2 * Copyright (C) 2018 Alyssa Rosenzweig
3 * Copyright (C) 2020 Collabora Ltd.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
25 #include "util/macros.h"
26 #include "util/u_prim.h"
27 #include "util/u_vbuf.h"
29 #include "panfrost-quirks.h"
33 #include "pan_cmdstream.h"
34 #include "pan_context.h"
37 /* If a BO is accessed for a particular shader stage, will it be in the primary
38 * batch (vertex/tiler) or the secondary batch (fragment)? Anything but
39 * fragment will be primary, e.g. compute jobs will be considered
40 * "vertex/tiler" by analogy */
42 static inline uint32_t
43 panfrost_bo_access_for_stage(enum pipe_shader_type stage
)
45 assert(stage
== PIPE_SHADER_FRAGMENT
||
46 stage
== PIPE_SHADER_VERTEX
||
47 stage
== PIPE_SHADER_COMPUTE
);
49 return stage
== PIPE_SHADER_FRAGMENT
?
50 PAN_BO_ACCESS_FRAGMENT
:
51 PAN_BO_ACCESS_VERTEX_TILER
;
55 panfrost_vt_emit_shared_memory(struct panfrost_context
*ctx
,
56 struct mali_vertex_tiler_postfix
*postfix
)
58 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
59 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
61 unsigned shift
= panfrost_get_stack_shift(batch
->stack_size
);
62 struct mali_shared_memory shared
= {
64 .scratchpad
= panfrost_batch_get_scratchpad(batch
, shift
, dev
->thread_tls_alloc
, dev
->core_count
)->gpu
,
65 .shared_workgroup_count
= ~0,
67 postfix
->shared_memory
= panfrost_pool_upload(&batch
->pool
, &shared
, sizeof(shared
));
71 panfrost_vt_attach_framebuffer(struct panfrost_context
*ctx
,
72 struct mali_vertex_tiler_postfix
*postfix
)
74 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
75 postfix
->shared_memory
= panfrost_batch_reserve_framebuffer(batch
);
79 panfrost_vt_update_rasterizer(struct panfrost_rasterizer
*rasterizer
,
80 struct mali_vertex_tiler_prefix
*prefix
,
81 struct mali_vertex_tiler_postfix
*postfix
)
83 postfix
->gl_enables
|= 0x7;
84 SET_BIT(postfix
->gl_enables
, MALI_FRONT_CCW_TOP
,
85 rasterizer
->base
.front_ccw
);
86 SET_BIT(postfix
->gl_enables
, MALI_CULL_FACE_FRONT
,
87 (rasterizer
->base
.cull_face
& PIPE_FACE_FRONT
));
88 SET_BIT(postfix
->gl_enables
, MALI_CULL_FACE_BACK
,
89 (rasterizer
->base
.cull_face
& PIPE_FACE_BACK
));
90 SET_BIT(prefix
->unknown_draw
, MALI_DRAW_FLATSHADE_FIRST
,
91 rasterizer
->base
.flatshade_first
);
95 panfrost_vt_update_primitive_size(struct panfrost_context
*ctx
,
96 struct mali_vertex_tiler_prefix
*prefix
,
97 union midgard_primitive_size
*primitive_size
)
99 struct panfrost_rasterizer
*rasterizer
= ctx
->rasterizer
;
101 if (!panfrost_writes_point_size(ctx
)) {
102 float val
= (prefix
->draw_mode
== MALI_DRAW_MODE_POINTS
) ?
103 rasterizer
->base
.point_size
:
104 rasterizer
->base
.line_width
;
106 primitive_size
->constant
= val
;
111 panfrost_vt_update_occlusion_query(struct panfrost_context
*ctx
,
112 struct mali_vertex_tiler_postfix
*postfix
)
114 SET_BIT(postfix
->gl_enables
, MALI_OCCLUSION_QUERY
, ctx
->occlusion_query
);
115 if (ctx
->occlusion_query
) {
116 postfix
->occlusion_counter
= ctx
->occlusion_query
->bo
->gpu
;
117 panfrost_batch_add_bo(ctx
->batch
, ctx
->occlusion_query
->bo
,
118 PAN_BO_ACCESS_SHARED
|
120 PAN_BO_ACCESS_FRAGMENT
);
122 postfix
->occlusion_counter
= 0;
127 panfrost_vt_init(struct panfrost_context
*ctx
,
128 enum pipe_shader_type stage
,
129 struct mali_vertex_tiler_prefix
*prefix
,
130 struct mali_vertex_tiler_postfix
*postfix
)
132 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
134 if (!ctx
->shader
[stage
])
137 memset(prefix
, 0, sizeof(*prefix
));
138 memset(postfix
, 0, sizeof(*postfix
));
140 if (device
->quirks
& IS_BIFROST
) {
141 postfix
->gl_enables
= 0x2;
142 panfrost_vt_emit_shared_memory(ctx
, postfix
);
144 postfix
->gl_enables
= 0x6;
145 panfrost_vt_attach_framebuffer(ctx
, postfix
);
148 if (stage
== PIPE_SHADER_FRAGMENT
) {
149 panfrost_vt_update_occlusion_query(ctx
, postfix
);
150 panfrost_vt_update_rasterizer(ctx
->rasterizer
, prefix
, postfix
);
155 panfrost_translate_index_size(unsigned size
)
159 return MALI_DRAW_INDEXED_UINT8
;
162 return MALI_DRAW_INDEXED_UINT16
;
165 return MALI_DRAW_INDEXED_UINT32
;
168 unreachable("Invalid index size");
172 /* Gets a GPU address for the associated index buffer. Only gauranteed to be
173 * good for the duration of the draw (transient), could last longer. Also get
174 * the bounds on the index buffer for the range accessed by the draw. We do
175 * these operations together because there are natural optimizations which
176 * require them to be together. */
179 panfrost_get_index_buffer_bounded(struct panfrost_context
*ctx
,
180 const struct pipe_draw_info
*info
,
181 unsigned *min_index
, unsigned *max_index
)
183 struct panfrost_resource
*rsrc
= pan_resource(info
->index
.resource
);
184 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
185 off_t offset
= info
->start
* info
->index_size
;
186 bool needs_indices
= true;
189 if (info
->max_index
!= ~0u) {
190 *min_index
= info
->min_index
;
191 *max_index
= info
->max_index
;
192 needs_indices
= false;
195 if (!info
->has_user_indices
) {
196 /* Only resources can be directly mapped */
197 panfrost_batch_add_bo(batch
, rsrc
->bo
,
198 PAN_BO_ACCESS_SHARED
|
200 PAN_BO_ACCESS_VERTEX_TILER
);
201 out
= rsrc
->bo
->gpu
+ offset
;
203 /* Check the cache */
204 needs_indices
= !panfrost_minmax_cache_get(rsrc
->index_cache
,
210 /* Otherwise, we need to upload to transient memory */
211 const uint8_t *ibuf8
= (const uint8_t *) info
->index
.user
;
212 out
= panfrost_pool_upload(&batch
->pool
, ibuf8
+ offset
,
219 u_vbuf_get_minmax_index(&ctx
->base
, info
, min_index
, max_index
);
221 if (!info
->has_user_indices
)
222 panfrost_minmax_cache_add(rsrc
->index_cache
,
223 info
->start
, info
->count
,
224 *min_index
, *max_index
);
231 panfrost_vt_set_draw_info(struct panfrost_context
*ctx
,
232 const struct pipe_draw_info
*info
,
233 enum mali_draw_mode draw_mode
,
234 struct mali_vertex_tiler_postfix
*vertex_postfix
,
235 struct mali_vertex_tiler_prefix
*tiler_prefix
,
236 struct mali_vertex_tiler_postfix
*tiler_postfix
,
237 unsigned *vertex_count
,
238 unsigned *padded_count
)
240 tiler_prefix
->draw_mode
= draw_mode
;
242 unsigned draw_flags
= 0;
244 if (panfrost_writes_point_size(ctx
))
245 draw_flags
|= MALI_DRAW_VARYING_SIZE
;
247 if (info
->primitive_restart
)
248 draw_flags
|= MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX
;
250 /* These doesn't make much sense */
252 draw_flags
|= 0x3000;
254 if (info
->index_size
) {
255 unsigned min_index
= 0, max_index
= 0;
257 tiler_prefix
->indices
= panfrost_get_index_buffer_bounded(ctx
,
262 /* Use the corresponding values */
263 *vertex_count
= max_index
- min_index
+ 1;
264 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= min_index
+ info
->index_bias
;
265 tiler_prefix
->offset_bias_correction
= -min_index
;
266 tiler_prefix
->index_count
= MALI_POSITIVE(info
->count
);
267 draw_flags
|= panfrost_translate_index_size(info
->index_size
);
269 tiler_prefix
->indices
= 0;
270 *vertex_count
= ctx
->vertex_count
;
271 tiler_postfix
->offset_start
= vertex_postfix
->offset_start
= info
->start
;
272 tiler_prefix
->offset_bias_correction
= 0;
273 tiler_prefix
->index_count
= MALI_POSITIVE(ctx
->vertex_count
);
276 tiler_prefix
->unknown_draw
= draw_flags
;
278 /* Encode the padded vertex count */
280 if (info
->instance_count
> 1) {
281 *padded_count
= panfrost_padded_vertex_count(*vertex_count
);
283 unsigned shift
= __builtin_ctz(ctx
->padded_count
);
284 unsigned k
= ctx
->padded_count
>> (shift
+ 1);
286 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= shift
;
287 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= k
;
289 *padded_count
= *vertex_count
;
291 /* Reset instancing state */
292 tiler_postfix
->instance_shift
= vertex_postfix
->instance_shift
= 0;
293 tiler_postfix
->instance_odd
= vertex_postfix
->instance_odd
= 0;
298 panfrost_shader_meta_init(struct panfrost_context
*ctx
,
299 enum pipe_shader_type st
,
300 struct mali_shader_meta
*meta
)
302 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
303 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
305 memset(meta
, 0, sizeof(*meta
));
306 meta
->shader
= (ss
->bo
? ss
->bo
->gpu
: 0) | ss
->first_tag
;
307 meta
->attribute_count
= ss
->attribute_count
;
308 meta
->varying_count
= ss
->varying_count
;
309 meta
->texture_count
= ctx
->sampler_view_count
[st
];
310 meta
->sampler_count
= ctx
->sampler_count
[st
];
312 if (dev
->quirks
& IS_BIFROST
) {
313 if (st
== PIPE_SHADER_VERTEX
)
314 meta
->bifrost1
.unk1
= 0x800000;
316 /* First clause ATEST |= 0x4000000.
