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"
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_pool_upload(&batch
->pool
, &shared
, sizeof(shared
));
71 panfrost_vt_attach_framebuffer(struct panfrost_context
*ctx
,
72 struct mali_vertex_tiler_postfix
*postfix
)
74 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
75 postfix
->shared_memory
= panfrost_batch_reserve_framebuffer(batch
);
79 panfrost_vt_update_rasterizer(struct panfrost_context
*ctx
,
80 struct mali_vertex_tiler_prefix
*prefix
,
81 struct mali_vertex_tiler_postfix
*postfix
)
83 struct panfrost_rasterizer
*rasterizer
= ctx
->rasterizer
;
85 postfix
->gl_enables
|= 0x7;
86 SET_BIT(postfix
->gl_enables
, MALI_FRONT_CCW_TOP
,
87 rasterizer
&& rasterizer
->base
.front_ccw
);
88 SET_BIT(postfix
->gl_enables
, MALI_CULL_FACE_FRONT
,
89 rasterizer
&& (rasterizer
->base
.cull_face
& PIPE_FACE_FRONT
));
90 SET_BIT(postfix
->gl_enables
, MALI_CULL_FACE_BACK
,
91 rasterizer
&& (rasterizer
->base
.cull_face
& PIPE_FACE_BACK
));
92 SET_BIT(prefix
->unknown_draw
, MALI_DRAW_FLATSHADE_FIRST
,
93 rasterizer
&& rasterizer
->base
.flatshade_first
);
97 panfrost_vt_update_primitive_size(struct panfrost_context
*ctx
,
98 struct mali_vertex_tiler_prefix
*prefix
,
99 union midgard_primitive_size
*primitive_size
)
101 struct panfrost_rasterizer
*rasterizer
= ctx
->rasterizer
;
103 if (!panfrost_writes_point_size(ctx
)) {
104 bool points
= prefix
->draw_mode
== MALI_POINTS
;
109 rasterizer
->base
.point_size
:
110 rasterizer
->base
.line_width
;
112 primitive_size
->constant
= val
;
117 panfrost_vt_update_occlusion_query(struct panfrost_context
*ctx
,
118 struct mali_vertex_tiler_postfix
*postfix
)
120 SET_BIT(postfix
->gl_enables
, MALI_OCCLUSION_QUERY
, ctx
->occlusion_query
);
121 if (ctx
->occlusion_query
) {
122 postfix
->occlusion_counter
= ctx
->occlusion_query
->bo
->gpu
;
123 panfrost_batch_add_bo(ctx
->batch
, ctx
->occlusion_query
->bo
,
124 PAN_BO_ACCESS_SHARED
|
126 PAN_BO_ACCESS_FRAGMENT
);
128 postfix
->occlusion_counter
= 0;
133 panfrost_vt_init(struct panfrost_context
*ctx
,
134 enum pipe_shader_type stage
,
135 struct mali_vertex_tiler_prefix
*prefix
,
136 struct mali_vertex_tiler_postfix
*postfix
)
138 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
140 if (!ctx
->shader
[stage
])
143 memset(prefix
, 0, sizeof(*prefix
));
144 memset(postfix
, 0, sizeof(*postfix
));
146 if (device
->quirks
& IS_BIFROST
) {
147 postfix
->gl_enables
= 0x2;
148 panfrost_vt_emit_shared_memory(ctx
, postfix
);
150 postfix
->gl_enables
= 0x6;
151 panfrost_vt_attach_framebuffer(ctx
, postfix
);
154 if (stage
== PIPE_SHADER_FRAGMENT
) {
155 panfrost_vt_update_occlusion_query(ctx
, postfix
);
156 panfrost_vt_update_rasterizer(ctx
, prefix
, postfix
);
161 panfrost_translate_index_size(unsigned size
)
165 return MALI_DRAW_INDEXED_UINT8
;
168 return MALI_DRAW_INDEXED_UINT16
;
171 return MALI_DRAW_INDEXED_UINT32
;
174 unreachable("Invalid index size");
178 /* Gets a GPU address for the associated index buffer. Only gauranteed to be
179 * good for the duration of the draw (transient), could last longer. Also get
180 * the bounds on the index buffer for the range accessed by the draw. We do
181 * these operations together because there are natural optimizations which
182 * require them to be together. */
185 panfrost_get_index_buffer_bounded(struct panfrost_context
*ctx
,
186 const struct pipe_draw_info
*info
,
187 unsigned *min_index
, unsigned *max_index
)
189 struct panfrost_resource
*rsrc
= pan_resource(info
->index
.resource
);
190 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
191 off_t offset
= info
->start
* info
->index_size
;
192 bool needs_indices
= true;
195 if (info
->max_index
!= ~0u) {
196 *min_index
= info
->min_index
;
197 *max_index
= info
->max_index
;
198 needs_indices
= false;
201 if (!info
->has_user_indices
) {
202 /* Only resources can be directly mapped */
203 panfrost_batch_add_bo(batch
, rsrc
->bo
,
204 PAN_BO_ACCESS_SHARED
|
206 PAN_BO_ACCESS_VERTEX_TILER
);
207 out
= rsrc
->bo
->gpu
+ offset
;
209 /* Check the cache */
210 needs_indices
= !panfrost_minmax_cache_get(rsrc
->index_cache
,
216 /* Otherwise, we need to upload to transient memory */
217 const uint8_t *ibuf8
= (const uint8_t *) info
->index
.user
;
218 out
= panfrost_pool_upload(&batch
->pool
, ibuf8
+ offset
,
225 u_vbuf_get_minmax_index(&ctx
->base
, info
, min_index
, max_index
);
227 if (!info
->has_user_indices
)
228 panfrost_minmax_cache_add(rsrc
->index_cache
,
229 info
->start
, info
->count
,
230 *min_index
, *max_index
);
237 panfrost_vt_set_draw_info(struct panfrost_context
*ctx
,
238 const struct pipe_draw_info
*info
,
239 enum mali_draw_mode draw_mode
,
240 struct mali_vertex_tiler_postfix
*vertex_postfix
,
241 struct mali_vertex_tiler_prefix
*tiler_prefix
,
242 struct mali_vertex_tiler_postfix
*tiler_postfix
,
243 unsigned *vertex_count
,
244 unsigned *padded_count
)
246 tiler_prefix
->draw_mode
= draw_mode
;
248 unsigned draw_flags
= 0;
250 if (panfrost_writes_point_size(ctx
))
251 draw_flags
|= MALI_DRAW_VARYING_SIZE
;
253 if (info
->primitive_restart
)
254 draw_flags
|= MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX
;
256 /* These doesn't make much sense */
258 draw_flags
|= 0x3000;
260 if (info
->index_size
) {
261 unsigned min_index
= 0, max_index
= 0;
263 tiler_prefix
->indices
= panfrost_get_index_buffer_bounded(ctx
,
268 /* Use the corresponding values */
269 *vertex_count
= max_index
- min_index
+ 1;
270 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= min_index
+ info
->index_bias
;
271 tiler_prefix
->offset_bias_correction
= -min_index
;
272 tiler_prefix
->index_count
= MALI_POSITIVE(info
->count
);
273 draw_flags
|= panfrost_translate_index_size(info
->index_size
);
275 tiler_prefix
->indices
= 0;
276 *vertex_count
= ctx
->vertex_count
;
277 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= info
->start
;
278 tiler_prefix
->offset_bias_correction
= 0;
279 tiler_prefix
->index_count
= MALI_POSITIVE(ctx
->vertex_count
);
282 tiler_prefix
->unknown_draw
= draw_flags
;
284 /* Encode the padded vertex count */
286 if (info
->instance_count
> 1) {
287 *padded_count
= panfrost_padded_vertex_count(*vertex_count
);
289 unsigned shift
= __builtin_ctz(ctx
->padded_count
);
290 unsigned k
= ctx
->padded_count
>> (shift
+ 1);
292 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= shift
;
293 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= k
;
295 *padded_count
= *vertex_count
;
297 /* Reset instancing state */
298 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= 0;
299 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= 0;
304 panfrost_shader_meta_init(struct panfrost_context
*ctx
,
305 enum pipe_shader_type st
,
306 struct mali_shader_meta
*meta
)
308 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
309 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
311 memset(meta
, 0, sizeof(*meta
));
312 meta
->shader
= (ss
->bo
? ss
->bo
->gpu
: 0) | ss
->first_tag
;
313 meta
->attribute_count
= ss
->attribute_count
;
314 meta
->varying_count
= ss
->varying_count
;
315 meta
->texture_count
= ctx
->sampler_view_count
[st
];
316 meta
->sampler_count
= ctx
->sampler_count
[st
];
318 if (dev
->quirks
& IS_BIFROST
) {
319 if (st
== PIPE_SHADER_VERTEX
)
320 meta
->bifrost1
.unk1
= 0x800000;
322 /* First clause ATEST |= 0x4000000.
