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 meta
->bifrost1
.unk1
= 0x800200;
331 meta
->bifrost1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
332 meta
->bifrost2
.preload_regs
= 0xC0;
333 meta
->bifrost2
.uniform_count
= MIN2(ss
->uniform_count
,
336 meta
->midgard1
.uniform_count
= MIN2(ss
->uniform_count
,
338 meta
->midgard1
.work_count
= ss
->work_reg_count
;
339 meta
->midgard1
.flags_hi
= 0x8; /* XXX */
340 meta
->midgard1
.flags_lo
= 0x220;
341 meta
->midgard1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
347 panfrost_translate_compare_func(enum pipe_compare_func in
)
350 case PIPE_FUNC_NEVER
:
351 return MALI_FUNC_NEVER
;
354 return MALI_FUNC_LESS
;
356 case PIPE_FUNC_EQUAL
:
357 return MALI_FUNC_EQUAL
;
359 case PIPE_FUNC_LEQUAL
:
360 return MALI_FUNC_LEQUAL
;
362 case PIPE_FUNC_GREATER
:
363 return MALI_FUNC_GREATER
;
365 case PIPE_FUNC_NOTEQUAL
:
366 return MALI_FUNC_NOTEQUAL
;
368 case PIPE_FUNC_GEQUAL
:
369 return MALI_FUNC_GEQUAL
;
371 case PIPE_FUNC_ALWAYS
:
372 return MALI_FUNC_ALWAYS
;
375 unreachable("Invalid func");
380 panfrost_translate_stencil_op(enum pipe_stencil_op in
)
383 case PIPE_STENCIL_OP_KEEP
:
384 return MALI_STENCIL_KEEP
;
386 case PIPE_STENCIL_OP_ZERO
:
387 return MALI_STENCIL_ZERO
;
389 case PIPE_STENCIL_OP_REPLACE
:
390 return MALI_STENCIL_REPLACE
;
392 case PIPE_STENCIL_OP_INCR
:
393 return MALI_STENCIL_INCR
;
395 case PIPE_STENCIL_OP_DECR
:
396 return MALI_STENCIL_DECR
;
398 case PIPE_STENCIL_OP_INCR_WRAP
:
399 return MALI_STENCIL_INCR_WRAP
;
401 case PIPE_STENCIL_OP_DECR_WRAP
:
402 return MALI_STENCIL_DECR_WRAP
;
404 case PIPE_STENCIL_OP_INVERT
:
405 return MALI_STENCIL_INVERT
;
408 unreachable("Invalid stencil op");
413 translate_tex_wrap(enum pipe_tex_wrap w
)
416 case PIPE_TEX_WRAP_REPEAT
:
417 return MALI_WRAP_REPEAT
;
419 case PIPE_TEX_WRAP_CLAMP
:
420 return MALI_WRAP_CLAMP
;
422 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
:
423 return MALI_WRAP_CLAMP_TO_EDGE
;
425 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
:
426 return MALI_WRAP_CLAMP_TO_BORDER
;
428 case PIPE_TEX_WRAP_MIRROR_REPEAT
:
429 return MALI_WRAP_MIRRORED_REPEAT
;
431 case PIPE_TEX_WRAP_MIRROR_CLAMP
:
432 return MALI_WRAP_MIRRORED_CLAMP
;
434 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
:
435 return MALI_WRAP_MIRRORED_CLAMP_TO_EDGE
;
437 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
:
438 return MALI_WRAP_MIRRORED_CLAMP_TO_BORDER
;
441 unreachable("Invalid wrap");
445 void panfrost_sampler_desc_init(const struct pipe_sampler_state
*cso
,
446 struct mali_sampler_descriptor
*hw
)
448 unsigned func
= panfrost_translate_compare_func(cso
->compare_func
);
449 bool min_nearest
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
;
450 bool mag_nearest
= cso
->mag_img_filter
== PIPE_TEX_FILTER_NEAREST
;
451 bool mip_linear
= cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
;
452 unsigned min_filter
= min_nearest
? MALI_SAMP_MIN_NEAREST
: 0;
453 unsigned mag_filter
= mag_nearest
? MALI_SAMP_MAG_NEAREST
: 0;
454 unsigned mip_filter
= mip_linear
?
455 (MALI_SAMP_MIP_LINEAR_1
| MALI_SAMP_MIP_LINEAR_2
) : 0;
456 unsigned normalized
= cso
->normalized_coords
? MALI_SAMP_NORM_COORDS
: 0;
458 *hw
= (struct mali_sampler_descriptor
) {
459 .filter_mode
= min_filter
| mag_filter
| mip_filter
|
461 .wrap_s
= translate_tex_wrap(cso
->wrap_s
),
462 .wrap_t
= translate_tex_wrap(cso
->wrap_t
),
463 .wrap_r
= translate_tex_wrap(cso
->wrap_r
),
464 .compare_func
= panfrost_flip_compare_func(func
),
466 cso
->border_color
.f
[0],
467 cso
->border_color
.f
[1],
468 cso
->border_color
.f
[2],
469 cso
->border_color
.f
[3]
471 .min_lod
= FIXED_16(cso
->min_lod
, false), /* clamp at 0 */
472 .max_lod
= FIXED_16(cso
->max_lod
, false),
473 .lod_bias
= FIXED_16(cso
->lod_bias
, true), /* can be negative */
474 .seamless_cube_map
= cso
->seamless_cube_map
,
477 /* If necessary, we disable mipmapping in the sampler descriptor by
478 * clamping the LOD as tight as possible (from 0 to epsilon,
479 * essentially -- remember these are fixed point numbers, so
482 if (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
483 hw
->max_lod
= hw
->min_lod
+ 1;
487 panfrost_make_stencil_state(const struct pipe_stencil_state
*in
,
488 struct mali_stencil_test
*out
)
490 out
->ref
= 0; /* Gallium gets it from elsewhere */
492 out
->mask
= in
->valuemask
;
493 out
->func
= panfrost_translate_compare_func(in
->func
);
494 out
->sfail
= panfrost_translate_stencil_op(in
->fail_op
);
495 out
->dpfail
= panfrost_translate_stencil_op(in
->zfail_op
);
496 out
->dppass
= panfrost_translate_stencil_op(in
->zpass_op
);
500 panfrost_frag_meta_rasterizer_update(struct panfrost_context
*ctx
,
501 struct mali_shader_meta
*fragmeta
)
503 if (!ctx
->rasterizer
) {
504 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, true);
505 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, false);
506 fragmeta
->depth_units
= 0.0f
;
507 fragmeta
->depth_factor
= 0.0f
;
508 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
, false);
509 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
, false);
513 bool msaa
= ctx
->rasterizer
->base
.multisample
;
515 /* TODO: Sample size */
516 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, msaa
);
517 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, !msaa
);
518 fragmeta
->depth_units
= ctx
->rasterizer
->base
.offset_units
* 2.0f
;
519 fragmeta
->depth_factor
= ctx
->rasterizer
->base
.offset_scale
;
521 /* XXX: Which bit is which? Does this maybe allow offseting not-tri? */
523 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
,
524 ctx
->rasterizer
->base
.offset_tri
);
525 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