317 * Less than 32 regs |= 0x200 */
318 meta
->bifrost1
.unk1
= 0x950020;
321 meta
->bifrost1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
322 if (st
== PIPE_SHADER_VERTEX
)
323 meta
->bifrost2
.preload_regs
= 0xC0;
325 meta
->bifrost2
.preload_regs
= 0x1;
326 SET_BIT(meta
->bifrost2
.preload_regs
, 0x10, ss
->reads_frag_coord
);
329 meta
->bifrost2
.uniform_count
= MIN2(ss
->uniform_count
,
332 meta
->midgard1
.uniform_count
= MIN2(ss
->uniform_count
,
334 meta
->midgard1
.work_count
= ss
->work_reg_count
;
336 /* TODO: This is not conformant on ES3 */
337 meta
->midgard1
.flags_hi
= MALI_SUPPRESS_INF_NAN
;
339 meta
->midgard1
.flags_lo
= 0x20;
340 meta
->midgard1
.uniform_buffer_count
= panfrost_ubo_count(ctx
, st
);
342 SET_BIT(meta
->midgard1
.flags_lo
, MALI_WRITES_GLOBAL
, ss
->writes_global
);
347 translate_tex_wrap(enum pipe_tex_wrap w
)
350 case PIPE_TEX_WRAP_REPEAT
: return MALI_WRAP_MODE_REPEAT
;
351 case PIPE_TEX_WRAP_CLAMP
: return MALI_WRAP_MODE_CLAMP
;
352 case PIPE_TEX_WRAP_CLAMP_TO_EDGE
: return MALI_WRAP_MODE_CLAMP_TO_EDGE
;
353 case PIPE_TEX_WRAP_CLAMP_TO_BORDER
: return MALI_WRAP_MODE_CLAMP_TO_BORDER
;
354 case PIPE_TEX_WRAP_MIRROR_REPEAT
: return MALI_WRAP_MODE_MIRRORED_REPEAT
;
355 case PIPE_TEX_WRAP_MIRROR_CLAMP
: return MALI_WRAP_MODE_MIRRORED_CLAMP
;
356 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE
: return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_EDGE
;
357 case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER
: return MALI_WRAP_MODE_MIRRORED_CLAMP_TO_BORDER
;
358 default: unreachable("Invalid wrap");
362 /* The hardware compares in the wrong order order, so we have to flip before
363 * encoding. Yes, really. */
365 static enum mali_func
366 panfrost_sampler_compare_func(const struct pipe_sampler_state
*cso
)
368 if (!cso
->compare_mode
)
369 return MALI_FUNC_NEVER
;
371 enum mali_func f
= panfrost_translate_compare_func(cso
->compare_func
);
372 return panfrost_flip_compare_func(f
);
375 static enum mali_mipmap_mode
376 pan_pipe_to_mipmode(enum pipe_tex_mipfilter f
)
379 case PIPE_TEX_MIPFILTER_NEAREST
: return MALI_MIPMAP_MODE_NEAREST
;
380 case PIPE_TEX_MIPFILTER_LINEAR
: return MALI_MIPMAP_MODE_TRILINEAR
;
381 case PIPE_TEX_MIPFILTER_NONE
: return MALI_MIPMAP_MODE_NONE
;
382 default: unreachable("Invalid");
386 void panfrost_sampler_desc_init(const struct pipe_sampler_state
*cso
,
387 struct mali_midgard_sampler_packed
*hw
)
389 pan_pack(hw
, MIDGARD_SAMPLER
, cfg
) {
390 cfg
.magnify_nearest
= cso
->mag_img_filter
== PIPE_TEX_FILTER_NEAREST
;
391 cfg
.minify_nearest
= cso
->min_img_filter
== PIPE_TEX_FILTER_NEAREST
;
392 cfg
.mipmap_mode
= (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_LINEAR
) ?
393 MALI_MIPMAP_MODE_TRILINEAR
: MALI_MIPMAP_MODE_NEAREST
;
394 cfg
.normalized_coordinates
= cso
->normalized_coords
;
396 cfg
.lod_bias
= FIXED_16(cso
->lod_bias
, true);
398 cfg
.minimum_lod
= FIXED_16(cso
->min_lod
, false);
400 /* If necessary, we disable mipmapping in the sampler descriptor by
401 * clamping the LOD as tight as possible (from 0 to epsilon,
402 * essentially -- remember these are fixed point numbers, so
405 cfg
.maximum_lod
= (cso
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
) ?
406 cfg
.minimum_lod
+ 1 :
407 FIXED_16(cso
->max_lod
, false);
409 cfg
.wrap_mode_s
= translate_tex_wrap(cso
->wrap_s
);
410 cfg
.wrap_mode_t
= translate_tex_wrap(cso
->wrap_t
);
411 cfg
.wrap_mode_r
= translate_tex_wrap(cso
->wrap_r
);
413 cfg
.compare_function
= panfrost_sampler_compare_func(cso
);
414 cfg
.seamless_cube_map
= cso
->seamless_cube_map
;
416 cfg
.border_color_r
= cso
->border_color
.f
[0];
417 cfg
.border_color_g
= cso
->border_color
.f
[1];
418 cfg
.border_color_b
= cso
->border_color
.f
[2];
419 cfg
.border_color_a
= cso
->border_color
.f
[3];
423 void panfrost_sampler_desc_init_bifrost(const struct pipe_sampler_state
*cso
,
424 struct mali_bifrost_sampler_packed
*hw
)
426 pan_pack(hw
, BIFROST_SAMPLER
, cfg
) {
427 cfg
.magnify_linear
= cso
->mag_img_filter
== PIPE_TEX_FILTER_LINEAR
;
428 cfg
.minify_linear
= cso
->min_img_filter
== PIPE_TEX_FILTER_LINEAR
;
429 cfg
.mipmap_mode
= pan_pipe_to_mipmode(cso
->min_mip_filter
);
430 cfg
.normalized_coordinates
= cso
->normalized_coords
;
432 cfg
.lod_bias
= FIXED_16(cso
->lod_bias
, true);
433 cfg
.minimum_lod
= FIXED_16(cso
->min_lod
, false);
434 cfg
.maximum_lod
= FIXED_16(cso
->max_lod
, false);
436 cfg
.wrap_mode_s
= translate_tex_wrap(cso
->wrap_s
);
437 cfg
.wrap_mode_t
= translate_tex_wrap(cso
->wrap_t
);
438 cfg
.wrap_mode_r
= translate_tex_wrap(cso
->wrap_r
);
440 cfg
.compare_function
= panfrost_sampler_compare_func(cso
);
441 cfg
.seamless_cube_map
= cso
->seamless_cube_map
;
446 panfrost_frag_meta_rasterizer_update(struct panfrost_context
*ctx
,
447 struct mali_shader_meta
*fragmeta
)
449 struct pipe_rasterizer_state
*rast
= &ctx
->rasterizer
->base
;
451 bool msaa
= rast
->multisample
;
453 /* TODO: Sample size */
454 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_MSAA
, msaa
);
455 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_MSAA
, !msaa
);
457 struct panfrost_shader_state
*fs
;
458 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
460 /* EXT_shader_framebuffer_fetch requires the shader to be run
461 * per-sample when outputs are read. */
462 bool per_sample
= ctx
->min_samples
> 1 || fs
->outputs_read
;
463 SET_BIT(fragmeta
->unknown2_3
, MALI_PER_SAMPLE
, msaa
&& per_sample
);
465 fragmeta
->depth_units
= rast
->offset_units
* 2.0f
;
466 fragmeta
->depth_factor
= rast
->offset_scale
;
468 /* XXX: Which bit is which? Does this maybe allow offseting not-tri? */
470 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_A
, rast
->offset_tri
);
471 SET_BIT(fragmeta
->unknown2_4
, MALI_DEPTH_RANGE_B
, rast
->offset_tri
);
473 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_NEAR
, rast
->depth_clip_near
);
474 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_CLIP_FAR
, rast
->depth_clip_far
);
478 panfrost_frag_meta_zsa_update(struct panfrost_context
*ctx
,
479 struct mali_shader_meta
*fragmeta
)
481 const struct panfrost_zsa_state
*so
= ctx
->depth_stencil
;
483 SET_BIT(fragmeta
->unknown2_4
, MALI_STENCIL_TEST
,
484 so
->base
.stencil
[0].enabled
);
486 fragmeta
->stencil_mask_front
= so
->stencil_mask_front
;
487 fragmeta
->stencil_mask_back
= so
->stencil_mask_back
;
489 /* Bottom bits for stencil ref, exactly one word */
490 fragmeta
->stencil_front
.opaque
[0] = so
->stencil_front
.opaque
[0] | ctx
->stencil_ref
.ref_value
[0];
492 /* If back-stencil is not enabled, use the front values */
494 if (so
->base
.stencil
[1].enabled
)
495 fragmeta
->stencil_back
.opaque
[0] = so
->stencil_back
.opaque
[0] | ctx
->stencil_ref
.ref_value
[1];
497 fragmeta
->stencil_back
= fragmeta
->stencil_front
;
499 SET_BIT(fragmeta
->unknown2_3
, MALI_DEPTH_WRITEMASK
,
500 so
->base
.depth
.writemask
);
502 fragmeta
->unknown2_3
&= ~MALI_DEPTH_FUNC_MASK
;
503 fragmeta
->unknown2_3
|= MALI_DEPTH_FUNC(panfrost_translate_compare_func(
504 so
->base
.depth
.enabled
? so
->base
.depth
.func
: PIPE_FUNC_ALWAYS
));
508 panfrost_fs_required(
509 struct panfrost_shader_state
*fs
,
510 struct panfrost_blend_final
*blend
,
513 /* If we generally have side effects */
517 /* If colour is written we need to execute */
518 for (unsigned i
= 0; i
< rt_count
; ++i
) {
519 if (!blend
[i
].no_colour
)
523 /* If depth is written and not implied we need to execute.
