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(&batch
->pool
, &shared
, sizeof(shared
));
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 out
= panfrost_pool_upload(&batch
->pool
, ibuf8
+ offset
,
227 u_vbuf_get_minmax_index(&ctx
->base
, info
, min_index
, max_index
);
229 if (!info
->has_user_indices
)
230 panfrost_minmax_cache_add(rsrc
->index_cache
,
231 info
->start
, info
->count
,
232 *min_index
, *max_index
);
239 panfrost_vt_set_draw_info(struct panfrost_context
*ctx
,
240 const struct pipe_draw_info
*info
,
241 enum mali_draw_mode draw_mode
,
242 struct mali_vertex_tiler_postfix
*vertex_postfix
,
243 struct mali_vertex_tiler_prefix
*tiler_prefix
,
244 struct mali_vertex_tiler_postfix
*tiler_postfix
,
245 unsigned *vertex_count
,
246 unsigned *padded_count
)
248 tiler_prefix
->draw_mode
= draw_mode
;
250 unsigned draw_flags
= 0;
252 if (panfrost_writes_point_size(ctx
))
253 draw_flags
|= MALI_DRAW_VARYING_SIZE
;
255 if (info
->primitive_restart
)
256 draw_flags
|= MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX
;
258 /* These doesn't make much sense */
260 draw_flags
|= 0x3000;
262 if (info
->index_size
) {
263 unsigned min_index
= 0, max_index
= 0;
265 tiler_prefix
->indices
= panfrost_get_index_buffer_bounded(ctx
,
270 /* Use the corresponding values */
271 *vertex_count
= max_index
- min_index
+ 1;
272 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= min_index
+ info
->index_bias
;
273 tiler_prefix
->offset_bias_correction
= -min_index
;
274 tiler_prefix
->index_count
= MALI_POSITIVE(info
->count
);
275 draw_flags
|= panfrost_translate_index_size(info
->index_size
);
277 tiler_prefix
->indices
= 0;
278 *vertex_count
= ctx
->vertex_count
;
279 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= info
->start
;
280 tiler_prefix
->offset_bias_correction
= 0;
281 tiler_prefix
->index_count
= MALI_POSITIVE(ctx
->vertex_count
);
284 tiler_prefix
->unknown_draw
= draw_flags
;
286 /* Encode the padded vertex count */
288 if (info
->instance_count
> 1) {
289 *padded_count
= panfrost_padded_vertex_count(*vertex_count
);
291 unsigned shift
= __builtin_ctz(ctx
->padded_count
);
292 unsigned k
= ctx
->padded_count
>> (shift
+ 1);
294 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= shift
;
295 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= k
;
297 *padded_count
= *vertex_count
;
299 /* Reset instancing state */
300 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= 0;
301 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= 0;
306 panfrost_shader_meta_init(struct panfrost_context
*ctx
,
307 enum pipe_shader_type st
,
308 struct mali_shader_meta
*meta
)
310 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
311 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
313 memset(meta
, 0, sizeof(*meta
));
314 meta
->shader
= (ss
->bo
? ss
->bo
->gpu
: 0) | ss
->first_tag
;
315 meta
->attribute_count
= ss
->attribute_count
;
316 meta
->varying_count
= ss
->varying_count
;
317 meta
->texture_count
= ctx
->sampler_view_count
[st
];
318 meta
->sampler_count
= ctx
->sampler_count
[st
];
320 if (dev
->quirks
& IS_BIFROST
) {
321 if (st
== PIPE_SHADER_VERTEX
)
322 meta
->bifrost1
.unk1
= 0x800000;
324 /* First clause ATEST |= 0x4000000.
325 * Less than 32 regs |= 0x200 */
326 meta
->bifrost1
.unk1
= 0x950020;
329 meta
->bifrost1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
330 if (st
== PIPE_SHADER_VERTEX
)
331 meta
->bifrost2
.preload_regs
= 0xC0;
333 meta
->bifrost2
.preload_regs
= 0x1;
334 SET_BIT(meta
->bifrost2
.preload_regs
, 0x10, ss
->reads_frag_coord
);
337 meta
->bifrost2
.uniform_count
= MIN2(ss
->uniform_count
,
340 meta
->midgard1
.uniform_count
= MIN2(ss
->uniform_count
,
342 meta
->midgard1
.work_count
= ss
->work_reg_count
;
344 /* TODO: This is not conformant on ES3 */
345 meta
->midgard1
.flags_hi
= MALI_SUPPRESS_INF_NAN
;
347 meta
->midgard1
.flags_lo
= 0x20;
348 meta
->midgard1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
350 SET_BIT(meta
->midgard1
.flags_lo
, MALI_WRITES_GLOBAL
, ss
->writes_global
);
355 translate_tex_wrap(enum pipe_tex_wrap w
)
358 case PIPE_TEX_WRAP_REPEAT
: return MALI_WRAP_MODE_REPEAT
;
359 case PIPE_TEX_WRAP_CLAMP
: return MALI_WRAP_MODE_CLAMP
;
360 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
: return MALI_WRAP_MODE_CLAMP_TO_EDGE
;
361 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
: return MALI_WRAP_MODE_CLAMP_TO_BORDER
;
362 case PIPE_TEX_WRAP_MIRROR_REPEAT
: return MALI_WRAP_MODE_MIRRORED_REPEAT
;
363 case PIPE_TEX_WRAP_MIRROR_CLAMP
: return MALI_WRAP_MODE_MIRRORED_CLAMP
;
364 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
: return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_EDGE
;
365 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
: return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_BORDER
;
366 default: unreachable("Invalid wrap");
370 /* The hardware compares in the wrong order order, so we have to flip before
371 * encoding. Yes, really. */
373 static enum mali_func
374 panfrost_sampler_compare_func(const struct pipe_sampler_state
*cso
)
376 if (!cso
->compare_mode
)
377 return MALI_FUNC_NEVER
;
379 enum mali_func f
= panfrost_translate_compare_func(cso
->compare_func
);
380 return panfrost_flip_compare_func(f
);
383 static enum mali_mipmap_mode
384 pan_pipe_to_mipmode(enum pipe_tex_mipfilter f
)
387 case PIPE_TEX_MIPFILTER_NEAREST
: return MALI_MIPMAP_MODE_NEAREST
;
388 case PIPE_TEX_MIPFILTER_LINEAR
: return MALI_MIPMAP_MODE_TRILINEAR
;
389 case PIPE_TEX_MIPFILTER_NONE
: return MALI_MIPMAP_MODE_NONE
;
390 default: unreachable("Invalid");
394 void panfrost_sampler_desc_init(const struct pipe_sampler_state
*cso
,
395 struct mali_midgard_sampler_packed
*hw
)
397 pan_pack(hw
, MIDGARD_SAMPLER
, cfg
) {
398 cfg
.magnify_nearest
= cso
->mag_img_filter
== PIPE_TEX_FILTER_NEAREST
;
399 cfg
.minify_nearest
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
;
400 cfg
.mipmap_mode
= (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) ?
401 MALI_MIPMAP_MODE_TRILINEAR
: MALI_MIPMAP_MODE_NEAREST
;
402 cfg
.normalized_coordinates
= cso
->normalized_coords
;
404 cfg
.lod_bias
= FIXED_16(cso
->lod_bias
, true);
406 cfg
.minimum_lod
= FIXED_16(cso
->min_lod
, false);
408 /* If necessary, we disable mipmapping in the sampler descriptor by
409 * clamping the LOD as tight as possible (from 0 to epsilon,
410 * essentially -- remember these are fixed point numbers, so
413 cfg
.maximum_lod
= (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
) ?
