
= rts
;
754 brts
[i
].flags
= flags
;
756 if (blend
[i
].is_shader
) {
757 /* The blend shader's address needs to be at
758 * the same top 32 bit as the fragment shader.
759 * TODO: Ensure that's always the case.
761 assert((blend
[i
].shader
.gpu
& (0xffffffffull
<< 32)) ==
762 (fs
->bo
->gpu
& (0xffffffffull
<< 32)));
763 brts
[i
].shader
= blend
[i
].shader
.gpu
;
765 } else if (ctx
->pipe_framebuffer
.nr_cbufs
> i
) {
766 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
767 const struct util_format_description
*format_desc
;
768 format_desc
= util_format_description(format
);
770 brts
[i
].equation
= *blend
[i
].equation
.equation
;
772 /* TODO: this is a bit more complicated */
773 brts
[i
].constant
= blend
[i
].equation
.constant
;
775 brts
[i
].format
= panfrost_format_to_bifrost_blend(format_desc
);
777 /* 0x19 disables blending and forces REPLACE
778 * mode (equivalent to rgb_mode = alpha_mode =
779 * x122, colour mask = 0xF). 0x1a allows
781 brts
[i
].unk2
= blend
[i
].no_blending
? 0x19 : 0x1a;
783 brts
[i
].shader_type
= fs
->blend_types
[i
];
785 /* Dummy attachment for depth-only */
787 brts
[i
].shader_type
= fs
->blend_types
[i
];
790 struct midgard_blend_rt
*mrts
= rts
;
791 mrts
[i
].flags
= flags
;
793 if (blend
[i
].is_shader
) {
794 mrts
[i
].blend
.shader
= blend
[i
].shader
.gpu
| blend
[i
].shader
.first_tag
;
796 mrts
[i
].blend
.equation
= *blend
[i
].equation
.equation
;
797 mrts
[i
].blend
.constant
= blend
[i
].equation
.constant
;
804 panfrost_frag_shader_meta_init(struct panfrost_context
*ctx
,
805 struct mali_shader_meta
*fragmeta
,
808 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
809 struct panfrost_shader_state
*fs
;
811 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
813 fragmeta
->alpha_coverage
= ~MALI_ALPHA_COVERAGE(0.000000);
814 fragmeta
->unknown2_3
= MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS
) | 0x3010;
815 fragmeta
->unknown2_4
= 0x4e0;
817 /* unknown2_4 has 0x10 bit set on T6XX and T720. We don't know why this
818 * is required (independent of 32-bit/64-bit descriptors), or why it's
819 * not used on later GPU revisions. Otherwise, all shader jobs fault on
820 * these earlier chips (perhaps this is a chicken bit of some kind).
821 * More investigation is needed. */
823 SET_BIT(fragmeta
->unknown2_4
, 0x10, dev
->quirks
& MIDGARD_SFBD
);
825 if (dev
->quirks
& IS_BIFROST
) {
828 /* Depending on whether it's legal to in the given shader, we try to
829 * enable early-z testing (or forward-pixel kill?) */
831 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_EARLY_Z
,
832 !fs
->can_discard
&& !fs
->writes_depth
);
834 /* Add the writes Z/S flags if needed. */
835 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_WRITES_Z
, fs
->writes_depth
);
836 SET_BIT(fragmeta
->midgard1
.flags_hi
, MALI_WRITES_S
, fs
->writes_stencil
);
838 /* Any time texturing is used, derivatives are implicitly calculated,
839 * so we need to enable helper invocations */
841 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_HELPER_INVOCATIONS
,
842 fs
->helper_invocations
);
844 const struct pipe_depth_stencil_alpha_state
*zsa
= ctx
->depth_stencil
;
846 bool depth_enabled
= fs
->writes_depth
||
847 (zsa
&& zsa
->depth
.enabled
&& zsa
->depth
.func
!= PIPE_FUNC_ALWAYS
);
849 SET_BIT(fragmeta
->midgard1
.flags_lo
, 0x400, !depth_enabled
&& fs
->can_discard
);
850 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_READS_ZS
, depth_enabled
&& fs
->can_discard
);
853 panfrost_frag_meta_rasterizer_update(ctx
, fragmeta
);
854 panfrost_frag_meta_zsa_update(ctx
, fragmeta
);
855 panfrost_frag_meta_blend_update(ctx
, fragmeta
, rts
);
859 panfrost_emit_shader_meta(struct panfrost_batch
*batch
,
860 enum pipe_shader_type st
,
861 struct mali_vertex_tiler_postfix
*postfix
)
863 struct panfrost_context
*ctx
= batch
->ctx
;
864 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
871 struct mali_shader_meta meta
;
873 panfrost_shader_meta_init(ctx
, st
, &meta
);
875 /* Add the shader BO to the batch. */
876 panfrost_batch_add_bo(batch
, ss
->bo
,
877 PAN_BO_ACCESS_PRIVATE
|
879 panfrost_bo_access_for_stage(st
));
883 if (st
== PIPE_SHADER_FRAGMENT
) {
884 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
885 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
886 size_t desc_size
= sizeof(meta
);
888 struct panfrost_transfer xfer
;
891 if (dev
->quirks
& MIDGARD_SFBD
)
893 else if (dev
->quirks
& IS_BIFROST
)
894 rt_size
= sizeof(struct bifrost_blend_rt
);
896 rt_size
= sizeof(struct midgard_blend_rt
);
898 desc_size
+= rt_size
* rt_count
;
901 rts
= rzalloc_size(ctx
, rt_size
* rt_count
);
903 panfrost_frag_shader_meta_init(ctx
, &meta
, rts
);
905 xfer
= panfrost_allocate_transient(batch
, desc_size
);
907 memcpy(xfer
.cpu
, &meta