,
526 ctx
->rasterizer
->base
.offset_tri
);
530 panfrost_frag_meta_zsa_update(struct panfrost_context
*ctx
,
531 struct mali_shader_meta
*fragmeta
)
533 const struct pipe_depth_stencil_alpha_state
*zsa
= ctx
->depth_stencil
;
534 int zfunc
= PIPE_FUNC_ALWAYS
;
537 struct pipe_stencil_state default_stencil
= {
539 .func
= PIPE_FUNC_ALWAYS
,
540 .fail_op
= MALI_STENCIL_KEEP
,
541 .zfail_op
= MALI_STENCIL_KEEP
,
542 .zpass_op
= MALI_STENCIL_KEEP
,
547 panfrost_make_stencil_state(&default_stencil
,
548 &fragmeta
->stencil_front
);
549 fragmeta
->stencil_mask_front
= default_stencil
.writemask
;
550 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
551 fragmeta
->stencil_mask_back
= default_stencil
.writemask
;
552 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
, false);
553 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
, false);
555 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
,
556 zsa
->stencil
[0].enabled
);
557 panfrost_make_stencil_state(&zsa
->stencil
[0],
558 &fragmeta
->stencil_front
);
559 fragmeta
->stencil_mask_front
= zsa
->stencil
[0].writemask
;
560 fragmeta
->stencil_front
.ref
= ctx
->stencil_ref
.ref_value
[0];
562 /* If back-stencil is not enabled, use the front values */
564 if (zsa
->stencil
[1].enabled
) {
565 panfrost_make_stencil_state(&zsa
->stencil
[1],
566 &fragmeta
->stencil_back
);
567 fragmeta
->stencil_mask_back
= zsa
->stencil
[1].writemask
;
568 fragmeta
->stencil_back
.ref
= ctx
->stencil_ref
.ref_value
[1];
570 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
571 fragmeta
->stencil_mask_back
= fragmeta
->stencil_mask_front
;
572 fragmeta
->stencil_back
.ref
= fragmeta
->stencil_front
.ref
;
575 if (zsa
->depth
.enabled
)
576 zfunc
= zsa
->depth
.func
;
578 /* Depth state (TODO: Refactor) */
580 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
,
581 zsa
->depth
.writemask
);
584 fragmeta
->unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
585 fragmeta
->unknown2_3
|= MALI_DEPTH_FUNC(panfrost_translate_compare_func(zfunc
));
589 panfrost_frag_meta_blend_update(struct panfrost_context
*ctx
,
590 struct mali_shader_meta
*fragmeta
,
591 struct midgard_blend_rt
*rts
)
593 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
595 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_DITHER
,
596 (dev
->quirks
& MIDGARD_SFBD
) && ctx
->blend
&&
597 !ctx
->blend
->base
.dither
);
599 /* Get blending setup */
600 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
602 struct panfrost_blend_final blend
[PIPE_MAX_COLOR_BUFS
];
603 unsigned shader_offset
= 0;
604 struct panfrost_bo
*shader_bo
= NULL
;
606 for (unsigned c
= 0; c
< rt_count
; ++c
)
607 blend
[c
] = panfrost_get_blend_for_context(ctx
, c
, &shader_bo
,
610 /* If there is a blend shader, work registers are shared. XXX: opt */
612 for (unsigned c
= 0; c
< rt_count
; ++c
) {
613 if (blend
[c
].is_shader
)
614 fragmeta
->midgard1
.work_count
= 16;
617 /* Even on MFBD, the shader descriptor gets blend shaders. It's *also*
618 * copied to the blend_meta appended (by convention), but this is the
619 * field actually read by the hardware. (Or maybe both are read...?).
620 * Specify the last RTi with a blend shader. */
622 fragmeta
->blend
.shader
= 0;
624 for (signed rt
= (rt_count
- 1); rt
>= 0; --rt
) {
625 if (!blend
[rt
].is_shader
)
628 fragmeta
->blend
.shader
= blend
[rt
].shader
.gpu
|
629 blend
[rt
].shader
.first_tag
;
633 if (dev
->quirks
& MIDGARD_SFBD
) {
634 /* When only a single render target platform is used, the blend
635 * information is inside the shader meta itself. We additionally
636 * need to signal CAN_DISCARD for nontrivial blend modes (so
637 * we're able to read back the destination buffer) */
639 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_BLEND_SHADER
,
642 if (!blend
[0].is_shader
) {
643 fragmeta
->blend
.equation
= *blend
[0].equation
.equation
;
644 fragmeta
->blend
.constant
= blend
[0].equation
.constant
;
647 SET_BIT(fragmeta
->unknown2_3
, MALI_CAN_DISCARD
,
648 !blend
[0].no_blending
);
652 /* Additional blend descriptor tacked on for jobs using MFBD */
654 for (unsigned i
= 0; i
< rt_count
; ++i
) {
655 rts
[i
].flags
= 0x200;
657 bool is_srgb
= (ctx
->pipe_framebuffer
.nr_cbufs
> i
) &&
658 (ctx
->pipe_framebuffer
.cbufs
[i
]) &&
659 util_format_is_srgb(ctx
->pipe_framebuffer
.cbufs
[i
]->format
);
661 SET_BIT(rts
[i
].flags
, MALI_BLEND_MRT_SHADER
, blend
[i
].is_shader
);
662 SET_BIT(rts
[i
].flags
, MALI_BLEND_LOAD_TIB
, !blend
[i
].no_blending
);
663 SET_BIT(rts
[i
].flags
, MALI_BLEND_SRGB
, is_srgb
);
664 SET_BIT(rts
[i
].flags
, MALI_BLEND_NO_DITHER
, !ctx
->blend
->base
.dither
);
666 if (blend
[i
].is_shader
) {
667 rts
[i
].blend
.shader
= blend
[i
].shader
.gpu
| blend
[i
].shader
.first_tag
;
669 rts
[i
].blend
.equation
= *blend
[i
].equation
.equation
;
670 rts
[i
].blend
.constant
= blend
[i
].equation
.constant
;
676 panfrost_frag_shader_meta_init(struct panfrost_context
*ctx
,
677 struct mali_shader_meta
*fragmeta
,
678 struct midgard_blend_rt
*rts
)
680 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
681 struct panfrost_shader_state
*fs
;
683 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
685 fragmeta
->alpha_coverage
= ~MALI_ALPHA_COVERAGE(0.000000);
686 fragmeta
->unknown2_3
= MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS
) | 0x3010;
687 fragmeta
->unknown2_4
= 0x4e0;
689 /* unknown2_4 has 0x10 bit set on T6XX and T720. We don't know why this
690 * is required (independent of 32-bit/64-bit descriptors), or why it's
691 * not used on later GPU revisions. Otherwise, all shader jobs fault on
692 * these earlier chips (perhaps this is a chicken bit of some kind).