414 cfg
.minimum_lod
+ 1 :
415 FIXED_16(cso
->max_lod
, false);
417 cfg
.wrap_mode_s
= translate_tex_wrap(cso
->wrap_s
);
418 cfg
.wrap_mode_t
= translate_tex_wrap(cso
->wrap_t
);
419 cfg
.wrap_mode_r
= translate_tex_wrap(cso
->wrap_r
);
421 cfg
.compare_function
= panfrost_sampler_compare_func(cso
);
422 cfg
.seamless_cube_map
= cso
->seamless_cube_map
;
424 cfg
.border_color_r
= cso
->border_color
.f
[0];
425 cfg
.border_color_g
= cso
->border_color
.f
[1];
426 cfg
.border_color_b
= cso
->border_color
.f
[2];
427 cfg
.border_color_a
= cso
->border_color
.f
[3];
431 void panfrost_sampler_desc_init_bifrost(const struct pipe_sampler_state
*cso
,
432 struct mali_bifrost_sampler_packed
*hw
)
434 pan_pack(hw
, BIFROST_SAMPLER
, cfg
) {
435 cfg
.magnify_linear
= cso
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
;
436 cfg
.minify_linear
= cso
->min_img_filter
== PIPE_TEX_FILTER_LINEAR
;
437 cfg
.mipmap_mode
= pan_pipe_to_mipmode(cso
->min_mip_filter
);
438 cfg
.normalized_coordinates
= cso
->normalized_coords
;
440 cfg
.lod_bias
= FIXED_16(cso
->lod_bias
, true);
441 cfg
.minimum_lod
= FIXED_16(cso
->min_lod
, false);
442 cfg
.maximum_lod
= FIXED_16(cso
->max_lod
, false);
444 cfg
.wrap_mode_s
= translate_tex_wrap(cso
->wrap_s
);
445 cfg
.wrap_mode_t
= translate_tex_wrap(cso
->wrap_t
);
446 cfg
.wrap_mode_r
= translate_tex_wrap(cso
->wrap_r
);
448 cfg
.compare_function
= panfrost_sampler_compare_func(cso
);
449 cfg
.seamless_cube_map
= cso
->seamless_cube_map
;
454 panfrost_frag_meta_rasterizer_update(struct panfrost_context
*ctx
,
455 struct mali_shader_meta
*fragmeta
)
457 struct pipe_rasterizer_state
*rast
= &ctx
->rasterizer
->base
;
459 bool msaa
= rast
->multisample
;
461 /* TODO: Sample size */
462 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, msaa
);
463 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, !msaa
);
465 struct panfrost_shader_state
*fs
;
466 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
468 /* EXT_shader_framebuffer_fetch requires the shader to be run
469 * per-sample when outputs are read. */
470 bool per_sample
= ctx
->min_samples
> 1 || fs
->outputs_read
;
471 SET_BIT(fragmeta
->unknown2_3
, MALI_PER_SAMPLE
, msaa
&& per_sample
);
473 fragmeta
->depth_units
= rast
->offset_units
* 2.0f
;
474 fragmeta
->depth_factor
= rast
->offset_scale
;
476 /* XXX: Which bit is which? Does this maybe allow offseting not-tri? */
478 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
, rast
->offset_tri
);
479 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
, rast
->offset_tri
);
481 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_NEAR
, rast
->depth_clip_near
);
482 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_FAR
, rast
->depth_clip_far
);
486 panfrost_frag_meta_zsa_update(struct panfrost_context
*ctx
,
487 struct mali_shader_meta
*fragmeta
)
489 const struct panfrost_zsa_state
*so
= ctx
->depth_stencil
;
491 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
,
492 so
->base
.stencil
[0].enabled
);
494 fragmeta
->stencil_mask_front
= so
->stencil_mask_front
;
495 fragmeta
->stencil_mask_back
= so
->stencil_mask_back
;
497 /* Bottom bits for stencil ref, exactly one word */
498 fragmeta
->stencil_front
.opaque
[0] = so
->stencil_front
.opaque
[0] | ctx
->stencil_ref
.ref_value
[0];
500 /* If back-stencil is not enabled, use the front values */
502 if (so
->base
.stencil
[1].enabled
)
503 fragmeta
->stencil_back
.opaque
[0] = so
->stencil_back
.opaque
[0] | ctx
->stencil_ref
.ref_value
[1];
505 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
507 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
,
508 so
->base
.depth
.writemask
);
510 fragmeta
->unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
511 fragmeta
->unknown2_3
|= MALI_DEPTH_FUNC(panfrost_translate_compare_func(
512 so
->base
.depth
.enabled
? so
->base
.depth
.func
: PIPE_FUNC_ALWAYS
));
516 panfrost_fs_required(
517 struct panfrost_shader_state
*fs
,
518 struct panfrost_blend_final
*blend
,
521 /* If we generally have side effects */
525 /* If colour is written we need to execute */
526 for (unsigned i
= 0; i
< rt_count
; ++i
) {
527 if (!blend
[i
].no_colour
)
531 /* If depth is written and not implied we need to execute.
532 * TODO: Predicate on Z/S writes being enabled */
533 return (fs
->writes_depth
|| fs
->writes_stencil
);
537 panfrost_frag_meta_blend_update(struct panfrost_context
*ctx
,
538 struct mali_shader_meta
*fragmeta
,
541 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
542 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
543 struct panfrost_shader_state
*fs
;
544 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
546 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_DITHER
,
547 (dev
->quirks
& MIDGARD_SFBD
) && ctx
->blend
&&
548 !ctx
->blend
->base
.dither
);
550 SET_BIT(fragmeta
->unknown2_4
, MALI_ALPHA_TO_COVERAGE
,
551 ctx
->blend
->base
.alpha_to_coverage
);
553 /* Get blending setup */
554 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
556 struct panfrost_blend_final blend
[PIPE_MAX_COLOR_BUFS
];
557 unsigned shader_offset
= 0;
558 struct panfrost_bo
*shader_bo
= NULL
;
560 for (unsigned c
= 0; c
< rt_count
; ++c
)
561 blend
[c
] = panfrost_get_blend_for_context(ctx
, c
, &shader_bo
,
564 /* Disable shader execution if we can */
565 if (dev
->quirks
& MIDGARD_SHADERLESS
566 && !panfrost_fs_required(fs
, blend
, rt_count
)) {
567 fragmeta
->shader
= 0;
568 fragmeta
->attribute_count
= 0;
569 fragmeta
->varying_count
= 0;
570 fragmeta
->texture_count
= 0;
571 fragmeta
->sampler_count
= 0;
573 /* This feature is not known to work on Bifrost */
574 fragmeta
->midgard1
.work_count
= 1;
575 fragmeta
->midgard1
.uniform_count
= 0;
576 fragmeta
->midgard1
.uniform_buffer_count
= 0;
579 /* If there is a blend shader, work registers are shared. We impose 8
580 * work registers as a limit for blend shaders. Should be lower XXX */
582 if (!(dev
->quirks
& IS_BIFROST
)) {
583 for (unsigned c
= 0; c
< rt_count
; ++c
) {
584 if (blend
[c
].is_shader
) {
585 fragmeta
->midgard1
.work_count
=
586 MAX2(fragmeta
->midgard1
.work_count
, 8);
591 /* Even on MFBD, the shader descriptor gets blend shaders. It's *also*
592 * copied to the blend_meta appended (by convention), but this is the
593 * field actually read by the hardware. (Or maybe both are read...?).
