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 struct mali_shared_memory shared
= {
62 .shared_workgroup_count
= ~0,
65 if (batch
->stack_size
) {
66 struct panfrost_bo
*stack
=
67 panfrost_batch_get_scratchpad(batch
, batch
->stack_size
,
68 dev
->thread_tls_alloc
,
71 shared
.stack_shift
= panfrost_get_stack_shift(batch
->stack_size
);
72 shared
.scratchpad
= stack
->gpu
;
75 postfix
->shared_memory
= panfrost_pool_upload_aligned(&batch
->pool
, &shared
, sizeof(shared
), 64);
79 panfrost_vt_attach_framebuffer(struct panfrost_context
*ctx
,
80 struct mali_vertex_tiler_postfix
*postfix
)
82 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
83 postfix
->shared_memory
= panfrost_batch_reserve_framebuffer(batch
);
87 panfrost_vt_update_rasterizer(struct panfrost_rasterizer
*rasterizer
,
88 struct mali_vertex_tiler_prefix
*prefix
,
89 struct mali_vertex_tiler_postfix
*postfix
)
91 postfix
->gl_enables
|= 0x7;
92 SET_BIT(postfix
->gl_enables
, MALI_FRONT_CCW_TOP
,
93 rasterizer
->base
.front_ccw
);
94 SET_BIT(postfix
->gl_enables
, MALI_CULL_FACE_FRONT
,
95 (rasterizer
->base
.cull_face
& PIPE_FACE_FRONT
));
96 SET_BIT(postfix
->gl_enables
, MALI_CULL_FACE_BACK
,
97 (rasterizer
->base
.cull_face
& PIPE_FACE_BACK
));
98 SET_BIT(prefix
->unknown_draw
, MALI_DRAW_FLATSHADE_FIRST
,
99 rasterizer
->base
.flatshade_first
);
103 panfrost_vt_update_primitive_size(struct panfrost_context
*ctx
,
104 struct mali_vertex_tiler_prefix
*prefix
,
105 union midgard_primitive_size
*primitive_size
)
107 struct panfrost_rasterizer
*rasterizer
= ctx
->rasterizer
;
109 if (!panfrost_writes_point_size(ctx
)) {
110 float val
= (prefix
->draw_mode
== MALI_DRAW_MODE_POINTS
) ?
111 rasterizer
->base
.point_size
:
112 rasterizer
->base
.line_width
;
114 primitive_size
->constant
= val
;
119 panfrost_vt_update_occlusion_query(struct panfrost_context
*ctx
,
120 struct mali_vertex_tiler_postfix
*postfix
)
122 SET_BIT(postfix
->gl_enables
, MALI_OCCLUSION_QUERY
, ctx
->occlusion_query
);
123 if (ctx
->occlusion_query
) {
124 postfix
->occlusion_counter
= ctx
->occlusion_query
->bo
->gpu
;
125 panfrost_batch_add_bo(ctx
->batch
, ctx
->occlusion_query
->bo
,
126 PAN_BO_ACCESS_SHARED
|
128 PAN_BO_ACCESS_FRAGMENT
);
130 postfix
->occlusion_counter
= 0;
135 panfrost_vt_init(struct panfrost_context
*ctx
,
136 enum pipe_shader_type stage
,
137 struct mali_vertex_tiler_prefix
*prefix
,
138 struct mali_vertex_tiler_postfix
*postfix
)
140 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
142 if (!ctx
->shader
[stage
])
145 memset(prefix
, 0, sizeof(*prefix
));
146 memset(postfix
, 0, sizeof(*postfix
));
148 if (device
->quirks
& IS_BIFROST
) {
149 postfix
->gl_enables
= 0x2;
150 panfrost_vt_emit_shared_memory(ctx
, postfix
);
152 postfix
->gl_enables
= 0x6;
153 panfrost_vt_attach_framebuffer(ctx
, postfix
);
156 if (stage
== PIPE_SHADER_FRAGMENT
) {
157 panfrost_vt_update_occlusion_query(ctx
, postfix
);
158 panfrost_vt_update_rasterizer(ctx
->rasterizer
, prefix
, postfix
);
163 panfrost_translate_index_size(unsigned size
)
167 return MALI_DRAW_INDEXED_UINT8
;
170 return MALI_DRAW_INDEXED_UINT16
;
173 return MALI_DRAW_INDEXED_UINT32
;
176 unreachable("Invalid index size");
180 /* Gets a GPU address for the associated index buffer. Only gauranteed to be
181 * good for the duration of the draw (transient), could last longer. Also get
182 * the bounds on the index buffer for the range accessed by the draw. We do
183 * these operations together because there are natural optimizations which
184 * require them to be together. */
187 panfrost_get_index_buffer_bounded(struct panfrost_context
*ctx
,
188 const struct pipe_draw_info
*info
,
189 unsigned *min_index
, unsigned *max_index
)
191 struct panfrost_resource
*rsrc
= pan_resource(info
->index
.resource
);
192 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
193 off_t offset
= info
->start
* info
->index_size
;
194 bool needs_indices
= true;
197 if (info
->max_index
!= ~0u) {
198 *min_index
= info
->min_index
;
199 *max_index
= info
->max_index
;
200 needs_indices
= false;
203 if (!info
->has_user_indices
) {
204 /* Only resources can be directly mapped */
205 panfrost_batch_add_bo(batch
, rsrc
->bo
,
206 PAN_BO_ACCESS_SHARED
|
208 PAN_BO_ACCESS_VERTEX_TILER
);
209 out
= rsrc
->bo
->gpu
+ offset
;
211 /* Check the cache */
212 needs_indices
= !panfrost_minmax_cache_get(rsrc
->index_cache
,
218 /* Otherwise, we need to upload to transient memory */
219 const uint8_t *ibuf8
= (const uint8_t *) info
->index
.user
;
220 struct panfrost_transfer T
=
221 panfrost_pool_alloc_aligned(&batch
->pool
,
222 info
->count
* info
->index_size
,
225 memcpy(T
.cpu
, ibuf8
+ offset
, info
->count
* info
->index_size
);
231 u_vbuf_get_minmax_index(&ctx
->base
, info
, min_index
, max_index
);
233 if (!info
->has_user_indices
)
234 panfrost_minmax_cache_add(rsrc
->index_cache
,
235 info
->start
, info
->count
,
236 *min_index
, *max_index
);
243 panfrost_vt_set_draw_info(struct panfrost_context
*ctx
,
244 const struct pipe_draw_info
*info
,
245 enum mali_draw_mode draw_mode
,
246 struct mali_vertex_tiler_postfix
*vertex_postfix
,
247 struct mali_vertex_tiler_prefix
*tiler_prefix
,
248 struct mali_vertex_tiler_postfix
*tiler_postfix
,
249 unsigned *vertex_count
,
250 unsigned *padded_count
)
252 tiler_prefix
->draw_mode
= draw_mode
;
254 unsigned draw_flags
= 0;
256 if (panfrost_writes_point_size(ctx
))
257 draw_flags
|= MALI_DRAW_VARYING_SIZE
;
259 if (info
->primitive_restart
)
260 draw_flags
|= MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX
;
262 /* These doesn't make much sense */
264 draw_flags
|= 0x3000;
266 if (info
->index_size
) {
267 unsigned min_index
= 0, max_index
= 0;
269 tiler_prefix
->indices
= panfrost_get_index_buffer_bounded(ctx
,
274 /* Use the corresponding values */
275 *vertex_count
= max_index
- min_index
+ 1;
276 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= min_index
+ info
->index_bias
;
277 tiler_prefix
->offset_bias_correction
= -min_index
;
278 tiler_prefix
->index_count
= MALI_POSITIVE(info
->count
);
279 draw_flags
|= panfrost_translate_index_size(info
->index_size
);
281 tiler_prefix
->indices
= 0;
282 *vertex_count
= ctx
->vertex_count
;
283 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= info
->start
;
284 tiler_prefix
->offset_bias_correction
= 0;
285 tiler_prefix
->index_count
= MALI_POSITIVE(ctx
->vertex_count
);
288 tiler_prefix
->unknown_draw
= draw_flags
;
290 /* Encode the padded vertex count */
292 if (info
->instance_count
> 1) {
293 *padded_count
= panfrost_padded_vertex_count(*vertex_count
);
295 unsigned shift
= __builtin_ctz(ctx
->padded_count
);
296 unsigned k
= ctx
->padded_count
>> (shift
+ 1);
298 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= shift
;
299 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= k
;
301 *padded_count
= *vertex_count
;
303 /* Reset instancing state */
304 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= 0;
305 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= 0;
310 panfrost_emit_compute_shader(struct panfrost_context
*ctx
,
311 enum pipe_shader_type st
,
312 struct mali_shader_meta
*meta
)
314 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
315 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
317 memset(meta
, 0, sizeof(*meta
));
318 meta
->shader
= ss
->shader
;
319 meta
->attribute_count
= ss
->attribute_count
;
320 meta
->varying_count
= ss
->varying_count
;
321 meta
->texture_count
= ctx
->sampler_view_count
[st
];
322 meta
->sampler_count
= ctx
->sampler_count
[st
];
324 if (dev
->quirks
& IS_BIFROST
) {
325 struct mali_bifrost_properties_packed prop
;
326 struct mali_preload_vertex_packed preload
;
328 pan_pack(&prop
, BIFROST_PROPERTIES
, cfg
) {
329 cfg
.unknown
= 0x800000; /* XXX */
330 cfg
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
333 /* TODO: True compute shaders */
334 pan_pack(&preload
, PRELOAD_VERTEX
, cfg
) {
335 cfg
.uniform_count
= ss
->uniform_count
;
336 cfg
.vertex_id
= true;
337 cfg
.instance_id
= true;
340 memcpy(&meta
->bifrost_props
, &prop
, sizeof(prop
));
341 memcpy(&meta
->bifrost_preload
, &preload
, sizeof(preload
));
343 struct mali_midgard_properties_packed prop
;
345 pan_pack(&prop
, MIDGARD_PROPERTIES
, cfg
) {
346 cfg
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
347 cfg
.uniform_count
= ss
->uniform_count
;
348 cfg
.work_register_count
= ss
->work_reg_count
;
349 cfg
.writes_globals
= ss
->writes_global
;
350 cfg
.suppress_inf_nan
= true; /* XXX */
353 memcpy(&meta
->midgard_props
, &prop
, sizeof(prop
));
358 translate_tex_wrap(enum pipe_tex_wrap w
)
361 case PIPE_TEX_WRAP_REPEAT
: return MALI_WRAP_MODE_REPEAT
;
362 case PIPE_TEX_WRAP_CLAMP
: return MALI_WRAP_MODE_CLAMP
;
363 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
: return MALI_WRAP_MODE_CLAMP_TO_EDGE
;
364 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
: return MALI_WRAP_MODE_CLAMP_TO_BORDER
;
365 case PIPE_TEX_WRAP_MIRROR_REPEAT
: return MALI_WRAP_MODE_MIRRORED_REPEAT
;
366 case PIPE_TEX_WRAP_MIRROR_CLAMP
: return MALI_WRAP_MODE_MIRRORED_CLAMP
;
367 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
: return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_EDGE
;
368 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
: return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_BORDER
;
369 default: unreachable("Invalid wrap");
373 /* The hardware compares in the wrong order order, so we have to flip before
374 * encoding. Yes, really. */
376 static enum mali_func
377 panfrost_sampler_compare_func(const struct pipe_sampler_state
*cso
)
379 if (!cso
->compare_mode
)
380 return MALI_FUNC_NEVER
;
382 enum mali_func f
= panfrost_translate_compare_func(cso
->compare_func
);
383 return panfrost_flip_compare_func(f
);
386 static enum mali_mipmap_mode
387 pan_pipe_to_mipmode(enum pipe_tex_mipfilter f
)
390 case PIPE_TEX_MIPFILTER_NEAREST
: return MALI_MIPMAP_MODE_NEAREST
;
391 case PIPE_TEX_MIPFILTER_LINEAR
: return MALI_MIPMAP_MODE_TRILINEAR
;
392 case PIPE_TEX_MIPFILTER_NONE
: return MALI_MIPMAP_MODE_NONE
;
393 default: unreachable("Invalid");
397 void panfrost_sampler_desc_init(const struct pipe_sampler_state
*cso
,
398 struct mali_midgard_sampler_packed
*hw
)
400 pan_pack(hw
, MIDGARD_SAMPLER
, cfg
) {
401 cfg
.magnify_nearest
= cso
->mag_img_filter
== PIPE_TEX_FILTER_NEAREST
;
402 cfg
.minify_nearest
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
;
403 cfg
.mipmap_mode
= (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) ?