693 * More investigation is needed. */
695 SET_BIT(fragmeta
->unknown2_4
, 0x10, dev
->quirks
& MIDGARD_SFBD
);
697 /* Depending on whether it's legal to in the given shader, we try to
698 * enable early-z testing (or forward-pixel kill?) */
700 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_EARLY_Z
,
701 !fs
->can_discard
&& !fs
->writes_depth
);
703 /* Add the writes Z/S flags if needed. */
704 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_WRITES_Z
, fs
->writes_depth
);
705 SET_BIT(fragmeta
->midgard1
.flags_hi
, MALI_WRITES_S
, fs
->writes_stencil
);
707 /* Any time texturing is used, derivatives are implicitly calculated,
708 * so we need to enable helper invocations */
710 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_HELPER_INVOCATIONS
,
711 fs
->helper_invocations
);
713 /* CAN_DISCARD should be set if the fragment shader possibly contains a
714 * 'discard' instruction. It is likely this is related to optimizations
715 * related to forward-pixel kill, as per "Mali Performance 3: Is
716 * EGL_BUFFER_PRESERVED a good thing?" by Peter Harris */
718 SET_BIT(fragmeta
->unknown2_3
, MALI_CAN_DISCARD
, fs
->can_discard
);
719 SET_BIT(fragmeta
->midgard1
.flags_lo
, 0x400, fs
->can_discard
);
721 panfrost_frag_meta_rasterizer_update(ctx
, fragmeta
);
722 panfrost_frag_meta_zsa_update(ctx
, fragmeta
);
723 panfrost_frag_meta_blend_update(ctx
, fragmeta
, rts
);
727 panfrost_emit_shader_meta(struct panfrost_batch
*batch
,
728 enum pipe_shader_type st
,
729 struct mali_vertex_tiler_postfix
*postfix
)
731 struct panfrost_context
*ctx
= batch
->ctx
;
732 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
739 struct mali_shader_meta meta
;
741 panfrost_shader_meta_init(ctx
, st
, &meta
);
743 /* Add the shader BO to the batch. */
744 panfrost_batch_add_bo(batch
, ss
->bo
,
745 PAN_BO_ACCESS_PRIVATE
|
747 panfrost_bo_access_for_stage(st
));
751 if (st
== PIPE_SHADER_FRAGMENT
) {
752 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
753 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
754 size_t desc_size
= sizeof(meta
);
755 struct midgard_blend_rt rts
[4];
756 struct panfrost_transfer xfer
;
758 assert(rt_count
<= ARRAY_SIZE(rts
));
760 panfrost_frag_shader_meta_init(ctx
, &meta
, rts
);
762 if (!(dev
->quirks
& MIDGARD_SFBD
))
763 desc_size
+= sizeof(*rts
) * rt_count
;
765 xfer
= panfrost_allocate_transient(batch
, desc_size
);
767 memcpy(xfer
.cpu
, &meta
, sizeof(meta
));
768 memcpy(xfer
.cpu
+ sizeof(meta
), rts
, sizeof(*rts
) * rt_count
);
770 shader_ptr
= xfer
.gpu
;
772 shader_ptr
= panfrost_upload_transient(batch
, &meta
,
776 postfix
->shader
= shader_ptr
;
780 panfrost_mali_viewport_init(struct panfrost_context
*ctx
,
781 struct mali_viewport
*mvp
)
783 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
785 /* Clip bounds are encoded as floats. The viewport itself is encoded as
786 * (somewhat) asymmetric ints. */
788 const struct pipe_scissor_state
*ss
= &ctx
->scissor
;
790 memset(mvp
, 0, sizeof(*mvp
));
792 /* By default, do no viewport clipping, i.e. clip to (-inf, inf) in
793 * each direction. Clipping to the viewport in theory should work, but
794 * in practice causes issues when we're not explicitly trying to
797 *mvp
= (struct mali_viewport
) {
798 .clip_minx
= -INFINITY
,
799 .clip_miny
= -INFINITY
,
800 .clip_maxx
= INFINITY
,
801 .clip_maxy
= INFINITY
,
804 /* Always scissor to the viewport by default. */
805 float vp_minx
= (int) (vp
->translate
[0] - fabsf(vp
->scale
[0]));
806 float vp_maxx
= (int) (vp
->translate
[0] + fabsf(vp
->scale
[0]));
808 float vp_miny
= (int) (vp
->translate
[1] - fabsf(vp
->scale
[1]));
809 float vp_maxy
= (int) (vp
->translate
[1] + fabsf(vp
->scale
[1]));
811 float minz
= (vp
->translate
[2] - fabsf(vp
->scale
[2]));
812 float maxz
= (vp
->translate
[2] + fabsf(vp
->scale
[2]));
814 /* Apply the scissor test */
816 unsigned minx
, miny
, maxx
, maxy
;
818 if (ss
&& ctx
->rasterizer
&& ctx
->rasterizer
->base
.scissor
) {
819 minx
= MAX2(ss
->minx
, vp_minx
);
820 miny
= MAX2(ss
->miny
, vp_miny
);
821 maxx
= MIN2(ss
->maxx
, vp_maxx
);
822 maxy
= MIN2(ss
->maxy
, vp_maxy
);
830 /* Hardware needs the min/max to be strictly ordered, so flip if we
831 * need to. The viewport transformation in the vertex shader will
832 * handle the negatives if we don't */
835 unsigned temp
= miny
;
841 unsigned temp
= minx
;
852 /* Clamp to the framebuffer size as a last check */
854 minx
= MIN2(ctx
->pipe_framebuffer
.width
, minx
);
855 maxx
= MIN2(ctx
->pipe_framebuffer
.width
, maxx
);
857 miny
= MIN2(ctx
->pipe_framebuffer
.height
, miny
);
858 maxy
= MIN2(ctx
->pipe_framebuffer
.height
, maxy
);
862 mvp
->viewport0
[0] = minx
;
863 mvp
->viewport1
[0] = MALI_POSITIVE(maxx
);
865 mvp
->viewport0
[1] = miny
;
866 mvp
->viewport1
[1] = MALI_POSITIVE(maxy
);
868 mvp
->clip_minz
= minz
;
869 mvp
->clip_maxz
= maxz
;
873 panfrost_emit_viewport(struct panfrost_batch
*batch
,
874 struct mali_vertex_tiler_postfix
*tiler_postfix
)
876 struct panfrost_context
*ctx
= batch
->ctx
;
877 struct mali_viewport mvp
;
879 panfrost_mali_viewport_init(batch
->ctx
, &mvp
);
881 /* Update the job, unless we're doing wallpapering (whose lack of
882 * scissor we can ignore, since if we "miss" a tile of wallpaper, it'll