323 * Less than 32 regs |= 0x200 */
324 meta
->bifrost1
.unk1
= 0x950020;
327 meta
->bifrost1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
328 if (st
== PIPE_SHADER_VERTEX
)
329 meta
->bifrost2
.preload_regs
= 0xC0;
331 meta
->bifrost2
.preload_regs
= 0x1;
332 SET_BIT(meta
->bifrost2
.preload_regs
, 0x10, ss
->reads_frag_coord
);
335 meta
->bifrost2
.uniform_count
= MIN2(ss
->uniform_count
,
338 meta
->midgard1
.uniform_count
= MIN2(ss
->uniform_count
,
340 meta
->midgard1
.work_count
= ss
->work_reg_count
;
342 /* TODO: This is not conformant on ES3 */
343 meta
->midgard1
.flags_hi
= MALI_SUPPRESS_INF_NAN
;
345 meta
->midgard1
.flags_lo
= 0x20;
346 meta
->midgard1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
348 SET_BIT(meta
->midgard1
.flags_hi
, MALI_WRITES_GLOBAL
, ss
->writes_global
);
353 panfrost_translate_compare_func(enum pipe_compare_func in
)
356 case PIPE_FUNC_NEVER
:
357 return MALI_FUNC_NEVER
;
360 return MALI_FUNC_LESS
;
362 case PIPE_FUNC_EQUAL
:
363 return MALI_FUNC_EQUAL
;
365 case PIPE_FUNC_LEQUAL
:
366 return MALI_FUNC_LEQUAL
;
368 case PIPE_FUNC_GREATER
:
369 return MALI_FUNC_GREATER
;
371 case PIPE_FUNC_NOTEQUAL
:
372 return MALI_FUNC_NOTEQUAL
;
374 case PIPE_FUNC_GEQUAL
:
375 return MALI_FUNC_GEQUAL
;
377 case PIPE_FUNC_ALWAYS
:
378 return MALI_FUNC_ALWAYS
;
381 unreachable("Invalid func");
386 panfrost_translate_stencil_op(enum pipe_stencil_op in
)
389 case PIPE_STENCIL_OP_KEEP
:
390 return MALI_STENCIL_KEEP
;
392 case PIPE_STENCIL_OP_ZERO
:
393 return MALI_STENCIL_ZERO
;
395 case PIPE_STENCIL_OP_REPLACE
:
396 return MALI_STENCIL_REPLACE
;
398 case PIPE_STENCIL_OP_INCR
:
399 return MALI_STENCIL_INCR
;
401 case PIPE_STENCIL_OP_DECR
:
402 return MALI_STENCIL_DECR
;
404 case PIPE_STENCIL_OP_INCR_WRAP
:
405 return MALI_STENCIL_INCR_WRAP
;
407 case PIPE_STENCIL_OP_DECR_WRAP
:
408 return MALI_STENCIL_DECR_WRAP
;
410 case PIPE_STENCIL_OP_INVERT
:
411 return MALI_STENCIL_INVERT
;
414 unreachable("Invalid stencil op");
419 translate_tex_wrap(enum pipe_tex_wrap w
)
422 case PIPE_TEX_WRAP_REPEAT
:
423 return MALI_WRAP_REPEAT
;
425 case PIPE_TEX_WRAP_CLAMP
:
426 return MALI_WRAP_CLAMP
;
428 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
429 return MALI_WRAP_CLAMP_TO_EDGE
;
431 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
432 return MALI_WRAP_CLAMP_TO_BORDER
;
434 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
435 return MALI_WRAP_MIRRORED_REPEAT
;
437 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
438 return MALI_WRAP_MIRRORED_CLAMP
;
440 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
441 return MALI_WRAP_MIRRORED_CLAMP_TO_EDGE
;
443 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
444 return MALI_WRAP_MIRRORED_CLAMP_TO_BORDER
;
447 unreachable("Invalid wrap");
451 void panfrost_sampler_desc_init(const struct pipe_sampler_state
*cso
,
452 struct mali_sampler_descriptor
*hw
)
454 unsigned func
= panfrost_translate_compare_func(cso
->compare_func
);
455 bool min_nearest
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
;
456 bool mag_nearest
= cso
->mag_img_filter
== PIPE_TEX_FILTER_NEAREST
;
457 bool mip_linear
= cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
;
458 unsigned min_filter
= min_nearest
? MALI_SAMP_MIN_NEAREST
: 0;
459 unsigned mag_filter
= mag_nearest
? MALI_SAMP_MAG_NEAREST
: 0;
460 unsigned mip_filter
= mip_linear
?
461 (MALI_SAMP_MIP_LINEAR_1
| MALI_SAMP_MIP_LINEAR_2
) : 0;
462 unsigned normalized
= cso
->normalized_coords
? MALI_SAMP_NORM_COORDS
: 0;
464 *hw
= (struct mali_sampler_descriptor
) {
465 .filter_mode
= min_filter
| mag_filter
| mip_filter
|
467 .wrap_s
= translate_tex_wrap(cso
->wrap_s
),
468 .wrap_t
= translate_tex_wrap(cso
->wrap_t
),
469 .wrap_r
= translate_tex_wrap(cso
->wrap_r
),
470 .compare_func
= cso
->compare_mode
?
471 panfrost_flip_compare_func(func
) :
474 cso
->border_color
.f
[0],
475 cso
->border_color
.f
[1],
476 cso
->border_color
.f
[2],
477 cso
->border_color
.f
[3]
479 .min_lod
= FIXED_16(cso
->min_lod
, false), /* clamp at 0 */
480 .max_lod
= FIXED_16(cso
->max_lod
, false),
481 .lod_bias
= FIXED_16(cso
->lod_bias
, true), /* can be negative */
482 .seamless_cube_map
= cso
->seamless_cube_map
,
485 /* If necessary, we disable mipmapping in the sampler descriptor by
486 * clamping the LOD as tight as possible (from 0 to epsilon,
487 * essentially -- remember these are fixed point numbers, so
490 if (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
491 hw
->max_lod
= hw
->min_lod
+ 1;
494 void panfrost_sampler_desc_init_bifrost(const struct pipe_sampler_state
*cso
,
495 struct bifrost_sampler_descriptor
*hw
)
497 *hw
= (struct bifrost_sampler_descriptor
) {
499 .wrap_s
= translate_tex_wrap(cso
->wrap_s
),
500 .wrap_t
= translate_tex_wrap(cso
->wrap_t
),
501 .wrap_r
= translate_tex_wrap(cso
->wrap_r
),
503 .min_filter
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
,
504 .norm_coords
= cso
->normalized_coords
,
505 .mip_filter
= cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
,
506 .mag_filter
= cso
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
,
507 .min_lod
= FIXED_16(cso
->min_lod
, false), /* clamp at 0 */
508 .max_lod
= FIXED_16(cso
->max_lod
, false),
511 /* If necessary, we disable mipmapping in the sampler descriptor by
512 * clamping the LOD as tight as possible (from 0 to epsilon,
513 * essentially -- remember these are fixed point numbers, so
516 if (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
517 hw
->max_lod
= hw
->min_lod
+ 1;
521 panfrost_make_stencil_state(const struct pipe_stencil_state
*in
,
522 struct mali_stencil_test
*out
)
524 out
->ref
= 0; /* Gallium gets it from elsewhere */
526 out
->mask
= in
->valuemask
;
527 out
->func
= panfrost_translate_compare_func(in
->func
);
528 out
->sfail
= panfrost_translate_stencil_op(in
->fail_op
);
529 out
->dpfail
= panfrost_translate_stencil_op(in
->zfail_op
);
530 out
->dppass
= panfrost_translate_stencil_op(in
->zpass_op
);
534 panfrost_frag_meta_rasterizer_update(struct panfrost_context
*ctx
,
535 struct mali_shader_meta
*fragmeta
)
537 if (!ctx
->rasterizer
) {
538 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, true);
539 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, false);
540 fragmeta
->depth_units
= 0.0f
;
541 fragmeta
->depth_factor
= 0.0f
;
542 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
, false);
543 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
, false);
544 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_NEAR
, true);
545 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_FAR
, true);
549 struct pipe_rasterizer_state
*rast
= &ctx
->rasterizer
->base
;
551 bool msaa
= rast
->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
);
557 struct panfrost_shader_state
*fs
;
558 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
560 /* EXT_shader_framebuffer_fetch requires the shader to be run
561 * per-sample when outputs are read. */
562 bool per_sample
= ctx
->min_samples
> 1 || fs
->outputs_read
;
563 SET_BIT(fragmeta
->unknown2_3
, MALI_PER_SAMPLE
, msaa
&& per_sample
);
565 fragmeta
->depth_units
= rast
->offset_units
* 2.0f
;
566 fragmeta
->depth_factor
= rast
->offset_scale
;
568 /* XXX: Which bit is which? Does this maybe allow offseting not-tri? */
570 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
, rast
->offset_tri
);
571 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
, rast
->offset_tri
);
573 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_NEAR
, rast
->depth_clip_near
);
574 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_FAR
, rast
->depth_clip_far
);
578 panfrost_frag_meta_zsa_update(struct panfrost_context
*ctx
,
579 struct mali_shader_meta
*fragmeta
)
581 const struct pipe_depth_stencil_alpha_state
*zsa
= ctx
->depth_stencil
;
582 int zfunc
= PIPE_FUNC_ALWAYS
;
585 struct pipe_stencil_state default_stencil
= {
587 .func
= PIPE_FUNC_ALWAYS
,
588 .fail_op
= MALI_STENCIL_KEEP
,
589 .zfail_op
= MALI_STENCIL_KEEP
,
590 .zpass_op
= MALI_STENCIL_KEEP
,
595 panfrost_make_stencil_state(&default_stencil
,
596 &fragmeta
->stencil_front
);
597 fragmeta
->stencil_mask_front
= default_stencil
.writemask
;
598 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
599 fragmeta
->stencil_mask_back
= default_stencil
.writemask
;
600 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
, false);
601 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
, false);
603 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
,
604 zsa
->stencil
[0].enabled
);
605 panfrost_make_stencil_state(&zsa
->stencil
[0],
606 &fragmeta
->stencil_front
);
607 fragmeta
->stencil_mask_front
= zsa
->stencil
[0].writemask
;
608 fragmeta
->stencil_front
.ref
= ctx
->stencil_ref
.ref_value
[0];
610 /* If back-stencil is not enabled, use the front values */
612 if (zsa
->stencil
[1].enabled
) {
613 panfrost_make_stencil_state(&zsa
->stencil
[1],
614 &fragmeta
->stencil_back
);
615 fragmeta
->stencil_mask_back
= zsa
->stencil
[1].writemask
;
616 fragmeta
->stencil_back
.ref
= ctx
->stencil_ref
.ref_value
[1];
618 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
619 fragmeta
->stencil_mask_back
= fragmeta
->stencil_mask_front
;
620 fragmeta
->stencil_back
.ref
= fragmeta
->stencil_front
.ref
;
623 if (zsa
->depth
.enabled
)
624 zfunc
= zsa
->depth
.func
;
626 /* Depth state (TODO: Refactor) */
628 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
,
629 zsa
->depth
.writemask
);
632 fragmeta
->unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
633 fragmeta
->unknown2_3
|= MALI_DEPTH_FUNC(panfrost_translate_compare_func(zfunc
));
637 panfrost_fs_required(
638 struct panfrost_shader_state
*fs
,
639 struct panfrost_blend_final
*blend
,
642 /* If we generally have side effects */
646 /* If colour is written we need to execute */
647 for (unsigned i
= 0; i
< rt_count
; ++i
) {
648 if (!blend
[i
].no_colour
)
652 /* If depth is written and not implied we need to execute.