524 * TODO: Predicate on Z/S writes being enabled */
525 return (fs
->writes_depth
|| fs
->writes_stencil
);
529 panfrost_frag_meta_blend_update(struct panfrost_context
*ctx
,
530 struct mali_shader_meta
*fragmeta
,
533 struct panfrost_batch
*batch
= panfrost_get_batch_for_fbo(ctx
);
534 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
535 struct panfrost_shader_state
*fs
;
536 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
538 SET_BIT(fragmeta
->unknown2_4
, MALI_NO_DITHER
,
539 (dev
->quirks
& MIDGARD_SFBD
) && ctx
->blend
&&
540 !ctx
->blend
->base
.dither
);
542 SET_BIT(fragmeta
->unknown2_4
, MALI_ALPHA_TO_COVERAGE
,
543 ctx
->blend
->base
.alpha_to_coverage
);
545 /* Get blending setup */
546 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
548 struct panfrost_blend_final blend
[PIPE_MAX_COLOR_BUFS
];
549 unsigned shader_offset
= 0;
550 struct panfrost_bo
*shader_bo
= NULL
;
552 for (unsigned c
= 0; c
< rt_count
; ++c
)
553 blend
[c
] = panfrost_get_blend_for_context(ctx
, c
, &shader_bo
,
556 /* Disable shader execution if we can */
557 if (dev
->quirks
& MIDGARD_SHADERLESS
558 && !panfrost_fs_required(fs
, blend
, rt_count
)) {
559 fragmeta
->shader
= 0;
560 fragmeta
->attribute_count
= 0;
561 fragmeta
->varying_count
= 0;
562 fragmeta
->texture_count
= 0;
563 fragmeta
->sampler_count
= 0;
565 /* This feature is not known to work on Bifrost */
566 fragmeta
->midgard1
.work_count
= 1;
567 fragmeta
->midgard1
.uniform_count
= 0;
568 fragmeta
->midgard1
.uniform_buffer_count
= 0;
571 /* If there is a blend shader, work registers are shared. We impose 8
572 * work registers as a limit for blend shaders. Should be lower XXX */
574 if (!(dev
->quirks
& IS_BIFROST
)) {
575 for (unsigned c
= 0; c
< rt_count
; ++c
) {
576 if (blend
[c
].is_shader
) {
577 fragmeta
->midgard1
.work_count
=
578 MAX2(fragmeta
->midgard1
.work_count
, 8);
583 /* Even on MFBD, the shader descriptor gets blend shaders. It's *also*
584 * copied to the blend_meta appended (by convention), but this is the
585 * field actually read by the hardware. (Or maybe both are read...?).
586 * Specify the last RTi with a blend shader. */
588 fragmeta
->blend
.shader
= 0;
590 for (signed rt
= (rt_count
- 1); rt
>= 0; --rt
) {
591 if (!blend
[rt
].is_shader
)
594 fragmeta
->blend
.shader
= blend
[rt
].shader
.gpu
|
595 blend
[rt
].shader
.first_tag
;
599 if (dev
->quirks
& MIDGARD_SFBD
) {
600 /* When only a single render target platform is used, the blend
601 * information is inside the shader meta itself. We additionally
602 * need to signal CAN_DISCARD for nontrivial blend modes (so
603 * we're able to read back the destination buffer) */
605 SET_BIT(fragmeta
->unknown2_3
, MALI_HAS_BLEND_SHADER
,
608 if (!blend
[0].is_shader
) {
609 fragmeta
->blend
.equation
= *blend
[0].equation
.equation
;
610 fragmeta
->blend
.constant
= blend
[0].equation
.constant
;
613 SET_BIT(fragmeta
->unknown2_3
, MALI_CAN_DISCARD
,
614 !blend
[0].no_blending
|| fs
->can_discard
);
616 batch
->draws
|= PIPE_CLEAR_COLOR0
;
620 if (dev
->quirks
& IS_BIFROST
) {
621 bool no_blend
= true;
623 for (unsigned i
= 0; i
< rt_count
; ++i
)
624 no_blend
&= (blend
[i
].no_blending
| blend
[i
].no_colour
);
626 SET_BIT(fragmeta
->bifrost1
.unk1
, MALI_BIFROST_EARLY_Z
,
627 !fs
->can_discard
&& !fs
->writes_depth
&& no_blend
);
630 /* Additional blend descriptor tacked on for jobs using MFBD */
632 for (unsigned i
= 0; i
< rt_count
; ++i
) {
635 if (ctx
->pipe_framebuffer
.nr_cbufs
> i
&& !blend
[i
].no_colour
) {
637 batch
->draws
|= (PIPE_CLEAR_COLOR0
<< i
);
639 bool is_srgb
= (ctx
->pipe_framebuffer
.nr_cbufs
> i
) &&
640 (ctx
->pipe_framebuffer
.cbufs
[i
]) &&
641 util_format_is_srgb(ctx
->pipe_framebuffer
.cbufs
[i
]->format
);
643 SET_BIT(flags
, MALI_BLEND_MRT_SHADER
, blend
[i
].is_shader
);
644 SET_BIT(flags
, MALI_BLEND_LOAD_TIB
, !blend
[i
].no_blending
);
645 SET_BIT(flags
, MALI_BLEND_SRGB
, is_srgb
);
646 SET_BIT(flags
, MALI_BLEND_NO_DITHER
, !ctx
->blend
->base
.dither
);
649 if (dev
->quirks
& IS_BIFROST
) {
650 struct bifrost_blend_rt
*brts
= rts
;
652 brts
[i
].flags
= flags
;
654 if (blend
[i
].is_shader
) {
655 /* The blend shader's address needs to be at
656 * the same top 32 bit as the fragment shader.
657 * TODO: Ensure that's always the case.
659 assert((blend
[i
].shader
.gpu
& (0xffffffffull
<< 32)) ==
660 (fs
->bo
->gpu
& (0xffffffffull
<< 32)));
661 brts
[i
].shader
= blend
[i
].shader
.gpu
;
663 } else if (ctx
->pipe_framebuffer
.nr_cbufs
> i
) {
664 enum pipe_format format
= ctx
->pipe_framebuffer
.cbufs
[i
]->format
;
665 const struct util_format_description
*format_desc
;
666 format_desc
= util_format_description(format
);
668 brts
[i
].equation
= *blend
[i
].equation
.equation
;
670 /* TODO: this is a bit more complicated */
671 brts
[i
].constant
= blend
[i
].equation
.constant
;
673 brts
[i
].format
= panfrost_format_to_bifrost_blend(format_desc
);
675 /* 0x19 disables blending and forces REPLACE
676 * mode (equivalent to rgb_mode = alpha_mode =
677 * x122, colour mask = 0xF). 0x1a allows
679 brts
[i
].unk2
= blend
[i
].no_blending
? 0x19 : 0x1a;
681 brts
[i
].shader_type
= fs
->blend_types
[i
];
683 /* Dummy attachment for depth-only */
685 brts
[i
].shader_type
= fs
->blend_types
[i
];
688 struct midgard_blend_rt
*mrts
= rts
;
689 mrts
[i
].flags
= flags
;
691 if (blend
[i
].is_shader
) {
692 mrts
[i
].blend
.shader
= blend
[i
].shader
.gpu
| blend
[i
].shader
.first_tag
;
694 mrts
[i
].blend
.equation
= *blend
[i
].equation
.equation
;
695 mrts
[i
].blend
.constant
= blend
[i
].equation
.constant
;
702 panfrost_frag_shader_meta_init(struct panfrost_context
*ctx
,
703 struct mali_shader_meta
*fragmeta
,
706 const struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
707 struct panfrost_shader_state
*fs
;
709 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
711 bool msaa
= ctx
->rasterizer
->base
.multisample
;
712 fragmeta
->coverage_mask
= msaa
? ctx
->sample_mask
: ~0;
714 fragmeta
->unknown2_3
= MALI_DEPTH_FUNC(MALI_FUNC_ALWAYS
) | 0x10;
715 fragmeta
->unknown2_4
= 0x4e0;
717 /* unknown2_4 has 0x10 bit set on T6XX and T720. We don't know why this
718 * is required (independent of 32-bit/64-bit descriptors), or why it's
719 * not used on later GPU revisions. Otherwise, all shader jobs fault on
720 * these earlier chips (perhaps this is a chicken bit of some kind).
721 * More investigation is needed. */
723 SET_BIT(fragmeta
->unknown2_4
, 0x10, dev
->quirks
& MIDGARD_SFBD
);
725 if (dev
->quirks
& IS_BIFROST
) {
728 /* Depending on whether it's legal to in the given shader, we try to
729 * enable early-z testing. TODO: respect e-z force */
731 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_EARLY_Z
,
732 !fs
->can_discard
&& !fs
->writes_global
&&
733 !fs
->writes_depth
&& !fs
->writes_stencil
&&
734 !ctx
->blend
->base
.alpha_to_coverage
);
736 /* Add the writes Z/S flags if needed. */
737 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_WRITES_Z
, fs
->writes_depth
);
738 SET_BIT(fragmeta
->midgard1
.flags_hi
, MALI_WRITES_S
, fs
->writes_stencil
);
740 /* Any time texturing is used, derivatives are implicitly calculated,
741 * so we need to enable helper invocations */
743 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_HELPER_INVOCATIONS
,
744 fs
->helper_invocations
);
746 /* If discard is enabled, which bit we set to convey this
747 * depends on if depth/stencil is used for the draw or not.