883 * just... be faster :) */
885 if (!ctx
->wallpaper_batch
)
886 panfrost_batch_union_scissor(batch
, mvp
.viewport0
[0],
888 mvp
.viewport1
[0] + 1,
889 mvp
.viewport1
[1] + 1);
891 tiler_postfix
->viewport
= panfrost_upload_transient(batch
, &mvp
,
896 panfrost_map_constant_buffer_gpu(struct panfrost_batch
*batch
,
897 enum pipe_shader_type st
,
898 struct panfrost_constant_buffer
*buf
,
901 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
902 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
905 panfrost_batch_add_bo(batch
, rsrc
->bo
,
906 PAN_BO_ACCESS_SHARED
|
908 panfrost_bo_access_for_stage(st
));
910 /* Alignment gauranteed by
911 * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
912 return rsrc
->bo
->gpu
+ cb
->buffer_offset
;
913 } else if (cb
->user_buffer
) {
914 return panfrost_upload_transient(batch
,
919 unreachable("No constant buffer");
923 struct sysval_uniform
{
933 panfrost_upload_viewport_scale_sysval(struct panfrost_batch
*batch
,
934 struct sysval_uniform
*uniform
)
936 struct panfrost_context
*ctx
= batch
->ctx
;
937 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
939 uniform
->f
[0] = vp
->scale
[0];
940 uniform
->f
[1] = vp
->scale
[1];
941 uniform
->f
[2] = vp
->scale
[2];
945 panfrost_upload_viewport_offset_sysval(struct panfrost_batch
*batch
,
946 struct sysval_uniform
*uniform
)
948 struct panfrost_context
*ctx
= batch
->ctx
;
949 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
951 uniform
->f
[0] = vp
->translate
[0];
952 uniform
->f
[1] = vp
->translate
[1];
953 uniform
->f
[2] = vp
->translate
[2];
956 static void panfrost_upload_txs_sysval(struct panfrost_batch
*batch
,
957 enum pipe_shader_type st
,
958 unsigned int sysvalid
,
959 struct sysval_uniform
*uniform
)
961 struct panfrost_context
*ctx
= batch
->ctx
;
962 unsigned texidx
= PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid
);
963 unsigned dim
= PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid
);
964 bool is_array
= PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid
);
965 struct pipe_sampler_view
*tex
= &ctx
->sampler_views
[st
][texidx
]->base
;
968 uniform
->i
[0] = u_minify(tex
->texture
->width0
, tex
->u
.tex
.first_level
);
971 uniform
->i
[1] = u_minify(tex
->texture
->height0
,
972 tex
->u
.tex
.first_level
);
975 uniform
->i
[2] = u_minify(tex
->texture
->depth0
,
976 tex
->u
.tex
.first_level
);
979 uniform
->i
[dim
] = tex
->texture
->array_size
;
983 panfrost_upload_ssbo_sysval(struct panfrost_batch
*batch
,
984 enum pipe_shader_type st
,
986 struct sysval_uniform
*uniform
)
988 struct panfrost_context
*ctx
= batch
->ctx
;
990 assert(ctx
->ssbo_mask
[st
] & (1 << ssbo_id
));
991 struct pipe_shader_buffer sb
= ctx
->ssbo
[st
][ssbo_id
];
993 /* Compute address */
994 struct panfrost_bo
*bo
= pan_resource(sb
.buffer
)->bo
;
996 panfrost_batch_add_bo(batch
, bo
,
997 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_RW
|
998 panfrost_bo_access_for_stage(st
));
1000 /* Upload address and size as sysval */
1001 uniform
->du
[0] = bo
->gpu
+ sb
.buffer_offset
;
1002 uniform
->u
[2] = sb
.buffer_size
;
1006 panfrost_upload_sampler_sysval(struct panfrost_batch
*batch
,
1007 enum pipe_shader_type st
,
1009 struct sysval_uniform
*uniform
)
1011 struct panfrost_context
*ctx
= batch
->ctx
;
1012 struct pipe_sampler_state
*sampl
= &ctx
->samplers
[st
][samp_idx
]->base
;
1014 uniform
->f
[0] = sampl
->min_lod
;
1015 uniform
->f
[1] = sampl
->max_lod
;
1016 uniform
->f
[2] = sampl
->lod_bias
;
1018 /* Even without any errata, Midgard represents "no mipmapping" as
1019 * fixing the LOD with the clamps; keep behaviour consistent. c.f.
1020 * panfrost_create_sampler_state which also explains our choice of
1021 * epsilon value (again to keep behaviour consistent) */
1023 if (sampl
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
1024 uniform
->f
[1] = uniform
->f
[0] + (1.0/256.0);
1028 panfrost_upload_num_work_groups_sysval(struct panfrost_batch
*batch
,
1029 struct sysval_uniform
*uniform
)
1031 struct panfrost_context
*ctx
= batch
->ctx
;
1033 uniform
->u
[0] = ctx
->compute_grid
->grid
[0];
1034 uniform
->u
[1] = ctx
->compute_grid
->grid
[1];
1035 uniform
->u
[2] = ctx
->compute_grid
->grid
[2];
1039 panfrost_upload_sysvals(struct panfrost_batch
*batch
, void *buf
,
1040 struct panfrost_shader_state
*ss
,
1041 enum pipe_shader_type st
)
1043 struct sysval_uniform
*uniforms
= (void *)buf
;
1045 for (unsigned i
= 0; i
< ss
->sysval_count
; ++i
) {
1046 int sysval
= ss
->sysval
[i
];
1048 switch (PAN_SYSVAL_TYPE(sysval
)) {
1049 case PAN_SYSVAL_VIEWPORT_SCALE
:
1050 panfrost_upload_viewport_scale_sysval(batch
,
1053 case PAN_SYSVAL_VIEWPORT_OFFSET
:
1054 panfrost_upload_viewport_offset_sysval(batch
,
1057 case PAN_SYSVAL_TEXTURE_SIZE
:
1058 panfrost_upload_txs_sysval(batch
, st
,
1059 PAN_SYSVAL_ID(sysval
),
1062 case PAN_SYSVAL_SSBO
:
1063 panfrost_upload_ssbo_sysval(batch
, st
,
1064 PAN_SYSVAL_ID(sysval
),
1067 case PAN_SYSVAL_NUM_WORK_GROUPS
:
1068 panfrost_upload_num_work_groups_sysval(batch
,
1071 case PAN_SYSVAL_SAMPLER
:
1072 panfrost_upload_sampler_sysval(batch
, st
,
1073 PAN_SYSVAL_ID(sysval
),
1083 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer
*buf
,
1086 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1087 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1090 return rsrc
->bo
->cpu
;
1091 else if (cb
->user_buffer
)
1092 return cb
->user_buffer
;
1094 unreachable("No constant buffer");
1098 panfrost_emit_const_buf(struct panfrost_batch
*batch
,
1099 enum pipe_shader_type stage
,
1100 struct mali_vertex_tiler_postfix
*postfix
)
1102 struct panfrost_context
*ctx
= batch
->ctx
;
1103 struct panfrost_shader_variants
*all
= ctx
->shader
[stage
];
1108 struct panfrost_constant_buffer
*buf
= &ctx
->constant_buffer
[stage
];
1110 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1112 /* Uniforms are implicitly UBO #0 */
1113 bool has_uniforms
= buf
->enabled_mask
& (1 << 0);
1115 /* Allocate room for the sysval and the uniforms */
1116 size_t sys_size
= sizeof(float) * 4 * ss
->sysval_count
;
1117 size_t uniform_size
= has_uniforms
? (buf
->cb
[0].buffer_size
) : 0;
1118 size_t size
= sys_size
+ uniform_size
;
1119 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1122 /* Upload sysvals requested by the shader */
1123 panfrost_upload_sysvals(batch
, transfer
.cpu
, ss
, stage
);
1125 /* Upload uniforms */
1126 if (has_uniforms
&& uniform_size
) {
1127 const void *cpu
= panfrost_map_constant_buffer_cpu(buf
, 0);
1128 memcpy(transfer
.cpu
+ sys_size
, cpu
, uniform_size
);
1131 /* Next up, attach UBOs. UBO #0 is the uniforms we just
1134 unsigned ubo_count
= panfrost_ubo_count(ctx
, stage
);
1135 assert(ubo_count
>= 1);
1137 size_t sz
= sizeof(uint64_t) * ubo_count
;
1138 uint64_t ubos
[PAN_MAX_CONST_BUFFERS
];
1139 int uniform_count
= ss
->uniform_count
;
1141 /* Upload uniforms as a UBO */
1142 ubos
[0] = MALI_MAKE_UBO(2 + uniform_count
, transfer
.gpu
);
1144 /* The rest are honest-to-goodness UBOs */
1146 for (unsigned ubo
= 1; ubo
< ubo_count
; ++ubo
) {
1147 size_t usz
= buf
->cb
[ubo
].buffer_size
;
1148 bool enabled
= buf
->enabled_mask
& (1 << ubo
);
1149 bool empty
= usz
== 0;
1151 if (!enabled
|| empty
) {
1152 /* Stub out disabled UBOs to catch accesses */
1153 ubos
[ubo
] = MALI_MAKE_UBO(0, 0xDEAD0000);
1157 mali_ptr gpu
= panfrost_map_constant_buffer_gpu(batch
, stage
,
1160 unsigned bytes_per_field
= 16;
1161 unsigned aligned
= ALIGN_POT(usz
, bytes_per_field
);
1162 ubos
[ubo
] = MALI_MAKE_UBO(aligned
/ bytes_per_field
, gpu
);
1165 mali_ptr ubufs
= panfrost_upload_transient(batch
, ubos
, sz
);
1166 postfix
->uniforms
= transfer
.gpu
;
1167 postfix
->uniform_buffers
= ubufs
;
1169 buf
->dirty_mask
= 0;
1173 panfrost_emit_shared_memory(struct panfrost_batch
*batch
,
1174 const struct pipe_grid_info
*info
,
1175 struct midgard_payload_vertex_tiler
*vtp
)
1177 struct panfrost_context
*ctx
= batch
->ctx
;
1178 struct panfrost_shader_variants
*all
= ctx
->shader
[PIPE_SHADER_COMPUTE
];
1179 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1180 unsigned single_size
= util_next_power_of_two(MAX2(ss
->shared_size
,
1182 unsigned shared_size
= single_size
* info
->grid
[0] * info
->grid
[1] *
1184 struct panfrost_bo
*bo
= panfrost_batch_get_shared_memory(batch
,
1188 struct mali_shared_memory shared
= {
1189 .shared_memory
= bo
->gpu
,
1190 .shared_workgroup_count
=
1191 util_logbase2_ceil(info
->grid
[0]) +
1192 util_logbase2_ceil(info
->grid
[1]) +
1193 util_logbase2_ceil(info
->grid
[2]),
1195 .shared_shift
= util_logbase2(single_size
) - 1
1198 vtp
->postfix
.shared_memory
= panfrost_upload_transient(batch
, &shared
,
1203 panfrost_get_tex_desc(struct panfrost_batch
*batch
,
1204 enum pipe_shader_type st
,
1205 struct panfrost_sampler_view
*view
)
1208 return (mali_ptr
) 0;
1210 struct pipe_sampler_view
*pview
= &view
->base
;
1211 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1213 /* Add the BO to the job so it's retained until the job is done. */
1215 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1216 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1217 panfrost_bo_access_for_stage(st
));
1219 panfrost_batch_add_bo(batch
, view
->bo
,
1220 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1221 panfrost_bo_access_for_stage(st
));
1223 return view
->bo
->gpu
;
1227 panfrost_emit_texture_descriptors(struct panfrost_batch
*batch
,
1228 enum pipe_shader_type stage
,
1229 struct mali_vertex_tiler_postfix
*postfix
)
1231 struct panfrost_context
*ctx
= batch
->ctx
;
1233 if (!ctx
->sampler_view_count
[stage
])
1236 uint64_t trampolines
[PIPE_MAX_SHADER_SAMPLER_VIEWS
];
1238 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
)
1239 trampolines
[i
] = panfrost_get_tex_desc(batch
, stage
,
1240 ctx
->sampler_views
[stage
][i
]);
1242 postfix
->texture_trampoline
= panfrost_upload_transient(batch
,
1245 ctx
->sampler_view_count
[stage
]);
1249 panfrost_emit_sampler_descriptors(struct panfrost_batch
*batch
,
1250 enum pipe_shader_type stage
,
1251 struct mali_vertex_tiler_postfix
*postfix
)
1253 struct panfrost_context
*ctx
= batch
->ctx
;
1255 if (!ctx
->sampler_count
[stage
])
1258 size_t desc_size
= sizeof(struct mali_sampler_descriptor
);
1259 size_t transfer_size
= desc_size
* ctx
->sampler_count
[stage
];
1260 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1262 struct mali_sampler_descriptor
*desc
= (struct mali_sampler_descriptor