653 * TODO: Predicate on Z/S writes being enabled */
654 return (fs
->writes_depth
|| fs
->writes_stencil
);
658 panfrost_frag_meta_blend_update(struct panfrost_context
*ctx
,
659 struct mali_shader_meta
*fragmeta
,
662 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
663 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
664 struct panfrost_shader_state
*fs
;
665 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
667 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_DITHER
,
668 (dev
->quirks
& MIDGARD_SFBD
) && ctx
->blend
&&
669 !ctx
->blend
->base
.dither
);
671 SET_BIT(fragmeta
->unknown2_4
, MALI_ALPHA_TO_COVERAGE
,
672 ctx
->blend
->base
.alpha_to_coverage
);
674 /* Get blending setup */
675 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
677 struct panfrost_blend_final blend
[PIPE_MAX_COLOR_BUFS
];
678 unsigned shader_offset
= 0;
679 struct panfrost_bo
*shader_bo
= NULL
;
681 for (unsigned c
= 0; c
< rt_count
; ++c
)
682 blend
[c
] = panfrost_get_blend_for_context(ctx
, c
, &shader_bo
,
685 /* Disable shader execution if we can */
686 if (dev
->quirks
& MIDGARD_SHADERLESS
687 && !panfrost_fs_required(fs
, blend
, rt_count
)) {
688 fragmeta
->shader
= 0;
689 fragmeta
->attribute_count
= 0;
690 fragmeta
->varying_count
= 0;
691 fragmeta
->texture_count
= 0;
692 fragmeta
->sampler_count
= 0;
694 /* This feature is not known to work on Bifrost */
695 fragmeta
->midgard1
.work_count
= 1;
696 fragmeta
->midgard1
.uniform_count
= 0;
697 fragmeta
->midgard1
.uniform_buffer_count
= 0;
700 /* If there is a blend shader, work registers are shared. We impose 8
701 * work registers as a limit for blend shaders. Should be lower XXX */
703 if (!(dev
->quirks
& IS_BIFROST
)) {
704 for (unsigned c
= 0; c
< rt_count
; ++c
) {
705 if (blend
[c
].is_shader
) {
706 fragmeta
->midgard1
.work_count
=
707 MAX2(fragmeta
->midgard1
.work_count
, 8);
712 /* Even on MFBD, the shader descriptor gets blend shaders. It's *also*
713 * copied to the blend_meta appended (by convention), but this is the
714 * field actually read by the hardware. (Or maybe both are read...?).
715 * Specify the last RTi with a blend shader. */
717 fragmeta
->blend
.shader
= 0;
719 for (signed rt
= (rt_count
- 1); rt
>= 0; --rt
) {
720 if (!blend
[rt
].is_shader
)
723 fragmeta
->blend
.shader
= blend
[rt
].shader
.gpu
|
724 blend
[rt
].shader
.first_tag
;
728 if (dev
->quirks
& MIDGARD_SFBD
) {
729 /* When only a single render target platform is used, the blend
730 * information is inside the shader meta itself. We additionally
731 * need to signal CAN_DISCARD for nontrivial blend modes (so
732 * we're able to read back the destination buffer) */
734 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_BLEND_SHADER
,
737 if (!blend
[0].is_shader
) {
738 fragmeta
->blend
.equation
= *blend
[0].equation
.equation
;
739 fragmeta
->blend
.constant
= blend
[0].equation
.constant
;
742 SET_BIT(fragmeta
->unknown2_3
, MALI_CAN_DISCARD
,
743 !blend
[0].no_blending
|| fs
->can_discard
);
745 batch
->draws
|= PIPE_CLEAR_COLOR0
;
749 if (dev
->quirks
& IS_BIFROST
) {
750 bool no_blend
= true;
752 for (unsigned i
= 0; i
< rt_count
; ++i
)
753 no_blend
&= (blend
[i
].no_blending
| blend
[i
].no_colour
);
755 SET_BIT(fragmeta
->bifrost1
.unk1
, MALI_BIFROST_EARLY_Z
,
756 !fs
->can_discard
&& !fs
->writes_depth
&& no_blend
);
759 /* Additional blend descriptor tacked on for jobs using MFBD */
761 for (unsigned i
= 0; i
< rt_count
; ++i
) {
764 if (ctx
->pipe_framebuffer
.nr_cbufs
> i
&& !blend
[i
].no_colour
) {
766 batch
->draws
|= (PIPE_CLEAR_COLOR0
<< i
);
768 bool is_srgb
= (ctx
->pipe_framebuffer
.nr_cbufs
> i
) &&
769 (ctx
->pipe_framebuffer
.cbufs
[i
]) &&
770 util_format_is_srgb(ctx
->pipe_framebuffer
.cbufs
[i
]->format
);
772 SET_BIT(flags
, MALI_BLEND_MRT_SHADER
, blend
[i
].is_shader
);
773 SET_BIT(flags
, MALI_BLEND_LOAD_TIB
, !blend
[i
].no_blending
);
774 SET_BIT(flags
, MALI_BLEND_SRGB
, is_srgb
);
775 SET_BIT(flags
, MALI_BLEND_NO_DITHER
, !ctx
->blend
->base
.dither
);
778 if (dev
->quirks
& IS_BIFROST
) {
779 struct bifrost_blend_rt
*brts
= rts
;
781 brts
[i
].flags
= flags
;
783 if (blend
[i
].is_shader
) {
784 /* The blend shader's address needs to be at
785 * the same top 32 bit as the fragment shader.
786 * TODO: Ensure that's always the case.
788 assert((blend
[i
].shader
.gpu
& (0xffffffffull
<< 32)) ==
789 (fs
->bo
->gpu
& (0xffffffffull
<< 32)));
790 brts
[i
].shader
= blend
[i
].shader
.gpu
;
792 } else if (ctx
->pipe_framebuffer
.nr_cbufs
> i
) {
793 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
794 const struct util_format_description
*format_desc
;
795 format_desc
= util_format_description(format
);
797 brts
[i
].equation
= *blend
[i
].equation
.equation
;
799 /* TODO: this is a bit more complicated */
800 brts
[i
].constant
= blend
[i
].equation
.constant
;
802 brts
[i
].format
= panfrost_format_to_bifrost_blend(format_desc
);
804 /* 0x19 disables blending and forces REPLACE
805 * mode (equivalent to rgb_mode = alpha_mode =
806 * x122, colour mask = 0xF). 0x1a allows
808 brts
[i
].unk2
= blend
[i
].no_blending
? 0x19 : 0x1a;
810 brts
[i
].shader_type
= fs
->blend_types
[i
];
812 /* Dummy attachment for depth-only */
814 brts
[i
].shader_type
= fs
->blend_types
[i
];
817 struct midgard_blend_rt
*mrts
= rts
;
818 mrts
[i
].flags
= flags
;
820 if (blend
[i
].is_shader
) {
821 mrts
[i
].blend
.shader
= blend
[i
].shader
.gpu
| blend
[i
].shader
.first_tag
;
823 mrts
[i
].blend
.equation
= *blend
[i
].equation
.equation
;
824 mrts
[i
].blend
.constant
= blend
[i
].equation
.constant
;
831 panfrost_frag_shader_meta_init(struct panfrost_context
*ctx
,
832 struct mali_shader_meta
*fragmeta
,
835 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
836 struct panfrost_shader_state
*fs
;
838 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
840 bool msaa
= ctx
->rasterizer
&& ctx
->rasterizer
->base
.multisample
;
841 fragmeta
->coverage_mask
= (msaa
? ctx
->sample_mask
: ~0) & 0xF;
843 fragmeta
->unknown2_3
= MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS
) | 0x10;
844 fragmeta
->unknown2_4
= 0x4e0;
846 /* unknown2_4 has 0x10 bit set on T6XX and T720. We don't know why this
847 * is required (independent of 32-bit/64-bit descriptors), or why it's
848 * not used on later GPU revisions. Otherwise, all shader jobs fault on
849 * these earlier chips (perhaps this is a chicken bit of some kind).