, sizeof(meta
));
908 memcpy(xfer
.cpu
+ sizeof(meta
), rts
, rt_size
* rt_count
);
913 shader_ptr
= xfer
.gpu
;
915 shader_ptr
= panfrost_upload_transient(batch
, &meta
,
919 postfix
->shader
= shader_ptr
;
923 panfrost_mali_viewport_init(struct panfrost_context
*ctx
,
924 struct mali_viewport
*mvp
)
926 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
928 /* Clip bounds are encoded as floats. The viewport itself is encoded as
929 * (somewhat) asymmetric ints. */
931 const struct pipe_scissor_state
*ss
= &ctx
->scissor
;
933 memset(mvp
, 0, sizeof(*mvp
));
935 /* By default, do no viewport clipping, i.e. clip to (-inf, inf) in
936 * each direction. Clipping to the viewport in theory should work, but
937 * in practice causes issues when we're not explicitly trying to
940 *mvp
= (struct mali_viewport
) {
941 .clip_minx
= -INFINITY
,
942 .clip_miny
= -INFINITY
,
943 .clip_maxx
= INFINITY
,
944 .clip_maxy
= INFINITY
,
947 /* Always scissor to the viewport by default. */
948 float vp_minx
= (int) (vp
->translate
[0] - fabsf(vp
->scale
[0]));
949 float vp_maxx
= (int) (vp
->translate
[0] + fabsf(vp
->scale
[0]));
951 float vp_miny
= (int) (vp
->translate
[1] - fabsf(vp
->scale
[1]));
952 float vp_maxy
= (int) (vp
->translate
[1] + fabsf(vp
->scale
[1]));
954 float minz
= (vp
->translate
[2] - fabsf(vp
->scale
[2]));
955 float maxz
= (vp
->translate
[2] + fabsf(vp
->scale
[2]));
957 /* Apply the scissor test */
959 unsigned minx
, miny
, maxx
, maxy
;
961 if (ss
&& ctx
->rasterizer
&& ctx
->rasterizer
->base
.scissor
) {
962 minx
= MAX2(ss
->minx
, vp_minx
);
963 miny
= MAX2(ss
->miny
, vp_miny
);
964 maxx
= MIN2(ss
->maxx
, vp_maxx
);
965 maxy
= MIN2(ss
->maxy
, vp_maxy
);
973 /* Hardware needs the min/max to be strictly ordered, so flip if we
974 * need to. The viewport transformation in the vertex shader will
975 * handle the negatives if we don't */
978 unsigned temp
= miny
;
984 unsigned temp
= minx
;
995 /* Clamp to the framebuffer size as a last check */
997 minx
= MIN2(ctx
->pipe_framebuffer
.width
, minx
);
998 maxx
= MIN2(ctx
->pipe_framebuffer
.width
, maxx
);
1000 miny
= MIN2(ctx
->pipe_framebuffer
.height
, miny
);
1001 maxy
= MIN2(ctx
->pipe_framebuffer
.height
, maxy
);
1005 mvp
->viewport0
[0] = minx
;
1006 mvp
->viewport1
[0] = MALI_POSITIVE(maxx
);
1008 mvp
->viewport0
[1] = miny
;
1009 mvp
->viewport1
[1] = MALI_POSITIVE(maxy
);
1011 mvp
->clip_minz
= minz
;
1012 mvp
->clip_maxz
= maxz
;
1016 panfrost_emit_viewport(struct panfrost_batch
*batch
,
1017 struct mali_vertex_tiler_postfix
*tiler_postfix
)
1019 struct panfrost_context
*ctx
= batch
->ctx
;
1020 struct mali_viewport mvp
;
1022 panfrost_mali_viewport_init(batch
->ctx
, &mvp
);
1024 /* Update the job, unless we're doing wallpapering (whose lack of
1025 * scissor we can ignore, since if we "miss" a tile of wallpaper, it'll
1026 * just... be faster :) */
1028 if (!ctx
->wallpaper_batch
)
1029 panfrost_batch_union_scissor(batch
, mvp
.viewport0
[0],
1031 mvp
.viewport1
[0] + 1,
1032 mvp
.viewport1
[1] + 1);
1034 tiler_postfix
->viewport
= panfrost_upload_transient(batch
, &mvp
,
1039 panfrost_map_constant_buffer_gpu(struct panfrost_batch
*batch
,
1040 enum pipe_shader_type st
,
1041 struct panfrost_constant_buffer
*buf
,
1044 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1045 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1048 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1049 PAN_BO_ACCESS_SHARED
|
1050 PAN_BO_ACCESS_READ
|
1051 panfrost_bo_access_for_stage(st
));
1053 /* Alignment gauranteed by
1054 * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
1055 return rsrc
->bo
->gpu
+ cb
->buffer_offset
;
1056 } else if (cb
->user_buffer
) {
1057 return panfrost_upload_transient(batch
,
1062 unreachable("No constant buffer");
1066 struct sysval_uniform
{
1076 panfrost_upload_viewport_scale_sysval(struct panfrost_batch
*batch
,
1077 struct sysval_uniform
*uniform
)
1079 struct panfrost_context
*ctx
= batch
->ctx
;
1080 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
1082 uniform
->f
[0] = vp
->scale
[0];
1083 uniform
->f
[1] = vp
->scale
[1];
1084 uniform
->f
[2] = vp
->scale
[2];
1088 panfrost_upload_viewport_offset_sysval(struct panfrost_batch
*batch
,
1089 struct sysval_uniform
*uniform
)
1091 struct panfrost_context
*ctx
= batch
->ctx
;
1092 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
1094 uniform
->f
[0] = vp
->translate
[0];
1095 uniform
->f
[1] = vp
->translate
[1];
1096 uniform
->f
[2] = vp
->translate
[2];
1099 static void panfrost_upload_txs_sysval(struct panfrost_batch
*batch
,
1100 enum pipe_shader_type st
,
1101 unsigned int sysvalid
,
1102 struct sysval_uniform
*uniform
)
1104 struct panfrost_context
*ctx
= batch
->ctx
;
1105 unsigned texidx
= PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid
);
1106 unsigned dim
= PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid
);
1107 bool is_array
= PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid
);
1108 struct pipe_sampler_view
*tex
= &ctx
->sampler_views
[st
][texidx
]->base
;
1111 uniform
->i
[0] = u_minify(tex
->texture
->width0
, tex
->u
.tex
.first_level
);
1114 uniform
->i
[1] = u_minify(tex
->texture
->height0
,
1115 tex
->u
.tex
.first_level
);
1118 uniform
->i
[2] = u_minify(tex
->texture
->depth0
,
1119 tex
->u
.tex
.first_level
);
1122 uniform
->i
[dim
] = tex
->texture
->array_size
;
1126 panfrost_upload_ssbo_sysval(struct panfrost_batch
*batch
,
1127 enum pipe_shader_type st
,
1129 struct sysval_uniform
*uniform
)
1131 struct panfrost_context
*ctx
= batch
->ctx
;
1133 assert(ctx
->ssbo_mask
[st
] & (1 << ssbo_id
));
1134 struct pipe_shader_buffer sb
= ctx
->ssbo
[st
][ssbo_id
];
1136 /* Compute address */
1137 struct panfrost_bo
*bo
= pan_resource(sb
.buffer
)->bo
;
1139 panfrost_batch_add_bo(batch
, bo
,
1140 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_RW
|
1141 panfrost_bo_access_for_stage(st
));
1143 /* Upload address and size as sysval */
1144 uniform
->du
[0] = bo
->gpu
+ sb
.buffer_offset
;
1145 uniform
->u
[2] = sb
.buffer_size
;
1149 panfrost_upload_sampler_sysval(struct panfrost_batch
*batch
,
1150 enum pipe_shader_type st
,
1152 struct sysval_uniform
*uniform
)
1154 struct panfrost_context
*ctx
= batch
->ctx
;
1155 struct pipe_sampler_state
*sampl
= &ctx
->samplers
[st
][samp_idx
]->base
;
1157 uniform
->f
[0] = sampl
->min_lod
;
1158 uniform
->f
[1] = sampl
->max_lod
;
1159 uniform
->f
[2] = sampl
->lod_bias
;
1161 /* Even without any errata, Midgard represents "no mipmapping" as
1162 * fixing the LOD with the clamps; keep behaviour consistent. c.f.
1163 * panfrost_create_sampler_state which also explains our choice of
1164 * epsilon value (again to keep behaviour consistent) */
1166 if (sampl
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
1167 uniform
->f
[1] = uniform
->f
[0] + (1.0/256.0);
1171 panfrost_upload_num_work_groups_sysval(struct panfrost_batch
*batch
,
1172 struct sysval_uniform
*uniform
)
1174 struct panfrost_context
*ctx
= batch
->ctx
;
1176 uniform
->u
[0] = ctx
->compute_grid
->grid
[0];
1177 uniform
->u
[1] = ctx
->compute_grid
->grid
[1];
1178 uniform
->u
[2] = ctx
->compute_grid
->grid
[2];
1182 panfrost_upload_sysvals(struct panfrost_batch
*batch
, void *buf
,
1183 struct panfrost_shader_state
*ss
,
1184 enum pipe_shader_type st
)
1186 struct sysval_uniform
*uniforms
= (void *)buf
;
1188 for (unsigned i
= 0; i
< ss
->sysval_count
; ++i
) {
1189 int sysval
= ss
->sysval
[i
];
1191 switch (PAN_SYSVAL_TYPE(sysval
)) {
1192 case PAN_SYSVAL_VIEWPORT_SCALE
:
1193 panfrost_upload_viewport_scale_sysval(batch
,
1196 case PAN_SYSVAL_VIEWPORT_OFFSET
:
1197 panfrost_upload_viewport_offset_sysval(batch
,
1200 case PAN_SYSVAL_TEXTURE_SIZE
:
1201 panfrost_upload_txs_sysval(batch
, st
,
1202 PAN_SYSVAL_ID(sysval
),
1205 case PAN_SYSVAL_SSBO
:
1206 panfrost_upload_ssbo_sysval(batch
, st
,
1207 PAN_SYSVAL_ID(sysval
),
1210 case PAN_SYSVAL_NUM_WORK_GROUPS
:
1211 panfrost_upload_num_work_groups_sysval(batch
,
1214 case PAN_SYSVAL_SAMPLER
:
1215 panfrost_upload_sampler_sysval(batch
, st
,
1216 PAN_SYSVAL_ID(sysval
),
1226 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer
*buf
,
1229 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1230 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1233 return rsrc
->bo
->cpu
;
1234 else if (cb
->user_buffer
)
1235 return cb
->user_buffer
;
1237 unreachable("No constant buffer");
1241 panfrost_emit_const_buf(struct panfrost_batch
*batch
,
1242 enum pipe_shader_type stage
,
1243 struct mali_vertex_tiler_postfix
*postfix
)
1245 struct panfrost_context
*ctx
= batch
->ctx
;
1246 struct panfrost_shader_variants
*all
= ctx
->shader
[stage
];
1251 struct panfrost_constant_buffer
*buf
= &ctx
->constant_buffer
[stage
];
1253 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1255 /* Uniforms are implicitly UBO #0 */
1256 bool has_uniforms
= buf
->enabled_mask
& (1 << 0);
1258 /* Allocate room for the sysval and the uniforms */
1259 size_t sys_size
= sizeof(float) * 4 * ss
->sysval_count
;
1260 size_t uniform_size
= has_uniforms
? (buf
->cb
[0].buffer_size
) : 0;
1261 size_t size
= sys_size
+ uniform_size
;
1262 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1265 /* Upload sysvals requested by the shader */
1266 panfrost_upload_sysvals(batch
, transfer
.cpu
, ss
, stage
);
1268 /* Upload uniforms */
1269 if (has_uniforms
&& uniform_size
) {
1270 const void *cpu
= panfrost_map_constant_buffer_cpu(buf
, 0);
1271 memcpy(transfer
.cpu
+ sys_size
, cpu
, uniform_size
);
1274 /* Next up, attach UBOs. UBO #0 is the uniforms we just
