2 * Copyright 2013 Advanced Micro Devices, Inc.
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
9 * license, and/or sell copies of the Software, and to permit persons to whom
10 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22 * USE OR OTHER DEALINGS IN THE SOFTWARE.
25 /* Resource binding slots and sampler states (each described with 8 or
26 * 4 dwords) are stored in lists in memory which is accessed by shaders
27 * using scalar load instructions.
29 * This file is responsible for managing such lists. It keeps a copy of all
30 * descriptors in CPU memory and re-uploads a whole list if some slots have
33 * This code is also reponsible for updating shader pointers to those lists.
35 * Note that CP DMA can't be used for updating the lists, because a GPU hang
36 * could leave the list in a mid-IB state and the next IB would get wrong
37 * descriptors and the whole context would be unusable at that point.
38 * (Note: The register shadowing can't be used due to the same reason)
40 * Also, uploading descriptors to newly allocated memory doesn't require
44 * Possible scenarios for one 16 dword image+sampler slot:
46 * | Image | w/ FMASK | Buffer | NULL
47 * [ 0: 3] Image[0:3] | Image[0:3] | Null[0:3] | Null[0:3]
48 * [ 4: 7] Image[4:7] | Image[4:7] | Buffer[0:3] | 0
49 * [ 8:11] Null[0:3] | Fmask[0:3] | Null[0:3] | Null[0:3]
50 * [12:15] Sampler[0:3] | Fmask[4:7] | Sampler[0:3] | Sampler[0:3]
52 * FMASK implies MSAA, therefore no sampler state.
53 * Sampler states are never unbound except when FMASK is bound.
59 #include "util/hash_table.h"
60 #include "util/u_idalloc.h"
61 #include "util/u_format.h"
62 #include "util/u_memory.h"
63 #include "util/u_upload_mgr.h"
66 /* NULL image and buffer descriptor for textures (alpha = 1) and images
69 * For images, all fields must be zero except for the swizzle, which
70 * supports arbitrary combinations of 0s and 1s. The texture type must be
71 * any valid type (e.g. 1D). If the texture type isn't set, the hw hangs.
73 * For buffers, all fields must be zero. If they are not, the hw hangs.
75 * This is the only reason why the buffer descriptor must be in words [4:7].
77 static uint32_t null_texture_descriptor
[8] = {
81 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_1
) |
82 S_008F1C_TYPE(V_008F1C_SQ_RSRC_IMG_1D
)
83 /* the rest must contain zeros, which is also used by the buffer
87 static uint32_t null_image_descriptor
[8] = {
91 S_008F1C_TYPE(V_008F1C_SQ_RSRC_IMG_1D
)
92 /* the rest must contain zeros, which is also used by the buffer
96 static uint64_t si_desc_extract_buffer_address(const uint32_t *desc
)
98 uint64_t va
= desc
[0] |
99 ((uint64_t)G_008F04_BASE_ADDRESS_HI(desc
[1]) << 32);
101 /* Sign-extend the 48-bit address. */
103 va
= (int64_t)va
>> 16;
107 static void si_init_descriptor_list(uint32_t *desc_list
,
108 unsigned element_dw_size
,
109 unsigned num_elements
,
110 const uint32_t *null_descriptor
)
114 /* Initialize the array to NULL descriptors if the element size is 8. */
115 if (null_descriptor
) {
116 assert(element_dw_size
% 8 == 0);
117 for (i
= 0; i
< num_elements
* element_dw_size
/ 8; i
++)
118 memcpy(desc_list
+ i
* 8, null_descriptor
, 8 * 4);
122 static void si_init_descriptors(struct si_descriptors
*desc
,
123 short shader_userdata_rel_index
,
124 unsigned element_dw_size
,
125 unsigned num_elements
)
127 desc
->list
= CALLOC(num_elements
, element_dw_size
* 4);
128 desc
->element_dw_size
= element_dw_size
;
129 desc
->num_elements
= num_elements
;
130 desc
->shader_userdata_offset
= shader_userdata_rel_index
* 4;
131 desc
->slot_index_to_bind_directly
= -1;
134 static void si_release_descriptors(struct si_descriptors
*desc
)
136 si_resource_reference(&desc
->buffer
, NULL
);
140 static bool si_upload_descriptors(struct si_context
*sctx
,
141 struct si_descriptors
*desc
)
143 unsigned slot_size
= desc
->element_dw_size
* 4;
144 unsigned first_slot_offset
= desc
->first_active_slot
* slot_size
;
145 unsigned upload_size
= desc
->num_active_slots
* slot_size
;
147 /* Skip the upload if no shader is using the descriptors. dirty_mask
148 * will stay dirty and the descriptors will be uploaded when there is
149 * a shader using them.