850 * More investigation is needed. */
852 SET_BIT(fragmeta
->unknown2_4
, 0x10, dev
->quirks
& MIDGARD_SFBD
);
854 if (dev
->quirks
& IS_BIFROST
) {
857 /* Depending on whether it's legal to in the given shader, we try to
858 * enable early-z testing. TODO: respect e-z force */
860 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_EARLY_Z
,
861 !fs
->can_discard
&& !fs
->writes_global
&&
862 !fs
->writes_depth
&& !fs
->writes_stencil
&&
863 !ctx
->blend
->base
.alpha_to_coverage
);
865 /* Add the writes Z/S flags if needed. */
866 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_WRITES_Z
, fs
->writes_depth
);
867 SET_BIT(fragmeta
->midgard1
.flags_hi
, MALI_WRITES_S
, fs
->writes_stencil
);
869 /* Any time texturing is used, derivatives are implicitly calculated,
870 * so we need to enable helper invocations */
872 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_HELPER_INVOCATIONS
,
873 fs
->helper_invocations
);
875 const struct pipe_depth_stencil_alpha_state
*zsa
= ctx
->depth_stencil
;
877 bool depth_enabled
= fs
->writes_depth
||
878 (zsa
&& zsa
->depth
.enabled
&& zsa
->depth
.func
!= PIPE_FUNC_ALWAYS
);
880 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_READS_TILEBUFFER
,
881 fs
->outputs_read
|| (!depth_enabled
&& fs
->can_discard
));
882 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_READS_ZS
, depth_enabled
&& fs
->can_discard
);
885 panfrost_frag_meta_rasterizer_update(ctx
, fragmeta
);
886 panfrost_frag_meta_zsa_update(ctx
, fragmeta
);
887 panfrost_frag_meta_blend_update(ctx
, fragmeta
, rts
);
891 panfrost_emit_shader_meta(struct panfrost_batch
*batch
,
892 enum pipe_shader_type st
,
893 struct mali_vertex_tiler_postfix
*postfix
)
895 struct panfrost_context
*ctx
= batch
->ctx
;
896 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
903 struct mali_shader_meta meta
;
905 panfrost_shader_meta_init(ctx
, st
, &meta
);
907 /* Add the shader BO to the batch. */
908 panfrost_batch_add_bo(batch
, ss
->bo
,
909 PAN_BO_ACCESS_PRIVATE
|
911 panfrost_bo_access_for_stage(st
));
915 if (st
== PIPE_SHADER_FRAGMENT
) {
916 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
917 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
918 size_t desc_size
= sizeof(meta
);
920 struct panfrost_transfer xfer
;
923 if (dev
->quirks
& MIDGARD_SFBD
)
925 else if (dev
->quirks
& IS_BIFROST
)
926 rt_size
= sizeof(struct bifrost_blend_rt
);
928 rt_size
= sizeof(struct midgard_blend_rt
);
930 desc_size
+= rt_size
* rt_count
;
933 rts
= rzalloc_size(ctx
, rt_size
* rt_count
);
935 panfrost_frag_shader_meta_init(ctx
, &meta
, rts
);
937 xfer
= panfrost_pool_alloc(&batch
->pool
, desc_size
);
939 memcpy(xfer
.cpu
, &meta
, sizeof(meta
));
940 memcpy(xfer
.cpu
+ sizeof(meta
), rts
, rt_size
* rt_count
);
945 shader_ptr
= xfer
.gpu
;
947 shader_ptr
= panfrost_pool_upload(&batch
->pool
, &meta
,
951 postfix
->shader
= shader_ptr
;
955 panfrost_mali_viewport_init(struct panfrost_context
*ctx
,
956 struct mali_viewport
*mvp
)
958 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
960 /* Clip bounds are encoded as floats. The viewport itself is encoded as
961 * (somewhat) asymmetric ints. */
963 const struct pipe_scissor_state
*ss
= &ctx
->scissor
;
965 memset(mvp
, 0, sizeof(*mvp
));
967 /* By default, do no viewport clipping, i.e. clip to (-inf, inf) in
968 * each direction. Clipping to the viewport in theory should work, but
969 * in practice causes issues when we're not explicitly trying to
972 *mvp
= (struct mali_viewport
) {
973 .clip_minx
= -INFINITY
,
974 .clip_miny
= -INFINITY
,
975 .clip_maxx
= INFINITY
,
976 .clip_maxy
= INFINITY
,
979 /* Always scissor to the viewport by default. */
980 float vp_minx
= (int) (vp
->translate
[0] - fabsf(vp
->scale
[0]));
981 float vp_maxx
= (int) (vp
->translate
[0] + fabsf(vp
->scale
[0]));
983 float vp_miny
= (int) (vp
->translate
[1] - fabsf(vp
->scale
[1]));
984 float vp_maxy
= (int) (vp
->translate
[1] + fabsf(vp
->scale
[1]));
986 float minz
= (vp
->translate
[2] - fabsf(vp
->scale
[2]));
987 float maxz
= (vp
->translate
[2] + fabsf(vp
->scale
[2]));
989 /* Apply the scissor test */
991 unsigned minx
, miny
, maxx
, maxy
;
993 if (ss
&& ctx
->rasterizer
&& ctx
->rasterizer
->base
.scissor
) {
994 minx
= MAX2(ss
->minx
, vp_minx
);
995 miny
= MAX2(ss
->miny
, vp_miny
);
996 maxx
= MIN2(ss
->maxx
, vp_maxx
);
997 maxy
= MIN2(ss
->maxy
, vp_maxy
);
1005 /* Hardware needs the min/max to be strictly ordered, so flip if we
1006 * need to. The viewport transformation in the vertex shader will
1007 * handle the negatives if we don't */
1010 unsigned temp
= miny
;
1016 unsigned temp
= minx
;
1027 /* Clamp to the framebuffer size as a last check */
1029 minx
= MIN2(ctx
->pipe_framebuffer
.width
, minx
);
1030 maxx
= MIN2(ctx
->pipe_framebuffer
.width
, maxx
);
1032 miny
= MIN2(ctx
->pipe_framebuffer
.height
, miny
);
1033 maxy
= MIN2(ctx
->pipe_framebuffer
.height
, maxy
);
1037 mvp
->viewport0
[0] = minx
;
1038 mvp
->viewport1
[0] = MALI_POSITIVE(maxx
);
1040 mvp
->viewport0
[1] = miny
;
1041 mvp
->viewport1
[1] = MALI_POSITIVE(maxy
);
1043 bool clip_near
= true;
1044 bool clip_far
= true;
1046 if (ctx
->rasterizer
) {
1047 clip_near
= ctx
->rasterizer
->base
.depth_clip_near
;
1048 clip_far
= ctx
->rasterizer
->base
.depth_clip_far
;
1051 mvp
->clip_minz
= clip_near
? minz
: -INFINITY
;
1052 mvp
->clip_maxz
= clip_far
? maxz
: INFINITY
;
1056 panfrost_emit_viewport(struct panfrost_batch
*batch
,
1057 struct mali_vertex_tiler_postfix
*tiler_postfix
)
1059 struct panfrost_context
*ctx
= batch
->ctx
;
1060 struct mali_viewport mvp
;
1062 panfrost_mali_viewport_init(batch
->ctx
, &mvp
);
1064 /* Update the job, unless we're doing wallpapering (whose lack of
1065 * scissor we can ignore, since if we "miss" a tile of wallpaper, it'll
1066 * just... be faster :) */
1068 if (!ctx
->wallpaper_batch
)
1069 panfrost_batch_union_scissor(batch
, mvp
.viewport0
[0],
1071 mvp
.viewport1
[0] + 1,
1072 mvp
.viewport1
[1] + 1);
1074 tiler_postfix
->viewport
= panfrost_pool_upload(&batch
->pool
, &mvp
,
1079 panfrost_map_constant_buffer_gpu(struct panfrost_batch
*batch
,
1080 enum pipe_shader_type st
,
1081 struct panfrost_constant_buffer
*buf
,
1084 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1085 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1088 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1089 PAN_BO_ACCESS_SHARED
|
1090 PAN_BO_ACCESS_READ
|
1091 panfrost_bo_access_for_stage(st
));
1093 /* Alignment gauranteed by
1094 * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
1095 return rsrc
->bo
->gpu
+ cb
->buffer_offset
;
1096 } else if (cb
->user_buffer
) {
1097 return panfrost_pool_upload(&batch
->pool
,
1102 unreachable("No constant buffer");
1106 struct sysval_uniform
{
1116 panfrost_upload_viewport_scale_sysval(struct panfrost_batch
*batch
,
1117 struct sysval_uniform
*uniform
)
1119 struct panfrost_context
*ctx
= batch
->ctx
;
1120 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
1122 uniform
->f
[0] = vp
->scale
[0];
1123 uniform
->f
[1] = vp
->scale
[1];
1124 uniform
->f
[2] = vp
->scale
[2];
1128 panfrost_upload_viewport_offset_sysval(struct panfrost_batch
*batch
,
1129 struct sysval_uniform
*uniform
)
1131 struct panfrost_context
*ctx
= batch
->ctx
;
1132 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
1134 uniform
->f
[0] = vp
->translate
[0];
1135 uniform
->f
[1] = vp
->translate
[1];
1136 uniform
->f
[2] = vp
->translate
[2];
1139 static void panfrost_upload_txs_sysval(struct panfrost_batch
*batch
,
1140 enum pipe_shader_type st
,
1141 unsigned int sysvalid
,
1142 struct sysval_uniform
*uniform
)
1144 struct panfrost_context
*ctx
= batch
->ctx
;
1145 unsigned texidx
= PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid
);
1146 unsigned dim
= PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid
);
1147 bool is_array
= PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid
);
1148 struct pipe_sampler_view
*tex
= &ctx
->sampler_views
[st
][texidx
]->base
;
1151 uniform
->i
[0] = u_minify(tex
->texture
->width0
, tex
->u
.tex
.first_level
);
1154 uniform
->i
[1] = u_minify(tex
->texture
->height0
,
1155 tex
->u
.tex
.first_level
);
1158 uniform
->i
[2] = u_minify(tex
->texture
->depth0
,
1159 tex
->u
.tex
.first_level
);
1162 uniform
->i
[dim
] = tex
->texture
->array_size
;
1166 panfrost_upload_ssbo_sysval(struct panfrost_batch
*batch
,
1167 enum pipe_shader_type st
,
1169 struct sysval_uniform
*uniform
)
1171 struct panfrost_context
*ctx
= batch
->ctx
;
1173 assert(ctx
->ssbo_mask
[st
] & (1 << ssbo_id
));
1174 struct pipe_shader_buffer sb
= ctx
->ssbo
[st
][ssbo_id
];
1176 /* Compute address */
1177 struct panfrost_bo
*bo
= pan_resource(sb
.buffer
)->bo
;
1179 panfrost_batch_add_bo(batch
, bo
,
1180 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_RW
|
1181 panfrost_bo_access_for_stage(st
));
1183 /* Upload address and size as sysval */
1184 uniform
->du
[0] = bo
->gpu
+ sb
.buffer_offset
;
1185 uniform
->u
[2] = sb
.buffer_size
;
1189 panfrost_upload_sampler_sysval(struct panfrost_batch
*batch
,
1190 enum pipe_shader_type st
,
1192 struct sysval_uniform
*uniform
)
1194 struct panfrost_context
*ctx
= batch
->ctx
;
1195 struct pipe_sampler_state
*sampl
= &ctx
->samplers
[st
][samp_idx
]->base
;
1197 uniform
->f
[0] = sampl
->min_lod
;
1198 uniform
->f
[1] = sampl
->max_lod
;
1199 uniform
->f
[2] = sampl
->lod_bias
;
1201 /* Even without any errata, Midgard represents "no mipmapping" as
1202 * fixing the LOD with the clamps; keep behaviour consistent. c.f.