1277 unsigned ubo_count
= panfrost_ubo_count(ctx
, stage
);
1278 assert(ubo_count
>= 1);
1280 size_t sz
= sizeof(uint64_t) * ubo_count
;
1281 uint64_t ubos
[PAN_MAX_CONST_BUFFERS
];
1282 int uniform_count
= ss
->uniform_count
;
1284 /* Upload uniforms as a UBO */
1285 ubos
[0] = MALI_MAKE_UBO(2 + uniform_count
, transfer
.gpu
);
1287 /* The rest are honest-to-goodness UBOs */
1289 for (unsigned ubo
= 1; ubo
< ubo_count
; ++ubo
) {
1290 size_t usz
= buf
->cb
[ubo
].buffer_size
;
1291 bool enabled
= buf
->enabled_mask
& (1 << ubo
);
1292 bool empty
= usz
== 0;
1294 if (!enabled
|| empty
) {
1295 /* Stub out disabled UBOs to catch accesses */
1296 ubos
[ubo
] = MALI_MAKE_UBO(0, 0xDEAD0000);
1300 mali_ptr gpu
= panfrost_map_constant_buffer_gpu(batch
, stage
,
1303 unsigned bytes_per_field
= 16;
1304 unsigned aligned
= ALIGN_POT(usz
, bytes_per_field
);
1305 ubos
[ubo
] = MALI_MAKE_UBO(aligned
/ bytes_per_field
, gpu
);
1308 mali_ptr ubufs
= panfrost_upload_transient(batch
, ubos
, sz
);
1309 postfix
->uniforms
= transfer
.gpu
;
1310 postfix
->uniform_buffers
= ubufs
;
1312 buf
->dirty_mask
= 0;
1316 panfrost_emit_shared_memory(struct panfrost_batch
*batch
,
1317 const struct pipe_grid_info
*info
,
1318 struct midgard_payload_vertex_tiler
*vtp
)
1320 struct panfrost_context
*ctx
= batch
->ctx
;
1321 struct panfrost_shader_variants
*all
= ctx
->shader
[PIPE_SHADER_COMPUTE
];
1322 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1323 unsigned single_size
= util_next_power_of_two(MAX2(ss
->shared_size
,
1325 unsigned shared_size
= single_size
* info
->grid
[0] * info
->grid
[1] *
1327 struct panfrost_bo
*bo
= panfrost_batch_get_shared_memory(batch
,
1331 struct mali_shared_memory shared
= {
1332 .shared_memory
= bo
->gpu
,
1333 .shared_workgroup_count
=
1334 util_logbase2_ceil(info
->grid
[0]) +
1335 util_logbase2_ceil(info
->grid
[1]) +
1336 util_logbase2_ceil(info
->grid
[2]),
1338 .shared_shift
= util_logbase2(single_size
) - 1
1341 vtp
->postfix
.shared_memory
= panfrost_upload_transient(batch
, &shared
,
1346 panfrost_get_tex_desc(struct panfrost_batch
*batch
,
1347 enum pipe_shader_type st
,
1348 struct panfrost_sampler_view
*view
)
1351 return (mali_ptr
) 0;
1353 struct pipe_sampler_view
*pview
= &view
->base
;
1354 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1356 /* Add the BO to the job so it's retained until the job is done. */
1358 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1359 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1360 panfrost_bo_access_for_stage(st
));
1362 panfrost_batch_add_bo(batch
, view
->midgard_bo
,
1363 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1364 panfrost_bo_access_for_stage(st
));
1366 return view
->midgard_bo
->gpu
;
1370 panfrost_emit_texture_descriptors(struct panfrost_batch
*batch
,
1371 enum pipe_shader_type stage
,
1372 struct mali_vertex_tiler_postfix
*postfix
)
1374 struct panfrost_context
*ctx
= batch
->ctx
;
1375 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1377 if (!ctx
->sampler_view_count
[stage
])
1380 if (device
->quirks
& IS_BIFROST
) {
1381 struct bifrost_texture_descriptor
*descriptors
;
1383 descriptors
= malloc(sizeof(struct bifrost_texture_descriptor
) *
1384 ctx
->sampler_view_count
[stage
]);
1386 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1387 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1388 struct pipe_sampler_view
*pview
= &view
->base
;
1389 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1391 /* Add the BOs to the job so they are retained until the job is done. */
1393 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1394 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1395 panfrost_bo_access_for_stage(stage
));
1397 panfrost_batch_add_bo(batch
, view
->bifrost_bo
,
1398 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1399 panfrost_bo_access_for_stage(stage
));
1401 memcpy(&descriptors
[i
], view
->bifrost_descriptor
, sizeof(*view
->bifrost_descriptor
));
1404 postfix
->textures
= panfrost_upload_transient(batch
,
1406 sizeof(struct bifrost_texture_descriptor
) *
1407 ctx
->sampler_view_count
[stage
]);
1411 uint64_t trampolines
[PIPE_MAX_SHADER_SAMPLER_VIEWS
];
1413 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
)
1414 trampolines
[i
] = panfrost_get_tex_desc(batch
, stage
,
1415 ctx
->sampler_views
[stage
][i
]);
1417 postfix
->textures
= panfrost_upload_transient(batch
,
1420 ctx
->sampler_view_count
[stage
]);
1425 panfrost_emit_sampler_descriptors(struct panfrost_batch
*batch
,
1426 enum pipe_shader_type stage
,
1427 struct mali_vertex_tiler_postfix
*postfix
)
1429 struct panfrost_context
*ctx
= batch
->ctx
;
1430 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1432 if (!ctx
->sampler_count
[stage
])
1435 if (device
->quirks