748 * Just one of depth OR stencil is enough to trigger this. */
750 const struct pipe_depth_stencil_alpha_state
*zsa
= &ctx
->depth_stencil
->base
;
752 fs
->writes_depth
|| fs
->writes_stencil
||
753 (zsa
->depth
.enabled
&& zsa
->depth
.func
!= PIPE_FUNC_ALWAYS
) ||
754 zsa
->stencil
[0].enabled
;
756 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_READS_TILEBUFFER
,
757 fs
->outputs_read
|| (!zs_enabled
&& fs
->can_discard
));
758 SET_BIT(fragmeta
->midgard1
.flags_lo
, MALI_READS_ZS
, zs_enabled
&& fs
->can_discard
);
761 panfrost_frag_meta_rasterizer_update(ctx
, fragmeta
);
762 panfrost_frag_meta_zsa_update(ctx
, fragmeta
);
763 panfrost_frag_meta_blend_update(ctx
, fragmeta
, rts
);
767 panfrost_emit_shader_meta(struct panfrost_batch
*batch
,
768 enum pipe_shader_type st
,
769 struct mali_vertex_tiler_postfix
*postfix
)
771 struct panfrost_context
*ctx
= batch
->ctx
;
772 struct panfrost_shader_state
*ss
= panfrost_get_shader_state(ctx
, st
);
779 struct mali_shader_meta meta
;
781 panfrost_shader_meta_init(ctx
, st
, &meta
);
783 /* Add the shader BO to the batch. */
784 panfrost_batch_add_bo(batch
, ss
->bo
,
785 PAN_BO_ACCESS_PRIVATE
|
787 panfrost_bo_access_for_stage(st
));
791 if (st
== PIPE_SHADER_FRAGMENT
) {
792 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
793 unsigned rt_count
= MAX2(ctx
->pipe_framebuffer
.nr_cbufs
, 1);
794 size_t desc_size
= sizeof(meta
);
796 struct panfrost_transfer xfer
;
799 if (dev
->quirks
& MIDGARD_SFBD
)
801 else if (dev
->quirks
& IS_BIFROST
)
802 rt_size
= sizeof(struct bifrost_blend_rt
);
804 rt_size
= sizeof(struct midgard_blend_rt
);
806 desc_size
+= rt_size
* rt_count
;
809 rts
= rzalloc_size(ctx
, rt_size
* rt_count
);
811 panfrost_frag_shader_meta_init(ctx
, &meta
, rts
);
813 xfer
= panfrost_pool_alloc(&batch
->pool
, desc_size
);
815 memcpy(xfer
.cpu
, &meta
, sizeof(meta
));
816 memcpy(xfer
.cpu
+ sizeof(meta
), rts
, rt_size
* rt_count
);
821 shader_ptr
= xfer
.gpu
;
823 shader_ptr
= panfrost_pool_upload(&batch
->pool
, &meta
,
827 postfix
->shader
= shader_ptr
;
831 panfrost_emit_viewport(struct panfrost_batch
*batch
,
832 struct mali_vertex_tiler_postfix
*tiler_postfix
)
834 struct panfrost_context
*ctx
= batch
->ctx
;
835 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
836 const struct pipe_scissor_state
*ss
= &ctx
->scissor
;
837 const struct pipe_rasterizer_state
*rast
= &ctx
->rasterizer
->base
;
838 const struct pipe_framebuffer_state
*fb
= &ctx
->pipe_framebuffer
;
840 /* Derive min/max from translate/scale. Note since |x| >= 0 by
841 * definition, we have that -|x| <= |x| hence translate - |scale| <=
842 * translate + |scale|, so the ordering is correct here. */
843 float vp_minx
= (int) (vp
->translate
[0] - fabsf(vp
->scale
[0]));
844 float vp_maxx
= (int) (vp
->translate
[0] + fabsf(vp
->scale
[0]));
845 float vp_miny
= (int) (vp
->translate
[1] - fabsf(vp
->scale
[1]));
846 float vp_maxy
= (int) (vp
->translate
[1] + fabsf(vp
->scale
[1]));
847 float minz
= (vp
->translate
[2] - fabsf(vp
->scale
[2]));
848 float maxz
= (vp
->translate
[2] + fabsf(vp
->scale
[2]));
850 /* Scissor to the intersection of viewport and to the scissor, clamped
851 * to the framebuffer */
853 unsigned minx
= MIN2(fb
->width
, vp_minx
);
854 unsigned maxx
= MIN2(fb
->width
, vp_maxx
);
855 unsigned miny
= MIN2(fb
->height
, vp_miny
);
856 unsigned maxy
= MIN2(fb
->height
, vp_maxy
);
858 if (ss
&& rast
->scissor
) {
859 minx
= MAX2(ss
->minx
, minx
);
860 miny
= MAX2(ss
->miny
, miny
);
861 maxx
= MIN2(ss
->maxx
, maxx
);
862 maxy
= MIN2(ss
->maxy
, maxy
);
865 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
, MALI_VIEWPORT_LENGTH
);
867 pan_pack(T
.cpu
, VIEWPORT
, cfg
) {
868 cfg
.scissor_minimum_x
= minx
;
869 cfg
.scissor_minimum_y
= miny
;
870 cfg
.scissor_maximum_x
= maxx
- 1;
871 cfg
.scissor_maximum_y
= maxy
- 1;
873 cfg
.minimum_z
= rast
->depth_clip_near
? minz
: -INFINITY
;
874 cfg
.maximum_z
= rast
->depth_clip_far
? maxz
: INFINITY
;
877 tiler_postfix
->viewport
= T
.gpu
;
878 panfrost_batch_union_scissor(batch
, minx
, miny
, maxx
, maxy
);
882 panfrost_map_constant_buffer_gpu(struct panfrost_batch
*batch
,
883 enum pipe_shader_type st
,
884 struct panfrost_constant_buffer
*buf
,
887 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
888 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
891 panfrost_batch_add_bo(batch
, rsrc
->bo
,
892 PAN_BO_ACCESS_SHARED
|
894 panfrost_bo_access_for_stage(st
));
896 /* Alignment gauranteed by
897 * PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT */
898 return rsrc
->bo
->gpu
+ cb
->buffer_offset
;
899 } else if (cb
->user_buffer
) {
900 return panfrost_pool_upload(&batch
->pool
,
905 unreachable("No constant buffer");
909 struct sysval_uniform
{
919 panfrost_upload_viewport_scale_sysval(struct panfrost_batch
*batch
,
920 struct sysval_uniform
*uniform
)
922 struct panfrost_context
*ctx
= batch
->ctx
;
923 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
925 uniform
->f
[0] = vp
->scale
[0];
926 uniform
->f
[1] = vp
->scale
[1];
927 uniform
->f
[2] = vp
->scale
[2];
931 panfrost_upload_viewport_offset_sysval(struct panfrost_batch
*batch
,
932 struct sysval_uniform
*uniform
)
934 struct panfrost_context
*ctx
= batch
->ctx
;
935 const struct pipe_viewport_state
*vp
= &ctx
->pipe_viewport
;
937 uniform
->f
[0] = vp
->translate
[0];
938 uniform
->f
[1] = vp
->translate
[1];
939 uniform
->f
[2] = vp
->translate
[2];
942 static void panfrost_upload_txs_sysval(struct panfrost_batch
*batch
,
943 enum pipe_shader_type st
,
944 unsigned int sysvalid
,
945 struct sysval_uniform
*uniform
)
947 struct panfrost_context
*ctx
= batch
->ctx
;
948 unsigned texidx
= PAN_SYSVAL_ID_TO_TXS_TEX_IDX(sysvalid
);
949 unsigned dim
= PAN_SYSVAL_ID_TO_TXS_DIM(sysvalid
);
950 bool is_array
= PAN_SYSVAL_ID_TO_TXS_IS_ARRAY(sysvalid
);
951 struct pipe_sampler_view
*tex
= &ctx
->sampler_views
[st
][texidx
]->base
;
954 uniform
->i
[0] = u_minify(tex
->texture
->width0
, tex
->u
.tex
.first_level
);
957 uniform
->i
[1] = u_minify(tex
->texture
->height0
,
958 tex
->u
.tex
.first_level
);
961 uniform
->i
[2] = u_minify(tex
->texture
->depth0
,
962 tex
->u
.tex
.first_level
);
965 uniform
->i
[dim
] = tex
->texture
->array_size
;
969 panfrost_upload_ssbo_sysval(struct panfrost_batch
*batch
,
970 enum pipe_shader_type st
,
972 struct sysval_uniform
*uniform
)
974 struct panfrost_context
*ctx
= batch
->ctx
;
976 assert(ctx
->ssbo_mask
[st
] & (1 << ssbo_id
));
977 struct pipe_shader_buffer sb
= ctx
->ssbo
[st
][ssbo_id
];
979 /* Compute address */
980 struct panfrost_bo
*bo
= pan_resource(sb
.buffer
)->bo
;
982 panfrost_batch_add_bo(batch
, bo
,
983 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_RW
|
984 panfrost_bo_access_for_stage(st
));
986 /* Upload address and size as sysval */
987 uniform
->du
[0] = bo
->gpu
+ sb
.buffer_offset
;
988 uniform
->u
[2] = sb
.buffer_size
;
992 panfrost_upload_sampler_sysval(struct panfrost_batch
*batch
,
993 enum pipe_shader_type st
,
995 struct sysval_uniform
*uniform
)
997 struct panfrost_context
*ctx
= batch
->ctx
;
998 struct pipe_sampler_state
*sampl
= &ctx
->samplers
[st
][samp_idx
]->base
;
1000 uniform
->f
[0] = sampl
->min_lod
;
1001 uniform
->f
[1] = sampl
->max_lod
;
1002 uniform
->f
[2] = sampl
->lod_bias
;
1004 /* Even without any errata, Midgard represents "no mipmapping" as
1005 * fixing the LOD with the clamps; keep behaviour consistent. c.f.