1203 * panfrost_create_sampler_state which also explains our choice of
1204 * epsilon value (again to keep behaviour consistent) */
1206 if (sampl
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
1207 uniform
->f
[1] = uniform
->f
[0] + (1.0/256.0);
1211 panfrost_upload_num_work_groups_sysval(struct panfrost_batch
*batch
,
1212 struct sysval_uniform
*uniform
)
1214 struct panfrost_context
*ctx
= batch
->ctx
;
1216 uniform
->u
[0] = ctx
->compute_grid
->grid
[0];
1217 uniform
->u
[1] = ctx
->compute_grid
->grid
[1];
1218 uniform
->u
[2] = ctx
->compute_grid
->grid
[2];
1222 panfrost_upload_sysvals(struct panfrost_batch
*batch
, void *buf
,
1223 struct panfrost_shader_state
*ss
,
1224 enum pipe_shader_type st
)
1226 struct sysval_uniform
*uniforms
= (void *)buf
;
1228 for (unsigned i
= 0; i
< ss
->sysval_count
; ++i
) {
1229 int sysval
= ss
->sysval
[i
];
1231 switch (PAN_SYSVAL_TYPE(sysval
)) {
1232 case PAN_SYSVAL_VIEWPORT_SCALE
:
1233 panfrost_upload_viewport_scale_sysval(batch
,
1236 case PAN_SYSVAL_VIEWPORT_OFFSET
:
1237 panfrost_upload_viewport_offset_sysval(batch
,
1240 case PAN_SYSVAL_TEXTURE_SIZE
:
1241 panfrost_upload_txs_sysval(batch
, st
,
1242 PAN_SYSVAL_ID(sysval
),
1245 case PAN_SYSVAL_SSBO
:
1246 panfrost_upload_ssbo_sysval(batch
, st
,
1247 PAN_SYSVAL_ID(sysval
),
1250 case PAN_SYSVAL_NUM_WORK_GROUPS
:
1251 panfrost_upload_num_work_groups_sysval(batch
,
1254 case PAN_SYSVAL_SAMPLER
:
1255 panfrost_upload_sampler_sysval(batch
, st
,
1256 PAN_SYSVAL_ID(sysval
),
1266 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer
*buf
,
1269 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1270 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1273 return rsrc
->bo
->cpu
;
1274 else if (cb
->user_buffer
)
1275 return cb
->user_buffer
;
1277 unreachable("No constant buffer");
1281 panfrost_emit_const_buf(struct panfrost_batch
*batch
,
1282 enum pipe_shader_type stage
,
1283 struct mali_vertex_tiler_postfix
*postfix
)
1285 struct panfrost_context
*ctx
= batch
->ctx
;
1286 struct panfrost_shader_variants
*all
= ctx
->shader
[stage
];
1291 struct panfrost_constant_buffer
*buf
= &ctx
->constant_buffer
[stage
];
1293 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1295 /* Uniforms are implicitly UBO #0 */
1296 bool has_uniforms
= buf
->enabled_mask
& (1 << 0);
1298 /* Allocate room for the sysval and the uniforms */
1299 size_t sys_size
= sizeof(float) * 4 * ss
->sysval_count
;
1300 size_t uniform_size
= has_uniforms
? (buf
->cb
[0].buffer_size
) : 0;
1301 size_t size
= sys_size
+ uniform_size
;
1302 struct panfrost_transfer transfer
= panfrost_pool_alloc(&batch
->pool
,
1305 /* Upload sysvals requested by the shader */
1306 panfrost_upload_sysvals(batch
, transfer
.cpu
, ss
, stage
);
1308 /* Upload uniforms */
1309 if (has_uniforms
&& uniform_size
) {
1310 const void *cpu
= panfrost_map_constant_buffer_cpu(buf
, 0);
1311 memcpy(transfer
.cpu
+ sys_size
, cpu
, uniform_size
);
1314 /* Next up, attach UBOs. UBO #0 is the uniforms we just
1317 unsigned ubo_count
= panfrost_ubo_count(ctx
, stage
);
1318 assert(ubo_count
>= 1);
1320 size_t sz
= sizeof(uint64_t) * ubo_count
;
1321 uint64_t ubos
[PAN_MAX_CONST_BUFFERS
];
1322 int uniform_count
= ss
->uniform_count
;
1324 /* Upload uniforms as a UBO */
1325 ubos
[0] = MALI_MAKE_UBO(2 + uniform_count
, transfer
.gpu
);
1327 /* The rest are honest-to-goodness UBOs */
1329 for (unsigned ubo
= 1; ubo
< ubo_count
; ++ubo
) {
1330 size_t usz
= buf
->cb
[ubo
].buffer_size
;
1331 bool enabled
= buf
->enabled_mask
& (1 << ubo
);
1332 bool empty
= usz
== 0;
1334 if (!enabled
|| empty
) {
1335 /* Stub out disabled UBOs to catch accesses */
1336 ubos
[ubo
] = MALI_MAKE_UBO(0, 0xDEAD0000);
1340 mali_ptr gpu
= panfrost_map_constant_buffer_gpu(batch
, stage
,
1343 unsigned bytes_per_field
= 16;
1344 unsigned aligned
= ALIGN_POT(usz
, bytes_per_field
);
1345 ubos
[ubo
] = MALI_MAKE_UBO(aligned
/ bytes_per_field
, gpu
);
1348 mali_ptr ubufs
= panfrost_pool_upload(&batch
->pool
, ubos
, sz
);
1349 postfix
->uniforms
= transfer
.gpu
;
1350 postfix
->uniform_buffers
= ubufs
;
1352 buf
->dirty_mask
= 0;
1356 panfrost_emit_shared_memory(struct panfrost_batch
*batch
,
1357 const struct pipe_grid_info
*info
,
1358 struct midgard_payload_vertex_tiler
*vtp
)
1360 struct panfrost_context
*ctx
= batch
->ctx
;
1361 struct panfrost_shader_variants
*all
= ctx
->shader
[PIPE_SHADER_COMPUTE
];
1362 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1363 unsigned single_size
= util_next_power_of_two(MAX2(ss
->shared_size
,
1365 unsigned shared_size
= single_size
* info
->grid
[0] * info
->grid
[1] *
1367 struct panfrost_bo
*bo
= panfrost_batch_get_shared_memory(batch
,
1371 struct mali_shared_memory shared
= {
1372 .shared_memory
= bo
->gpu
,
1373 .shared_workgroup_count
=
1374 util_logbase2_ceil(info
->grid
[0]) +
1375 util_logbase2_ceil(info
->grid
[1]) +
1376 util_logbase2_ceil(info
->grid
[2]),
1378 .shared_shift
= util_logbase2(single_size
) - 1
1381 vtp
->postfix
.shared_memory
= panfrost_pool_upload(&batch
->pool
, &shared
,
1386 panfrost_get_tex_desc(struct panfrost_batch
*batch
,
1387 enum pipe_shader_type st
,
1388 struct panfrost_sampler_view
*view
)
1391 return (mali_ptr
) 0;
1393 struct pipe_sampler_view
*pview
= &view
->base
;
1394 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1396 /* Add the BO to the job so it's retained until the job is done. */
1398 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1399 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1400 panfrost_bo_access_for_stage(st
));
1402 panfrost_batch_add_bo(batch
, view
->bo
,
1403 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1404 panfrost_bo_access_for_stage(st
));
1406 return view
->bo
->gpu
;
1410 panfrost_update_sampler_view(struct panfrost_sampler_view
*view
,
1411 struct pipe_context
*pctx
)
1413 struct panfrost_resource
*rsrc
= pan_resource(view
->base
.texture
);
1414 if (view
->texture_bo
!= rsrc
->bo
->gpu
||
1415 view
->layout
!= rsrc
->layout
) {
1416 panfrost_bo_unreference(view
->bo
);
1417 panfrost_create_sampler_view_bo(view
, pctx
, &rsrc
->base
);
1422 panfrost_emit_texture_descriptors(struct panfrost_batch
*batch
,
1423 enum pipe_shader_type stage
,
1424 struct mali_vertex_tiler_postfix
*postfix
)
1426 struct panfrost_context
*ctx
= batch
->ctx
;
1427 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1429 if (!ctx
->sampler_view_count
[stage
])
1432 if (device
->quirks
& IS_BIFROST
) {
1433 struct bifrost_texture_descriptor
*descriptors
;
1435 descriptors
= malloc(sizeof(struct bifrost_texture_descriptor
) *
1436 ctx
->sampler_view_count
[stage
]);
1438 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1439 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1440 struct pipe_sampler_view
*pview
= &view
->base
;
1441 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1442 panfrost_update_sampler_view(view
, &ctx
->base
);
1444 /* Add the BOs to the job so they are retained until the job is done. */
1446 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1447 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1448 panfrost_bo_access_for_stage(stage
));
1450 panfrost_batch_add_bo(batch
, view
->bo
,