594 * Specify the last RTi with a blend shader. */
596 fragmeta
->blend
.shader
= 0;
598 for (signed rt
= (rt_count
- 1); rt
>= 0; --rt
) {
599 if (!blend
[rt
].is_shader
)
602 fragmeta
->blend
.shader
= blend
[rt
].shader
.gpu
|
603 blend
[rt
].shader
.first_tag
;
607 if (dev
->quirks
& MIDGARD_SFBD
) {
608 /* When only a single render target platform is used, the blend
609 * information is inside the shader meta itself. We additionally
610 * need to signal CAN_DISCARD for nontrivial blend modes (so
611 * we're able to read back the destination buffer) */
613 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_BLEND_SHADER
,
616 if (!blend
[0].is_shader
) {
617 fragmeta
->blend
.equation
= *blend
[0].equation
.equation
;
618 fragmeta
->blend
.constant
= blend
[0].equation
.constant
;
621 SET_BIT(fragmeta
->unknown2_3
, MALI_CAN_DISCARD
,
622 !blend
[0].no_blending
|| fs
->can_discard
);
624 batch
->draws
|= PIPE_CLEAR_COLOR0
;
628 if (dev
->quirks
& IS_BIFROST
) {
629 bool no_blend
= true;
631 for (unsigned i
= 0; i
< rt_count
; ++i
)
632 no_blend
&= (blend
[i
].no_blending
| blend
[i
].no_colour
);
634 SET_BIT(fragmeta
->bifrost1
.unk1
, MALI_BIFROST_EARLY_Z
,
635 !fs
->can_discard
&& !fs
->writes_depth
&& no_blend
);
638 /* Additional blend descriptor tacked on for jobs using MFBD */
640 for (unsigned i
= 0; i
< rt_count
; ++i
) {
643 if (ctx
->pipe_framebuffer
.nr_cbufs
> i
&& !blend
[i
].no_colour
) {
645 batch
->draws
|= (PIPE_CLEAR_COLOR0
<< i
);
647 bool is_srgb
= (ctx
->pipe_framebuffer
.nr_cbufs
> i
) &&
648 (ctx
->pipe_framebuffer
.cbufs
[i
]) &&
649 util_format_is_srgb(ctx
->pipe_framebuffer
.cbufs
[i
]->format
);
651 SET_BIT(flags
, MALI_BLEND_MRT_SHADER
, blend
[i
].is_shader
);
652 SET_BIT(flags
, MALI_BLEND_LOAD_TIB
, !blend
[i
].no_blending
);
653 SET_BIT(flags
, MALI_BLEND_SRGB
, is_srgb
);
654 SET_BIT(flags
, MALI_BLEND_NO_DITHER
, !ctx
->blend
->base
.dither
);
657 if (dev
->quirks
& IS_BIFROST
) {
658 struct bifrost_blend_rt
*brts
= rts
;
660 brts
[i
].flags
= flags
;
662 if (blend
[i
].is_shader
) {
663 /* The blend shader's address needs to be at
664 * the same top 32 bit as the fragment shader.
665 * TODO: Ensure that's always the case.
667 assert((blend
[i
].shader
.gpu
& (0xffffffffull
<< 32)) ==
668 (fs
->bo
->gpu
& (0xffffffffull
<< 32)));
669 brts
[i
].shader
= blend
[i
].shader
.gpu
;
671 } else if (ctx
->pipe_framebuffer
.nr_cbufs
> i
) {
672 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
673 const struct util_format_description
*format_desc
;
674 format_desc
= util_format_description(format
);
676 brts
[i
].equation
= *blend
[i
].equation
.equation
;
678 /* TODO: this is a bit more complicated */
679 brts
[i
].constant
= blend
[i
].equation
.constant
;
681 brts
[i
].format
= panfrost_format_to_bifrost_blend(format_desc
);
683 /* 0x19 disables blending and forces REPLACE
684 * mode (equivalent to rgb_mode = alpha_mode =
685 * x122, colour mask = 0xF). 0x1a allows
687 brts
[i
].unk2
= blend
[i
].no_blending
? 0x19 : 0x1a;
689 brts
[i
].shader_type
= fs
->blend_types
[i
];
691 /* Dummy attachment for depth-only */
693 brts
[i
].shader_type
= fs
->blend_types
[i
];
696 struct midgard_blend_rt
*mrts
= rts
;
697 mrts
[i
].flags
= flags
;
699 if (blend
[i
].is_shader
) {
700 mrts
[i
].blend
.shader
= blend
[i
].shader
.gpu
| blend
[i
].shader
.first_tag
;
702 mrts
[i
].blend
.equation
= *blend
[i
].equation
.equation
;
703 mrts
[i
].blend
.constant
= blend
[i
].equation
.constant
;
710 panfrost_frag_shader_meta_init(struct panfrost_context
*ctx
,
711 struct mali_shader_meta
*fragmeta
,
714 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
715 struct panfrost_shader_state
*fs
;
717 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
719 bool msaa
= ctx
->rasterizer
->base
.multisample
;
720 fragmeta
->coverage_mask
= msaa
? ctx
->sample_mask
: ~0;
722 fragmeta
->unknown2_3
= MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS
) | 0x10;
723 fragmeta
->unknown2_4
= 0x4e0;
725 /* unknown2_4 has 0x10 bit set on T6XX and T720. We don't know why this
726 * is required (independent of 32-bit/64-bit descriptors), or why it's
727 * not used on later GPU revisions. Otherwise, all shader jobs fault on
728 * these earlier chips (perhaps this is a chicken bit of some kind).
729 * More investigation is needed. */
731 SET_BIT(fragmeta
->unknown2_4
, 0x10, dev
->quirks
& MIDGARD_SFBD
);
733 if (dev
->quirks
& IS_BIFROST
) {
736 /* Depending on whether it's legal to in the given shader, we try to
737 * enable early-z testing. TODO: respect e-z force */
739 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_EARLY_Z
,
740 !fs
->can_discard
&& !fs
->writes_global
&&
741 !fs
->writes_depth
&& !fs
->writes_stencil
&&
742 !ctx
->blend
->base
.alpha_to_coverage
);
744 /* Add the writes Z/S flags if needed. */
745 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_WRITES_Z
, fs
->writes_depth
);
746 SET_BIT(fragmeta
->midgard1
.flags_hi
, MALI_WRITES_S
, fs
->writes_stencil
);
748 /* Any time texturing is used, derivatives are implicitly calculated,
749 * so we need to enable helper invocations */
751 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_HELPER_INVOCATIONS
,
752 fs
->helper_invocations
);
754 /* If discard is enabled, which bit we set to convey this
755 * depends on if depth/stencil is used for the draw or not.