*)transfer
.cpu
;
1264 for (int i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1265 desc
[i
] = ctx
->samplers
[stage
][i
]->hw
;
1267 postfix
->sampler_descriptor
= transfer
.gpu
;
1271 panfrost_emit_vertex_attr_meta(struct panfrost_batch
*batch
,
1272 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1274 struct panfrost_context
*ctx
= batch
->ctx
;
1279 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1281 panfrost_vertex_state_upd_attr_offs(ctx
, vertex_postfix
);
1282 vertex_postfix
->attribute_meta
= panfrost_upload_transient(batch
, so
->hw
,
1288 panfrost_emit_vertex_data(struct panfrost_batch
*batch
,
1289 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1291 struct panfrost_context
*ctx
= batch
->ctx
;
1292 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1294 /* Staged mali_attr, and index into them. i =/= k, depending on the
1295 * vertex buffer mask and instancing. Twice as much room is allocated,
1296 * for a worst case of NPOT_DIVIDEs which take up extra slot */
1297 union mali_attr attrs
[PIPE_MAX_ATTRIBS
* 2];
1300 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1301 /* We map a mali_attr to be 1:1 with the mali_attr_meta, which
1302 * means duplicating some vertex buffers (who cares? aside from
1303 * maybe some caching implications but I somehow doubt that
1306 struct pipe_vertex_element
*elem
= &so
->pipe
[i
];
1307 unsigned vbi
= elem
->vertex_buffer_index
;
1309 /* The exception to 1:1 mapping is that we can have multiple
1310 * entries (NPOT divisors), so we fixup anyways */
1312 so
->hw
[i
].index
= k
;
1314 if (!(ctx
->vb_mask
& (1 << vbi
)))
1317 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1318 struct panfrost_resource
*rsrc
;
1320 rsrc
= pan_resource(buf
->buffer
.resource
);
1324 /* Align to 64 bytes by masking off the lower bits. This
1325 * will be adjusted back when we fixup the src_offset in
1328 mali_ptr raw_addr
= rsrc
->bo
->gpu
+ buf
->buffer_offset
;
1329 mali_ptr addr
= raw_addr
& ~63;
1330 unsigned chopped_addr
= raw_addr
- addr
;
1332 /* Add a dependency of the batch on the vertex buffer */
1333 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1334 PAN_BO_ACCESS_SHARED
|
1335 PAN_BO_ACCESS_READ
|
1336 PAN_BO_ACCESS_VERTEX_TILER
);
1338 /* Set common fields */
1339 attrs
[k
].elements
= addr
;
1340 attrs
[k
].stride
= buf
->stride
;
1342 /* Since we advanced the base pointer, we shrink the buffer
1344 attrs
[k
].size
= rsrc
->base
.width0
- buf
->buffer_offset
;
1346 /* We need to add the extra size we masked off (for
1347 * correctness) so the data doesn't get clamped away */
1348 attrs
[k
].size
+= chopped_addr
;
1350 /* For non-instancing make sure we initialize */
1351 attrs
[k
].shift
= attrs
[k
].extra_flags
= 0;
1353 /* Instancing uses a dramatically different code path than
1354 * linear, so dispatch for the actual emission now that the
1355 * common code is finished */
1357 unsigned divisor
= elem
->instance_divisor
;
1359 if (divisor
&& ctx
->instance_count
== 1) {
1360 /* Silly corner case where there's a divisor(=1) but
1361 * there's no legitimate instancing. So we want *every*
1362 * attribute to be the same. So set stride to zero so
1363 * we don't go anywhere. */
1365 attrs
[k
].size
= attrs
[k
].stride
+ chopped_addr
;
1366 attrs
[k
].stride
= 0;
1367 attrs
[k
++].elements
|= MALI_ATTR_LINEAR
;
1368 } else if (ctx
->instance_count
<= 1) {
1369 /* Normal, non-instanced attributes */
1370 attrs
[k
++].elements
|= MALI_ATTR_LINEAR
;
1372 unsigned instance_shift
= vertex_postfix
->instance_shift
;
1373 unsigned instance_odd
= vertex_postfix
->instance_odd
;
1375 k
+= panfrost_vertex_instanced(ctx
->padded_count
,
1378 divisor
, &attrs
[k
]);
1382 /* Add special gl_VertexID/gl_InstanceID buffers */
1384 panfrost_vertex_id(ctx
->padded_count
, &attrs
[k
]);
1385 so
->hw
[PAN_VERTEX_ID
].index
= k
++;
1386 panfrost_instance_id(ctx
->padded_count
, &attrs
[k
]);
1387 so
->hw
[PAN_INSTANCE_ID
].index
= k
++;
1389 /* Upload whatever we emitted and go */
1391 vertex_postfix
->attributes
= panfrost_upload_transient(batch
, attrs
,
1392 k
* sizeof(*attrs
));
1396 panfrost_emit_varyings(struct panfrost_batch
*batch
, union mali_attr
*slot
,
1397 unsigned stride
, unsigned count
)
1399 /* Fill out the descriptor */
1400 slot
->stride
= stride
;
1401 slot
->size
= stride
* count
;
1402 slot
->shift
= slot
->extra_flags
= 0;
1404 struct panfrost_transfer transfer
= panfrost_allocate_transient(batch
,
1407 slot
->elements
= transfer
.gpu
| MALI_ATTR_LINEAR
;
1409 return transfer
.gpu
;
1413 panfrost_emit_streamout(struct panfrost_batch
*batch
, union mali_attr
*slot
,
1414 unsigned stride
, unsigned offset
, unsigned count
,
1415 struct pipe_stream_output_target
*target
)
1417 /* Fill out the descriptor */
1418 slot
->stride
= stride
* 4;
1419 slot
->shift
= slot
->extra_flags
= 0;
1421 unsigned max_size
= target
->buffer_size
;
1422 unsigned expected_size
= slot
->stride
* count
;
1424 slot
->size
= MIN2(max_size
, expected_size
);
1426 /* Grab the BO and bind it to the batch */
1427 struct panfrost_bo
*bo
= pan_resource(target
->buffer
)->bo
;
1429 /* Varyings are WRITE from the perspective of the VERTEX but READ from
1430 * the perspective of the TILER and FRAGMENT.