1451 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1452 panfrost_bo_access_for_stage(stage
));
1454 memcpy(&descriptors
[i
], view
->bifrost_descriptor
, sizeof(*view
->bifrost_descriptor
));
1457 postfix
->textures
= panfrost_pool_upload(&batch
->pool
,
1459 sizeof(struct bifrost_texture_descriptor
) *
1460 ctx
->sampler_view_count
[stage
]);
1464 uint64_t trampolines
[PIPE_MAX_SHADER_SAMPLER_VIEWS
];
1466 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1467 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1469 panfrost_update_sampler_view(view
, &ctx
->base
);
1471 trampolines
[i
] = panfrost_get_tex_desc(batch
, stage
, view
);
1474 postfix
->textures
= panfrost_pool_upload(&batch
->pool
,
1477 ctx
->sampler_view_count
[stage
]);
1482 panfrost_emit_sampler_descriptors(struct panfrost_batch
*batch
,
1483 enum pipe_shader_type stage
,
1484 struct mali_vertex_tiler_postfix
*postfix
)
1486 struct panfrost_context
*ctx
= batch
->ctx
;
1487 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1489 if (!ctx
->sampler_count
[stage
])
1492 if (device
->quirks
& IS_BIFROST
) {
1493 size_t desc_size
= sizeof(struct bifrost_sampler_descriptor
);
1494 size_t transfer_size
= desc_size
* ctx
->sampler_count
[stage
];
1495 struct panfrost_transfer transfer
= panfrost_pool_alloc(&batch
->pool
,
1497 struct bifrost_sampler_descriptor
*desc
= (struct bifrost_sampler_descriptor
*)transfer
.cpu
;
1499 for (int i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1500 desc
[i
] = ctx
->samplers
[stage
][i
]->bifrost_hw
;
1502 postfix
->sampler_descriptor
= transfer
.gpu
;
1504 size_t desc_size
= sizeof(struct mali_sampler_descriptor
);
1505 size_t transfer_size
= desc_size
* ctx
->sampler_count
[stage
];
1506 struct panfrost_transfer transfer
= panfrost_pool_alloc(&batch
->pool
,
1508 struct mali_sampler_descriptor
*desc
= (struct mali_sampler_descriptor
*)transfer
.cpu
;
1510 for (int i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1511 desc
[i
] = ctx
->samplers
[stage
][i
]->midgard_hw
;
1513 postfix
->sampler_descriptor
= transfer
.gpu
;
1518 panfrost_emit_vertex_attr_meta(struct panfrost_batch
*batch
,
1519 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1521 struct panfrost_context
*ctx
= batch
->ctx
;
1526 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1528 panfrost_vertex_state_upd_attr_offs(ctx
, vertex_postfix
);
1529 vertex_postfix
->attribute_meta
= panfrost_pool_upload(&batch
->pool
, so
->hw
,
1535 panfrost_emit_vertex_data(struct panfrost_batch
*batch
,
1536 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1538 struct panfrost_context
*ctx
= batch
->ctx
;
1539 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1541 /* Staged mali_attr, and index into them. i =/= k, depending on the
1542 * vertex buffer mask and instancing. Twice as much room is allocated,
1543 * for a worst case of NPOT_DIVIDEs which take up extra slot */
1544 union mali_attr attrs
[PIPE_MAX_ATTRIBS
* 2];
1547 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1548 /* We map a mali_attr to be 1:1 with the mali_attr_meta, which
1549 * means duplicating some vertex buffers (who cares? aside from
1550 * maybe some caching implications but I somehow doubt that
1553 struct pipe_vertex_element
*elem
= &so
->pipe
[i
];
1554 unsigned vbi
= elem
->vertex_buffer_index
;
1556 /* The exception to 1:1 mapping is that we can have multiple
1557 * entries (NPOT divisors), so we fixup anyways */
1559 so
->hw
[i
].index
= k
;
1561 if (!(ctx
->vb_mask
& (1 << vbi
)))
1564 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1565 struct panfrost_resource
*rsrc
;
1567 rsrc
= pan_resource(buf
->buffer
.resource
);
1571 /* Align to 64 bytes by masking off the lower bits. This
1572 * will be adjusted back when we fixup the src_offset in
1575 mali_ptr raw_addr
= rsrc
->bo
->gpu
+ buf
->buffer_offset
;
1576 mali_ptr addr
= raw_addr
& ~63;
1577 unsigned chopped_addr
= raw_addr
- addr
;
1579 /* Add a dependency of the batch on the vertex buffer */
1580 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1581 PAN_BO_ACCESS_SHARED
|
1582 PAN_BO_ACCESS_READ
|
1583 PAN_BO_ACCESS_VERTEX_TILER
);
1585 /* Set common fields */
1586 attrs
[k
].elements
= addr
;
1587 attrs
[k
].stride
= buf
->stride
;
1589 /* Since we advanced the base pointer, we shrink the buffer
1591 attrs
[k
].size
= rsrc
->base
.width0
- buf
->buffer_offset
;
1593 /* We need to add the extra size we masked off (for
1594 * correctness) so the data doesn't get clamped away */
1595 attrs
[k
].size
+= chopped_addr
;
1597 /* For non-instancing make sure we initialize */
1598 attrs
[k
].shift
= attrs
[k
].extra_flags
= 0;
1600 /* Instancing uses a dramatically different code path than
1601 * linear, so dispatch for the actual emission now that the
1602 * common code is finished */
1604 unsigned divisor
= elem
->instance_divisor
;
1606 if (divisor
&& ctx
->instance_count
== 1) {
1607 /* Silly corner case where there's a divisor(=1) but
1608 * there's no legitimate instancing. So we want *every*
1609 * attribute to be the same. So set stride to zero so
1610 * we don't go anywhere. */
1612 attrs
[k
].size
= attrs
[k
].stride
+ chopped_addr
;
1613 attrs
[k
].stride
= 0;
1614 attrs
[k
++].elements
|= MALI_ATTR_LINEAR
;
1615 } else if (ctx
->instance_count
<= 1) {
1616 /* Normal, non-instanced attributes */
1617 attrs
[k
++].elements
|= MALI_ATTR_LINEAR
;
1619 unsigned instance_shift
= vertex_postfix
->instance_shift
;
1620 unsigned instance_odd
= vertex_postfix
->instance_odd
;
1622 k
+= panfrost_vertex_instanced(ctx
->padded_count
,
1625 divisor
, &attrs
[k
]);
1629 /* Add special gl_VertexID/gl_InstanceID buffers */
1631 panfrost_vertex_id(ctx
->padded_count
, &attrs
[k
]);
1632 so
->hw
[PAN_VERTEX_ID
].index
= k
++;
1633 panfrost_instance_id(ctx
->padded_count
, &attrs
[k
]);
1634 so
->hw
[PAN_INSTANCE_ID
].index
= k
++;
1636 /* Upload whatever we emitted and go */
1638 vertex_postfix
->attributes
= panfrost_pool_upload(&batch
->pool
, attrs
,
1639 k
* sizeof(*attrs
));
1643 panfrost_emit_varyings(struct panfrost_batch
*batch
, union mali_attr
*slot
,
1644 unsigned stride
, unsigned count
)
1646 /* Fill out the descriptor */
1647 slot
->stride
= stride
;
1648 slot
->size
= stride
* count
;
1649 slot
->shift
= slot
->extra_flags
= 0;
1651 struct panfrost_transfer transfer
= panfrost_pool_alloc(&batch
->pool
,
1654 slot
->elements
= transfer
.gpu
| MALI_ATTR_LINEAR
;
1656 return transfer
.gpu
;
1660 panfrost_streamout_offset(unsigned stride
, unsigned offset
,
1661 struct pipe_stream_output_target
*target
)
1663 return (target
->buffer_offset
+ (offset
* stride
* 4)) & 63;
1667 panfrost_emit_streamout(struct panfrost_batch
*batch
, union mali_attr
*slot
,
1668 unsigned stride
, unsigned offset
, unsigned count
,
1669 struct pipe_stream_output_target
*target
)
1671 /* Fill out the descriptor */
1672 slot
->stride
= stride
* 4;
1673 slot
->shift
= slot
->extra_flags
= 0;
1675 unsigned max_size
= target
->buffer_size
;
1676 unsigned expected_size
= slot
->stride
* count
;
1678 /* Grab the BO and bind it to the batch */
1679 struct panfrost_bo
*bo
= pan_resource(target
->buffer
)->bo
;
1681 /* Varyings are WRITE from the perspective of the VERTEX but READ from
1682 * the perspective of the TILER and FRAGMENT.