404 MALI_MIPMAP_MODE_TRILINEAR
: MALI_MIPMAP_MODE_NEAREST
;
405 cfg
.normalized_coordinates
= cso
->normalized_coords
;
407 cfg
.lod_bias
= FIXED_16(cso
->lod_bias
, true);
409 cfg
.minimum_lod
= FIXED_16(cso
->min_lod
, false);
411 /* If necessary, we disable mipmapping in the sampler descriptor by
412 * clamping the LOD as tight as possible (from 0 to epsilon,
413 * essentially -- remember these are fixed point numbers, so
416 cfg
.maximum_lod
= (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
) ?
417 cfg
.minimum_lod
+ 1 :
418 FIXED_16(cso
->max_lod
, false);
420 cfg
.wrap_mode_s
= translate_tex_wrap(cso
->wrap_s
);
421 cfg
.wrap_mode_t
= translate_tex_wrap(cso
->wrap_t
);
422 cfg
.wrap_mode_r
= translate_tex_wrap(cso
->wrap_r
);
424 cfg
.compare_function
= panfrost_sampler_compare_func(cso
);
425 cfg
.seamless_cube_map
= cso
->seamless_cube_map
;
427 cfg
.border_color_r
= cso
->border_color
.f
[0];
428 cfg
.border_color_g
= cso
->border_color
.f
[1];
429 cfg
.border_color_b
= cso
->border_color
.f
[2];
430 cfg
.border_color_a
= cso
->border_color
.f
[3];
434 void panfrost_sampler_desc_init_bifrost(const struct pipe_sampler_state
*cso
,
435 struct mali_bifrost_sampler_packed
*hw
)
437 pan_pack(hw
, BIFROST_SAMPLER
, cfg
) {
438 cfg
.magnify_linear
= cso
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
;
439 cfg
.minify_linear
= cso
->min_img_filter
== PIPE_TEX_FILTER_LINEAR
;
440 cfg
.mipmap_mode
= pan_pipe_to_mipmode(cso
->min_mip_filter
);
441 cfg
.normalized_coordinates
= cso
->normalized_coords
;
443 cfg
.lod_bias
= FIXED_16(cso
->lod_bias
, true);
444 cfg
.minimum_lod
= FIXED_16(cso
->min_lod
, false);
445 cfg
.maximum_lod
= FIXED_16(cso
->max_lod
, false);
447 cfg
.wrap_mode_s
= translate_tex_wrap(cso
->wrap_s
);
448 cfg
.wrap_mode_t
= translate_tex_wrap(cso
->wrap_t
);
449 cfg
.wrap_mode_r
= translate_tex_wrap(cso
->wrap_r
);
451 cfg
.compare_function
= panfrost_sampler_compare_func(cso
);
452 cfg
.seamless_cube_map
= cso
->seamless_cube_map
;
457 panfrost_fs_required(
458 struct panfrost_shader_state
*fs
,
459 struct panfrost_blend_final
*blend
,
462 /* If we generally have side effects */
466 /* If colour is written we need to execute */
467 for (unsigned i
= 0; i
< rt_count
; ++i
) {
468 if (!blend
[i
].no_colour
)
472 /* If depth is written and not implied we need to execute.
473 * TODO: Predicate on Z/S writes being enabled */
474 return (fs
->writes_depth
|| fs
->writes_stencil
);
478 panfrost_emit_blend(struct panfrost_batch
*batch
, void *rts
,
479 struct panfrost_blend_final
*blend
)
481 const struct panfrost_device
*dev
= pan_device(batch
->ctx
->base
.screen
);
482 struct panfrost_shader_state
*fs
= panfrost_get_shader_state(batch
->ctx
, PIPE_SHADER_FRAGMENT
);
483 unsigned rt_count
= batch
->key
.nr_cbufs
;
485 struct bifrost_blend_rt
*brts
= rts
;
486 struct midgard_blend_rt
*mrts
= rts
;
488 /* Disable blending for depth-only on Bifrost */
490 if (rt_count
== 0 && dev
->quirks
& IS_BIFROST
)
493 for (unsigned i
= 0; i
< rt_count
; ++i
) {
496 pan_pack(&flags
, BLEND_FLAGS
, cfg
) {
497 if (blend
[i
].no_colour
) {
502 batch
->draws
|= (PIPE_CLEAR_COLOR0
<< i
);
504 cfg
.srgb
= util_format_is_srgb(batch
->key
.cbufs
[i
]->format
);
505 cfg
.load_destination
= blend
[i
].load_dest
;
506 cfg
.dither_disable
= !batch
->ctx
->blend
->base
.dither
;
508 if (!(dev
->quirks
& IS_BIFROST
))
509 cfg
.midgard_blend_shader
= blend
[i
].is_shader
;
512 if (dev
->quirks
& IS_BIFROST
) {
513 brts
[i
].flags
= flags
;
515 if (blend
[i
].is_shader
) {
516 /* The blend shader's address needs to be at
517 * the same top 32 bit as the fragment shader.
518 * TODO: Ensure that's always the case.