1006 * panfrost_create_sampler_state which also explains our choice of
1007 * epsilon value (again to keep behaviour consistent) */
1009 if (sampl
->min_mip_filter
== PIPE_TEX_MIPFILTER_NONE
)
1010 uniform
->f
[1] = uniform
->f
[0] + (1.0/256.0);
1014 panfrost_upload_num_work_groups_sysval(struct panfrost_batch
*batch
,
1015 struct sysval_uniform
*uniform
)
1017 struct panfrost_context
*ctx
= batch
->ctx
;
1019 uniform
->u
[0] = ctx
->compute_grid
->grid
[0];
1020 uniform
->u
[1] = ctx
->compute_grid
->grid
[1];
1021 uniform
->u
[2] = ctx
->compute_grid
->grid
[2];
1025 panfrost_upload_sysvals(struct panfrost_batch
*batch
, void *buf
,
1026 struct panfrost_shader_state
*ss
,
1027 enum pipe_shader_type st
)
1029 struct sysval_uniform
*uniforms
= (void *)buf
;
1031 for (unsigned i
= 0; i
< ss
->sysval_count
; ++i
) {
1032 int sysval
= ss
->sysval
[i
];
1034 switch (PAN_SYSVAL_TYPE(sysval
)) {
1035 case PAN_SYSVAL_VIEWPORT_SCALE
:
1036 panfrost_upload_viewport_scale_sysval(batch
,
1039 case PAN_SYSVAL_VIEWPORT_OFFSET
:
1040 panfrost_upload_viewport_offset_sysval(batch
,
1043 case PAN_SYSVAL_TEXTURE_SIZE
:
1044 panfrost_upload_txs_sysval(batch
, st
,
1045 PAN_SYSVAL_ID(sysval
),
1048 case PAN_SYSVAL_SSBO
:
1049 panfrost_upload_ssbo_sysval(batch
, st
,
1050 PAN_SYSVAL_ID(sysval
),
1053 case PAN_SYSVAL_NUM_WORK_GROUPS
:
1054 panfrost_upload_num_work_groups_sysval(batch
,
1057 case PAN_SYSVAL_SAMPLER
:
1058 panfrost_upload_sampler_sysval(batch
, st
,
1059 PAN_SYSVAL_ID(sysval
),
1069 panfrost_map_constant_buffer_cpu(struct panfrost_constant_buffer
*buf
,
1072 struct pipe_constant_buffer
*cb
= &buf
->cb
[index
];
1073 struct panfrost_resource
*rsrc
= pan_resource(cb
->buffer
);
1076 return rsrc
->bo
->cpu
;
1077 else if (cb
->user_buffer
)
1078 return cb
->user_buffer
;
1080 unreachable("No constant buffer");
1084 panfrost_emit_const_buf(struct panfrost_batch
*batch
,
1085 enum pipe_shader_type stage
,
1086 struct mali_vertex_tiler_postfix
*postfix
)
1088 struct panfrost_context
*ctx
= batch
->ctx
;
1089 struct panfrost_shader_variants
*all
= ctx
->shader
[stage
];
1094 struct panfrost_constant_buffer
*buf
= &ctx
->constant_buffer
[stage
];
1096 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1098 /* Uniforms are implicitly UBO #0 */
1099 bool has_uniforms
= buf
->enabled_mask
& (1 << 0);
1101 /* Allocate room for the sysval and the uniforms */
1102 size_t sys_size
= sizeof(float) * 4 * ss
->sysval_count
;
1103 size_t uniform_size
= has_uniforms
? (buf
->cb
[0].buffer_size
) : 0;
1104 size_t size
= sys_size
+ uniform_size
;
1105 struct panfrost_transfer transfer
= panfrost_pool_alloc(&batch
->pool
,
1108 /* Upload sysvals requested by the shader */
1109 panfrost_upload_sysvals(batch
, transfer
.cpu
, ss
, stage
);
1111 /* Upload uniforms */
1112 if (has_uniforms
&& uniform_size
) {
1113 const void *cpu
= panfrost_map_constant_buffer_cpu(buf
, 0);
1114 memcpy(transfer
.cpu
+ sys_size
, cpu
, uniform_size
);
1117 /* Next up, attach UBOs. UBO #0 is the uniforms we just
1120 unsigned ubo_count
= panfrost_ubo_count(ctx
, stage
);
1121 assert(ubo_count
>= 1);
1123 size_t sz
= MALI_UNIFORM_BUFFER_LENGTH
* ubo_count
;
1124 struct panfrost_transfer ubos
= panfrost_pool_alloc(&batch
->pool
, sz
);
1125 uint64_t *ubo_ptr
= (uint64_t *) ubos
.cpu
;
1127 /* Upload uniforms as a UBO */
1129 if (ss
->uniform_count
) {
1130 pan_pack(ubo_ptr
, UNIFORM_BUFFER
, cfg
) {
1131 cfg
.entries
= ss
->uniform_count
;
1132 cfg
.pointer
= transfer
.gpu
;
1138 /* The rest are honest-to-goodness UBOs */
1140 for (unsigned ubo
= 1; ubo
< ubo_count
; ++ubo
) {
1141 size_t usz
= buf
->cb
[ubo
].buffer_size
;
1142 bool enabled
= buf
->enabled_mask
& (1 << ubo
);
1143 bool empty
= usz
== 0;
1145 if (!enabled
|| empty
) {
1150 pan_pack(ubo_ptr
+ ubo
, UNIFORM_BUFFER
, cfg
) {
1151 cfg
.entries
= DIV_ROUND_UP(usz
, 16);
1152 cfg
.pointer
= panfrost_map_constant_buffer_gpu(batch
,
1157 postfix
->uniforms
= transfer
.gpu
;
1158 postfix
->uniform_buffers
= ubos
.gpu
;
1160 buf
->dirty_mask
= 0;
1164 panfrost_emit_shared_memory(struct panfrost_batch
*batch
,
1165 const struct pipe_grid_info
*info
,
1166 struct midgard_payload_vertex_tiler
*vtp
)
1168 struct panfrost_context
*ctx
= batch
->ctx
;
1169 struct panfrost_shader_variants
*all
= ctx
->shader
[PIPE_SHADER_COMPUTE
];
1170 struct panfrost_shader_state
*ss
= &all
->variants
[all
->active_variant
];
1171 unsigned single_size
= util_next_power_of_two(MAX2(ss
->shared_size
,
1173 unsigned shared_size
= single_size
* info
->grid
[0] * info
->grid
[1] *
1175 struct panfrost_bo
*bo
= panfrost_batch_get_shared_memory(batch
,
1179 struct mali_shared_memory shared
= {
1180 .shared_memory
= bo
->gpu
,
1181 .shared_workgroup_count
=
1182 util_logbase2_ceil(info
->grid
[0]) +
1183 util_logbase2_ceil(info
->grid
[1]) +
1184 util_logbase2_ceil(info
->grid
[2]),
1186 .shared_shift
= util_logbase2(single_size
) - 1
1189 vtp
->postfix
.shared_memory
= panfrost_pool_upload(&batch
->pool
, &shared
,
1194 panfrost_get_tex_desc(struct panfrost_batch
*batch
,
1195 enum pipe_shader_type st
,
1196 struct panfrost_sampler_view
*view
)
1199 return (mali_ptr
) 0;
1201 struct pipe_sampler_view
*pview
= &view
->base
;
1202 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1204 /* Add the BO to the job so it's retained until the job is done. */
1206 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1207 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1208 panfrost_bo_access_for_stage(st
));
1210 panfrost_batch_add_bo(batch
, view
->bo
,
1211 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1212 panfrost_bo_access_for_stage(st
));
1214 return view
->bo
->gpu
;
1218 panfrost_update_sampler_view(struct panfrost_sampler_view
*view
,
1219 struct pipe_context
*pctx
)
1221 struct panfrost_resource
*rsrc
= pan_resource(view
->base
.texture
);
1222 if (view
->texture_bo
!= rsrc
->bo
->gpu
||
1223 view
->modifier
!= rsrc
->modifier
) {
1224 panfrost_bo_unreference(view
->bo
);
1225 panfrost_create_sampler_view_bo(view
, pctx
, &rsrc
->base
);
1230 panfrost_emit_texture_descriptors(struct panfrost_batch
*batch
,
1231 enum pipe_shader_type stage
,
1232 struct mali_vertex_tiler_postfix
*postfix
)
1234 struct panfrost_context
*ctx
= batch
->ctx
;
1235 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
1237 if (!ctx
->sampler_view_count
[stage
])
1240 if (device
->quirks
& IS_BIFROST
) {
1241 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
,
1242 MALI_BIFROST_TEXTURE_LENGTH
*
1243 ctx
->sampler_view_count
[stage
]);
1245 struct mali_bifrost_texture_packed
*out
=
1246 (struct mali_bifrost_texture_packed
*) T
.cpu
;
1248 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1249 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1250 struct pipe_sampler_view
*pview
= &view
->base
;
1251 struct panfrost_resource
*rsrc
= pan_resource(pview
->texture
);
1253 panfrost_update_sampler_view(view
, &ctx
->base
);
1254 out
[i
] = view
->bifrost_descriptor
;
1256 /* Add the BOs to the job so they are retained until the job is done. */
1258 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1259 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1260 panfrost_bo_access_for_stage(stage
));
1262 panfrost_batch_add_bo(batch
, view
->bo
,
1263 PAN_BO_ACCESS_SHARED
| PAN_BO_ACCESS_READ
|
1264 panfrost_bo_access_for_stage(stage
));
1267 postfix
->textures
= T
.gpu
;
1269 uint64_t trampolines
[PIPE_MAX_SHADER_SAMPLER_VIEWS
];
1271 for (int i
= 0; i
< ctx
->sampler_view_count
[stage
]; ++i
) {
1272 struct panfrost_sampler_view