756 * Just one of depth OR stencil is enough to trigger this. */
758 const struct pipe_depth_stencil_alpha_state
*zsa
= &ctx
->depth_stencil
->base
;
760 fs
->writes_depth
|| fs
->writes_stencil
||
761 (zsa
->depth
.enabled
&& zsa
->depth
.func
!= PIPE_FUNC_ALWAYS
) ||
762 zsa
->stencil
[0].enabled
;
764 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_READS_TILEBUFFER
,
765 fs
->outputs_read
|| (!zs_enabled
&& fs
->can_discard
));
766 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_READS_ZS
, zs_enabled
&& fs
->can_discard
);
769 panfrost_frag_meta_rasterizer_update(ctx
, fragmeta
);
770 panfrost_frag_meta_zsa_update(ctx
, fragmeta
);
771 panfrost_frag_meta_blend_update(ctx
, fragmeta
, rts
);
775 panfrost_emit_shader_meta(struct panfrost_batch
*batch
,
776 enum pipe_shader_type st
,
777 struct mali_vertex_tiler_postfix
*postfix
)
779 struct panfrost_context
*ctx
= batch
->ctx
;
780 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
787 struct mali_shader_meta meta
;
789 panfrost_shader_meta_init(ctx
, st
, &meta
);
791 /* Add the shader BO to the batch. */
792 panfrost_batch_add_bo(batch
, ss
->bo
,
793 PAN_BO_ACCESS_PRIVATE
|
795 panfrost_bo_access_for_stage(st
));
799 if (st
== PIPE_SHADER_FRAGMENT
) {
800 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
801 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
802 size_t desc_size
= sizeof(meta
);
804 struct panfrost_transfer xfer
;
807 if (dev
->quirks
& MIDGARD_SFBD
)
809 else if (dev
->quirks
& IS_BIFROST
)
810 rt_size
= sizeof(struct bifrost_blend_rt
);
812 rt_size
= sizeof(struct midgard_blend_rt
);
814 desc_size
+= rt_size
* rt_count
;
817 rts
= rzalloc_size(ctx
, rt_size
* rt_count
);
819 panfrost_frag_shader_meta_init(ctx
, &meta
, rts
);
821 xfer
= panfrost_pool_alloc(&batch
->pool
, desc_size
);
823 memcpy(xfer
.cpu
, &meta
, sizeof(meta
));
824 memcpy(xfer
.cpu
+ sizeof(meta
), rts
, rt_size
* rt_count
);
829 shader_ptr
= xfer
.gpu
;
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(&batch
->pool
,
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
= panfrost_pool_alloc(&batch
->pool
,
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
= panfrost_pool_alloc(&batch
->pool
, sz
);
1133 uint64_t *ubo_ptr
= (uint64_t *) ubos
.cpu
;
1135 /* Upload uniforms as a UBO */
1137 if (ss
->uniform_count
) {
1138 pan_pack(ubo_ptr
, UNIFORM_BUFFER
, cfg
) {
1139 cfg
.entries
= ss
->uniform_count
;
1140 cfg
.pointer
= transfer
.gpu
;
1146 /* The rest are honest-to-goodness UBOs */
1148 for (unsigned ubo
= 1; ubo
< ubo_count
; ++ubo
) {
1149 size_t usz
= buf
->cb
[ubo
].buffer_size
;
1150 bool enabled
= buf
->enabled_mask
& (1 << ubo
);
1151 bool empty
= usz
== 0;
1153 if (!enabled
|| empty
) {
1158 pan_pack(ubo_ptr
+ ubo
, UNIFORM_BUFFER
, cfg
) {
1159 cfg
.entries
= DIV_ROUND_UP(usz
, 16);
1160 cfg
.pointer
= panfrost_map_constant_buffer_gpu(batch
,
1165 postfix
->uniforms
= transfer
.gpu
;
1166 postfix
->uniform_buffers
= ubos
.gpu
;
1168 buf
->dirty_mask
= 0;
1172 panfrost_emit_shared_memory(struct panfrost_batch
*batch
,
1173 const struct pipe_grid_info
*info
,
1174 struct midgard_payload_vertex_tiler
*vtp
)
1176 struct panfrost_context
*ctx
= batch
->ctx
;
1177 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
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
,
1183 unsigned log2_instances
=
1184 util_logbase2_ceil(info
->grid
[0]) +
1185 util_logbase2_ceil(info
->grid
[1]) +
1186 util_logbase2_ceil(info
->grid
[2]);
1188 unsigned shared_size
= single_size
* (1 << log2_instances
) * dev
->core_count
;
1189 struct panfrost_bo
*bo
= panfrost_batch_get_shared_memory(batch
,
1193 struct mali_shared_memory shared
= {
1194 .shared_memory
= bo
->gpu
,
1195 .shared_workgroup_count
= log2_instances
,
1196 .shared_shift
= util_logbase2(single_size
) + 1
1199 vtp
->postfix
.shared_memory
= panfrost_pool_upload(&batch
->pool
, &shared
,
1204 panfrost_get_tex_desc(struct panfrost_batch
*batch
,
1205 enum pipe_shader_type st
,
1206 struct panfrost_sampler_view
*view
)
1209 return (mali_ptr
) 0;
1211 struct pipe_sampler_view
*pview
= &view
->base
;
1212 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1214 /* Add the BO to the job so it's retained until the job is done. */
1216 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1217 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1218 panfrost_bo_access_for_stage(st
));
1220 panfrost_batch_add_bo(batch
, view
->bo
,
1221 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1222 panfrost_bo_access_for_stage(st
));
1224 return view
->bo
->gpu
;
1228 panfrost_update_sampler_view(struct panfrost_sampler_view
*view
,
1229 struct pipe_context
*pctx
)
1231 struct panfrost_resource
*rsrc
= pan_resource(view
->base
.texture
);
1232 if (view
->texture_bo
!= rsrc
->bo
->gpu
||
1233 view
->modifier
!= rsrc
->modifier
) {
1234 panfrost_bo_unreference(view
->bo
);
1235 panfrost_create_sampler_view_bo(view
, pctx
, &rsrc
->base
);
1240 panfrost_emit_texture_descriptors(struct panfrost_batch
*batch
,
1241 enum pipe_shader_type stage
,
1242 struct mali_vertex_tiler_postfix
*postfix
)
1244 struct panfrost_context
*ctx
= batch
->ctx
;
1245 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1247 if (!ctx
->sampler_view_count
[stage
])
1250 if (device
->quirks
& IS_BIFROST
) {
1251 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
,
1252 MALI_BIFROST_TEXTURE_LENGTH
*
1253 ctx
->sampler_view_count
[stage
]);
1255 struct mali_bifrost_texture_packed
*out
=
1256 (struct mali_bifrost_texture_packed
*) T
.cpu
;
1258 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1259 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1260 struct pipe_sampler_view
*pview
= &view
->base
;
1261 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1263 panfrost_update_sampler_view(view
, &ctx
->base
);
1264 out
[i
] = view
->bifrost_descriptor
;
1266 /* Add the BOs to the job so they are retained until the job is done. */
1268 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1269 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1270 panfrost_bo_access_for_stage(stage
));
1272 panfrost_batch_add_bo(batch
, view
->bo
,
1273 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1274 panfrost_bo_access_for_stage(stage
));
1277 postfix
->textures
= T
.gpu
;
1279 uint64_t trampolines
[PIPE_MAX_SHADER_SAMPLER_VIEWS
];
1281 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1282 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1284 panfrost_update_sampler_view(view
, &ctx
->base
);
1286 trampolines
[i
] = panfrost_get_tex_desc(batch
, stage
, view
);
1289 postfix
->textures
= panfrost_pool_upload(&batch
->pool
,
1292 ctx
->sampler_view_count
[stage
]);
1297 panfrost_emit_sampler_descriptors(struct panfrost_batch
*batch
,
1298 enum pipe_shader_type stage
,
1299 struct mali_vertex_tiler_postfix
*postfix
)
1301 struct panfrost_context
*ctx
= batch
->ctx
;
1303 if (!ctx
->sampler_count
[stage
])
1306 size_t desc_size
= MALI_BIFROST_SAMPLER_LENGTH