520 assert((blend
[i
].shader
.gpu
& (0xffffffffull
<< 32)) ==
521 (fs
->bo
->gpu
& (0xffffffffull
<< 32)));
522 brts
[i
].shader
= blend
[i
].shader
.gpu
;
525 enum pipe_format format
= batch
->key
.cbufs
[i
]->format
;
526 const struct util_format_description
*format_desc
;
527 format_desc
= util_format_description(format
);
529 brts
[i
].equation
= blend
[i
].equation
.equation
;
531 /* TODO: this is a bit more complicated */
532 brts
[i
].constant
= blend
[i
].equation
.constant
;
534 brts
[i
].format
= panfrost_format_to_bifrost_blend(format_desc
);
536 /* 0x19 disables blending and forces REPLACE
537 * mode (equivalent to rgb_mode = alpha_mode =
538 * x122, colour mask = 0xF). 0x1a allows
540 brts
[i
].unk2
= blend
[i
].opaque
? 0x19 : 0x1a;
542 brts
[i
].shader_type
= fs
->blend_types
[i
];
545 memcpy(&mrts
[i
].flags
, &flags
, sizeof(flags
));
547 if (blend
[i
].is_shader
) {
548 mrts
[i
].blend
.shader
= blend
[i
].shader
.gpu
| blend
[i
].shader
.first_tag
;
550 mrts
[i
].blend
.equation
= blend
[i
].equation
.equation
;
551 mrts
[i
].blend
.constant
= blend
[i
].equation
.constant
;
558 panfrost_emit_frag_shader(struct panfrost_context
*ctx
,
559 struct mali_shader_meta
*fragmeta
,
560 struct panfrost_blend_final
*blend
)
562 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
563 struct panfrost_shader_state
*fs
;
565 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
567 struct pipe_rasterizer_state
*rast
= &ctx
->rasterizer
->base
;
568 const struct panfrost_zsa_state
*zsa
= ctx
->depth_stencil
;
569 unsigned rt_count
= ctx
->pipe_framebuffer
.nr_cbufs
;
571 memset(fragmeta
, 0, sizeof(*fragmeta
));
573 fragmeta
->shader
= fs
->shader
;
574 fragmeta
->attribute_count
= fs
->attribute_count
;
575 fragmeta
->varying_count
= fs
->varying_count
;
576 fragmeta
->texture_count
= ctx
->sampler_view_count
[PIPE_SHADER_FRAGMENT
];
577 fragmeta
->sampler_count
= ctx
->sampler_count
[PIPE_SHADER_FRAGMENT
];
579 if (dev
->quirks
& IS_BIFROST
) {
580 struct mali_bifrost_properties_packed prop
;
581 struct mali_preload_fragment_packed preload
;
583 bool no_blend
= true;
585 for (unsigned i
= 0; i
< rt_count
; ++i
)
586 no_blend
&= (!blend
[i
].load_dest
| blend
[i
].no_colour
);
588 pan_pack(&prop
, BIFROST_PROPERTIES
, cfg
) {
589 cfg
.unknown
= 0x950020; /* XXX */
590 cfg
.uniform_buffer_count
= panfrost_ubo_count(ctx
, PIPE_SHADER_FRAGMENT
);
591 cfg
.early_z_enable
= !fs
->can_discard
&& !fs
->writes_depth
&& no_blend
;
594 pan_pack(&preload
, PRELOAD_FRAGMENT
, cfg
) {
595 cfg
.uniform_count
= fs
->uniform_count
;
596 cfg
.fragment_position
= fs
->reads_frag_coord
;
599 memcpy(&fragmeta
->bifrost_props
, &prop
, sizeof(prop
));
600 memcpy(&fragmeta
->bifrost_preload
, &preload
, sizeof(preload
));
602 struct mali_midgard_properties_packed prop
;
604 /* Reasons to disable early-Z from a shader perspective */
605 bool late_z
= fs
->can_discard
|| fs
->writes_global
||
606 fs
->writes_depth
|| fs
->writes_stencil
;
608 /* Reasons to disable early-Z from a CSO perspective */
609 bool alpha_to_coverage
= ctx
->blend
->base
.alpha_to_coverage
;
611 /* If either depth or stencil is enabled, discard matters */
613 (zsa
->base
.depth
.enabled
&& zsa
->base
.depth
.func
!= PIPE_FUNC_ALWAYS
) ||
614 zsa
->base
.stencil
[0].enabled
;
616 bool has_blend_shader
= false;
618 for (unsigned c
= 0; c
< rt_count
; ++c
)
619 has_blend_shader
|= blend
[c
].is_shader
;
621 pan_pack(&prop
, MIDGARD_PROPERTIES
, cfg
) {
622 cfg
.uniform_buffer_count
= panfrost_ubo_count(ctx
, PIPE_SHADER_FRAGMENT
);
623 cfg
.uniform_count
= fs
->uniform_count
;
624 cfg
.work_register_count
= fs
->work_reg_count
;
625 cfg
.writes_globals
= fs
->writes_global
;
626 cfg
.suppress_inf_nan
= true; /* XXX */
628 /* TODO: Reduce this limit? */
629 if (has_blend_shader
)
630 cfg
.work_register_count
= MAX2(cfg
.work_register_count
, 8);
632 cfg
.stencil_from_shader
= fs
->writes_stencil
;
633 cfg
.helper_invocation_enable
= fs
->helper_invocations
;
634 cfg
.depth_source
= fs
->writes_depth
?
635 MALI_DEPTH_SOURCE_SHADER
:
636 MALI_DEPTH_SOURCE_FIXED_FUNCTION
;
638 /* Depend on other state */
639 cfg
.early_z_enable
= !(late_z
|| alpha_to_coverage
);
640 cfg
.reads_tilebuffer
= fs
->outputs_read
|| (!zs_enabled
&& fs
->can_discard
);
641 cfg
.reads_depth_stencil
= zs_enabled
&& fs
->can_discard
;
644 memcpy(&fragmeta
->midgard_props
, &prop
, sizeof(prop
));
647 bool msaa
= rast
->multisample
;
648 fragmeta
->coverage_mask
= msaa
? ctx
->sample_mask
: ~0;
650 fragmeta
->unknown2_3
= MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS
) | 0x10;
651 fragmeta
->unknown2_4
= 0x4e0;
653 /* TODO: Sample size */
654 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, msaa
);
655 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, !msaa
);
657 /* EXT_shader_framebuffer_fetch requires the shader to be run
658 * per-sample when outputs are read. */
659 bool per_sample
= ctx
->min_samples
> 1 || fs
->outputs_read
;
660 SET_BIT(fragmeta
->unknown2_3
, MALI_PER_SAMPLE
, msaa
&& per_sample
);
662 fragmeta
->depth_units
= rast
->offset_units
* 2.0f
;
663 fragmeta
->depth_factor
= rast
->offset_scale
;
665 /* XXX: Which bit is which? Does this maybe allow offseting not-tri? */
667 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
, rast
->offset_tri
);
668 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
, rast
->offset_tri
);
670 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_NEAR
, rast
->depth_clip_near
);
671 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_FAR
, rast
->depth_clip_far
);
673 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
,
674 zsa
->base
.stencil
[0].enabled
);
676 fragmeta
->stencil_mask_front
= zsa
->stencil_mask_front
;
677 fragmeta
->stencil_mask_back
= zsa
->stencil_mask_back
;
679 /* Bottom bits for stencil ref, exactly one word */
680 fragmeta
->stencil_front
.opaque
[0] = zsa
->stencil_front
.opaque
[0] | ctx
->stencil_ref
.ref_value
[0];
682 /* If back-stencil is not enabled, use the front values */
684 if (zsa
->base
.stencil
[1].enabled
)
685 fragmeta
->stencil_back
.opaque
[0] = zsa
->stencil_back
.opaque
[0] | ctx
->stencil_ref
.ref_value
[1];
687 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
689 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
,
690 zsa
->base
.depth
.writemask
);
692 fragmeta
->unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
693 fragmeta
->unknown2_3
|= MALI_DEPTH_FUNC(panfrost_translate_compare_func(
694 zsa
->base
.depth
.enabled
? zsa
->base
.depth
.func
: PIPE_FUNC_ALWAYS
));
696 SET_BIT(fragmeta
->unknown2_4
, MALI_ALPHA_TO_COVERAGE
,
697 ctx
->blend
->base
.alpha_to_coverage
);
699 /* Disable shader execution if we can */
700 if (dev
->quirks
& MIDGARD_SHADERLESS
701 && !panfrost_fs_required(fs
, blend
, rt_count
)) {
702 fragmeta
->shader
= 0x1;
703 fragmeta
->attribute_count
= 0;
704 fragmeta
->varying_count
= 0;
705 fragmeta
->texture_count
= 0;
706 fragmeta
->sampler_count
= 0;
708 /* This feature is not known to work on Bifrost */
709 struct mali_midgard_properties_packed prop
;
711 pan_pack(&prop
, MIDGARD_PROPERTIES
, cfg
) {
712 cfg
.work_register_count
= 1;
713 cfg
.depth_source
= MALI_DEPTH_SOURCE_FIXED_FUNCTION
;
714 cfg
.early_z_enable
= true;
717 memcpy(&fragmeta
->midgard_props
, &prop
, sizeof(prop
));
720 if (dev
->quirks
& MIDGARD_SFBD
) {
721 /* When only a single render target platform is used, the blend
722 * information is inside the shader meta itself. We additionally
723 * need to signal CAN_DISCARD for nontrivial blend modes (so
724 * we're able to read back the destination buffer) */
726 if (blend
[0].no_colour
)
729 fragmeta
->unknown2_4
|= MALI_SFBD_ENABLE
;
731 SET_BIT(fragmeta
->unknown2_4
, MALI_SFBD_SRGB
,
732 util_format_is_srgb(ctx
->pipe_framebuffer
.cbufs
[0]->format
));
734 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_BLEND_SHADER
,