1432 panfrost_batch_add_bo(batch
, bo
,
1433 PAN_BO_ACCESS_SHARED
|
1435 PAN_BO_ACCESS_VERTEX_TILER
|
1436 PAN_BO_ACCESS_FRAGMENT
);
1438 mali_ptr addr
= bo
->gpu
+ target
->buffer_offset
+ (offset
* slot
->stride
);
1439 slot
->elements
= addr
;
1442 /* Given a shader and buffer indices, link varying metadata together */
1445 is_special_varying(gl_varying_slot loc
)
1448 case VARYING_SLOT_POS
:
1449 case VARYING_SLOT_PSIZ
:
1450 case VARYING_SLOT_PNTC
:
1451 case VARYING_SLOT_FACE
:
1459 panfrost_emit_varying_meta(void *outptr
, struct panfrost_shader_state
*ss
,
1460 signed general
, signed gl_Position
,
1461 signed gl_PointSize
, signed gl_PointCoord
,
1462 signed gl_FrontFacing
)
1464 struct mali_attr_meta
*out
= (struct mali_attr_meta
*) outptr
;
1466 for (unsigned i
= 0; i
< ss
->varying_count
; ++i
) {
1467 gl_varying_slot location
= ss
->varyings_loc
[i
];
1471 case VARYING_SLOT_POS
:
1472 index
= gl_Position
;
1474 case VARYING_SLOT_PSIZ
:
1475 index
= gl_PointSize
;
1477 case VARYING_SLOT_PNTC
:
1478 index
= gl_PointCoord
;
1480 case VARYING_SLOT_FACE
:
1481 index
= gl_FrontFacing
;
1489 out
[i
].index
= index
;
1494 has_point_coord(unsigned mask
, gl_varying_slot loc
)
1496 if ((loc
>= VARYING_SLOT_TEX0
) && (loc
<= VARYING_SLOT_TEX7
))
1497 return (mask
& (1 << (loc
- VARYING_SLOT_TEX0
)));
1498 else if (loc
== VARYING_SLOT_PNTC
)
1499 return (mask
& (1 << 8));
1504 /* Helpers for manipulating stream out information so we can pack varyings
1505 * accordingly. Compute the src_offset for a given captured varying */
1507 static struct pipe_stream_output
*
1508 pan_get_so(struct pipe_stream_output_info
*info
, gl_varying_slot loc
)
1510 for (unsigned i
= 0; i
< info
->num_outputs
; ++i
) {
1511 if (info
->output
[i
].register_index
== loc
)
1512 return &info
->output
[i
];
1515 unreachable("Varying not captured");
1518 /* TODO: Integers */
1519 static enum mali_format
1520 pan_xfb_format(unsigned nr_components
)
1522 switch (nr_components
) {
1523 case 1: return MALI_R32F
;
1524 case 2: return MALI_RG32F
;
1525 case 3: return MALI_RGB32F
;
1526 case 4: return MALI_RGBA32F
;
1527 default: unreachable("Invalid format");
1532 panfrost_emit_varying_descriptor(struct panfrost_batch
*batch
,
1533 unsigned vertex_count
,
1534 struct mali_vertex_tiler_postfix
*vertex_postfix
,
1535 struct mali_vertex_tiler_postfix
*tiler_postfix
,
1536 union midgard_primitive_size
*primitive_size
)
1538 /* Load the shaders */
1539 struct panfrost_context
*ctx
= batch
->ctx
;
1540 struct panfrost_shader_state
*vs
, *fs
;
1541 unsigned int num_gen_varyings
= 0;
1542 size_t vs_size
, fs_size
;
1544 /* Allocate the varying descriptor */
1546 vs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_VERTEX
);
1547 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
1548 vs_size
= sizeof(struct mali_attr_meta
) * vs
->varying_count
;
1549 fs_size
= sizeof(struct mali_attr_meta
) * fs
->varying_count
;
1551 struct panfrost_transfer trans
= panfrost_allocate_transient(batch
,
1555 struct pipe_stream_output_info
*so
= &vs
->stream_output
;
1557 /* Check if this varying is linked by us. This is the case for
1558 * general-purpose, non-captured varyings. If it is, link it. If it's
1559 * not, use the provided stream out information to determine the
1560 * offset, since it was already linked for us. */
1562 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1563 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1565 bool special
= is_special_varying(loc
);
1566 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1569 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1571 unsigned dst_offset
= o
->dst_offset
* 4; /* dwords */
1572 vs
->varyings
[i
].src_offset
= dst_offset
;
1573 } else if (!special
) {
1574 vs
->varyings
[i
].src_offset
= 16 * (num_gen_varyings
++);
1578 /* Conversely, we need to set src_offset for the captured varyings.
1579 * Here, the layout is defined by the stream out info, not us */
1581 /* Link up with fragment varyings */
1582 bool reads_point_coord
= fs
->reads_point_coord
;
1584 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1585 gl_varying_slot loc
= fs
->varyings_loc
[i
];
1586 unsigned src_offset
;
1590 for (unsigned j
= 0; j
< vs
->varying_count
; ++j
) {
1591 if (vs
->varyings_loc
[j
] == loc
) {
1597 /* Either assign or reuse */
1599 src_offset
= vs
->varyings
[vs_idx
].src_offset
;
1601 src_offset
= 16 * (num_gen_varyings
++);
1603 fs
->varyings
[i
].src_offset
= src_offset
;
1605 if (has_point_coord(fs
->point_sprite_mask
, loc
))
1606 reads_point_coord
= true;
1609 memcpy(trans
.cpu
, vs
->varyings
, vs_size
);
1610 memcpy(trans
.cpu
+ vs_size
, fs
->varyings
, fs_size
);
1612 union mali_attr varyings
[PIPE_MAX_ATTRIBS
] = {0};
1614 /* Figure out how many streamout buffers could be bound */
1615 unsigned so_count
= ctx
->streamout
.num_targets
;
1616 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1617 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1619 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1620 if (!captured
) continue;
1622 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1623 so_count
= MAX2(so_count
, o
->output_buffer
+ 1);
1626 signed idx
= so_count
;
1627 signed general
= idx
++;
1628 signed gl_Position
= idx
++;
1629 signed gl_PointSize
= vs
->writes_point_size
? (idx
++) : -1;
1630 signed gl_PointCoord
= reads_point_coord
? (idx
++) : -1;
1631 signed gl_FrontFacing
= fs
->reads_face
? (idx
++) : -1;
1632 signed gl_FragCoord
= fs
->reads_frag_coord
? (idx
++) : -1;
1634 /* Emit the stream out buffers */
1636 unsigned out_count
= u_stream_outputs_for_vertices(ctx
->active_prim
,
1639 for (unsigned i
= 0; i
< so_count
; ++i
) {
1640 if (i
< ctx
->streamout
.num_targets
) {
1641 panfrost_emit_streamout(batch
, &varyings
[i
],
1643 ctx
->streamout
.offsets
[i
],
1645 ctx
->streamout
.targets
[i
]);
1647 /* Emit a dummy buffer */
1648 panfrost_emit_varyings(batch
, &varyings
[i
],
1652 /* Clear the attribute type */
1653 varyings
[i
].elements
&= ~0xF;
1657 panfrost_emit_varyings(batch
, &varyings
[general
],
1658 num_gen_varyings
* 16,
1661 mali_ptr varyings_p
;
1663 /* fp32 vec4 gl_Position */
1664 varyings_p
= panfrost_emit_varyings(batch
, &varyings
[gl_Position
],
1665 sizeof(float) * 4, vertex_count
);
1666 tiler_postfix
->position_varying
= varyings_p
;
1669 if (panfrost_writes_point_size(ctx
)) {
1670 varyings_p
= panfrost_emit_varyings(batch
,
1671 &varyings
[gl_PointSize
],
1673 primitive_size
->pointer
= varyings_p
;
1676 if (reads_point_coord
)
1677 varyings
[gl_PointCoord
].elements
= MALI_VARYING_POINT_COORD
;
1680 varyings
[gl_FrontFacing
].elements
= MALI_VARYING_FRONT_FACING
;
1682 if (fs
->reads_frag_coord
)
1683 varyings
[gl_FragCoord
].elements
= MALI_VARYING_FRAG_COORD
;
1685 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1686 assert(!(device
->quirks
& IS_BIFROST
) || !(reads_point_coord
|| fs
->reads_face
|| fs
->reads_frag_coord
));
1688 /* Let's go ahead and link varying meta to the buffer in question, now
1689 * that that information is available. VARYING_SLOT_POS is mapped to
1690 * gl_FragCoord for fragment shaders but gl_Positionf or vertex shaders
1693 panfrost_emit_varying_meta(trans
.cpu
, vs
, general
, gl_Position
,
1694 gl_PointSize
, gl_PointCoord
,
1697 panfrost_emit_varying_meta(trans
.cpu
+ vs_size
, fs
, general
,
1698 gl_FragCoord
, gl_PointSize
,
1699 gl_PointCoord
, gl_FrontFacing
);
1701 /* Replace streamout */
1703 struct mali_attr_meta
*ovs
= (struct mali_attr_meta
*)trans
.cpu
;
1704 struct mali_attr_meta
*ofs
= ovs
+ vs
->varying_count
;
1706 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1707 gl_varying_slot loc
= vs
->varyings_loc
[i
];
1709 bool captured
= ((vs
->so_mask
& (1ll << loc
)) ? true : false);
1713 struct pipe_stream_output
*o
= pan_get_so(so
, loc
);
1714 ovs
[i
].index
= o
->output_buffer
;
1716 /* Set the type appropriately. TODO: Integer varyings XXX */
1717 assert(o
->stream
== 0);
1718 ovs
[i
].format
= pan_xfb_format(o
->num_components
);
1719 ovs
[i
].swizzle
= panfrost_get_default_swizzle(o
->num_components
);
1721 /* Link to the fragment */
1725 for (unsigned j
= 0; j
< fs
->varying_count
; ++j
) {
1726 if (fs
->varyings_loc
[j
] == loc
) {
1733 ofs
[fs_idx
].index
= ovs
[i
].index
;
1734 ofs
[fs_idx
].format
= ovs
[i
].format
;
1735 ofs
[fs_idx
].swizzle
= ovs
[i
].swizzle
;
1739 /* Replace point sprite */
1740 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1741 /* If we have a point sprite replacement, handle that here. We
1742 * have to translate location first. TODO: Flip y in shader.