;
1307 assert(MALI_BIFROST_SAMPLER_LENGTH
== MALI_MIDGARD_SAMPLER_LENGTH
);
1309 size_t sz
= desc_size
* ctx
->sampler_count
[stage
];
1310 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
, sz
);
1311 struct mali_midgard_sampler_packed
*out
= (struct mali_midgard_sampler_packed
*) T
.cpu
;
1313 for (unsigned i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1314 out
[i
] = ctx
->samplers
[stage
][i
]->hw
;
1316 postfix
->sampler_descriptor
= T
.gpu
;
1320 panfrost_emit_vertex_data(struct panfrost_batch
*batch
,
1321 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1323 struct panfrost_context
*ctx
= batch
->ctx
;
1324 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1326 unsigned instance_shift
= vertex_postfix
->instance_shift
;
1327 unsigned instance_odd
= vertex_postfix
->instance_odd
;
1329 /* Worst case: everything is NPOT */
1331 struct panfrost_transfer S
= panfrost_pool_alloc(&batch
->pool
,
1332 MALI_ATTRIBUTE_LENGTH
* PIPE_MAX_ATTRIBS
* 2);
1334 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
,
1335 MALI_ATTRIBUTE_LENGTH
* (PAN_INSTANCE_ID
+ 1));
1337 struct mali_attribute_buffer_packed
*bufs
=
1338 (struct mali_attribute_buffer_packed
*) S
.cpu
;
1340 struct mali_attribute_packed
*out
=
1341 (struct mali_attribute_packed
*) T
.cpu
;
1343 unsigned attrib_to_buffer
[PIPE_MAX_ATTRIBS
] = { 0 };
1346 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1347 /* We map buffers 1:1 with the attributes, which
1348 * means duplicating some vertex buffers (who cares? aside from
1349 * maybe some caching implications but I somehow doubt that
1352 struct pipe_vertex_element
*elem
= &so
->pipe
[i
];
1353 unsigned vbi
= elem
->vertex_buffer_index
;
1354 attrib_to_buffer
[i
] = k
;
1356 if (!(ctx
->vb_mask
& (1 << vbi
)))
1359 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1360 struct panfrost_resource
*rsrc
;
1362 rsrc
= pan_resource(buf
->buffer
.resource
);
1366 /* Add a dependency of the batch on the vertex buffer */
1367 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1368 PAN_BO_ACCESS_SHARED
|
1369 PAN_BO_ACCESS_READ
|
1370 PAN_BO_ACCESS_VERTEX_TILER
);
1372 /* Mask off lower bits, see offset fixup below */
1373 mali_ptr raw_addr
= rsrc
->bo
->gpu
+ buf
->buffer_offset
;
1374 mali_ptr addr
= raw_addr
& ~63;
1376 /* Since we advanced the base pointer, we shrink the buffer
1377 * size, but add the offset we subtracted */
1378 unsigned size
= rsrc
->base
.width0
+ (raw_addr
- addr
)
1379 - buf
->buffer_offset
;
1381 /* When there is a divisor, the hardware-level divisor is
1382 * the product of the instance divisor and the padded count */
1383 unsigned divisor
= elem
->instance_divisor
;
1384 unsigned hw_divisor
= ctx
->padded_count
* divisor
;
1385 unsigned stride
= buf
->stride
;
1387 /* If there's a divisor(=1) but no instancing, we want every
1388 * attribute to be the same */
1390 if (divisor
&& ctx
->instance_count
== 1)
1393 if (!divisor
|| ctx
->instance_count
<= 1) {
1394 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1395 if (ctx
->instance_count
> 1)
1396 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_MODULUS
;
1399 cfg
.stride
= stride
;
1401 cfg
.divisor_r
= instance_shift
;
1402 cfg
.divisor_p
= instance_odd
;
1404 } else if (util_is_power_of_two_or_zero(hw_divisor
)) {
1405 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1406 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_POT_DIVISOR
;
1408 cfg
.stride
= stride
;
1410 cfg
.divisor_r
= __builtin_ctz(hw_divisor
);
1414 unsigned shift
= 0, extra_flags
= 0;
1416 unsigned magic_divisor
=
1417 panfrost_compute_magic_divisor(hw_divisor
, &shift
, &extra_flags
);
1419 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1420 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR
;
1422 cfg
.stride
= stride
;
1425 cfg
.divisor_r
= shift
;
1426 cfg
.divisor_e
= extra_flags
;
1429 pan_pack(bufs
+ k
+ 1, ATTRIBUTE_BUFFER_CONTINUATION_NPOT
, cfg
) {
1430 cfg
.divisor_numerator
= magic_divisor
;
1431 cfg
.divisor
= divisor
;
1440 /* Add special gl_VertexID/gl_InstanceID buffers */
1442 panfrost_vertex_id(ctx
->padded_count
, &bufs
[k
], ctx
->instance_count
> 1);
1444 pan_pack(out
+ PAN_VERTEX_ID
, ATTRIBUTE
, cfg
) {
1445 cfg
.buffer_index
= k
++;
1446 cfg
.format
= so
->formats
[PAN_VERTEX_ID
];
1449 panfrost_instance_id(ctx
->padded_count
, &bufs
[k
], ctx
->instance_count
> 1);
1451 pan_pack(out
+ PAN_INSTANCE_ID
, ATTRIBUTE
, cfg
) {
1452 cfg
.buffer_index
= k
++;
1453 cfg
.format
= so
->formats
[PAN_INSTANCE_ID
];
1456 /* Attribute addresses require 64-byte alignment, so let:
1458 * base' = base & ~63 = base - (base & 63)
1459 * offset' = offset + (base & 63)
1461 * Since base' + offset' = base + offset, these are equivalent
1462 * addressing modes and now base is 64 aligned.
1465 unsigned start
= vertex_postfix
->offset_start
;
1467 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1468 unsigned vbi
= so
->pipe
[i
].vertex_buffer_index
;
1469 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1471 /* Adjust by the masked off bits of the offset. Make sure we
1472 * read src_offset from so->hw (which is not GPU visible)
1473 * rather than target (which is) due to caching effects */
1475 unsigned src_offset
= so
->pipe
[i
].src_offset
;
1477 /* BOs aligned to 4k so guaranteed aligned to 64 */
1478 src_offset
+= (buf
->buffer_offset
& 63);
1480 /* Also, somewhat obscurely per-instance data needs to be
1481 * offset in response to a delayed start in an indexed draw */
1483 if (so
->pipe
[i
].instance_divisor
&& ctx
->instance_count
> 1 && start
)
1484 src_offset
-= buf
->stride
* start
;
1486 pan_pack(out
+ i
, ATTRIBUTE
, cfg
) {
1487 cfg
.buffer_index
= attrib_to_buffer
[i
];
1488 cfg
.format
= so
->formats
[i
];
1489 cfg
.offset
= src_offset
;
1493 vertex_postfix
->attributes
= S
.gpu
;
1494 vertex_postfix
->attribute_meta
= T
.gpu
;
1498 panfrost_emit_varyings(struct panfrost_batch
*batch
,
1499 struct mali_attribute_buffer_packed
*slot
,
1500 unsigned stride
, unsigned count
)
1502 unsigned size
= stride
* count
;
1503 mali_ptr ptr
= panfrost_pool_alloc(&batch
->invisible_pool
, size
).gpu
;
1505 pan_pack(slot
, ATTRIBUTE_BUFFER
, cfg
) {
1506 cfg
.stride
= stride
;
1515 panfrost_streamout_offset(unsigned stride
, unsigned offset
,
1516 struct pipe_stream_output_target
*target
)
1518 return (target
->buffer_offset
+ (offset
* stride
* 4)) & 63;
1522 panfrost_emit_streamout(struct panfrost_batch
*batch
,
1523 struct mali_attribute_buffer_packed
*slot
,
1524 unsigned stride_words
, unsigned offset
, unsigned count
,
1525 struct pipe_stream_output_target
*target
)
1527 unsigned stride
= stride_words
* 4;
1528 unsigned max_size
= target
->buffer_size
;
1529 unsigned expected_size
= stride
* count
;
1531 /* Grab the BO and bind it to the batch */
1532 struct panfrost_bo
*bo
= pan_resource(target
->buffer
)->bo
;
1534 /* Varyings are WRITE from the perspective of the VERTEX but READ from
1535 * the perspective of the TILER and FRAGMENT.