737 if (blend
[0].is_shader
) {
738 fragmeta
->blend
.shader
= blend
[0].shader
.gpu
|
739 blend
[0].shader
.first_tag
;
741 fragmeta
->blend
.equation
= blend
[0].equation
.equation
;
742 fragmeta
->blend
.constant
= blend
[0].equation
.constant
;
745 SET_BIT(fragmeta
->unknown2_3
, MALI_CAN_DISCARD
,
748 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_DITHER
, !ctx
->blend
->base
.dither
);
749 } else if (!(dev
->quirks
& IS_BIFROST
)) {
750 /* Bug where MRT-capable hw apparently reads the last blend
751 * shader from here instead of the usual location? */
753 for (signed rt
= ((signed) rt_count
- 1); rt
>= 0; --rt
) {
754 if (!blend
[rt
].is_shader
)
757 fragmeta
->blend
.shader
= blend
[rt
].shader
.gpu
|
758 blend
[rt
].shader
.first_tag
;
765 panfrost_emit_shader_meta(struct panfrost_batch
*batch
,
766 enum pipe_shader_type st
,
767 struct mali_vertex_tiler_postfix
*postfix
)
769 struct panfrost_context
*ctx
= batch
->ctx
;
770 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
777 struct mali_shader_meta meta
;
779 /* Add the shader BO to the batch. */
780 panfrost_batch_add_bo(batch
, ss
->bo
,
781 PAN_BO_ACCESS_PRIVATE
|
783 panfrost_bo_access_for_stage(st
));
787 if (st
== PIPE_SHADER_FRAGMENT
) {
788 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
789 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
790 size_t desc_size
= sizeof(meta
);
792 struct panfrost_transfer xfer
;
795 if (dev
->quirks
& MIDGARD_SFBD
)
797 else if (dev
->quirks
& IS_BIFROST
)
798 rt_size
= sizeof(struct bifrost_blend_rt
);
800 rt_size
= sizeof(struct midgard_blend_rt
);
802 desc_size
+= rt_size
* rt_count
;
805 rts
= rzalloc_size(ctx
, rt_size
* rt_count
);
807 struct panfrost_blend_final blend
[PIPE_MAX_COLOR_BUFS
];
809 for (unsigned c
= 0; c
< ctx
->pipe_framebuffer
.nr_cbufs
; ++c
)
810 blend
[c
] = panfrost_get_blend_for_context(ctx
, c
);
812 panfrost_emit_frag_shader(ctx
, &meta
, blend
);
814 if (!(dev
->quirks
& MIDGARD_SFBD
))
815 panfrost_emit_blend(batch
, rts
, blend
);
817 batch
->draws
|= PIPE_CLEAR_COLOR0
;
819 xfer
= panfrost_pool_alloc_aligned(&batch
->pool
, desc_size
, sizeof(meta
));
821 memcpy(xfer
.cpu
, &meta
, sizeof(meta
));
822 memcpy(xfer
.cpu
+ sizeof(meta
), rts
, rt_size
* rt_count
);
827 shader_ptr
= xfer
.gpu
;
829 panfrost_emit_compute_shader(ctx
, st
, &meta
);
831 shader_ptr
= panfrost_pool_upload(&batch
->pool
, &meta
,
835 postfix
->shader
= shader_ptr
;
839 panfrost_emit_viewport(struct panfrost_batch
*batch
,
840 struct mali_vertex_tiler_postfix
*tiler_postfix
)
842 struct panfrost_context
*ctx
= batch
->ctx
;
843 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
844 const struct pipe_scissor_state
*ss
= &ctx
->scissor
;
845 const struct pipe_rasterizer_state
*rast
= &ctx
->rasterizer
->base
;
846 const struct pipe_framebuffer_state
*fb
= &ctx
->pipe_framebuffer
;
848 /* Derive min/max from translate/scale. Note since |x| >= 0 by
849 * definition, we have that -|x| <= |x| hence translate - |scale| <=
850 * translate + |scale|, so the ordering is correct here. */
851 float vp_minx
= (int) (vp
->translate
[0] - fabsf(vp
->scale
[0]));
852 float vp_maxx
= (int) (vp
->translate
[0] + fabsf(vp
->scale
[0]));
853 float vp_miny
= (int) (vp
->translate
[1] - fabsf(vp
->scale
[1]));
854 float vp_maxy
= (int) (vp
->translate
[1] + fabsf(vp
->scale
[1]));
855 float minz
= (vp
->translate
[2] - fabsf(vp
->scale
[2]));
856 float maxz
= (vp
->translate
[2] + fabsf(vp
->scale
[2]));
858 /* Scissor to the intersection of viewport and to the scissor, clamped
859 * to the framebuffer */
861 unsigned minx
= MIN2(fb
->width
, vp_minx
);
862 unsigned maxx
= MIN2(fb
->width
, vp_maxx
);
863 unsigned miny
= MIN2(fb
->height
, vp_miny
);
864 unsigned maxy
= MIN2(fb
->height
, vp_maxy
);
866 if (ss
&& rast
->scissor
) {
867 minx
= MAX2(ss
->minx
, minx
);
868 miny
= MAX2(ss
->miny
, miny
);
869 maxx
= MIN2(ss
->maxx
, maxx
);
870 maxy
= MIN2(ss
->maxy
, maxy
);
873 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
, MALI_VIEWPORT_LENGTH
);
875 pan_pack(T
.cpu
, VIEWPORT
, cfg
) {
876 cfg
.scissor_minimum_x
= minx
;
877 cfg
.scissor_minimum_y
= miny
;
878 cfg
.scissor_maximum_x
= maxx
- 1;
879 cfg
.scissor_maximum_y
= maxy
- 1;
881 cfg
.minimum_z
= rast
->depth_clip_near
? minz
: -INFINITY
;
882 cfg
.maximum_z
= rast
->depth_clip_far
? maxz
: INFINITY
;
885 tiler_postfix
->viewport
= T
.gpu
;
886 panfrost_batch_union_scissor(batch
, minx
, miny
, maxx
, maxy
);
890 panfrost_map_constant_buffer_gpu(struct panfrost_batch
*batch
,
891 enum pipe_shader_type st
,
892 struct panfrost_constant_buffer
*buf
,
895 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
896 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
899 panfrost_batch_add_bo(batch
, rsrc
->bo
,
900 PAN_BO_ACCESS_SHARED
|
902 panfrost_bo_access_for_stage(st
));
904 /* Alignment gauranteed by
905 * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
906 return rsrc
->bo
->gpu
+ cb
->buffer_offset
;
907 } else if (cb
->user_buffer
) {
908 return panfrost_pool_upload_aligned(&batch
->pool
,
911 cb
->buffer_size
, 16);
913 unreachable("No constant buffer");
917 struct sysval_uniform
{
927 panfrost_upload_viewport_scale_sysval(struct panfrost_batch
*batch
,
928 struct sysval_uniform
*uniform
)
930 struct panfrost_context
*ctx
= batch
->ctx
;
931 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
933 uniform
->f
[0] = vp
->scale
[0];
934 uniform
->f
[1] = vp
->scale
[1];
935 uniform
->f
[2] = vp
->scale
[2];
939 panfrost_upload_viewport_offset_sysval(struct panfrost_batch
*batch
,
940 struct sysval_uniform
*uniform
)
942 struct panfrost_context
*ctx
= batch
->ctx
;
943 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
945 uniform
->f
[0] = vp
->translate
[0];
946 uniform
->f
[1] = vp
->translate
[1];
947 uniform
->f
[2] = vp
->translate
[2];
950 static void panfrost_upload_txs_sysval(struct panfrost_batch
*batch
,
951 enum pipe_shader_type st
,
952 unsigned int sysvalid
,
953 struct sysval_uniform
*uniform
)
955 struct panfrost_context
*ctx
= batch
->ctx
;
956 unsigned texidx
= PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid
);
957 unsigned dim
= PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid
);
958 bool is_array
= PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid
);
959 struct pipe_sampler_view
*tex
= &ctx
->sampler_views
[st
][texidx
]->base
;
962 uniform
->i
[0] = u_minify(tex
->texture
->width0
, tex
->u
.tex
.first_level
);
965 uniform
->i
[1] = u_minify(tex
->texture
->height0
,
966 tex
->u
.tex
.first_level
);
969 uniform
->i
[2] = u_minify(tex
->texture
->depth0
,
970 tex
->u
.tex
.first_level
);
973 uniform
->i
[dim
] = tex
->texture
->array_size
;
977 panfrost_upload_ssbo_sysval(struct panfrost_batch
*batch
,
978 enum pipe_shader_type st
,
980 struct sysval_uniform
*uniform
)
982 struct panfrost_context
*ctx
= batch
->ctx
;
984 assert(ctx
->ssbo_mask
[st
] & (1 << ssbo_id
));
985 struct pipe_shader_buffer sb
= ctx
->ssbo
[st
][ssbo_id
];
987 /* Compute address */
988 struct panfrost_bo
*bo
= pan_resource(sb
.buffer
)->bo
;
990 panfrost_batch_add_bo(batch
, bo
,
991 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_RW
|
992 panfrost_bo_access_for_stage(st
));
994 /* Upload address and size as sysval */
995 uniform
->du
[0] = bo
->gpu
+ sb
.buffer_offset
;
996 uniform
->u
[2] = sb
.buffer_size
;
1000 panfrost_upload_sampler_sysval(struct panfrost_batch
*batch
,
1001 enum pipe_shader_type st
,
1003 struct sysval_uniform
*uniform
)
1005 struct panfrost_context
*ctx
= batch
->ctx
;
1006 struct pipe_sampler_state
*sampl
= &ctx
->samplers
[st
][samp_idx
]->base
;
1008 uniform
->f
[0] = sampl
->min_lod
;
1009 uniform
->f
[1] = sampl
->max_lod
;
1010 uniform
->f
[2] = sampl
->lod_bias
;
1012 /* Even without any errata, Midgard represents "no mipmapping" as
1013 * fixing the LOD with the clamps; keep behaviour consistent. c.f.