& IS_BIFROST
) {
1436 size_t desc_size
= sizeof(struct bifrost_sampler_descriptor
);
1437 size_t transfer_size
= desc_size
* ctx
->sampler_count
[stage
];
1438 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1440 struct bifrost_sampler_descriptor
*desc
= (struct bifrost_sampler_descriptor
*)transfer
.cpu
;
1442 for (int i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1443 desc
[i
] = ctx
->samplers
[stage
][i
]->bifrost_hw
;
1445 postfix
->sampler_descriptor
= transfer
.gpu
;
1447 size_t desc_size
= sizeof(struct mali_sampler_descriptor
);
1448 size_t transfer_size
= desc_size
* ctx
->sampler_count
[stage
];
1449 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1451 struct mali_sampler_descriptor
*desc
= (struct mali_sampler_descriptor
*)transfer
.cpu
;
1453 for (int i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1454 desc
[i
] = ctx
->samplers
[stage
][i
]->midgard_hw
;
1456 postfix
->sampler_descriptor
= transfer
.gpu
;
1461 panfrost_emit_vertex_attr_meta(struct panfrost_batch
*batch
,
1462 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1464 struct panfrost_context
*ctx
= batch
->ctx
;
1469 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1471 panfrost_vertex_state_upd_attr_offs(ctx
, vertex_postfix
);
1472 vertex_postfix
->attribute_meta
= panfrost_upload_transient(batch
, so
->hw
,
1478 panfrost_emit_vertex_data(struct panfrost_batch
*batch
,
1479 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1481 struct panfrost_context
*ctx
= batch
->ctx
;
1482 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1484 /* Staged mali_attr, and index into them. i =/= k, depending on the
1485 * vertex buffer mask and instancing. Twice as much room is allocated,
1486 * for a worst case of NPOT_DIVIDEs which take up extra slot */
1487 union mali_attr attrs
[PIPE_MAX_ATTRIBS
* 2];
1490 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1491 /* We map a mali_attr to be 1:1 with the mali_attr_meta, which
1492 * means duplicating some vertex buffers (who cares? aside from
1493 * maybe some caching implications but I somehow doubt that
1496 struct pipe_vertex_element
*elem
= &so
->pipe
[i
];
1497 unsigned vbi
= elem
->vertex_buffer_index
;
1499 /* The exception to 1:1 mapping is that we can have multiple
1500 * entries (NPOT divisors), so we fixup anyways */
1502 so
->hw
[i
].index
= k
;
1504 if (!(ctx
->vb_mask
& (1 << vbi
)))
1507 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1508 struct panfrost_resource
*rsrc
;
1510 rsrc
= pan_resource(buf
->buffer
.resource
);
1514 /* Align to 64 bytes by masking off the lower bits. This
1515 * will be adjusted back when we fixup the src_offset in
1518 mali_ptr raw_addr
= rsrc
->bo
->gpu
+ buf
->buffer_offset
;
1519 mali_ptr addr
= raw_addr
& ~63;
1520 unsigned chopped_addr
= raw_addr
- addr
;
1522 /* Add a dependency of the batch on the vertex buffer */
1523 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1524 PAN_BO_ACCESS_SHARED
|
1525 PAN_BO_ACCESS_READ
|
1526 PAN_BO_ACCESS_VERTEX_TILER
);
1528 /* Set common fields */
1529 attrs
[k
].elements
= addr
;
1530 attrs
[k
].stride
= buf
->stride
;
1532 /* Since we advanced the base pointer, we shrink the buffer
1534 attrs
[k
].size
= rsrc
->base
.width0
- buf
->buffer_offset
;
1536 /* We need to add the extra size we masked off (for
1537 * correctness) so the data doesn't get clamped away */
1538 attrs
[k
].size
+= chopped_addr
;
1540 /* For non-instancing make sure we initialize */
1541 attrs
[k
].shift
= attrs
[k
].extra_flags
= 0;
1543 /* Instancing uses a dramatically different code path than
1544 * linear, so dispatch for the actual emission now that the
1545 * common code is finished */
1547 unsigned divisor
= elem
->instance_divisor
;
1549 if (divisor
&& ctx
->instance_count
== 1) {
1550 /* Silly corner case where there's a divisor(=1) but
1551 * there's no legitimate instancing. So we want *every*
1552 * attribute to be the same. So set stride to zero so
1553 * we don't go anywhere. */
1555 attrs
[k
].size
= attrs
[k
].stride
+ chopped_addr
;
1556 attrs
[k
].stride
= 0;
1557 attrs
[k
++].elements
|= MALI_ATTR_LINEAR
;
1558 } else if (ctx
->instance_count
<= 1) {
1559 /* Normal, non-instanced attributes */
1560 attrs
[k
++].elements
|= MALI_ATTR_LINEAR
;
1562 unsigned instance_shift
= vertex_postfix
->instance_shift
;
1563 unsigned instance_odd
= vertex_postfix
->instance_odd
;
1565 k
+= panfrost_vertex_instanced(ctx
->padded_count
,
1568 divisor
, &attrs
[k
]);
1572 /* Add special gl_VertexID/gl_InstanceID buffers */
1574 panfrost_vertex_id(ctx
->padded_count
, &attrs
[k
]);
1575 so
->hw
[PAN_VERTEX_ID
].index
= k
++;
1576 panfrost_instance_id(ctx
->padded_count
, &attrs
[k
]);
1577 so
->hw
[PAN_INSTANCE_ID
].index
= k
++;
1579 /* Upload whatever we emitted and go */