1537 panfrost_batch_add_bo(batch
, bo
,
1538 PAN_BO_ACCESS_SHARED
|
1540 PAN_BO_ACCESS_VERTEX_TILER
|
1541 PAN_BO_ACCESS_FRAGMENT
);
1543 /* We will have an offset applied to get alignment */
1544 mali_ptr addr
= bo
->gpu
+ target
->buffer_offset
+ (offset
* stride
);
1546 pan_pack(slot
, ATTRIBUTE_BUFFER
, cfg
) {
1547 cfg
.pointer
= (addr
& ~63);
1548 cfg
.stride
= stride
;
1549 cfg
.size
= MIN2(max_size
, expected_size
) + (addr
& 63);
1554 has_point_coord(unsigned mask
, gl_varying_slot loc
)
1556 if ((loc
>= VARYING_SLOT_TEX0
) && (loc
<= VARYING_SLOT_TEX7
))
1557 return (mask
& (1 << (loc
- VARYING_SLOT_TEX0
)));
1558 else if (loc
== VARYING_SLOT_PNTC
)
1559 return (mask
& (1 << 8));
1564 /* Helpers for manipulating stream out information so we can pack varyings
1565 * accordingly. Compute the src_offset for a given captured varying */
1567 static struct pipe_stream_output
*
1568 pan_get_so(struct pipe_stream_output_info
*info
, gl_varying_slot loc
)
1570 for (unsigned i
= 0; i
< info
->num_outputs
; ++i
) {
1571 if (info
->output
[i
].register_index
== loc
)
1572 return &info
->output
[i
];
1575 unreachable("Varying not captured");
1579 pan_varying_size(enum mali_format fmt
)
1581 unsigned type
= MALI_EXTRACT_TYPE(fmt
);
1582 unsigned chan
= MALI_EXTRACT_CHANNELS(fmt
);
1583 unsigned bits
= MALI_EXTRACT_BITS(fmt
);
1586 if (bits
== MALI_CHANNEL_FLOAT
) {
1588 bool fp16
= (type
== MALI_FORMAT_SINT
);
1589 assert(fp16
|| (type
== MALI_FORMAT_UNORM
));
1593 assert(type
>= MALI_FORMAT_SNORM
&& type
<= MALI_FORMAT_SINT
);
1604 /* Indices for named (non-XFB) varyings that are present. These are packed
1605 * tightly so they correspond to a bitfield present (P) indexed by (1 <<
1606 * PAN_VARY_*). This has the nice property that you can lookup the buffer index
1607 * of a given special field given a shift S by:
1609 * idx = popcount(P & ((1 << S) - 1))
1611 * That is... look at all of the varyings that come earlier and count them, the
1612 * count is the new index since plus one. Likewise, the total number of special
1613 * buffers required is simply popcount(P)
1616 enum pan_special_varying
{
1617 PAN_VARY_GENERAL
= 0,
1618 PAN_VARY_POSITION
= 1,
1620 PAN_VARY_PNTCOORD
= 3,
1622 PAN_VARY_FRAGCOORD
= 5,
1628 /* Given a varying, figure out which index it correpsonds to */
1630 static inline unsigned
1631 pan_varying_index(unsigned present
, enum pan_special_varying v
)
1633 unsigned mask
= (1 << v
) - 1;
1634 return util_bitcount(present
& mask
);
1637 /* Get the base offset for XFB buffers, which by convention come after
1638 * everything else. Wrapper function for semantic reasons; by construction this
1639 * is just popcount. */
1641 static inline unsigned
1642 pan_xfb_base(unsigned present
)
1644 return util_bitcount(present
);
1647 /* Computes the present mask for varyings so we can start emitting varying records */
1649 static inline unsigned
1650 pan_varying_present(
1651 struct panfrost_shader_state
*vs
,
1652 struct panfrost_shader_state
*fs
,
1655 /* At the moment we always emit general and position buffers. Not
1656 * strictly necessary but usually harmless */
1658 unsigned present
= (1 << PAN_VARY_GENERAL
) | (1 << PAN_VARY_POSITION
);
1660 /* Enable special buffers by the shader info */
1662 if (vs
->writes_point_size
)
1663 present
|= (1 << PAN_VARY_PSIZ
);
1665 if (fs
->reads_point_coord
)
1666 present
|= (1 << PAN_VARY_PNTCOORD
);
1669 present
|= (1 << PAN_VARY_FACE
);
1671 if (fs
->reads_frag_coord
&& !(quirks
& IS_BIFROST
))
1672 present
|= (1 << PAN_VARY_FRAGCOORD
);
1674 /* Also, if we have a point sprite, we need a point coord buffer */
1676 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1677 gl_varying_slot loc
= fs
->varyings_loc
[i
];
1679 if (has_point_coord(fs
->point_sprite_mask
, loc
))
1680 present
|= (1 << PAN_VARY_PNTCOORD
);
1686 /* Emitters for varying records */
1689 pan_emit_vary(struct mali_attribute_packed
*out
,
1690 unsigned present
, enum pan_special_varying buf
,
1691 unsigned quirks
, enum mali_format format
,
1694 unsigned nr_channels
= MALI_EXTRACT_CHANNELS(format
);
1695 unsigned swizzle
= quirks
& HAS_SWIZZLES
?
1696 panfrost_get_default_swizzle(nr_channels
) :
1697 panfrost_bifrost_swizzle(nr_channels
);
1699 pan_pack(out
, ATTRIBUTE
, cfg
) {
1700 cfg
.buffer_index
= pan_varying_index(present
, buf
);
1701 cfg
.unknown
= quirks
& IS_BIFROST
? 0x0 : 0x1;
1702 cfg
.format
= (format
<< 12) | swizzle
;
1703 cfg
.offset
= offset
;
1707 /* General varying that is unused */
1710 pan_emit_vary_only(struct mali_attribute_packed
*out
,
1711 unsigned present
, unsigned quirks
)
1713 pan_emit_vary(out
, present
, 0, quirks
, MALI_VARYING_DISCARD
, 0);
1716 /* Special records */
1718 static const enum mali_format pan_varying_formats
[PAN_VARY_MAX
] = {
1719 [PAN_VARY_POSITION
] = MALI_VARYING_POS
,
1720 [PAN_VARY_PSIZ
] = MALI_R16F
,
1721 [PAN_VARY_PNTCOORD
] = MALI_R16F
,
1722 [PAN_VARY_FACE
] = MALI_R32I
,
1723 [PAN_VARY_FRAGCOORD
] = MALI_RGBA32F
1727 pan_emit_vary_special(struct mali_attribute_packed
*out
,
1728 unsigned present
, enum pan_special_varying buf
,
1731 assert(buf
< PAN_VARY_MAX
);
1732 pan_emit_vary(out
, present
, buf
, quirks
, pan_varying_formats
[buf
], 0);
1735 static enum mali_format
1736 pan_xfb_format(enum mali_format format
, unsigned nr
)
1738 if (MALI_EXTRACT_BITS(format
) == MALI_CHANNEL_FLOAT
)
1739 return MALI_R32F
| MALI_NR_CHANNELS(nr
);
1741 return MALI_EXTRACT_TYPE(format
) | MALI_NR_CHANNELS(nr
) | MALI_CHANNEL_32
;
1744 /* Transform feedback records. Note struct pipe_stream_output is (if packed as
1745 * a bitfield) 32-bit, smaller than a 64-bit pointer, so may as well pass by
1749 pan_emit_vary_xfb(struct mali_attribute_packed
*out
,
1752 unsigned *streamout_offsets
,
1754 enum mali_format format
,
1755 struct pipe_stream_output o
)
1757 unsigned swizzle
= quirks
& HAS_SWIZZLES
?