1684 panfrost_batch_add_bo(batch
, bo
,
1685 PAN_BO_ACCESS_SHARED
|
1687 PAN_BO_ACCESS_VERTEX_TILER
|
1688 PAN_BO_ACCESS_FRAGMENT
);
1690 /* We will have an offset applied to get alignment */
1691 mali_ptr addr
= bo
->gpu
+ target
->buffer_offset
+ (offset
* slot
->stride
);
1692 slot
->elements
= (addr
& ~63) | MALI_ATTR_LINEAR
;
1693 slot
->size
= MIN2(max_size
, expected_size
) + (addr
& 63);
1697 has_point_coord(unsigned mask
, gl_varying_slot loc
)
1699 if ((loc
>= VARYING_SLOT_TEX0
) && (loc
<= VARYING_SLOT_TEX7
))
1700 return (mask
& (1 << (loc
- VARYING_SLOT_TEX0
)));
1701 else if (loc
== VARYING_SLOT_PNTC
)
1702 return (mask
& (1 << 8));
1707 /* Helpers for manipulating stream out information so we can pack varyings
1708 * accordingly. Compute the src_offset for a given captured varying */
1710 static struct pipe_stream_output
*
1711 pan_get_so(struct pipe_stream_output_info
*info
, gl_varying_slot loc
)
1713 for (unsigned i
= 0; i
< info
->num_outputs
; ++i
) {
1714 if (info
->output
[i
].register_index
== loc
)
1715 return &info
->output
[i
];
1718 unreachable("Varying not captured");
1722 pan_varying_size(enum mali_format fmt
)
1724 unsigned type
= MALI_EXTRACT_TYPE(fmt
);
1725 unsigned chan
= MALI_EXTRACT_CHANNELS(fmt
);
1726 unsigned bits
= MALI_EXTRACT_BITS(fmt
);
1729 if (bits
== MALI_CHANNEL_FLOAT
) {
1731 bool fp16
= (type
== MALI_FORMAT_SINT
);
1732 assert(fp16
|| (type
== MALI_FORMAT_UNORM
));
1736 assert(type
>= MALI_FORMAT_SNORM
&& type
<= MALI_FORMAT_SINT
);
1747 /* Indices for named (non-XFB) varyings that are present. These are packed
1748 * tightly so they correspond to a bitfield present (P) indexed by (1 <<
1749 * PAN_VARY_*). This has the nice property that you can lookup the buffer index
1750 * of a given special field given a shift S by:
1752 * idx = popcount(P & ((1 << S) - 1))
1754 * That is... look at all of the varyings that come earlier and count them, the
1755 * count is the new index since plus one. Likewise, the total number of special
1756 * buffers required is simply popcount(P)
1759 enum pan_special_varying
{
1760 PAN_VARY_GENERAL
= 0,
1761 PAN_VARY_POSITION
= 1,
1763 PAN_VARY_PNTCOORD
= 3,
1765 PAN_VARY_FRAGCOORD
= 5,
1771 /* Given a varying, figure out which index it correpsonds to */
1773 static inline unsigned
1774 pan_varying_index(unsigned present
, enum pan_special_varying v
)
1776 unsigned mask
= (1 << v
) - 1;
1777 return util_bitcount(present
& mask
);
1780 /* Get the base offset for XFB buffers, which by convention come after
1781 * everything else. Wrapper function for semantic reasons; by construction this
1782 * is just popcount. */
1784 static inline unsigned
1785 pan_xfb_base(unsigned present
)
1787 return util_bitcount(present
);
1790 /* Computes the present mask for varyings so we can start emitting varying records */
1792 static inline unsigned
1793 pan_varying_present(
1794 struct panfrost_shader_state
*vs
,
1795 struct panfrost_shader_state
*fs
,
1798 /* At the moment we always emit general and position buffers. Not
1799 * strictly necessary but usually harmless */
1801 unsigned present
= (1 << PAN_VARY_GENERAL
) | (1 << PAN_VARY_POSITION
);
1803 /* Enable special buffers by the shader info */
1805 if (vs
->writes_point_size
)
1806 present
|= (1 << PAN_VARY_PSIZ
);
1808 if (fs
->reads_point_coord
)
1809 present
|= (1 << PAN_VARY_PNTCOORD
);
1812 present
|= (1 << PAN_VARY_FACE
);
1814 if (fs
->reads_frag_coord
&& !(quirks
& IS_BIFROST
))
1815 present
|= (1 << PAN_VARY_FRAGCOORD
);
1817 /* Also, if we have a point sprite, we need a point coord buffer */
1819 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1820 gl_varying_slot loc
= fs
->varyings_loc
[i
];
1822 if (has_point_coord(fs
->point_sprite_mask
, loc
))
1823 present
|= (1 << PAN_VARY_PNTCOORD
);
1829 /* Emitters for varying records */
1831 static struct mali_attr_meta
1832 pan_emit_vary(unsigned present
, enum pan_special_varying buf
,
1833 unsigned quirks
, enum mali_format format
,
1836 unsigned nr_channels
= MALI_EXTRACT_CHANNELS(format
);
1838 struct mali_attr_meta meta
= {
1839 .index
= pan_varying_index(present
, buf
),
1840 .unknown1
= quirks
& IS_BIFROST
? 0x0 : 0x2,
1841 .swizzle
= quirks
& HAS_SWIZZLES
?
1842 panfrost_get_default_swizzle(nr_channels
) :
1843 panfrost_bifrost_swizzle(nr_channels
),
1845 .src_offset
= offset
1851 /* General varying that is unused */
1853 static struct mali_attr_meta
1854 pan_emit_vary_only(unsigned present
, unsigned quirks
)
1856 return pan_emit_vary(present
, 0, quirks
, MALI_VARYING_DISCARD
, 0);
1859 /* Special records */
1861 static const enum mali_format pan_varying_formats
[PAN_VARY_MAX
] = {
1862 [PAN_VARY_POSITION
] = MALI_VARYING_POS
,
1863 [PAN_VARY_PSIZ
] = MALI_R16F
,
1864 [PAN_VARY_PNTCOORD
] = MALI_R16F
,
1865 [PAN_VARY_FACE
] = MALI_R32I
,
1866 [PAN_VARY_FRAGCOORD
] = MALI_RGBA32F
1869 static struct mali_attr_meta
1870 pan_emit_vary_special(unsigned present
, enum pan_special_varying buf
,
1873 assert(buf
< PAN_VARY_MAX
);
1874 return pan_emit_vary(present
, buf
, quirks
, pan_varying_formats
[buf
], 0);
1877 static enum mali_format
1878 pan_xfb_format(enum mali_format format
, unsigned nr
)
1880 if (MALI_EXTRACT_BITS(format
) == MALI_CHANNEL_FLOAT
)
1881 return MALI_R32F
| MALI_NR_CHANNELS(nr
);
1883 return MALI_EXTRACT_TYPE(format
) | MALI_NR_CHANNELS(nr
) | MALI_CHANNEL_32
;
1886 /* Transform feedback records. Note struct pipe_stream_output is (if packed as
1887 * a bitfield) 32-bit, smaller than a 64-bit pointer, so may as well pass by
1890 static struct mali_attr_meta
1891 pan_emit_vary_xfb(unsigned present
,
1893 unsigned *streamout_offsets
,
1895 enum mali_format format
,
1896 struct pipe_stream_output o
)
1898 /* Otherwise construct a record for it */
1899 struct mali_attr_meta meta
= {
1900 /* XFB buffers come after everything else */
1901 .index
= pan_xfb_base(present
) + o
.output_buffer
,
1903 /* As usual unknown bit */
1904 .unknown1
= quirks
& IS_BIFROST
? 0x0 : 0x2,
1906 /* Override swizzle with number of channels */
1907 .swizzle
= quirks
& HAS_SWIZZLES
?
1908 panfrost_get_default_swizzle(o
.num_components
) :
1909 panfrost_bifrost_swizzle(o
.num_components
),
1911 /* Override number of channels and precision to highp */
1912 .format
= pan_xfb_format(format
, o
.num_components
),
1914 /* Apply given offsets together */
1915 .src_offset
= (o
.dst_offset
* 4) /* dwords */
1916 + streamout_offsets
[o
.output_buffer
]
1922 /* Determine if we should capture a varying for XFB. This requires actually
1923 * having a buffer for it. If we don't capture it, we'll fallback to a general
1924 * varying path (linked or unlinked, possibly discarding the write) */
1927 panfrost_xfb_captured(struct panfrost_shader_state
*xfb
,
1928 unsigned loc
, unsigned max_xfb
)
1930 if (!(xfb
->so_mask
& (1ll << loc
)))
1933 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1934 return o
->output_buffer
< max_xfb
;
1937 /* Higher-level wrapper around all of the above, classifying a varying into one
1938 * of the above types */
1940 static struct mali_attr_meta
1941 panfrost_emit_varying(
1942 struct panfrost_shader_state
*stage
,
1943 struct panfrost_shader_state
*other
,
1944 struct panfrost_shader_state
*xfb
,
1947 unsigned *streamout_offsets
,
1949 unsigned *gen_offsets
,
1950 enum mali_format
*gen_formats
,
1951 unsigned *gen_stride
,
1956 gl_varying_slot loc
= stage
->varyings_loc
[idx
];
1957 enum mali_format format
= stage
->varyings
[idx
];
1959 /* Override format to match linkage */
1960 if (!should_alloc
&& gen_formats
[idx
])
1961 format
= gen_formats
[idx
];
1963 if (has_point_coord(stage
->point_sprite_mask
, loc
)) {
1964 return pan_emit_vary_special(present
, PAN_VARY_PNTCOORD
, quirks
);
1965 } else if (panfrost_xfb_captured(xfb
, loc
, max_xfb
)) {
1966 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1967 return pan_emit_vary_xfb(present
, max_xfb
, streamout_offsets
, quirks
, format
, *o
);
1968 } else if (loc
== VARYING_SLOT_POS
) {
1970 return pan_emit_vary_special(present
, PAN_VARY_FRAGCOORD
, quirks
);
1972 return pan_emit_vary_special(present
, PAN_VARY_POSITION
, quirks
);
1973 } else if (loc
== VARYING_SLOT_PSIZ
) {
1974 return pan_emit_vary_special(present
, PAN_VARY_PSIZ
, quirks
);
1975 } else if (loc
== VARYING_SLOT_PNTC
) {
1976 return pan_emit_vary_special(present
, PAN_VARY_PNTCOORD
, quirks
);
1977 } else if (loc
== VARYING_SLOT_FACE
) {
1978 return pan_emit_vary_special(present