1014 * panfrost_create_sampler_state which also explains our choice of
1015 * epsilon value (again to keep behaviour consistent) */
1017 if (sampl
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
1018 uniform
->f
[1] = uniform
->f
[0] + (1.0/256.0);
1022 panfrost_upload_num_work_groups_sysval(struct panfrost_batch
*batch
,
1023 struct sysval_uniform
*uniform
)
1025 struct panfrost_context
*ctx
= batch
->ctx
;
1027 uniform
->u
[0] = ctx
->compute_grid
->grid
[0];
1028 uniform
->u
[1] = ctx
->compute_grid
->grid
[1];
1029 uniform
->u
[2] = ctx
->compute_grid
->grid
[2];
1033 panfrost_upload_sysvals(struct panfrost_batch
*batch
, void *buf
,
1034 struct panfrost_shader_state
*ss
,
1035 enum pipe_shader_type st
)
1037 struct sysval_uniform
*uniforms
= (void *)buf
;
1039 for (unsigned i
= 0; i
< ss
->sysval_count
; ++i
) {
1040 int sysval
= ss
->sysval
[i
];
1042 switch (PAN_SYSVAL_TYPE(sysval
)) {
1043 case PAN_SYSVAL_VIEWPORT_SCALE
:
1044 panfrost_upload_viewport_scale_sysval(batch
,
1047 case PAN_SYSVAL_VIEWPORT_OFFSET
:
1048 panfrost_upload_viewport_offset_sysval(batch
,
1051 case PAN_SYSVAL_TEXTURE_SIZE
:
1052 panfrost_upload_txs_sysval(batch
, st
,
1053 PAN_SYSVAL_ID(sysval
),
1056 case PAN_SYSVAL_SSBO
:
1057 panfrost_upload_ssbo_sysval(batch
, st
,
1058 PAN_SYSVAL_ID(sysval
),
1061 case PAN_SYSVAL_NUM_WORK_GROUPS
:
1062 panfrost_upload_num_work_groups_sysval(batch
,
1065 case PAN_SYSVAL_SAMPLER
:
1066 panfrost_upload_sampler_sysval(batch
, st
,
1067 PAN_SYSVAL_ID(sysval
),
1077 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer
*buf
,
1080 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1081 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1084 return rsrc
->bo
->cpu
;
1085 else if (cb
->user_buffer
)
1086 return cb
->user_buffer
;
1088 unreachable("No constant buffer");
1092 panfrost_emit_const_buf(struct panfrost_batch
*batch
,
1093 enum pipe_shader_type stage
,
1094 struct mali_vertex_tiler_postfix
*postfix
)
1096 struct panfrost_context
*ctx
= batch
->ctx
;
1097 struct panfrost_shader_variants
*all
= ctx
->shader
[stage
];
1102 struct panfrost_constant_buffer
*buf
= &ctx
->constant_buffer
[stage
];
1104 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1106 /* Uniforms are implicitly UBO #0 */
1107 bool has_uniforms
= buf
->enabled_mask
& (1 << 0);
1109 /* Allocate room for the sysval and the uniforms */
1110 size_t sys_size
= sizeof(float) * 4 * ss
->sysval_count
;
1111 size_t uniform_size
= has_uniforms
? (buf
->cb
[0].buffer_size
) : 0;
1112 size_t size
= sys_size
+ uniform_size
;
1113 struct panfrost_transfer transfer
=
1114 panfrost_pool_alloc_aligned(&batch
->pool
, size
, 16);
1116 /* Upload sysvals requested by the shader */
1117 panfrost_upload_sysvals(batch
, transfer
.cpu
, ss
, stage
);
1119 /* Upload uniforms */
1120 if (has_uniforms
&& uniform_size
) {
1121 const void *cpu
= panfrost_map_constant_buffer_cpu(buf
, 0);
1122 memcpy(transfer
.cpu
+ sys_size
, cpu
, uniform_size
);
1125 /* Next up, attach UBOs. UBO #0 is the uniforms we just
1128 unsigned ubo_count
= panfrost_ubo_count(ctx
, stage
);
1129 assert(ubo_count
>= 1);
1131 size_t sz
= MALI_UNIFORM_BUFFER_LENGTH
* ubo_count
;
1132 struct panfrost_transfer ubos
=
1133 panfrost_pool_alloc_aligned(&batch
->pool
, sz
,
1134 MALI_UNIFORM_BUFFER_LENGTH
);
1136 uint64_t *ubo_ptr
= (uint64_t *) ubos
.cpu
;
1138 /* Upload uniforms as a UBO */
1141 pan_pack(ubo_ptr
, UNIFORM_BUFFER
, cfg
) {
1142 cfg
.entries
= DIV_ROUND_UP(size
, 16);
1143 cfg
.pointer
= transfer
.gpu
;
1149 /* The rest are honest-to-goodness UBOs */
1151 for (unsigned ubo
= 1; ubo
< ubo_count
; ++ubo
) {
1152 size_t usz
= buf
->cb
[ubo
].buffer_size
;
1153 bool enabled
= buf
->enabled_mask
& (1 << ubo
);
1154 bool empty
= usz
== 0;
1156 if (!enabled
|| empty
) {
1161 pan_pack(ubo_ptr
+ ubo
, UNIFORM_BUFFER
, cfg
) {
1162 cfg
.entries
= DIV_ROUND_UP(usz
, 16);
1163 cfg
.pointer
= panfrost_map_constant_buffer_gpu(batch
,
1168 postfix
->uniforms
= transfer
.gpu
;
1169 postfix
->uniform_buffers
= ubos
.gpu
;
1171 buf
->dirty_mask
= 0;
1175 panfrost_emit_shared_memory(struct panfrost_batch
*batch
,
1176 const struct pipe_grid_info
*info
,
1177 struct midgard_payload_vertex_tiler
*vtp
)
1179 struct panfrost_context
*ctx
= batch
->ctx
;
1180 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
1181 struct panfrost_shader_variants
*all
= ctx
->shader
[PIPE_SHADER_COMPUTE
];
1182 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1183 unsigned single_size
= util_next_power_of_two(MAX2(ss
->shared_size
,
1186 unsigned log2_instances
=
1187 util_logbase2_ceil(info
->grid
[0]) +
1188 util_logbase2_ceil(info
->grid
[1]) +
1189 util_logbase2_ceil(info
->grid
[2]);
1191 unsigned shared_size
= single_size
* (1 << log2_instances
) * dev
->core_count
;
1192 struct panfrost_bo
*bo
= panfrost_batch_get_shared_memory(batch
,
1196 struct mali_shared_memory shared
= {
1197 .shared_memory
= bo
->gpu
,
1198 .shared_workgroup_count
= log2_instances
,
1199 .shared_shift
= util_logbase2(single_size
) + 1
1202 vtp
->postfix
.shared_memory
= panfrost_pool_upload_aligned(&batch
->pool
, &shared
,
1203 sizeof(shared
), 64);
1207 panfrost_get_tex_desc(struct panfrost_batch
*batch
,
1208 enum pipe_shader_type st
,
1209 struct panfrost_sampler_view
*view
)
1212 return (mali_ptr
) 0;
1214 struct pipe_sampler_view
*pview
= &view
->base
;
1215 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1217 /* Add the BO to the job so it's retained until the job is done. */
1219 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1220 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1221 panfrost_bo_access_for_stage(st
));
1223 panfrost_batch_add_bo(batch
, view
->bo
,
1224 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1225 panfrost_bo_access_for_stage(st
));
1227 return view
->bo
->gpu
;
1231 panfrost_update_sampler_view(struct panfrost_sampler_view
*view
,
1232 struct pipe_context
*pctx
)
1234 struct panfrost_resource
*rsrc
= pan_resource(view
->base
.texture
);
1235 if (view
->texture_bo
!= rsrc
->bo
->gpu
||
1236 view
->modifier
!= rsrc
->modifier
) {
1237 panfrost_bo_unreference(view
->bo
);
1238 panfrost_create_sampler_view_bo(view
, pctx
, &rsrc
->base
);
1243 panfrost_emit_texture_descriptors(struct panfrost_batch
*batch
,
1244 enum pipe_shader_type stage
,
1245 struct mali_vertex_tiler_postfix
*postfix
)
1247 struct panfrost_context
*ctx
= batch
->ctx
;
1248 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1250 if (!ctx
->sampler_view_count
[stage
])
1253 if (device
->quirks
& IS_BIFROST
) {
1254 struct panfrost_transfer T
= panfrost_pool_alloc_aligned(&batch
->pool
,
1255 MALI_BIFROST_TEXTURE_LENGTH
*
1256 ctx
->sampler_view_count
[stage
],
1257 MALI_BIFROST_TEXTURE_LENGTH
);
1259 struct mali_bifrost_texture_packed
*out
=
1260 (struct mali_bifrost_texture_packed
*) T
.cpu
;
1262 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1263 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1264 struct pipe_sampler_view
*pview
= &view
->base
;
1265 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1267 panfrost_update_sampler_view(view
, &ctx
->base
);
1268 out
[i
] = view
->bifrost_descriptor
;
1270 /* Add the BOs to the job so they are retained until the job is done. */
1272 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1273 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1274 panfrost_bo_access_for_stage(stage
));
1276 panfrost_batch_add_bo(batch
, view
->bo
,
1277 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1278 panfrost_bo_access_for_stage(stage
));
1281 postfix
->textures
= T
.gpu
;
1283 uint64_t trampolines
[PIPE_MAX_SHADER_SAMPLER_VIEWS
];
1285 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1286 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1288 panfrost_update_sampler_view(view
, &ctx
->base
);
1290 trampolines
[i
] = panfrost_get_tex_desc(batch
, stage