1743 * We're already keying ... just time crunch .. */
1745 if (has_point_coord(fs
->point_sprite_mask
,
1746 fs
->varyings_loc
[i
])) {
1747 ofs
[i
].index
= gl_PointCoord
;
1749 /* Swizzle out the z/w to 0/1 */
1750 ofs
[i
].format
= MALI_RG16F
;
1751 ofs
[i
].swizzle
= panfrost_get_default_swizzle(2);
1755 /* Fix up unaligned addresses */
1756 for (unsigned i
= 0; i
< so_count
; ++i
) {
1757 if (varyings
[i
].elements
< MALI_RECORD_SPECIAL
)
1760 unsigned align
= (varyings
[i
].elements
& 63);
1762 /* While we're at it, the SO buffers are linear */
1765 varyings
[i
].elements
|= MALI_ATTR_LINEAR
;
1769 /* We need to adjust alignment */
1770 varyings
[i
].elements
&= ~63;
1771 varyings
[i
].elements
|= MALI_ATTR_LINEAR
;
1772 varyings
[i
].size
+= align
;
1774 for (unsigned v
= 0; v
< vs
->varying_count
; ++v
) {
1775 if (ovs
[v
].index
!= i
)
1778 ovs
[v
].src_offset
= vs
->varyings
[v
].src_offset
+ align
;
1781 for (unsigned f
= 0; f
< fs
->varying_count
; ++f
) {
1782 if (ofs
[f
].index
!= i
)
1785 ofs
[f
].src_offset
= fs
->varyings
[f
].src_offset
+ align
;
1789 varyings_p
= panfrost_upload_transient(batch
, varyings
,
1790 idx
* sizeof(*varyings
));
1791 vertex_postfix
->varyings
= varyings_p
;
1792 tiler_postfix
->varyings
= varyings_p
;
1794 vertex_postfix
->varying_meta
= trans
.gpu
;
1795 tiler_postfix
->varying_meta
= trans
.gpu
+ vs_size
;
1799 panfrost_emit_vertex_tiler_jobs(struct panfrost_batch
*batch
,
1800 struct mali_vertex_tiler_prefix
*vertex_prefix
,
1801 struct mali_vertex_tiler_postfix
*vertex_postfix
,
1802 struct mali_vertex_tiler_prefix
*tiler_prefix
,
1803 struct mali_vertex_tiler_postfix
*tiler_postfix
,
1804 union midgard_primitive_size
*primitive_size
)
1806 struct panfrost_context
*ctx
= batch
->ctx
;
1807 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1808 bool wallpapering
= ctx
->wallpaper_batch
&& batch
->tiler_dep
;
1809 struct bifrost_payload_vertex bifrost_vertex
= {0,};
1810 struct bifrost_payload_tiler bifrost_tiler
= {0,};
1811 struct midgard_payload_vertex_tiler midgard_vertex
= {0,};
1812 struct midgard_payload_vertex_tiler midgard_tiler
= {0,};
1814 size_t vp_size
, tp_size
;
1816 if (device
->quirks
& IS_BIFROST
) {
1817 bifrost_vertex
.prefix
= *vertex_prefix
;
1818 bifrost_vertex
.postfix
= *vertex_postfix
;
1819 vp
= &bifrost_vertex
;
1820 vp_size
= sizeof(bifrost_vertex
);
1822 bifrost_tiler
.prefix
= *tiler_prefix
;
1823 bifrost_tiler
.tiler
.primitive_size
= *primitive_size
;
1824 bifrost_tiler
.tiler
.tiler_meta
= panfrost_batch_get_tiler_meta(batch
, ~0);
1825 bifrost_tiler
.postfix
= *tiler_postfix
;
1826 tp
= &bifrost_tiler
;
1827 tp_size
= sizeof(bifrost_tiler
);
1829 midgard_vertex
.prefix
= *vertex_prefix
;
1830 midgard_vertex
.postfix
= *vertex_postfix
;
1831 vp
= &midgard_vertex
;
1832 vp_size
= sizeof(midgard_vertex
);
1834 midgard_tiler
.prefix
= *tiler_prefix
;
1835 midgard_tiler
.postfix
= *tiler_postfix
;
1836 midgard_tiler
.primitive_size
= *primitive_size
;
1837 tp
= &midgard_tiler
;
1838 tp_size
= sizeof(midgard_tiler
);
1842 /* Inject in reverse order, with "predicted" job indices.
1843 * THIS IS A HACK XXX */
1844 panfrost_new_job(batch
, JOB_TYPE_TILER
, false,
1845 batch
->job_index
+ 2, tp
, tp_size
, true);
1846 panfrost_new_job(batch
, JOB_TYPE_VERTEX
, false, 0,
1851 /* If rasterizer discard is enable, only submit the vertex */
1853 bool rasterizer_discard
= ctx
->rasterizer
&&
1854 ctx
->rasterizer
->base
.rasterizer_discard
;
1856 unsigned vertex
= panfrost_new_job(batch
, JOB_TYPE_VERTEX
, false, 0,
1857 vp
, vp_size
, false);
1859 if (rasterizer_discard
)
1862 panfrost_new_job(batch
, JOB_TYPE_TILER
, false, vertex
, tp
, tp_size
,
1866 /* TODO: stop hardcoding this */
1868 panfrost_emit_sample_locations(struct panfrost_batch
*batch
)
1870 uint16_t locations
[] = {
1921 return panfrost_upload_transient(batch
, locations
, 96 * sizeof(uint16_t));