*view
= ctx
->sampler_views
[stage
][i
];
1274 panfrost_update_sampler_view(view
, &ctx
->base
);
1276 trampolines
[i
] = panfrost_get_tex_desc(batch
, stage
, view
);
1279 postfix
->textures
= panfrost_pool_upload(&batch
->pool
,
1282 ctx
->sampler_view_count
[stage
]);
1287 panfrost_emit_sampler_descriptors(struct panfrost_batch
*batch
,
1288 enum pipe_shader_type stage
,
1289 struct mali_vertex_tiler_postfix
*postfix
)
1291 struct panfrost_context
*ctx
= batch
->ctx
;
1293 if (!ctx
->sampler_count
[stage
])
1296 size_t desc_size
= MALI_BIFROST_SAMPLER_LENGTH
;
1297 assert(MALI_BIFROST_SAMPLER_LENGTH
== MALI_MIDGARD_SAMPLER_LENGTH
);
1299 size_t sz
= desc_size
* ctx
->sampler_count
[stage
];
1300 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
, sz
);
1301 struct mali_midgard_sampler_packed
*out
= (struct mali_midgard_sampler_packed
*) T
.cpu
;
1303 for (unsigned i
= 0; i
< ctx
->sampler_count
[stage
]; ++i
)
1304 out
[i
] = ctx
->samplers
[stage
][i
]->hw
;
1306 postfix
->sampler_descriptor
= T
.gpu
;
1310 panfrost_emit_vertex_data(struct panfrost_batch
*batch
,
1311 struct mali_vertex_tiler_postfix
*vertex_postfix
)
1313 struct panfrost_context
*ctx
= batch
->ctx
;
1314 struct panfrost_vertex_state
*so
= ctx
->vertex
;
1316 unsigned instance_shift
= vertex_postfix
->instance_shift
;
1317 unsigned instance_odd
= vertex_postfix
->instance_odd
;
1319 /* Worst case: everything is NPOT */
1321 struct panfrost_transfer S
= panfrost_pool_alloc(&batch
->pool
,
1322 MALI_ATTRIBUTE_LENGTH
* PIPE_MAX_ATTRIBS
* 2);
1324 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
,
1325 MALI_ATTRIBUTE_LENGTH
* (PAN_INSTANCE_ID
+ 1));
1327 struct mali_attribute_buffer_packed
*bufs
=
1328 (struct mali_attribute_buffer_packed
*) S
.cpu
;
1330 struct mali_attribute_packed
*out
=
1331 (struct mali_attribute_packed
*) T
.cpu
;
1333 unsigned attrib_to_buffer
[PIPE_MAX_ATTRIBS
] = { 0 };
1336 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1337 /* We map buffers 1:1 with the attributes, which
1338 * means duplicating some vertex buffers (who cares? aside from
1339 * maybe some caching implications but I somehow doubt that
1342 struct pipe_vertex_element
*elem
= &so
->pipe
[i
];
1343 unsigned vbi
= elem
->vertex_buffer_index
;
1344 attrib_to_buffer
[i
] = k
;
1346 if (!(ctx
->vb_mask
& (1 << vbi
)))
1349 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1350 struct panfrost_resource
*rsrc
;
1352 rsrc
= pan_resource(buf
->buffer
.resource
);
1356 /* Add a dependency of the batch on the vertex buffer */
1357 panfrost_batch_add_bo(batch
, rsrc
->bo
,
1358 PAN_BO_ACCESS_SHARED
|
1359 PAN_BO_ACCESS_READ
|
1360 PAN_BO_ACCESS_VERTEX_TILER
);
1362 /* Mask off lower bits, see offset fixup below */
1363 mali_ptr raw_addr
= rsrc
->bo
->gpu
+ buf
->buffer_offset
;
1364 mali_ptr addr
= raw_addr
& ~63;
1366 /* Since we advanced the base pointer, we shrink the buffer
1367 * size, but add the offset we subtracted */
1368 unsigned size
= rsrc
->base
.width0
+ (raw_addr
- addr
)
1369 - buf
->buffer_offset
;
1371 /* When there is a divisor, the hardware-level divisor is
1372 * the product of the instance divisor and the padded count */
1373 unsigned divisor
= elem
->instance_divisor
;
1374 unsigned hw_divisor
= ctx
->padded_count
* divisor
;
1375 unsigned stride
= buf
->stride
;
1377 /* If there's a divisor(=1) but no instancing, we want every
1378 * attribute to be the same */
1380 if (divisor
&& ctx
->instance_count
== 1)
1383 if (!divisor
|| ctx
->instance_count
<= 1) {
1384 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1385 if (ctx
->instance_count
> 1)
1386 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_MODULUS
;
1389 cfg
.stride
= stride
;
1391 cfg
.divisor_r
= instance_shift
;
1392 cfg
.divisor_p
= instance_odd
;
1394 } else if (util_is_power_of_two_or_zero(hw_divisor
)) {
1395 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1396 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_POT_DIVISOR
;
1398 cfg
.stride
= stride
;
1400 cfg
.divisor_r
= __builtin_ctz(hw_divisor
);
1404 unsigned shift
= 0, extra_flags
= 0;
1406 unsigned magic_divisor
=
1407 panfrost_compute_magic_divisor(hw_divisor
, &shift
, &extra_flags
);
1409 pan_pack(bufs
+ k
, ATTRIBUTE_BUFFER
, cfg
) {
1410 cfg
.type
= MALI_ATTRIBUTE_TYPE_1D_NPOT_DIVISOR
;
1412 cfg
.stride
= stride
;
1415 cfg
.divisor_r
= shift
;
1416 cfg
.divisor_e
= extra_flags
;
1419 pan_pack(bufs
+ k
+ 1, ATTRIBUTE_BUFFER_CONTINUATION_NPOT
, cfg
) {
1420 cfg
.divisor_numerator
= magic_divisor
;
1421 cfg
.divisor
= divisor
;
1430 /* Add special gl_VertexID/gl_InstanceID buffers */
1432 panfrost_vertex_id(ctx
->padded_count
, &bufs
[k
], ctx
->instance_count
> 1);
1434 pan_pack(out
+ PAN_VERTEX_ID
, ATTRIBUTE
, cfg
) {
1435 cfg
.buffer_index
= k
++;
1436 cfg
.format
= so
->formats
[PAN_VERTEX_ID
];
1439 panfrost_instance_id(ctx
->padded_count
, &bufs
[k
], ctx
->instance_count
> 1);
1441 pan_pack(out
+ PAN_INSTANCE_ID
, ATTRIBUTE
, cfg
) {
1442 cfg
.buffer_index
= k
++;
1443 cfg
.format
= so
->formats
[PAN_INSTANCE_ID
];
1446 /* Attribute addresses require 64-byte alignment, so let:
1448 * base' = base & ~63 = base - (base & 63)
1449 * offset' = offset + (base & 63)
1451 * Since base' + offset' = base + offset, these are equivalent
1452 * addressing modes and now base is 64 aligned.
1455 unsigned start
= vertex_postfix
->offset_start
;
1457 for (unsigned i
= 0; i
< so
->num_elements
; ++i
) {
1458 unsigned vbi
= so
->pipe
[i
].vertex_buffer_index
;
1459 struct pipe_vertex_buffer
*buf
= &ctx
->vertex_buffers
[vbi
];
1461 /* Adjust by the masked off bits of the offset. Make sure we
1462 * read src_offset from so->hw (which is not GPU visible)
1463 * rather than target (which is) due to caching effects */
1465 unsigned src_offset
= so
->pipe
[i
].src_offset
;
1467 /* BOs aligned to 4k so guaranteed aligned to 64 */
1468 src_offset
+= (buf
->buffer_offset
& 63);
1470 /* Also, somewhat obscurely per-instance data needs to be
1471 * offset in response to a delayed start in an indexed draw */
1473 if (so
->pipe
[i
].instance_divisor
&& ctx
->instance_count
> 1 && start
)
1474 src_offset
-= buf
->stride
* start
;
1476 pan_pack(out
+ i
, ATTRIBUTE
, cfg
) {
1477 cfg
.buffer_index
= attrib_to_buffer
[i
];
1478 cfg
.format
= so
->formats
[i
];
1479 cfg
.offset
= src_offset
;
1483 vertex_postfix
->attributes
= S
.gpu
;
1484 vertex_postfix
->attribute_meta
= T
.gpu
;
1488 panfrost_emit_varyings(struct panfrost_batch
*batch
,
1489 struct mali_attribute_buffer_packed
*slot
,
1490 unsigned stride
, unsigned count
)
1492 unsigned size
= stride
* count
;
1493 mali_ptr ptr
= panfrost_pool_alloc(&batch
->pool
, size
).gpu
;
1495 pan_pack(slot
, ATTRIBUTE_BUFFER
, cfg
) {
1496 cfg
.stride
= stride
;
1505 panfrost_streamout_offset(unsigned stride
, unsigned offset
,
1506 struct pipe_stream_output_target
*target
)
1508 return (target
->buffer_offset
+ (offset
* stride
* 4)) & 63;
1512 panfrost_emit_streamout(struct panfrost_batch
*batch
,
1513 struct mali_attribute_buffer_packed
*slot
,
1514 unsigned stride_words
, unsigned offset
, unsigned count
,
1515 struct pipe_stream_output_target
*target
)
1517 unsigned stride
= stride_words
* 4;
1518 unsigned max_size
= target
->buffer_size
;
1519 unsigned expected_size
= stride
* count
;
1521 /* Grab the BO and bind it to the batch */
1522 struct panfrost_bo
*bo
= pan_resource(target
->buffer
)->bo
;
1524 /* Varyings are WRITE from the perspective of the VERTEX but READ from
1525 * the perspective of the TILER and FRAGMENT.