, view
);
1293 postfix
->textures
= panfrost_pool_upload_aligned(&batch
->pool
,
1296 ctx
->sampler_view_count
[stage
],
1302 panfrost_emit_sampler_descriptors(struct panfrost_batch
*batch
,
1303 enum pipe_shader_type stage
,
1304 struct mali_vertex_tiler_postfix
*postfix
)
1306 struct panfrost_context
*ctx
= batch
->ctx
;
1308 if (!ctx
->sampler_count
[stage
])
1311 size_t desc_size
= MALI_BIFROST_SAMPLER_LENGTH
;
1312 assert(MALI_BIFROST_SAMPLER_LENGTH
== MALI_MIDGARD_SAMPLER_LENGTH
);
1314 size_t sz
= desc_size
* ctx
->sampler_count
[stage
];
1315 struct panfrost_transfer T
= panfrost_pool_alloc_aligned(&batch
->pool
, sz
, desc_size
);
1316 struct mali_midgard_sampler_packed
*out
= (struct mali_midgard_sampler_packed
*) T
.cpu
;
1318 for (unsigned i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1319 out
[i
] = ctx
->samplers
[stage
][i
]->hw
;
1321 postfix
->sampler_descriptor
= T
.gpu
;
1325 panfrost_emit_vertex_data(struct panfrost_batch
*batch
,
1326 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1328 struct panfrost_context
*ctx
= batch
->ctx
;
1329 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1330 struct panfrost_shader_state
*vs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_VERTEX
);
1332 unsigned instance_shift
= vertex_postfix
->instance_shift
;
1333 unsigned instance_odd
= vertex_postfix
->instance_odd
;
1335 /* Worst case: everything is NPOT, which is only possible if instancing
1336 * is enabled. Otherwise single record is gauranteed */
1337 bool could_npot
= instance_shift
|| instance_odd
;
1339 struct panfrost_transfer S
= panfrost_pool_alloc_aligned(&batch
->pool
,
1340 MALI_ATTRIBUTE_BUFFER_LENGTH
* vs
->attribute_count
*
1341 (could_npot
? 2 : 1),
1342 MALI_ATTRIBUTE_BUFFER_LENGTH
* 2);
1344 struct panfrost_transfer T
= panfrost_pool_alloc_aligned(&batch
->pool
,
1345 MALI_ATTRIBUTE_LENGTH
* vs
->attribute_count
,
1346 MALI_ATTRIBUTE_LENGTH
);
1348 struct mali_attribute_buffer_packed
*bufs
=
1349 (struct mali_attribute_buffer_packed
*) S
.cpu
;
1351 struct mali_attribute_packed
*out
=
1352 (struct mali_attribute_packed
*) T
.cpu
;
1354 unsigned attrib_to_buffer
[PIPE_MAX_ATTRIBS
] = { 0 };
1357 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1358 /* We map buffers 1:1 with the attributes, which
1359 * means duplicating some vertex buffers (who cares? aside from
1360 * maybe some caching implications but I somehow doubt that
1363 struct pipe_vertex_element
*elem
= &so
->pipe
[i
];
1364 unsigned vbi
= elem
->vertex_buffer_index
;
1365 attrib_to_buffer
[i
] = k
;
1367 if (!(ctx
->vb_mask
& (1 << vbi
)))
1370 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1371 struct panfrost_resource
*rsrc
;
1373 rsrc
= pan_resource(buf
->buffer
.resource
);
1377 /* Add a dependency of the batch on the vertex buffer */
1378 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1379 PAN_BO_ACCESS_SHARED
|
1380 PAN_BO_ACCESS_READ
|
1381 PAN_BO_ACCESS_VERTEX_TILER
);
1383 /* Mask off lower bits, see offset fixup below */
1384 mali_ptr raw_addr
= rsrc
->bo
->gpu
+ buf
->buffer_offset
;
1385 mali_ptr addr
= raw_addr
& ~63;
1387 /* Since we advanced the base pointer, we shrink the buffer
1388 * size, but add the offset we subtracted */
1389 unsigned size
= rsrc
->base
.width0
+ (raw_addr
- addr
)
1390 - buf
->buffer_offset
;
1392 /* When there is a divisor, the hardware-level divisor is
1393 * the product of the instance divisor and the padded count */
1394 unsigned divisor
= elem
->instance_divisor
;
1395 unsigned hw_divisor
= ctx
->padded_count
* divisor
;
1396 unsigned stride
= buf
->stride
;
1398 /* If there's a divisor(=1) but no instancing, we want every
1399 * attribute to be the same */
1401 if (divisor
&& ctx
->instance_count
== 1)
1404 if (!divisor
|| ctx
->instance_count
<= 1) {
1405 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1406 if (ctx
->instance_count
> 1)
1407 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_MODULUS
;
1410 cfg
.stride
= stride
;
1412 cfg
.divisor_r
= instance_shift
;
1413 cfg
.divisor_p
= instance_odd
;
1415 } else if (util_is_power_of_two_or_zero(hw_divisor
)) {
1416 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1417 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_POT_DIVISOR
;
1419 cfg
.stride
= stride
;
1421 cfg
.divisor_r
= __builtin_ctz(hw_divisor
);
1425 unsigned shift
= 0, extra_flags
= 0;
1427 unsigned magic_divisor
=
1428 panfrost_compute_magic_divisor(hw_divisor
, &shift
, &extra_flags
);
1430 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1431 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR
;
1433 cfg
.stride
= stride
;
1436 cfg
.divisor_r
= shift
;
1437 cfg
.divisor_e
= extra_flags
;
1440 pan_pack(bufs
+ k
+ 1, ATTRIBUTE_BUFFER_CONTINUATION_NPOT
, cfg
) {
1441 cfg
.divisor_numerator
= magic_divisor
;
1442 cfg
.divisor
= divisor
;
1451 /* Add special gl_VertexID/gl_InstanceID buffers */
1453 if (unlikely(vs
->attribute_count
>= PAN_VERTEX_ID
)) {
1454 panfrost_vertex_id(ctx
->padded_count
, &bufs
[k
], ctx
->instance_count
> 1);
1456 pan_pack(out
+ PAN_VERTEX_ID
, ATTRIBUTE
, cfg
) {
1457 cfg
.buffer_index
= k
++;
1458 cfg
.format
= so
->formats
[PAN_VERTEX_ID
];
1461 panfrost_instance_id(ctx
->padded_count
, &bufs
[k
], ctx
->instance_count
> 1);
1463 pan_pack(out
+ PAN_INSTANCE_ID
, ATTRIBUTE
, cfg
) {
1464 cfg
.buffer_index
= k
++;
1465 cfg
.format
= so
->formats
[PAN_INSTANCE_ID
];
1469 /* Attribute addresses require 64-byte alignment, so let:
1471 * base' = base & ~63 = base - (base & 63)
1472 * offset' = offset + (base & 63)
1474 * Since base' + offset' = base + offset, these are equivalent
1475 * addressing modes and now base is 64 aligned.
1478 unsigned start
= vertex_postfix
->offset_start
;
1480 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1481 unsigned vbi
= so
->pipe
[i
].vertex_buffer_index
;
1482 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1484 /* Adjust by the masked off bits of the offset. Make sure we
1485 * read src_offset from so->hw (which is not GPU visible)
1486 * rather than target (which is) due to caching effects */
1488 unsigned src_offset
= so
->pipe
[i
].src_offset
;
1490 /* BOs aligned to 4k so guaranteed aligned to 64 */
1491 src_offset
+= (buf
->buffer_offset
& 63);
1493 /* Also, somewhat obscurely per-instance data needs to be
1494 * offset in response to a delayed start in an indexed draw */
1496 if (so
->pipe
[i
].instance_divisor
&& ctx
->instance_count
> 1 && start
)
1497 src_offset
-= buf
->stride
* start
;
1499 pan_pack(out
+ i
, ATTRIBUTE
, cfg
) {
1500 cfg
.buffer_index
= attrib_to_buffer
[i
];
1501 cfg
.format
= so
->formats
[i
];
1502 cfg
.offset
= src_offset
;
1506 vertex_postfix
->attributes
= S
.gpu
;
1507 vertex_postfix
->attribute_meta
= T
.gpu
;
1511 panfrost_emit_varyings(struct panfrost_batch
*batch
,
1512 struct mali_attribute_buffer_packed
*slot
,
1513 unsigned stride
, unsigned count
)
1515 unsigned size
= stride
* count
;
1516 mali_ptr ptr
= panfrost_pool_alloc_aligned(&batch
->invisible_pool
, size
, 64).gpu
;
1518 pan_pack(slot
, ATTRIBUTE_BUFFER
, cfg
) {
1519 cfg
.stride
= stride
;
1528 panfrost_streamout_offset(unsigned stride
, unsigned offset
,
1529 struct pipe_stream_output_target
*target
)
1531 return (target
->buffer_offset
+ (offset
* stride
* 4)) & 63;
1535 panfrost_emit_streamout(struct panfrost_batch
*batch
,
1536 struct mali_attribute_buffer_packed
*slot
,
1537 unsigned stride_words
, unsigned offset
, unsigned count
,
1538 struct pipe_stream_output_target
*target
)
1540 unsigned stride
= stride_words
* 4;
1541 unsigned max_size
= target
->buffer_size
;
1542 unsigned expected_size
= stride
* count
;
1544 /* Grab the BO and bind it to the batch */
1545 struct panfrost_bo
*bo
= pan_resource(target
->buffer
)->bo
;
1547 /* Varyings are WRITE from the perspective of the VERTEX but READ from
1548 * the perspective of the TILER and FRAGMENT.