, PAN_VARY_FACE
, quirks
);
1981 /* We've exhausted special cases, so it's otherwise a general varying. Check if we're linked */
1982 signed other_idx
= -1;
1984 for (unsigned j
= 0; j
< other
->varying_count
; ++j
) {
1985 if (other
->varyings_loc
[j
] == loc
) {
1992 return pan_emit_vary_only(present
, quirks
);
1994 unsigned offset
= gen_offsets
[other_idx
];
1997 /* We're linked, so allocate a space via a watermark allocation */
1998 enum mali_format alt
= other
->varyings
[other_idx
];
2000 /* Do interpolation at minimum precision */
2001 unsigned size_main
= pan_varying_size(format
);
2002 unsigned size_alt
= pan_varying_size(alt
);
2003 unsigned size
= MIN2(size_main
, size_alt
);
2005 /* If a varying is marked for XFB but not actually captured, we
2006 * should match the format to the format that would otherwise
2007 * be used for XFB, since dEQP checks for invariance here. It's
2008 * unclear if this is required by the spec. */
2010 if (xfb
->so_mask
& (1ull << loc
)) {
2011 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
2012 format
= pan_xfb_format(format
, o
->num_components
);
2013 size
= pan_varying_size(format
);
2014 } else if (size
== size_alt
) {
2018 gen_offsets
[idx
] = *gen_stride
;
2019 gen_formats
[other_idx
] = format
;
2020 offset
= *gen_stride
;
2021 *gen_stride
+= size
;
2024 return pan_emit_vary(present
, PAN_VARY_GENERAL
,
2025 quirks
, format
, offset
);
2029 pan_emit_special_input(union mali_attr
*varyings
,
2031 enum pan_special_varying v
,
2034 if (present
& (1 << v
)) {
2035 /* Ensure we write exactly once for performance and with fields
2036 * zeroed appropriately to avoid flakes */
2038 union mali_attr s
= {
2042 varyings
[pan_varying_index(present
, v
)] = s
;
2047 panfrost_emit_varying_descriptor(struct panfrost_batch
*batch
,
2048 unsigned vertex_count
,
2049 struct mali_vertex_tiler_postfix
*vertex_postfix
,
2050 struct mali_vertex_tiler_postfix
*tiler_postfix
,
2051 union midgard_primitive_size
*primitive_size
)
2053 /* Load the shaders */
2054 struct panfrost_context
*ctx
= batch
->ctx
;
2055 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
2056 struct panfrost_shader_state
*vs
, *fs
;
2057 size_t vs_size
, fs_size
;
2059 /* Allocate the varying descriptor */
2061 vs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_VERTEX
);
2062 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
2063 vs_size
= sizeof(struct mali_attr_meta
) * vs
->varying_count
;
2064 fs_size
= sizeof(struct mali_attr_meta
) * fs
->varying_count
;
2066 struct panfrost_transfer trans
= panfrost_pool_alloc(&batch
->pool
,
2070 struct pipe_stream_output_info
*so
= &vs
->stream_output
;
2071 unsigned present
= pan_varying_present(vs
, fs
, dev
->quirks
);
2073 /* Check if this varying is linked by us. This is the case for
2074 * general-purpose, non-captured varyings. If it is, link it. If it's
2075 * not, use the provided stream out information to determine the
2076 * offset, since it was already linked for us. */
2078 unsigned gen_offsets
[32];
2079 enum mali_format gen_formats
[32];
2080 memset(gen_offsets
, 0, sizeof(gen_offsets
));
2081 memset(gen_formats
, 0, sizeof(gen_formats
));
2083 unsigned gen_stride
= 0;
2084 assert(vs
->varying_count
< ARRAY_SIZE(gen_offsets
));
2085 assert(fs
->varying_count
< ARRAY_SIZE(gen_offsets
));
2087 unsigned streamout_offsets
[32];
2089 for (unsigned i
= 0; i
< ctx
->streamout
.num_targets
; ++i
) {
2090 streamout_offsets
[i
] = panfrost_streamout_offset(
2092 ctx
->streamout
.offsets
[i
],
2093 ctx
->streamout
.targets
[i
]);
2096 struct mali_attr_meta
*ovs
= (struct mali_attr_meta
*)trans
.cpu
;
2097 struct mali_attr_meta
*ofs
= ovs
+ vs
->varying_count
;
2099 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
2100 ovs
[i
] = panfrost_emit_varying(vs
, fs
, vs
, present
,
2101 ctx
->streamout
.num_targets
, streamout_offsets
,
2103 gen_offsets
, gen_formats
, &gen_stride
, i
, true, false);
2106 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
2107 ofs
[i
] = panfrost_emit_varying(fs
, vs
, vs
, present
,
2108 ctx
->streamout
.num_targets
, streamout_offsets
,
2110 gen_offsets
, gen_formats
, &gen_stride
, i
, false, true);
2113 unsigned xfb_base
= pan_xfb_base(present
);
2114 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
,
2115 sizeof(union mali_attr
) * (xfb_base
+ ctx
->streamout
.num_targets
));
2116 union mali_attr
*varyings
= (union mali_attr
*) T
.cpu
;
2118 /* Emit the stream out buffers */
2120 unsigned out_count
= u_stream_outputs_for_vertices(ctx
->active_prim
,
2123 for (unsigned i
= 0; i
< ctx
->streamout
.num_targets
; ++i
) {
2124 panfrost_emit_streamout(batch
, &varyings
[xfb_base
+ i
],
2126 ctx
->streamout
.offsets
[i
],
2128 ctx
->streamout
.targets
[i
]);
2131 panfrost_emit_varyings(batch
,
2132 &varyings
[pan_varying_index(present
, PAN_VARY_GENERAL
)],
2133 gen_stride
, vertex_count
);
2135 /* fp32 vec4 gl_Position */
2136 tiler_postfix
->position_varying
= panfrost_emit_varyings(batch
,
2137 &varyings
[pan_varying_index(present
, PAN_VARY_POSITION
)],
2138 sizeof(float) * 4, vertex_count
);
2140 if (present
& (1 << PAN_VARY_PSIZ
)) {
2141 primitive_size
->pointer
= panfrost_emit_varyings(batch
,
2142 &varyings
[pan_varying_index(present
, PAN_VARY_PSIZ
)],
2146 pan_emit_special_input(varyings
, present
, PAN_VARY_PNTCOORD
, MALI_VARYING_POINT_COORD
);
2147 pan_emit_special_input(varyings
, present
, PAN_VARY_FACE
, MALI_VARYING_FRONT_FACING
);
2148 pan_emit_special_input(varyings
, present
, PAN_VARY_FRAGCOORD
, MALI_VARYING_FRAG_COORD
);
2150 vertex_postfix
->varyings
= T
.gpu
;
2151 tiler_postfix
->varyings
= T
.gpu
;
2153 vertex_postfix
->varying_meta
= trans
.gpu
;
2154 tiler_postfix
->varying_meta
= trans
.gpu
+ vs_size
;
2158 panfrost_emit_vertex_tiler_jobs(struct panfrost_batch
*batch
,
2159 struct mali_vertex_tiler_prefix
*vertex_prefix
,
2160 struct mali_vertex_tiler_postfix
*vertex_postfix
,
2161 struct mali_vertex_tiler_prefix
*tiler_prefix
,
2162 struct mali_vertex_tiler_postfix
*tiler_postfix
,
2163 union midgard_primitive_size
*primitive_size
)
2165 struct panfrost_context
*ctx
= batch
->ctx
;
2166 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
2167 bool wallpapering
= ctx
->wallpaper_batch
&& batch
->scoreboard
.tiler_dep
;
2168 struct bifrost_payload_vertex bifrost_vertex
= {0,};
2169 struct bifrost_payload_tiler bifrost_tiler
= {0,};
2170 struct midgard_payload_vertex_tiler midgard_vertex
= {0,};
2171 struct midgard_payload_vertex_tiler midgard_tiler
= {0,};
2173 size_t vp_size
, tp_size
;
2175 if (device
->quirks
& IS_BIFROST
) {
2176 bifrost_vertex
.prefix
= *vertex_prefix
;
2177 bifrost_vertex
.postfix
= *vertex_postfix
;
2178 vp
= &bifrost_vertex
;
2179 vp_size
= sizeof(bifrost_vertex
);
2181 bifrost_tiler
.prefix
= *tiler_prefix
;
2182 bifrost_tiler
.tiler
.primitive_size
= *primitive_size
;
2183 bifrost_tiler
.tiler
.tiler_meta
= panfrost_batch_get_tiler_meta(batch
, ~0);
2184 bifrost_tiler
.postfix
= *tiler_postfix
;
2185 tp
= &bifrost_tiler
;
2186 tp_size
= sizeof(bifrost_tiler
);
2188 midgard_vertex
.prefix
= *vertex_prefix
;
2189 midgard_vertex
.postfix
= *vertex_postfix
;
2190 vp
= &midgard_vertex
;
2191 vp_size
= sizeof(midgard_vertex
);
2193 midgard_tiler
.prefix
= *tiler_prefix
;
2194 midgard_tiler
.postfix
= *tiler_postfix
;
2195 midgard_tiler
.primitive_size
= *primitive_size
;
2196 tp
= &midgard_tiler
;
2197 tp_size
= sizeof(midgard_tiler
);
2201 /* Inject in reverse order, with "predicted" job indices.
2202 * THIS IS A HACK XXX */
2203 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, JOB_TYPE_TILER
, false,
2204 batch
->scoreboard
.job_index
+ 2, tp
, tp_size
, true);
2205 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, JOB_TYPE_VERTEX
, false, 0,
2210 /* If rasterizer discard is enable, only submit the vertex */
2212 bool rasterizer_discard
= ctx
->rasterizer
&&
2213 ctx
->rasterizer
->base
.rasterizer_discard
;
2215 unsigned vertex
= panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, JOB_TYPE_VERTEX
, false, 0,
2216 vp
, vp_size
, false);
2218 if (rasterizer_discard
)
2221 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, JOB_TYPE_TILER
, false, vertex
, tp
, tp_size
,
2225 /* TODO: stop hardcoding this */
2227 panfrost_emit_sample_locations(struct panfrost_batch
*batch
)
2229 uint16_t locations
[] = {
2280 return panfrost_pool_upload(&batch
->pool
, locations
, 96 * sizeof(uint16_t));