1758 panfrost_get_default_swizzle(o
.num_components
) :
1759 panfrost_bifrost_swizzle(o
.num_components
);
1761 pan_pack(out
, ATTRIBUTE
, cfg
) {
1762 /* XFB buffers come after everything else */
1763 cfg
.buffer_index
= pan_xfb_base(present
) + o
.output_buffer
;
1764 cfg
.unknown
= quirks
& IS_BIFROST
? 0x0 : 0x1;
1766 /* Override number of channels and precision to highp */
1767 cfg
.format
= (pan_xfb_format(format
, o
.num_components
) << 12) | swizzle
;
1769 /* Apply given offsets together */
1770 cfg
.offset
= (o
.dst_offset
* 4) /* dwords */
1771 + streamout_offsets
[o
.output_buffer
];
1775 /* Determine if we should capture a varying for XFB. This requires actually
1776 * having a buffer for it. If we don't capture it, we'll fallback to a general
1777 * varying path (linked or unlinked, possibly discarding the write) */
1780 panfrost_xfb_captured(struct panfrost_shader_state
*xfb
,
1781 unsigned loc
, unsigned max_xfb
)
1783 if (!(xfb
->so_mask
& (1ll << loc
)))
1786 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1787 return o
->output_buffer
< max_xfb
;
1791 pan_emit_general_varying(struct mali_attribute_packed
*out
,
1792 struct panfrost_shader_state
*other
,
1793 struct panfrost_shader_state
*xfb
,
1794 gl_varying_slot loc
,
1795 enum mali_format format
,
1798 unsigned *gen_offsets
,
1799 enum mali_format
*gen_formats
,
1800 unsigned *gen_stride
,
1804 /* Check if we're linked */
1805 signed other_idx
= -1;
1807 for (unsigned j
= 0; j
< other
->varying_count
; ++j
) {
1808 if (other
->varyings_loc
[j
] == loc
) {
1814 if (other_idx
< 0) {
1815 pan_emit_vary_only(out
, present
, quirks
);
1819 unsigned offset
= gen_offsets
[other_idx
];
1822 /* We're linked, so allocate a space via a watermark allocation */
1823 enum mali_format alt
= other
->varyings
[other_idx
];
1825 /* Do interpolation at minimum precision */
1826 unsigned size_main
= pan_varying_size(format
);
1827 unsigned size_alt
= pan_varying_size(alt
);
1828 unsigned size
= MIN2(size_main
, size_alt
);
1830 /* If a varying is marked for XFB but not actually captured, we
1831 * should match the format to the format that would otherwise
1832 * be used for XFB, since dEQP checks for invariance here. It's
1833 * unclear if this is required by the spec. */
1835 if (xfb
->so_mask
& (1ull << loc
)) {
1836 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1837 format
= pan_xfb_format(format
, o
->num_components
);
1838 size
= pan_varying_size(format
);
1839 } else if (size
== size_alt
) {
1843 gen_offsets
[idx
] = *gen_stride
;
1844 gen_formats
[other_idx
] = format
;
1845 offset
= *gen_stride
;
1846 *gen_stride
+= size
;
1849 pan_emit_vary(out
, present
, PAN_VARY_GENERAL
, quirks
, format
, offset
);
1852 /* Higher-level wrapper around all of the above, classifying a varying into one
1853 * of the above types */
1856 panfrost_emit_varying(
1857 struct mali_attribute_packed
*out
,
1858 struct panfrost_shader_state
*stage
,
1859 struct panfrost_shader_state
*other
,
1860 struct panfrost_shader_state
*xfb
,
1863 unsigned *streamout_offsets
,
1865 unsigned *gen_offsets
,
1866 enum mali_format
*gen_formats
,
1867 unsigned *gen_stride
,
1872 gl_varying_slot loc
= stage
->varyings_loc
[idx
];
1873 enum mali_format format
= stage
->varyings
[idx
];
1875 /* Override format to match linkage */
1876 if (!should_alloc
&& gen_formats
[idx
])
1877 format
= gen_formats
[idx
];
1879 if (has_point_coord(stage
->point_sprite_mask
, loc
)) {
1880 pan_emit_vary_special(out
, present
, PAN_VARY_PNTCOORD
, quirks
);
1881 } else if (panfrost_xfb_captured(xfb
, loc
, max_xfb
)) {
1882 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1883 pan_emit_vary_xfb(out
, present
, max_xfb
, streamout_offsets
, quirks
, format
, *o
);
1884 } else if (loc
== VARYING_SLOT_POS
) {
1886 pan_emit_vary_special(out
, present
, PAN_VARY_FRAGCOORD
, quirks
);
1888 pan_emit_vary_special(out
, present
, PAN_VARY_POSITION
, quirks
);
1889 } else if (loc
== VARYING_SLOT_PSIZ
) {
1890 pan_emit_vary_special(out
, present
, PAN_VARY_PSIZ
, quirks
);
1891 } else if (loc
== VARYING_SLOT_PNTC
) {
1892 pan_emit_vary_special(out
, present
, PAN_VARY_PNTCOORD
, quirks
);
1893 } else if (loc
== VARYING_SLOT_FACE
) {
1894 pan_emit_vary_special(out
, present
, PAN_VARY_FACE
, quirks
);
1896 pan_emit_general_varying(out
, other
, xfb
, loc
, format
, present
,
1897 quirks
, gen_offsets
, gen_formats
, gen_stride
,
1903 pan_emit_special_input(struct mali_attribute_buffer_packed
*out
,
1905 enum pan_special_varying v
,
1908 if (present
& (1 << v
)) {
1909 unsigned idx
= pan_varying_index(present
, v
);
1911 pan_pack(out
+ idx
, ATTRIBUTE_BUFFER
, cfg
) {
1912 cfg
.special
= special
;
1919 panfrost_emit_varying_descriptor(struct panfrost_batch
*batch
,
1920 unsigned vertex_count
,
1921 struct mali_vertex_tiler_postfix
*vertex_postfix
,
1922 struct mali_vertex_tiler_postfix
*tiler_postfix
,
1923 union midgard_primitive_size
*primitive_size
)
1925 /* Load the shaders */
1926 struct panfrost_context
*ctx
= batch
->ctx
;
1927 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
1928 struct panfrost_shader_state
*vs
, *fs
;
1929 size_t vs_size
, fs_size
;
1931 /* Allocate the varying descriptor */
1933 vs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_VERTEX
);
1934 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
1935 vs_size
= MALI_ATTRIBUTE_LENGTH
* vs
->varying_count
;
1936 fs_size
= MALI_ATTRIBUTE_LENGTH
* fs
->varying_count
;
1938 struct panfrost_transfer trans
= panfrost_pool_alloc(&batch
->pool
,
1942 struct pipe_stream_output_info
*so
= &vs
->stream_output
;
1943 unsigned present
= pan_varying_present(vs
, fs
, dev
->quirks
);