1550 panfrost_batch_add_bo(batch
, bo
,
1551 PAN_BO_ACCESS_SHARED
|
1553 PAN_BO_ACCESS_VERTEX_TILER
|
1554 PAN_BO_ACCESS_FRAGMENT
);
1556 /* We will have an offset applied to get alignment */
1557 mali_ptr addr
= bo
->gpu
+ target
->buffer_offset
+ (offset
* stride
);
1559 pan_pack(slot
, ATTRIBUTE_BUFFER
, cfg
) {
1560 cfg
.pointer
= (addr
& ~63);
1561 cfg
.stride
= stride
;
1562 cfg
.size
= MIN2(max_size
, expected_size
) + (addr
& 63);
1567 has_point_coord(unsigned mask
, gl_varying_slot loc
)
1569 if ((loc
>= VARYING_SLOT_TEX0
) && (loc
<= VARYING_SLOT_TEX7
))
1570 return (mask
& (1 << (loc
- VARYING_SLOT_TEX0
)));
1571 else if (loc
== VARYING_SLOT_PNTC
)
1572 return (mask
& (1 << 8));
1577 /* Helpers for manipulating stream out information so we can pack varyings
1578 * accordingly. Compute the src_offset for a given captured varying */
1580 static struct pipe_stream_output
*
1581 pan_get_so(struct pipe_stream_output_info
*info
, gl_varying_slot loc
)
1583 for (unsigned i
= 0; i
< info
->num_outputs
; ++i
) {
1584 if (info
->output
[i
].register_index
== loc
)
1585 return &info
->output
[i
];
1588 unreachable("Varying not captured");
1592 pan_varying_size(enum mali_format fmt
)
1594 unsigned type
= MALI_EXTRACT_TYPE(fmt
);
1595 unsigned chan
= MALI_EXTRACT_CHANNELS(fmt
);
1596 unsigned bits
= MALI_EXTRACT_BITS(fmt
);
1599 if (bits
== MALI_CHANNEL_FLOAT
) {
1601 bool fp16
= (type
== MALI_FORMAT_SINT
);
1602 assert(fp16
|| (type
== MALI_FORMAT_UNORM
));
1606 assert(type
>= MALI_FORMAT_SNORM
&& type
<= MALI_FORMAT_SINT
);
1617 /* Indices for named (non-XFB) varyings that are present. These are packed
1618 * tightly so they correspond to a bitfield present (P) indexed by (1 <<
1619 * PAN_VARY_*). This has the nice property that you can lookup the buffer index
1620 * of a given special field given a shift S by:
1622 * idx = popcount(P & ((1 << S) - 1))
1624 * That is... look at all of the varyings that come earlier and count them, the
1625 * count is the new index since plus one. Likewise, the total number of special
1626 * buffers required is simply popcount(P)
1629 enum pan_special_varying
{
1630 PAN_VARY_GENERAL
= 0,
1631 PAN_VARY_POSITION
= 1,
1633 PAN_VARY_PNTCOORD
= 3,
1635 PAN_VARY_FRAGCOORD
= 5,
1641 /* Given a varying, figure out which index it correpsonds to */
1643 static inline unsigned
1644 pan_varying_index(unsigned present
, enum pan_special_varying v
)
1646 unsigned mask
= (1 << v
) - 1;
1647 return util_bitcount(present
& mask
);
1650 /* Get the base offset for XFB buffers, which by convention come after
1651 * everything else. Wrapper function for semantic reasons; by construction this
1652 * is just popcount. */
1654 static inline unsigned
1655 pan_xfb_base(unsigned present
)
1657 return util_bitcount(present
);
1660 /* Computes the present mask for varyings so we can start emitting varying records */
1662 static inline unsigned
1663 pan_varying_present(
1664 struct panfrost_shader_state
*vs
,
1665 struct panfrost_shader_state
*fs
,
1668 /* At the moment we always emit general and position buffers. Not
1669 * strictly necessary but usually harmless */
1671 unsigned present
= (1 << PAN_VARY_GENERAL
) | (1 << PAN_VARY_POSITION
);
1673 /* Enable special buffers by the shader info */
1675 if (vs
->writes_point_size
)
1676 present
|= (1 << PAN_VARY_PSIZ
);
1678 if (fs
->reads_point_coord
)
1679 present
|= (1 << PAN_VARY_PNTCOORD
);
1682 present
|= (1 << PAN_VARY_FACE
);
1684 if (fs
->reads_frag_coord
&& !(quirks
& IS_BIFROST
))
1685 present
|= (1 << PAN_VARY_FRAGCOORD
);
1687 /* Also, if we have a point sprite, we need a point coord buffer */
1689 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1690 gl_varying_slot loc
= fs
->varyings_loc
[i
];
1692 if (has_point_coord(fs
->point_sprite_mask
, loc
))
1693 present
|= (1 << PAN_VARY_PNTCOORD
);
1699 /* Emitters for varying records */
1702 pan_emit_vary(struct mali_attribute_packed
*out
,
1703 unsigned present
, enum pan_special_varying buf
,
1704 unsigned quirks
, enum mali_format format
,
1707 unsigned nr_channels
= MALI_EXTRACT_CHANNELS(format
);
1708 unsigned swizzle
= quirks
& HAS_SWIZZLES
?
1709 panfrost_get_default_swizzle(nr_channels
) :
1710 panfrost_bifrost_swizzle(nr_channels
);
1712 pan_pack(out
, ATTRIBUTE
, cfg
) {
1713 cfg
.buffer_index
= pan_varying_index(present
, buf
);
1714 cfg
.unknown
= quirks
& IS_BIFROST
? 0x0 : 0x1;
1715 cfg
.format
= (format
<< 12) | swizzle
;
1716 cfg
.offset
= offset
;
1720 /* General varying that is unused */
1723 pan_emit_vary_only(struct mali_attribute_packed
*out
,
1724 unsigned present
, unsigned quirks
)
1726 pan_emit_vary(out
, present
, 0, quirks
, MALI_VARYING_DISCARD
, 0);
1729 /* Special records */
1731 static const enum mali_format pan_varying_formats
[PAN_VARY_MAX
] = {
1732 [PAN_VARY_POSITION
] = MALI_VARYING_POS
,
1733 [PAN_VARY_PSIZ
] = MALI_R16F
,
1734 [PAN_VARY_PNTCOORD
] = MALI_R16F
,
1735 [PAN_VARY_FACE
] = MALI_R32I
,
1736 [PAN_VARY_FRAGCOORD
] = MALI_RGBA32F
1740 pan_emit_vary_special(struct mali_attribute_packed
*out
,
1741 unsigned present
, enum pan_special_varying buf
,
1744 assert(buf
< PAN_VARY_MAX
);
1745 pan_emit_vary(out
, present
, buf
, quirks
, pan_varying_formats
[buf
], 0);
1748 static enum mali_format
1749 pan_xfb_format(enum mali_format format
, unsigned nr
)
1751 if (MALI_EXTRACT_BITS(format
) == MALI_CHANNEL_FLOAT
)
1752 return MALI_R32F
| MALI_NR_CHANNELS(nr
);
1754 return MALI_EXTRACT_TYPE(format
) | MALI_NR_CHANNELS(nr
) | MALI_CHANNEL_32
;
1757 /* Transform feedback records. Note struct pipe_stream_output is (if packed as
1758 * a bitfield) 32-bit, smaller than a 64-bit pointer, so may as well pass by
1762 pan_emit_vary_xfb(struct mali_attribute_packed
*out
,
1765 unsigned *streamout_offsets
,
1767 enum mali_format format
,
1768 struct pipe_stream_output o
)
1770 unsigned swizzle
= quirks
& HAS_SWIZZLES
?
1771 panfrost_get_default_swizzle(o
.num_components
) :
1772 panfrost_bifrost_swizzle(o
.num_components
);
1774 pan_pack(out
, ATTRIBUTE
, cfg
) {
1775 /* XFB buffers come after everything else */
1776 cfg
.buffer_index
= pan_xfb_base(present
) + o
.output_buffer
;
1777 cfg
.unknown
= quirks
& IS_BIFROST
? 0x0 : 0x1;
1779 /* Override number of channels and precision to highp */
1780 cfg
.format
= (pan_xfb_format(format
, o
.num_components
) << 12) | swizzle
;
1782 /* Apply given offsets together */
1783 cfg
.offset
= (o
.dst_offset
* 4) /* dwords */
1784 + streamout_offsets
[o
.output_buffer
];
1788 /* Determine if we should capture a varying for XFB. This requires actually
1789 * having a buffer for it. If we don't capture it, we'll fallback to a general
1790 * varying path (linked or unlinked, possibly discarding the write) */
1793 panfrost_xfb_captured(struct panfrost_shader_state
*xfb
,
1794 unsigned loc
, unsigned max_xfb
)
1796 if (!(xfb
->so_mask
& (1ll << loc
)))
1799 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1800 return o
->output_buffer
< max_xfb
;
1804 pan_emit_general_varying(struct mali_attribute_packed
*out
,
1805 struct panfrost_shader_state
*other
,
1806 struct panfrost_shader_state
*xfb
,
1807 gl_varying_slot loc
,
1808 enum mali_format format
,
1811 unsigned *gen_offsets
,
1812 enum mali_format
*gen_formats
,
1813 unsigned *gen_stride
,
1817 /* Check if we're linked */
1818 signed other_idx
= -1;
1820 for (unsigned j
= 0; j
< other
->varying_count
; ++j
) {
1821 if (other
->varyings_loc
[j
] == loc
) {
1827 if (other_idx
< 0) {
1828 pan_emit_vary_only(out
, present
, quirks
);
1832 unsigned offset
= gen_offsets
[other_idx
];
1835 /* We're linked, so allocate a space via a watermark allocation */
1836 enum mali_format alt
= other
->varyings
[other_idx
];
1838 /* Do interpolation at minimum precision */
1839 unsigned size_main
= pan_varying_size(format
);
1840 unsigned size_alt
= pan_varying_size(alt
);
1841 unsigned size
= MIN2(size_main
, size_alt
);
1843 /* If a varying is marked for XFB but not actually captured, we
1844 * should match the format to the format that would otherwise
1845 * be used for XFB, since dEQP checks for invariance here. It's
1846 * unclear if this is required by the spec. */
1848 if (xfb
->so_mask
& (1ull << loc
)) {
1849 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1850 format
= pan_xfb_format(format
, o
->num_components
);
1851 size
= pan_varying_size(format
);
1852 } else if (size
== size_alt
) {
1856 gen_offsets
[idx
] = *gen_stride
;
1857 gen_formats
[other_idx
] = format
;
1858 offset
= *gen_stride
;
1859 *gen_stride
+= size
;
1862 pan_emit_vary(out
, present
, PAN_VARY_GENERAL
, quirks
, format
, offset
);
1865 /* Higher-level wrapper around all of the above, classifying a varying into one