1527 panfrost_batch_add_bo(batch
, bo
,
1528 PAN_BO_ACCESS_SHARED
|
1530 PAN_BO_ACCESS_VERTEX_TILER
|
1531 PAN_BO_ACCESS_FRAGMENT
);
1533 /* We will have an offset applied to get alignment */
1534 mali_ptr addr
= bo
->gpu
+ target
->buffer_offset
+ (offset
* stride
);
1536 pan_pack(slot
, ATTRIBUTE_BUFFER
, cfg
) {
1537 cfg
.pointer
= (addr
& ~63);
1538 cfg
.stride
= stride
;
1539 cfg
.size
= MIN2(max_size
, expected_size
) + (addr
& 63);
1544 has_point_coord(unsigned mask
, gl_varying_slot loc
)
1546 if ((loc
>= VARYING_SLOT_TEX0
) && (loc
<= VARYING_SLOT_TEX7
))
1547 return (mask
& (1 << (loc
- VARYING_SLOT_TEX0
)));
1548 else if (loc
== VARYING_SLOT_PNTC
)
1549 return (mask
& (1 << 8));
1554 /* Helpers for manipulating stream out information so we can pack varyings
1555 * accordingly. Compute the src_offset for a given captured varying */
1557 static struct pipe_stream_output
*
1558 pan_get_so(struct pipe_stream_output_info
*info
, gl_varying_slot loc
)
1560 for (unsigned i
= 0; i
< info
->num_outputs
; ++i
) {
1561 if (info
->output
[i
].register_index
== loc
)
1562 return &info
->output
[i
];
1565 unreachable("Varying not captured");
1569 pan_varying_size(enum mali_format fmt
)
1571 unsigned type
= MALI_EXTRACT_TYPE(fmt
);
1572 unsigned chan
= MALI_EXTRACT_CHANNELS(fmt
);
1573 unsigned bits
= MALI_EXTRACT_BITS(fmt
);
1576 if (bits
== MALI_CHANNEL_FLOAT
) {
1578 bool fp16
= (type
== MALI_FORMAT_SINT
);
1579 assert(fp16
|| (type
== MALI_FORMAT_UNORM
));
1583 assert(type
>= MALI_FORMAT_SNORM
&& type
<= MALI_FORMAT_SINT
);
1594 /* Indices for named (non-XFB) varyings that are present. These are packed
1595 * tightly so they correspond to a bitfield present (P) indexed by (1 <<
1596 * PAN_VARY_*). This has the nice property that you can lookup the buffer index
1597 * of a given special field given a shift S by:
1599 * idx = popcount(P & ((1 << S) - 1))
1601 * That is... look at all of the varyings that come earlier and count them, the
1602 * count is the new index since plus one. Likewise, the total number of special
1603 * buffers required is simply popcount(P)
1606 enum pan_special_varying
{
1607 PAN_VARY_GENERAL
= 0,
1608 PAN_VARY_POSITION
= 1,
1610 PAN_VARY_PNTCOORD
= 3,
1612 PAN_VARY_FRAGCOORD
= 5,
1618 /* Given a varying, figure out which index it correpsonds to */
1620 static inline unsigned
1621 pan_varying_index(unsigned present
, enum pan_special_varying v
)
1623 unsigned mask
= (1 << v
) - 1;
1624 return util_bitcount(present
& mask
);
1627 /* Get the base offset for XFB buffers, which by convention come after
1628 * everything else. Wrapper function for semantic reasons; by construction this
1629 * is just popcount. */
1631 static inline unsigned
1632 pan_xfb_base(unsigned present
)
1634 return util_bitcount(present
);
1637 /* Computes the present mask for varyings so we can start emitting varying records */
1639 static inline unsigned
1640 pan_varying_present(
1641 struct panfrost_shader_state
*vs
,
1642 struct panfrost_shader_state
*fs
,
1645 /* At the moment we always emit general and position buffers. Not
1646 * strictly necessary but usually harmless */
1648 unsigned present
= (1 << PAN_VARY_GENERAL
) | (1 << PAN_VARY_POSITION
);
1650 /* Enable special buffers by the shader info */
1652 if (vs
->writes_point_size
)
1653 present
|= (1 << PAN_VARY_PSIZ
);
1655 if (fs
->reads_point_coord
)
1656 present
|= (1 << PAN_VARY_PNTCOORD
);
1659 present
|= (1 << PAN_VARY_FACE
);
1661 if (fs
->reads_frag_coord
&& !(quirks
& IS_BIFROST
))
1662 present
|= (1 << PAN_VARY_FRAGCOORD
);
1664 /* Also, if we have a point sprite, we need a point coord buffer */
1666 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1667 gl_varying_slot loc
= fs
->varyings_loc
[i
];
1669 if (has_point_coord(fs
->point_sprite_mask
, loc
))
1670 present
|= (1 << PAN_VARY_PNTCOORD
);
1676 /* Emitters for varying records */
1679 pan_emit_vary(struct mali_attribute_packed
*out
,
1680 unsigned present
, enum pan_special_varying buf
,
1681 unsigned quirks
, enum mali_format format
,
1684 unsigned nr_channels
= MALI_EXTRACT_CHANNELS(format
);
1685 unsigned swizzle
= quirks
& HAS_SWIZZLES
?
1686 panfrost_get_default_swizzle(nr_channels
) :
1687 panfrost_bifrost_swizzle(nr_channels
);
1689 pan_pack(out
, ATTRIBUTE
, cfg
) {
1690 cfg
.buffer_index
= pan_varying_index(present
, buf
);
1691 cfg
.unknown
= quirks
& IS_BIFROST
? 0x0 : 0x1;
1692 cfg
.format
= (format
<< 12) | swizzle
;
1693 cfg
.offset
= offset
;
1697 /* General varying that is unused */
1700 pan_emit_vary_only(struct mali_attribute_packed
*out
,
1701 unsigned present
, unsigned quirks
)
1703 pan_emit_vary(out
, present
, 0, quirks
, MALI_VARYING_DISCARD
, 0);
1706 /* Special records */
1708 static const enum mali_format pan_varying_formats
[PAN_VARY_MAX
] = {
1709 [PAN_VARY_POSITION
] = MALI_VARYING_POS
,
1710 [PAN_VARY_PSIZ
] = MALI_R16F
,
1711 [PAN_VARY_PNTCOORD
] = MALI_R16F
,
1712 [PAN_VARY_FACE
] = MALI_R32I
,
1713 [PAN_VARY_FRAGCOORD
] = MALI_RGBA32F
1717 pan_emit_vary_special(struct mali_attribute_packed
*out
,
1718 unsigned present
, enum pan_special_varying buf
,
1721 assert(buf
< PAN_VARY_MAX
);
1722 pan_emit_vary(out
, present
, buf
, quirks
, pan_varying_formats
[buf
], 0);
1725 static enum mali_format
1726 pan_xfb_format(enum mali_format format
, unsigned nr
)
1728 if (MALI_EXTRACT_BITS(format
) == MALI_CHANNEL_FLOAT
)
1729 return MALI_R32F
| MALI_NR_CHANNELS(nr
);
1731 return MALI_EXTRACT_TYPE(format
) | MALI_NR_CHANNELS(nr
) | MALI_CHANNEL_32
;
1734 /* Transform feedback records. Note struct pipe_stream_output is (if packed as
1735 * a bitfield) 32-bit, smaller than a 64-bit pointer, so may as well pass by
1739 pan_emit_vary_xfb(struct mali_attribute_packed
*out
,
1742 unsigned *streamout_offsets
,
1744 enum mali_format format
,
1745 struct pipe_stream_output o
)
1747 unsigned swizzle
= quirks
& HAS_SWIZZLES
?
1748 panfrost_get_default_swizzle(o
.num_components
) :
1749 panfrost_bifrost_swizzle(o
.num_components
);
1751 pan_pack(out
, ATTRIBUTE
, cfg
) {
1752 /* XFB buffers come after everything else */
1753 cfg
.buffer_index
= pan_xfb_base(present
) + o
.output_buffer
;
1754 cfg
.unknown
= quirks
& IS_BIFROST
? 0x0 : 0x1;
1756 /* Override number of channels and precision to highp */
1757 cfg
.format
= (pan_xfb_format(format
, o
.num_components
) << 12) | swizzle
;
1759 /* Apply given offsets together */
1760 cfg
.offset
= (o
.dst_offset
* 4) /* dwords */
1761 + streamout_offsets
[o
.output_buffer
];
1765 /* Determine if we should capture a varying for XFB. This requires actually
1766 * having a buffer for it. If we don't capture it, we'll fallback to a general
1767 * varying path (linked or unlinked, possibly discarding the write) */
1770 panfrost_xfb_captured(struct panfrost_shader_state
*xfb
,
1771 unsigned loc
, unsigned max_xfb
)
1773 if (!(xfb
->so_mask
& (1ll << loc
)))
1776 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1777 return o
->output_buffer
< max_xfb
;
1781 pan_emit_general_varying(struct mali_attribute_packed
*out
,
1782 struct panfrost_shader_state
*other
,
1783 struct panfrost_shader_state
*xfb
,
1784 gl_varying_slot loc
,
1785 enum mali_format format
,
1788 unsigned *gen_offsets
,
1789 enum mali_format
*gen_formats
,
1790 unsigned *gen_stride
,
1794 /* Check if we're linked */
1795 signed other_idx
= -1;
1797 for (unsigned j
= 0; j
< other
->varying_count
; ++j
) {
1798 if (other
->varyings_loc
[j
] == loc
) {
1804 if (other_idx
< 0) {
1805 pan_emit_vary_only(out
, present
, quirks
);
1809 unsigned offset
= gen_offsets
[other_idx
];
1812 /* We're linked, so allocate a space via a watermark allocation */
1813 enum mali_format alt
= other
->varyings
[other_idx
];
1815 /* Do interpolation at minimum precision */
1816 unsigned size_main
= pan_varying_size(format
);
1817 unsigned size_alt
= pan_varying_size(alt
);
1818 unsigned size
= MIN2(size_main
, size_alt
);
1820 /* If a varying is marked for XFB but not actually captured, we
1821 * should match the format to the format that would otherwise
1822 * be used for XFB, since dEQP checks for invariance here. It's
1823 * unclear if this is required by the spec. */
1825 if (xfb
->so_mask
& (1ull << loc
)) {
1826 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1827 format
= pan_xfb_format(format
, o
->num_components
);
1828 size
= pan_varying_size(format
);
1829 } else if (size
== size_alt
) {
1833 gen_offsets
[idx
] = *gen_stride
;
1834 gen_formats
[other_idx
] = format
;
1835 offset
= *gen_stride
;
1836 *gen_stride
+= size
;
1839 pan_emit_vary(out
, present
, PAN_VARY_GENERAL
, quirks
, format
, offset
);
1842 /* Higher-level wrapper around all of the above, classifying a varying into one