1581 vertex_postfix
->attributes
= panfrost_upload_transient(batch
, attrs
,
1582 k
* sizeof(*attrs
));
1586 panfrost_emit_varyings(struct panfrost_batch
*batch
, union mali_attr
*slot
,
1587 unsigned stride
, unsigned count
)
1589 /* Fill out the descriptor */
1590 slot
->stride
= stride
;
1591 slot
->size
= stride
* count
;
1592 slot
->shift
= slot
->extra_flags
= 0;
1594 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1597 slot
->elements
= transfer
.gpu
| MALI_ATTR_LINEAR
;
1599 return transfer
.gpu
;
1603 panfrost_emit_streamout(struct panfrost_batch
*batch
, union mali_attr
*slot
,
1604 unsigned stride
, unsigned offset
, unsigned count
,
1605 struct pipe_stream_output_target
*target
)
1607 /* Fill out the descriptor */
1608 slot
->stride
= stride
* 4;
1609 slot
->shift
= slot
->extra_flags
= 0;
1611 unsigned max_size
= target
->buffer_size
;
1612 unsigned expected_size
= slot
->stride
* count
;
1614 slot
->size
= MIN2(max_size
, expected_size
);
1616 /* Grab the BO and bind it to the batch */
1617 struct panfrost_bo
*bo
= pan_resource(target
->buffer
)->bo
;
1619 /* Varyings are WRITE from the perspective of the VERTEX but READ from
1620 * the perspective of the TILER and FRAGMENT.
1622 panfrost_batch_add_bo(batch
, bo
,
1623 PAN_BO_ACCESS_SHARED
|
1625 PAN_BO_ACCESS_VERTEX_TILER
|
1626 PAN_BO_ACCESS_FRAGMENT
);
1628 mali_ptr addr
= bo
->gpu
+ target
->buffer_offset
+ (offset
* slot
->stride
);
1629 slot
->elements
= addr
;
1632 /* Given a shader and buffer indices, link varying metadata together */
1635 is_special_varying(gl_varying_slot loc
)
1638 case VARYING_SLOT_POS
:
1639 case VARYING_SLOT_PSIZ
:
1640 case VARYING_SLOT_PNTC
:
1641 case VARYING_SLOT_FACE
:
1649 panfrost_emit_varying_meta(void *outptr
, struct panfrost_shader_state
*ss
,
1650 signed general
, signed gl_Position
,
1651 signed gl_PointSize
, signed gl_PointCoord
,
1652 signed gl_FrontFacing
)
1654 struct mali_attr_meta
*out
= (struct mali_attr_meta
*) outptr
;
1656 for (unsigned i
= 0; i
< ss
->varying_count
; ++i
) {
1657 gl_varying_slot location
= ss
->varyings_loc
[i
];
1661 case VARYING_SLOT_POS
:
1662 index
= gl_Position
;
1664 case VARYING_SLOT_PSIZ
:
1665 index
= gl_PointSize
;
1667 case VARYING_SLOT_PNTC
:
1668 index
= gl_PointCoord
;
1670 case VARYING_SLOT_FACE
:
1671 index
= gl_FrontFacing
;
1679 out
[i
].index
= index
;
1684 has_point_coord(unsigned mask
, gl_varying_slot loc
)
1686 if ((loc
>= VARYING_SLOT_TEX0
) && (loc
<= VARYING_SLOT_TEX7
))
1687 return (mask
& (1 << (loc
- VARYING_SLOT_TEX0
)));
1688 else if (loc
== VARYING_SLOT_PNTC
)
1689 return (mask
& (1 << 8));
1694 /* Helpers for manipulating stream out information so we can pack varyings
1695 * accordingly. Compute the src_offset for a given captured varying */
1697 static struct pipe_stream_output
*
1698 pan_get_so(struct pipe_stream_output_info
*info
, gl_varying_slot loc
)
1700 for (unsigned i
= 0; i
< info
->num_outputs
; ++i
) {
1701 if (info
->output
[i
].register_index
== loc
)
1702 return &info
->output
[i
];
1705 unreachable("Varying not captured");
1709 panfrost_emit_varying_descriptor(struct panfrost_batch
*batch
,
1710 unsigned vertex_count
,
1711 struct mali_vertex_tiler_postfix
*vertex_postfix
,
1712 struct mali_vertex_tiler_postfix
*tiler_postfix
,
1713 union midgard_primitive_size
*primitive_size
)
1715 /* Load the shaders */
1716 struct panfrost_context
*ctx
= batch
->ctx
;
1717 struct panfrost_shader_state
*vs
, *fs
;
1718 unsigned int num_gen_varyings
= 0;
1719 size_t vs_size
, fs_size
;
1721 /* Allocate the varying descriptor */
1723 vs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_VERTEX
);
1724 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
1725 vs_size
= sizeof(struct mali_attr_meta
) * vs
->varying_count
;
1726 fs_size
= sizeof(struct mali_attr_meta
) * fs
->varying_count
;
1728 struct panfrost_transfer trans
= panfrost_allocate_transient(batch
,
1732 struct pipe_stream_output_info
*so
= &vs
->stream_output
;
1734 /* Check if this varying is linked by us. This is the case for
1735 * general-purpose, non-captured varyings. If it is, link it. If it's
1736 * not, use the provided stream out information to determine the
1737 * offset, since it was already linked for us. */
1739 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1740 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1742 bool special
= is_special_varying(loc
);
1743 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1746 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1748 unsigned dst_offset
= o
->dst_offset
* 4; /* dwords */
1749 vs
->varyings
[i
].src_offset
= dst_offset
;
1750 } else if (!special
) {
1751 vs
->varyings
[i
].src_offset
= 16 * (num_gen_varyings
++);
1755 /* Conversely, we need to set src_offset for the captured varyings.