1866 * of the above types */
1869 panfrost_emit_varying(
1870 struct mali_attribute_packed
*out
,
1871 struct panfrost_shader_state
*stage
,
1872 struct panfrost_shader_state
*other
,
1873 struct panfrost_shader_state
*xfb
,
1876 unsigned *streamout_offsets
,
1878 unsigned *gen_offsets
,
1879 enum mali_format
*gen_formats
,
1880 unsigned *gen_stride
,
1885 gl_varying_slot loc
= stage
->varyings_loc
[idx
];
1886 enum mali_format format
= stage
->varyings
[idx
];
1888 /* Override format to match linkage */
1889 if (!should_alloc
&& gen_formats
[idx
])
1890 format
= gen_formats
[idx
];
1892 if (has_point_coord(stage
->point_sprite_mask
, loc
)) {
1893 pan_emit_vary_special(out
, present
, PAN_VARY_PNTCOORD
, quirks
);
1894 } else if (panfrost_xfb_captured(xfb
, loc
, max_xfb
)) {
1895 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1896 pan_emit_vary_xfb(out
, present
, max_xfb
, streamout_offsets
, quirks
, format
, *o
);
1897 } else if (loc
== VARYING_SLOT_POS
) {
1899 pan_emit_vary_special(out
, present
, PAN_VARY_FRAGCOORD
, quirks
);
1901 pan_emit_vary_special(out
, present
, PAN_VARY_POSITION
, quirks
);
1902 } else if (loc
== VARYING_SLOT_PSIZ
) {
1903 pan_emit_vary_special(out
, present
, PAN_VARY_PSIZ
, quirks
);
1904 } else if (loc
== VARYING_SLOT_PNTC
) {
1905 pan_emit_vary_special(out
, present
, PAN_VARY_PNTCOORD
, quirks
);
1906 } else if (loc
== VARYING_SLOT_FACE
) {
1907 pan_emit_vary_special(out
, present
, PAN_VARY_FACE
, quirks
);
1909 pan_emit_general_varying(out
, other
, xfb
, loc
, format
, present
,
1910 quirks
, gen_offsets
, gen_formats
, gen_stride
,
1916 pan_emit_special_input(struct mali_attribute_buffer_packed
*out
,
1918 enum pan_special_varying v
,
1921 if (present
& (1 << v
)) {
1922 unsigned idx
= pan_varying_index(present
, v
);
1924 pan_pack(out
+ idx
, ATTRIBUTE_BUFFER
, cfg
) {
1925 cfg
.special
= special
;
1932 panfrost_emit_varying_descriptor(struct panfrost_batch
*batch
,
1933 unsigned vertex_count
,
1934 struct mali_vertex_tiler_postfix
*vertex_postfix
,
1935 struct mali_vertex_tiler_postfix
*tiler_postfix
,
1936 union midgard_primitive_size
*primitive_size
)
1938 /* Load the shaders */
1939 struct panfrost_context
*ctx
= batch
->ctx
;
1940 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
1941 struct panfrost_shader_state
*vs
, *fs
;
1942 size_t vs_size
, fs_size
;
1944 /* Allocate the varying descriptor */
1946 vs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_VERTEX
);
1947 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
1948 vs_size
= MALI_ATTRIBUTE_LENGTH
* vs
->varying_count
;
1949 fs_size
= MALI_ATTRIBUTE_LENGTH
* fs
->varying_count
;
1951 struct panfrost_transfer trans
= panfrost_pool_alloc_aligned(
1952 &batch
->pool
, vs_size
+ fs_size
, MALI_ATTRIBUTE_LENGTH
);
1954 struct pipe_stream_output_info
*so
= &vs
->stream_output
;
1955 unsigned present
= pan_varying_present(vs
, fs
, dev
->quirks
);
1957 /* Check if this varying is linked by us. This is the case for
1958 * general-purpose, non-captured varyings. If it is, link it. If it's
1959 * not, use the provided stream out information to determine the
1960 * offset, since it was already linked for us. */
1962 unsigned gen_offsets
[32];
1963 enum mali_format gen_formats
[32];
1964 memset(gen_offsets
, 0, sizeof(gen_offsets
));
1965 memset(gen_formats
, 0, sizeof(gen_formats
));
1967 unsigned gen_stride
= 0;
1968 assert(vs
->varying_count
< ARRAY_SIZE(gen_offsets
));
1969 assert(fs
->varying_count
< ARRAY_SIZE(gen_offsets
));
1971 unsigned streamout_offsets
[32];
1973 for (unsigned i
= 0; i
< ctx
->streamout
.num_targets
; ++i
) {
1974 streamout_offsets
[i
] = panfrost_streamout_offset(
1976 ctx
->streamout
.offsets
[i
],
1977 ctx
->streamout
.targets
[i
]);
1980 struct mali_attribute_packed
*ovs
= (struct mali_attribute_packed
*)trans
.cpu
;
1981 struct mali_attribute_packed
*ofs
= ovs
+ vs
->varying_count
;
1983 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1984 panfrost_emit_varying(ovs
+ i
, vs
, fs
, vs
, present
,
1985 ctx
->streamout
.num_targets
, streamout_offsets
,
1987 gen_offsets
, gen_formats
, &gen_stride
, i
, true, false);
1990 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1991 panfrost_emit_varying(ofs
+ i
, fs
, vs
, vs
, present
,
1992 ctx
->streamout
.num_targets
, streamout_offsets
,
1994 gen_offsets
, gen_formats
, &gen_stride
, i
, false, true);
1997 unsigned xfb_base
= pan_xfb_base(present
);
1998 struct panfrost_transfer T
= panfrost_pool_alloc_aligned(&batch
->pool
,
1999 MALI_ATTRIBUTE_BUFFER_LENGTH
* (xfb_base
+ ctx
->streamout
.num_targets
),
2000 MALI_ATTRIBUTE_BUFFER_LENGTH
* 2);
2001 struct mali_attribute_buffer_packed
*varyings
=
2002 (struct mali_attribute_buffer_packed
*) T
.cpu
;
2004 /* Emit the stream out buffers */
2006 unsigned out_count
= u_stream_outputs_for_vertices(ctx
->active_prim
,
2009 for (unsigned i
= 0; i
< ctx
->streamout
.num_targets
; ++i
) {
2010 panfrost_emit_streamout(batch
, &varyings
[xfb_base
+ i
],
2012 ctx
->streamout
.offsets
[i
],
2014 ctx
->streamout
.targets
[i
]);
2017 panfrost_emit_varyings(batch
,
2018 &varyings
[pan_varying_index(present
, PAN_VARY_GENERAL
)],
2019 gen_stride
, vertex_count
);
2021 /* fp32 vec4 gl_Position */
2022 tiler_postfix
->position_varying
= panfrost_emit_varyings(batch
,
2023 &varyings
[pan_varying_index(present
, PAN_VARY_POSITION
)],
2024 sizeof(float) * 4, vertex_count
);
2026 if (present
& (1 << PAN_VARY_PSIZ
)) {
2027 primitive_size
->pointer
= panfrost_emit_varyings(batch
,
2028 &varyings
[pan_varying_index(present
, PAN_VARY_PSIZ
)],
2032 pan_emit_special_input(varyings
, present
, PAN_VARY_PNTCOORD
, MALI_ATTRIBUTE_SPECIAL_POINT_COORD
);
2033 pan_emit_special_input(varyings
, present
, PAN_VARY_FACE
, MALI_ATTRIBUTE_SPECIAL_FRONT_FACING
);
2034 pan_emit_special_input(varyings
, present
, PAN_VARY_FRAGCOORD
, MALI_ATTRIBUTE_SPECIAL_FRAG_COORD
);
2036 vertex_postfix
->varyings
= T
.gpu
;
2037 tiler_postfix
->varyings
= T
.gpu
;
2039 vertex_postfix
->varying_meta
= trans
.gpu
;
2040 tiler_postfix
->varying_meta
= trans
.gpu
+ vs_size
;
2044 panfrost_emit_vertex_tiler_jobs(struct panfrost_batch
*batch
,
2045 struct mali_vertex_tiler_prefix
*vertex_prefix
,
2046 struct mali_vertex_tiler_postfix
*vertex_postfix
,
2047 struct mali_vertex_tiler_prefix
*tiler_prefix
,
2048 struct mali_vertex_tiler_postfix
*tiler_postfix
,
2049 union midgard_primitive_size
*primitive_size
)
2051 struct panfrost_context
*ctx
= batch
->ctx
;
2052 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
2053 bool wallpapering
= ctx
->wallpaper_batch
&& batch
->scoreboard
.tiler_dep
;
2054 struct bifrost_payload_vertex bifrost_vertex
= {0,};
2055 struct bifrost_payload_tiler bifrost_tiler
= {0,};
2056 struct midgard_payload_vertex_tiler midgard_vertex
= {0,};
2057 struct midgard_payload_vertex_tiler midgard_tiler
= {0,};
2059 size_t vp_size
, tp_size
;
2061 if (device
->quirks
& IS_BIFROST
) {
2062 bifrost_vertex
.prefix
= *vertex_prefix
;
2063 bifrost_vertex
.postfix
= *vertex_postfix
;
2064 vp
= &bifrost_vertex
;
2065 vp_size
= sizeof(bifrost_vertex
);
2067 bifrost_tiler
.prefix
= *tiler_prefix
;
2068 bifrost_tiler
.tiler
.primitive_size
= *primitive_size
;
2069 bifrost_tiler
.tiler
.tiler_meta
= panfrost_batch_get_tiler_meta(batch
, ~0);
2070 bifrost_tiler
.postfix
= *tiler_postfix
;
2071 tp
= &bifrost_tiler
;
2072 tp_size
= sizeof(bifrost_tiler
);
2074 midgard_vertex
.prefix
= *vertex_prefix
;
2075 midgard_vertex
.postfix
= *vertex_postfix
;
2076 vp
= &midgard_vertex
;
2077 vp_size
= sizeof(midgard_vertex
);
2079 midgard_tiler
.prefix
= *tiler_prefix
;
2080 midgard_tiler
.postfix
= *tiler_postfix
;
2081 midgard_tiler
.primitive_size
= *primitive_size
;
2082 tp
= &midgard_tiler
;
2083 tp_size
= sizeof(midgard_tiler
);
2087 /* Inject in reverse order, with "predicted" job indices.
2088 * THIS IS A HACK XXX */
2089 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_TILER
, false,
2090 batch
->scoreboard
.job_index
+ 2, tp
, tp_size
, true);
2091 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_VERTEX
, false, 0,
2096 /* If rasterizer discard is enable, only submit the vertex */
2098 unsigned vertex
= panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_VERTEX
, false, 0,
2099 vp
, vp_size
, false);
2101 if (ctx
->rasterizer
->base
.rasterizer_discard
)
2104 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_TILER
, false, vertex
, tp
, tp_size
,
2108 /* TODO: stop hardcoding this */
2110 panfrost_emit_sample_locations(struct panfrost_batch
*batch
)
2112 uint16_t locations
[] = {
2163 return panfrost_pool_upload_aligned(&batch
->pool
, locations
, 96 * sizeof(uint16_t), 64);