1843 * of the above types */
1846 panfrost_emit_varying(
1847 struct mali_attribute_packed
*out
,
1848 struct panfrost_shader_state
*stage
,
1849 struct panfrost_shader_state
*other
,
1850 struct panfrost_shader_state
*xfb
,
1853 unsigned *streamout_offsets
,
1855 unsigned *gen_offsets
,
1856 enum mali_format
*gen_formats
,
1857 unsigned *gen_stride
,
1862 gl_varying_slot loc
= stage
->varyings_loc
[idx
];
1863 enum mali_format format
= stage
->varyings
[idx
];
1865 /* Override format to match linkage */
1866 if (!should_alloc
&& gen_formats
[idx
])
1867 format
= gen_formats
[idx
];
1869 if (has_point_coord(stage
->point_sprite_mask
, loc
)) {
1870 pan_emit_vary_special(out
, present
, PAN_VARY_PNTCOORD
, quirks
);
1871 } else if (panfrost_xfb_captured(xfb
, loc
, max_xfb
)) {
1872 struct pipe_stream_output
*o
= pan_get_so(&xfb
->stream_output
, loc
);
1873 pan_emit_vary_xfb(out
, present
, max_xfb
, streamout_offsets
, quirks
, format
, *o
);
1874 } else if (loc
== VARYING_SLOT_POS
) {
1876 pan_emit_vary_special(out
, present
, PAN_VARY_FRAGCOORD
, quirks
);
1878 pan_emit_vary_special(out
, present
, PAN_VARY_POSITION
, quirks
);
1879 } else if (loc
== VARYING_SLOT_PSIZ
) {
1880 pan_emit_vary_special(out
, present
, PAN_VARY_PSIZ
, quirks
);
1881 } else if (loc
== VARYING_SLOT_PNTC
) {
1882 pan_emit_vary_special(out
, present
, PAN_VARY_PNTCOORD
, quirks
);
1883 } else if (loc
== VARYING_SLOT_FACE
) {
1884 pan_emit_vary_special(out
, present
, PAN_VARY_FACE
, quirks
);
1886 pan_emit_general_varying(out
, other
, xfb
, loc
, format
, present
,
1887 quirks
, gen_offsets
, gen_formats
, gen_stride
,
1893 pan_emit_special_input(struct mali_attribute_buffer_packed
*out
,
1895 enum pan_special_varying v
,
1898 if (present
& (1 << v
)) {
1899 unsigned idx
= pan_varying_index(present
, v
);
1901 pan_pack(out
+ idx
, ATTRIBUTE_BUFFER
, cfg
) {
1902 cfg
.special
= special
;
1909 panfrost_emit_varying_descriptor(struct panfrost_batch
*batch
,
1910 unsigned vertex_count
,
1911 struct mali_vertex_tiler_postfix
*vertex_postfix
,
1912 struct mali_vertex_tiler_postfix
*tiler_postfix
,
1913 union midgard_primitive_size
*primitive_size
)
1915 /* Load the shaders */
1916 struct panfrost_context
*ctx
= batch
->ctx
;
1917 struct panfrost_device
*dev
= pan_device(ctx
->base
.screen
);
1918 struct panfrost_shader_state
*vs
, *fs
;
1919 size_t vs_size
, fs_size
;
1921 /* Allocate the varying descriptor */
1923 vs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_VERTEX
);
1924 fs
= panfrost_get_shader_state(ctx
, PIPE_SHADER_FRAGMENT
);
1925 vs_size
= MALI_ATTRIBUTE_LENGTH
* vs
->varying_count
;
1926 fs_size
= MALI_ATTRIBUTE_LENGTH
* fs
->varying_count
;
1928 struct panfrost_transfer trans
= panfrost_pool_alloc(&batch
->pool
,
1932 struct pipe_stream_output_info
*so
= &vs
->stream_output
;
1933 unsigned present
= pan_varying_present(vs
, fs
, dev
->quirks
);
1935 /* Check if this varying is linked by us. This is the case for
1936 * general-purpose, non-captured varyings. If it is, link it. If it's
1937 * not, use the provided stream out information to determine the
1938 * offset, since it was already linked for us. */
1940 unsigned gen_offsets
[32];
1941 enum mali_format gen_formats
[32];
1942 memset(gen_offsets
, 0, sizeof(gen_offsets
));
1943 memset(gen_formats
, 0, sizeof(gen_formats
));
1945 unsigned gen_stride
= 0;
1946 assert(vs
->varying_count
< ARRAY_SIZE(gen_offsets
));
1947 assert(fs
->varying_count
< ARRAY_SIZE(gen_offsets
));
1949 unsigned streamout_offsets
[32];
1951 for (unsigned i
= 0; i
< ctx
->streamout
.num_targets
; ++i
) {
1952 streamout_offsets
[i
] = panfrost_streamout_offset(
1954 ctx
->streamout
.offsets
[i
],
1955 ctx
->streamout
.targets
[i
]);
1958 struct mali_attribute_packed
*ovs
= (struct mali_attribute_packed
*)trans
.cpu
;
1959 struct mali_attribute_packed
*ofs
= ovs
+ vs
->varying_count
;
1961 for (unsigned i
= 0; i
< vs
->varying_count
; i
++) {
1962 panfrost_emit_varying(ovs
+ i
, vs
, fs
, vs
, present
,
1963 ctx
->streamout
.num_targets
, streamout_offsets
,
1965 gen_offsets
, gen_formats
, &gen_stride
, i
, true, false);
1968 for (unsigned i
= 0; i
< fs
->varying_count
; i
++) {
1969 panfrost_emit_varying(ofs
+ i
, fs
, vs
, vs
, present
,
1970 ctx
->streamout
.num_targets
, streamout_offsets
,
1972 gen_offsets
, gen_formats
, &gen_stride
, i
, false, true);
1975 unsigned xfb_base
= pan_xfb_base(present
);
1976 struct panfrost_transfer T
= panfrost_pool_alloc(&batch
->pool
,
1977 MALI_ATTRIBUTE_BUFFER_LENGTH
* (xfb_base
+ ctx
->streamout
.num_targets
));
1978 struct mali_attribute_buffer_packed
*varyings
=
1979 (struct mali_attribute_buffer_packed
*) T
.cpu
;
1981 /* Emit the stream out buffers */
1983 unsigned out_count
= u_stream_outputs_for_vertices(ctx
->active_prim
,
1986 for (unsigned i
= 0; i
< ctx
->streamout
.num_targets
; ++i
) {
1987 panfrost_emit_streamout(batch
, &varyings
[xfb_base
+ i
],
1989 ctx
->streamout
.offsets
[i
],
1991 ctx
->streamout
.targets
[i
]);
1994 panfrost_emit_varyings(batch
,
1995 &varyings
[pan_varying_index(present
, PAN_VARY_GENERAL
)],
1996 gen_stride
, vertex_count
);
1998 /* fp32 vec4 gl_Position */
1999 tiler_postfix
->position_varying
= panfrost_emit_varyings(batch
,
2000 &varyings
[pan_varying_index(present
, PAN_VARY_POSITION
)],
2001 sizeof(float) * 4, vertex_count
);
2003 if (present
& (1 << PAN_VARY_PSIZ
)) {
2004 primitive_size
->pointer
= panfrost_emit_varyings(batch
,
2005 &varyings
[pan_varying_index(present
, PAN_VARY_PSIZ
)],
2009 pan_emit_special_input(varyings
, present
, PAN_VARY_PNTCOORD
, MALI_ATTRIBUTE_SPECIAL_POINT_COORD
);
2010 pan_emit_special_input(varyings
, present
, PAN_VARY_FACE
, MALI_ATTRIBUTE_SPECIAL_FRONT_FACING
);
2011 pan_emit_special_input(varyings
, present
, PAN_VARY_FRAGCOORD
, MALI_ATTRIBUTE_SPECIAL_FRAG_COORD
);
2013 vertex_postfix
->varyings
= T
.gpu
;
2014 tiler_postfix
->varyings
= T
.gpu
;
2016 vertex_postfix
->varying_meta
= trans
.gpu
;
2017 tiler_postfix
->varying_meta
= trans
.gpu
+ vs_size
;
2021 panfrost_emit_vertex_tiler_jobs(struct panfrost_batch
*batch
,
2022 struct mali_vertex_tiler_prefix
*vertex_prefix
,
2023 struct mali_vertex_tiler_postfix
*vertex_postfix
,
2024 struct mali_vertex_tiler_prefix
*tiler_prefix
,
2025 struct mali_vertex_tiler_postfix
*tiler_postfix
,
2026 union midgard_primitive_size
*primitive_size
)
2028 struct panfrost_context
*ctx
= batch
->ctx
;
2029 struct panfrost_device
*device
= pan_device(ctx
->base
.screen
);
2030 bool wallpapering
= ctx
->wallpaper_batch
&& batch
->scoreboard
.tiler_dep
;
2031 struct bifrost_payload_vertex bifrost_vertex
= {0,};
2032 struct bifrost_payload_tiler bifrost_tiler
= {0,};
2033 struct midgard_payload_vertex_tiler midgard_vertex
= {0,};
2034 struct midgard_payload_vertex_tiler midgard_tiler
= {0,};
2036 size_t vp_size
, tp_size
;
2038 if (device
->quirks
& IS_BIFROST
) {
2039 bifrost_vertex
.prefix
= *vertex_prefix
;
2040 bifrost_vertex
.postfix
= *vertex_postfix
;
2041 vp
= &bifrost_vertex
;
2042 vp_size
= sizeof(bifrost_vertex
);
2044 bifrost_tiler
.prefix
= *tiler_prefix
;
2045 bifrost_tiler
.tiler
.primitive_size
= *primitive_size
;
2046 bifrost_tiler
.tiler
.tiler_meta
= panfrost_batch_get_tiler_meta(batch
, ~0);
2047 bifrost_tiler
.postfix
= *tiler_postfix
;
2048 tp
= &bifrost_tiler
;
2049 tp_size
= sizeof(bifrost_tiler
);
2051 midgard_vertex
.prefix
= *vertex_prefix
;
2052 midgard_vertex
.postfix
= *vertex_postfix
;
2053 vp
= &midgard_vertex
;
2054 vp_size
= sizeof(midgard_vertex
);
2056 midgard_tiler
.prefix
= *tiler_prefix
;
2057 midgard_tiler
.postfix
= *tiler_postfix
;
2058 midgard_tiler
.primitive_size
= *primitive_size
;
2059 tp
= &midgard_tiler
;
2060 tp_size
= sizeof(midgard_tiler
);
2064 /* Inject in reverse order, with "predicted" job indices.
2065 * THIS IS A HACK XXX */
2066 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_TILER
, false,
2067 batch
->scoreboard
.job_index
+ 2, tp
, tp_size
, true);
2068 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_VERTEX
, false, 0,
2073 /* If rasterizer discard is enable, only submit the vertex */
2075 unsigned vertex
= panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_VERTEX
, false, 0,
2076 vp
, vp_size
, false);
2078 if (ctx
->rasterizer
->base
.rasterizer_discard
)
2081 panfrost_new_job(&batch
->pool
, &batch
->scoreboard
, MALI_JOB_TYPE_TILER
, false, vertex
, tp
, tp_size
,
2085 /* TODO: stop hardcoding this */
2087 panfrost_emit_sample_locations(struct panfrost_batch
*batch
)
2089 uint16_t locations
[] = {
2140 return panfrost_pool_upload(&batch
->pool
, locations
, 96 * sizeof(uint16_t));