1945 /* Check if this varying is linked by us. This is the case for
1946 * general-purpose, non-captured varyings. If it is, link it. If it's
1947 * not, use the provided stream out information to determine the
1948 * offset, since it was already linked for us. */
1950 unsigned gen_offsets
[32];
1951 enum mali_format gen_formats
[32];
1952 memset(gen_offsets
, 0, sizeof(gen_offsets
));
1953 memset(gen_formats
, 0, sizeof(gen_formats
));
1955 unsigned gen_stride
= 0;
1956 assert(vs
->varying_count
< ARRAY_SIZE(gen_offsets
));
1957 assert(fs
->varying_count
< ARRAY_SIZE(gen_offsets
));
1959 unsigned streamout_offsets
[32];
1961 for (unsigned i
= 0; i
< ctx
->streamout
.num_targets
; ++i
) {
1962 streamout_offsets
[i
] = panfrost_streamout_offset(
1964 ctx
->streamout
.offsets
[i
],
1965 ctx
->streamout
.targets
[i
]);
1968 struct mali_attribute_packed
*ovs
= (struct mali_attribute_packed
*)trans
.cpu
;
1969 struct mali_attribute_packed
*ofs
= ovs
+ vs
->varying_count
;
1971 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1972 panfrost_emit_varying(ovs
+ i
, vs
, fs
, vs
, present
,
1973 ctx
->streamout
.num_targets
, streamout_offsets
,
1975 gen_offsets
, gen_formats
, &gen_stride
, i
, true, false);
1978 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1979 panfrost_emit_varying(ofs
+ i
, fs
, vs
, vs
, present
,
1980 ctx
->streamout
.num_targets
, streamout_offsets
,
1982 gen_offsets
, gen_formats
, &gen_stride
, i
, false, true);
1985 unsigned xfb_base
= pan_xfb_base(present
);
1986 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
,
1987 MALI_ATTRIBUTE_BUFFER_LENGTH
* (xfb_base
+ ctx
->streamout
.num_targets
));
1988 struct mali_attribute_buffer_packed
*varyings
=
1989 (struct mali_attribute_buffer_packed
*) T
.cpu
;
1991 /* Emit the stream out buffers */
1993 unsigned out_count
= u_stream_outputs_for_vertices(ctx
->active_prim
,
1996 for (unsigned i
= 0; i
< ctx
->streamout
.num_targets
; ++i
) {
1997 panfrost_emit_streamout(batch
, &varyings
[xfb_base
+ i
],
1999 ctx
->streamout
.offsets
[i
],
2001 ctx
->streamout
.targets
[i
]);
2004 panfrost_emit_varyings(batch
,
2005 &varyings
[pan_varying_index(present
, PAN_VARY_GENERAL
)],
2006 gen_stride
, vertex_count
);
2008 /* fp32 vec4 gl_Position */
2009 tiler_postfix
->position_varying
= panfrost_emit_varyings(batch
,
2010 &varyings
[pan_varying_index(present
, PAN_VARY_POSITION
)],
2011 sizeof(float) * 4, vertex_count
);
2013 if (present
& (1 << PAN_VARY_PSIZ
)) {
2014 primitive_size
->pointer
= panfrost_emit_varyings(batch
,
2015 &varyings
[pan_varying_index(present
, PAN_VARY_PSIZ
)],
2019 pan_emit_special_input(varyings
, present
, PAN_VARY_PNTCOORD
, MALI_ATTRIBUTE_SPECIAL_POINT_COORD
);
2020 pan_emit_special_input(varyings
, present
, PAN_VARY_FACE
, MALI_ATTRIBUTE_SPECIAL_FRONT_FACING
);
2021 pan_emit_special_input(varyings
, present
, PAN_VARY_FRAGCOORD
, MALI_ATTRIBUTE_SPECIAL_FRAG_COORD
);
2023 vertex_postfix
->varyings
= T
.gpu
;
2024 tiler_postfix
->varyings
= T
.gpu
;
2026 vertex_postfix
->varying_meta
= trans
.gpu
;
2027 tiler_postfix
->varying_meta
= trans
.gpu
+ vs_size
;
2031 panfrost_emit_vertex_tiler_jobs(struct panfrost_batch
*batch
,
2032 struct mali_vertex_tiler_prefix
*vertex_prefix
,
2033 struct mali_vertex_tiler_postfix
*vertex_postfix
,
2034 struct mali_vertex_tiler_prefix
*tiler_prefix
,
2035 struct mali_vertex_tiler_postfix
*tiler_postfix
,
2036 union midgard_primitive_size
*primitive_size
)
2038 struct panfrost_context
*ctx
= batch
->ctx
;
2039 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
2040 bool wallpapering
= ctx
->wallpaper_batch
&& batch
->scoreboard
.tiler_dep
;
2041 struct bifrost_payload_vertex bifrost_vertex
= {0,};
2042 struct bifrost_payload_tiler bifrost_tiler
= {0,};
2043 struct midgard_payload_vertex_tiler midgard_vertex
= {0,};
2044 struct midgard_payload_vertex_tiler midgard_tiler
= {0,};
2046 size_t vp_size
, tp_size
;
2048 if (device
->quirks
& IS_BIFROST
) {
2049 bifrost_vertex
.prefix
= *vertex_prefix
;
2050 bifrost_vertex
.postfix
= *vertex_postfix
;
2051 vp
= &bifrost_vertex
;
2052 vp_size
= sizeof(bifrost_vertex
);
2054 bifrost_tiler
.prefix
= *tiler_prefix
;
2055 bifrost_tiler
.tiler
.primitive_size
= *primitive_size
;
2056 bifrost_tiler
.tiler
.tiler_meta
= panfrost_batch_get_tiler_meta(batch
, ~0);
2057 bifrost_tiler
.postfix
= *tiler_postfix
;
2058 tp
= &bifrost_tiler
;
2059 tp_size
= sizeof(bifrost_tiler
);
2061 midgard_vertex
.prefix
= *vertex_prefix
;
2062 midgard_vertex
.postfix
= *vertex_postfix
;
2063 vp
= &midgard_vertex
;
2064 vp_size
= sizeof(midgard_vertex
);
2066 midgard_tiler
.prefix
= *tiler_prefix
;
2067 midgard_tiler
.postfix
= *tiler_postfix
;
2068 midgard_tiler
.primitive_size
= *primitive_size
;
2069 tp
= &midgard_tiler
;
2070 tp_size
= sizeof(midgard_tiler
);
2074 /* Inject in reverse order, with "predicted" job indices.
2075 * THIS IS A HACK XXX */
2076 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_TILER
, false,
2077 batch
->scoreboard
.job_index
+ 2, tp
, tp_size
, true);
2078 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_VERTEX
, false, 0,
2083 /* If rasterizer discard is enable, only submit the vertex */
2085 unsigned vertex
= panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_VERTEX
, false, 0,
2086 vp
, vp_size
, false);
2088 if (ctx
->rasterizer
->base
.rasterizer_discard
)
2091 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_TILER
, false, vertex
, tp
, tp_size
,
2095 /* TODO: stop hardcoding this */
2097 panfrost_emit_sample_locations(struct panfrost_batch
*batch
)
2099 uint16_t locations
[] = {
2150 return panfrost_pool_upload(&batch
->pool
, locations
, 96 * sizeof(uint16_t));