1756 * Here, the layout is defined by the stream out info, not us */
1758 /* Link up with fragment varyings */
1759 bool reads_point_coord
= fs
->reads_point_coord
;
1761 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1762 gl_varying_slot loc
= fs
->varyings_loc
[i
];
1763 unsigned src_offset
;
1767 for (unsigned j
= 0; j
< vs
->varying_count
; ++j
) {
1768 if (vs
->varyings_loc
[j
] == loc
) {
1774 /* Either assign or reuse */
1776 src_offset
= vs
->varyings
[vs_idx
].src_offset
;
1778 src_offset
= 16 * (num_gen_varyings
++);
1780 fs
->varyings
[i
].src_offset
= src_offset
;
1782 if (has_point_coord(fs
->point_sprite_mask
, loc
))
1783 reads_point_coord
= true;
1786 memcpy(trans
.cpu
, vs
->varyings
, vs_size
);
1787 memcpy(trans
.cpu
+ vs_size
, fs
->varyings
, fs_size
);
1789 union mali_attr varyings
[PIPE_MAX_ATTRIBS
] = {0};
1791 /* Figure out how many streamout buffers could be bound */
1792 unsigned so_count
= ctx
->streamout
.num_targets
;
1793 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1794 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1796 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1797 if (!captured
) continue;
1799 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1800 so_count
= MAX2(so_count
, o
->output_buffer
+ 1);
1803 signed idx
= so_count
;
1804 signed general
= idx
++;
1805 signed gl_Position
= idx
++;
1806 signed gl_PointSize
= vs
->writes_point_size
? (idx
++) : -1;
1807 signed gl_PointCoord
= reads_point_coord
? (idx
++) : -1;
1808 signed gl_FrontFacing
= fs
->reads_face
? (idx
++) : -1;
1809 signed gl_FragCoord
= fs
->reads_frag_coord
? (idx
++) : -1;
1811 /* Emit the stream out buffers */
1813 unsigned out_count
= u_stream_outputs_for_vertices(ctx
->active_prim
,
1816 for (unsigned i
= 0; i
< so_count
; ++i
) {
1817 if (i
< ctx
->streamout
.num_targets
) {
1818 panfrost_emit_streamout(batch
, &varyings
[i
],
1820 ctx
->streamout
.offsets
[i
],
1822 ctx
->streamout
.targets
[i
]);
1824 /* Emit a dummy buffer */
1825 panfrost_emit_varyings(batch
, &varyings
[i
],
1829 /* Clear the attribute type */
1830 varyings
[i
].elements
&= ~0xF;
1834 panfrost_emit_varyings(batch
, &varyings
[general
],
1835 num_gen_varyings
* 16,
1838 mali_ptr varyings_p
;
1840 /* fp32 vec4 gl_Position */
1841 varyings_p
= panfrost_emit_varyings(batch
, &varyings
[gl_Position
],
1842 sizeof(float) * 4, vertex_count
);
1843 tiler_postfix
->position_varying
= varyings_p
;
1846 if (panfrost_writes_point_size(ctx
)) {
1847 varyings_p
= panfrost_emit_varyings(batch
,
1848 &varyings
[gl_PointSize
],
1850 primitive_size
->pointer
= varyings_p
;
1853 if (reads_point_coord
)
1854 varyings
[gl_PointCoord
].elements
= MALI_VARYING_POINT_COORD
;
1857 varyings
[gl_FrontFacing
].elements
= MALI_VARYING_FRONT_FACING
;
1859 if (fs
->reads_frag_coord
)
1860 varyings
[gl_FragCoord
].elements
= MALI_VARYING_FRAG_COORD
;
1862 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1863 assert(!(device
->quirks
& IS_BIFROST
) || !(reads_point_coord
));
1865 /* Let's go ahead and link varying meta to the buffer in question, now
1866 * that that information is available. VARYING_SLOT_POS is mapped to
1867 * gl_FragCoord for fragment shaders but gl_Positionf or vertex shaders
1870 panfrost_emit_varying_meta(trans
.cpu
, vs
, general
, gl_Position
,
1871 gl_PointSize
, gl_PointCoord
,
1874 panfrost_emit_varying_meta(trans
.cpu
+ vs_size
, fs
, general
,
1875 gl_FragCoord
, gl_PointSize
,
1876 gl_PointCoord
, gl_FrontFacing
);
1878 /* Replace streamout */
1880 struct mali_attr_meta
*ovs
= (struct mali_attr_meta
*)trans
.cpu
;
1881 struct mali_attr_meta
*ofs
= ovs
+ vs
->varying_count
;
1883 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1884 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1886 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1890 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1891 ovs
[i
].index
= o
->output_buffer
;
1893 assert(o
->stream
== 0);
1894 ovs
[i
].format
= (vs
->varyings
[i
].format
& ~MALI_NR_CHANNELS(4))
1895 | MALI_NR_CHANNELS(o
->num_components
);
1897 if (device
->quirks
& HAS_SWIZZLES
)
1898 ovs
[i
].swizzle
= panfrost_get_default_swizzle(o
->num_components
);
1900 ovs
[i
].swizzle
= panfrost_bifrost_swizzle(o
->num_components
);
1902 /* Link to the fragment */
1906 for (unsigned j
= 0; j
< fs
->varying_count
; ++j
) {
1907 if (fs
->varyings_loc
[j
] == loc
) {
1914 ofs
[fs_idx
].index
= ovs
[i
].index
;
1915 ofs
[fs_idx
].format
= ovs
[i
].format
;
1916 ofs
[fs_idx
].swizzle
= ovs
[i
].swizzle
;
1920 /* Replace point sprite */
1921 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1922 /* If we have a point sprite replacement, handle that here. We
1923 * have to translate location first. TODO: Flip y in shader.
1924 * We're already keying ... just time crunch .. */
1926 if (has_point_coord(fs
->point_sprite_mask
,
1927 fs
->varyings_loc
[i
])) {
1928 ofs
[i
].index
= gl_PointCoord
;
1930 /* Swizzle out the z/w to 0/1 */
1931 ofs
[i
].format
= MALI_RG16F
;
1932 ofs
[i
].swizzle
= panfrost_get_default_swizzle(2);
1936 /* Fix up unaligned addresses */
1937 for (unsigned i
= 0; i
< so_count
; ++i
) {
1938 if (varyings
[i
].elements
< MALI_RECORD_SPECIAL
)
1941 unsigned align
= (varyings
[i
].elements
& 63);
1943 /* While we're at it, the SO buffers are linear */
1946 varyings
[i
].elements
|= MALI_ATTR_LINEAR
;
1950 /* We need to adjust alignment */
1951 varyings
[i
].elements
&= ~63;
1952 varyings
[i
].elements
|= MALI_ATTR_LINEAR
;
1953 varyings
[i
].size
+= align
;
1955 for (unsigned v
= 0; v
< vs
->varying_count
; ++v
) {
1956 if (ovs
[v
].index
!= i
)
1959 ovs
[v
].src_offset
= vs
->varyings
[v
].src_offset
+ align
;
1962 for (unsigned f
= 0; f
< fs
->varying_count
; ++f
) {
1963 if (ofs
[f
].index
!= i
)
1966 ofs
[f
].src_offset
= fs
->varyings
[f
].src_offset
+ align
;
1970 varyings_p
= panfrost_upload_transient(batch
, varyings
,
1971 idx
* sizeof(*varyings
));
1972 vertex_postfix
->varyings
= varyings_p
;
1973 tiler_postfix
->varyings
= varyings_p
;
1975 vertex_postfix
->varying_meta
= trans
.gpu
;
1976 tiler_postfix
->varying_meta
= trans
.gpu
+ vs_size
;
1980 panfrost_emit_vertex_tiler_jobs(struct panfrost_batch
*batch
,
1981 struct mali_vertex_tiler_prefix
*vertex_prefix
,
1982 struct mali_vertex_tiler_postfix
*vertex_postfix
,
1983 struct mali_vertex_tiler_prefix
*tiler_prefix
,
1984 struct mali_vertex_tiler_postfix
*tiler_postfix
,
1985 union midgard_primitive_size
*primitive_size
)
1987 struct panfrost_context
*ctx
= batch
->ctx
;
1988 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1989 bool wallpapering
= ctx
->wallpaper_batch
&& batch
->tiler_dep
;
1990 struct bifrost_payload_vertex bifrost_vertex
= {0,};
1991 struct bifrost_payload_tiler bifrost_tiler
= {0,};
1992 struct midgard_payload_vertex_tiler midgard_vertex
= {0,};
1993 struct midgard_payload_vertex_tiler midgard_tiler
= {0,};
1995 size_t vp_size
, tp_size
;
1997 if (device
->quirks
& IS_BIFROST
) {
1998 bifrost_vertex
.prefix
= *vertex_prefix
;
1999 bifrost_vertex
.postfix
= *vertex_postfix
;
2000 vp
= &bifrost_vertex
;
2001 vp_size
= sizeof(bifrost_vertex
);
2003 bifrost_tiler
.prefix
= *tiler_prefix
;
2004 bifrost_tiler
.tiler
.primitive_size
= *primitive_size
;
2005 bifrost_tiler
.tiler
.tiler_meta
= panfrost_batch_get_tiler_meta(batch
, ~0);
2006 bifrost_tiler
.postfix
= *tiler_postfix
;
2007 tp
= &bifrost_tiler
;
2008 tp_size
= sizeof(bifrost_tiler
);
2010 midgard_vertex
.prefix
= *vertex_prefix
;
2011 midgard_vertex
.postfix
= *vertex_postfix
;
2012 vp
= &midgard_vertex
;
2013 vp_size
= sizeof(midgard_vertex
);
2015 midgard_tiler
.prefix
= *tiler_prefix
;
2016 midgard_tiler
.postfix
= *tiler_postfix
;
2017 midgard_tiler
.primitive_size
= *primitive_size
;
2018 tp
= &midgard_tiler
;
2019 tp_size
= sizeof(midgard_tiler
);
2023 /* Inject in reverse order, with "predicted" job indices.
2024 * THIS IS A HACK XXX */
2025 panfrost_new_job(batch
, JOB_TYPE_TILER
, false,
2026 batch
->job_index
+ 2, tp
, tp_size
, true);
2027 panfrost_new_job(batch
, JOB_TYPE_VERTEX
, false, 0,
2032 /* If rasterizer discard is enable, only submit the vertex */
2034 bool rasterizer_discard
= ctx
->rasterizer
&&
2035 ctx
->rasterizer
->base
.rasterizer_discard
;
2037 unsigned vertex
= panfrost_new_job(batch
, JOB_TYPE_VERTEX
, false, 0,
2038 vp
, vp_size
, false);
2040 if (rasterizer_discard
)
2043 panfrost_new_job(batch
, JOB_TYPE_TILER
, false, vertex
, tp
, tp_size
,
2047 /* TODO: stop hardcoding this */
2049 panfrost_emit_sample_locations(struct panfrost_batch
*batch
)
2051 uint16_t locations
[] = {
2102 return panfrost_upload_transient(batch
, locations
, 96 * sizeof(uint16_t));