154 /* If there is just one active descriptor, bind it directly. */
155 if ((int)desc
->first_active_slot
== desc
->slot_index_to_bind_directly
&&
156 desc
->num_active_slots
== 1) {
157 uint32_t *descriptor
= &desc
->list
[desc
->slot_index_to_bind_directly
*
158 desc
->element_dw_size
];
160 /* The buffer is already in the buffer list. */
161 si_resource_reference(&desc
->buffer
, NULL
);
162 desc
->gpu_list
= NULL
;
163 desc
->gpu_address
= si_desc_extract_buffer_address(descriptor
);
164 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.shader_pointers
);
169 unsigned buffer_offset
;
170 u_upload_alloc(sctx
->b
.const_uploader
, first_slot_offset
, upload_size
,
171 si_optimal_tcc_alignment(sctx
, upload_size
),
172 &buffer_offset
, (struct pipe_resource
**)&desc
->buffer
,
175 desc
->gpu_address
= 0;
176 return false; /* skip the draw call */
179 util_memcpy_cpu_to_le32(ptr
, (char*)desc
->list
+ first_slot_offset
,
181 desc
->gpu_list
= ptr
- first_slot_offset
/ 4;
183 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
, desc
->buffer
,
184 RADEON_USAGE_READ
, RADEON_PRIO_DESCRIPTORS
);
186 /* The shader pointer should point to slot 0. */
187 buffer_offset
-= first_slot_offset
;
188 desc
->gpu_address
= desc
->buffer
->gpu_address
+ buffer_offset
;
190 assert(desc
->buffer
->flags
& RADEON_FLAG_32BIT
);
191 assert((desc
->buffer
->gpu_address
>> 32) == sctx
->screen
->info
.address32_hi
);
192 assert((desc
->gpu_address
>> 32) == sctx
->screen
->info
.address32_hi
);
194 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.shader_pointers
);
199 si_descriptors_begin_new_cs(struct si_context
*sctx
, struct si_descriptors
*desc
)
204 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
, desc
->buffer
,
205 RADEON_USAGE_READ
, RADEON_PRIO_DESCRIPTORS
);
210 static inline enum radeon_bo_priority
211 si_get_sampler_view_priority(struct si_resource
*res
)
213 if (res
->b
.b
.target
== PIPE_BUFFER
)
214 return RADEON_PRIO_SAMPLER_BUFFER
;
216 if (res
->b
.b
.nr_samples
> 1)
217 return RADEON_PRIO_SAMPLER_TEXTURE_MSAA
;
219 return RADEON_PRIO_SAMPLER_TEXTURE
;
222 static struct si_descriptors
*
223 si_sampler_and_image_descriptors(struct si_context
*sctx
, unsigned shader
)
225 return &sctx
->descriptors
[si_sampler_and_image_descriptors_idx(shader
)];
228 static void si_release_sampler_views(struct si_samplers
*samplers
)
232 for (i
= 0; i
< ARRAY_SIZE(samplers
->views
); i
++) {
233 pipe_sampler_view_reference(&samplers
->views
[i
], NULL
);
237 static void si_sampler_view_add_buffer(struct si_context
*sctx
,
238 struct pipe_resource
*resource
,
239 enum radeon_bo_usage usage
,
240 bool is_stencil_sampler
,
243 struct si_texture
*tex
= (struct si_texture
*)resource
;
244 enum radeon_bo_priority priority
;
249 /* Use the flushed depth texture if direct sampling is unsupported. */
250 if (resource
->target
!= PIPE_BUFFER
&&
251 tex
->is_depth
&& !si_can_sample_zs(tex
, is_stencil_sampler
))
252 tex
= tex
->flushed_depth_texture
;
254 priority
= si_get_sampler_view_priority(&tex
->buffer
);
255 radeon_add_to_gfx_buffer_list_check_mem(sctx
, &tex
->buffer
, usage
, priority
,
258 if (resource
->target
== PIPE_BUFFER
)
261 /* Add separate DCC. */
262 if (tex
->dcc_separate_buffer
) {
263 radeon_add_to_gfx_buffer_list_check_mem(sctx
, tex
->dcc_separate_buffer
,
264 usage
, RADEON_PRIO_SEPARATE_META
, check_mem
);
268 static void si_sampler_views_begin_new_cs(struct si_context
*sctx
,
269 struct si_samplers
*samplers
)
271 unsigned mask
= samplers
->enabled_mask
;
273 /* Add buffers to the CS. */
275 int i
= u_bit_scan(&mask
);
276 struct si_sampler_view
*sview
= (struct si_sampler_view
*)samplers
->views
[i
];
278 si_sampler_view_add_buffer(sctx
, sview
->base
.texture
,
280 sview
->is_stencil_sampler
, false);
284 /* Set buffer descriptor fields that can be changed by reallocations. */
285 static void si_set_buf_desc_address(struct si_resource
*buf
,
286 uint64_t offset
, uint32_t *state
)
288 uint64_t va
= buf
->gpu_address
+ offset
;
291 state
[1] &= C_008F04_BASE_ADDRESS_HI
;
292 state
[1] |= S_008F04_BASE_ADDRESS_HI(va
>> 32);
295 /* Set texture descriptor fields that can be changed by reallocations.
298 * \param base_level_info information of the level of BASE_ADDRESS
299 * \param base_level the level of BASE_ADDRESS
300 * \param first_level pipe_sampler_view.u.tex.first_level
301 * \param block_width util_format_get_blockwidth()
302 * \param is_stencil select between separate Z & Stencil
303 * \param state descriptor to update
305 void si_set_mutable_tex_desc_fields(struct si_screen
*sscreen
,
306 struct si_texture
*tex
,
307 const struct legacy_surf_level
*base_level_info
,
308 unsigned base_level
, unsigned first_level
,
309 unsigned block_width
, bool is_stencil
,
312 uint64_t va
, meta_va
= 0;
314 if (tex
->is_depth
&& !si_can_sample_zs(tex
, is_stencil
)) {
315 tex
= tex
->flushed_depth_texture
;
319 va
= tex
->buffer
.gpu_address
;
321 if (sscreen
->info
.chip_class
>= GFX9
) {
322 /* Only stencil_offset needs to be added here. */
324 va
+= tex
->surface
.u
.gfx9
.stencil_offset
;
326 va
+= tex
->surface
.u
.gfx9
.surf_offset
;
328 va
+= base_level_info
->offset
;
332 state
[1] &= C_008F14_BASE_ADDRESS_HI
;
333 state
[1] |= S_008F14_BASE_ADDRESS_HI(va
>> 40);
335 /* Only macrotiled modes can set tile swizzle.
336 * GFX9 doesn't use (legacy) base_level_info.
338 if (sscreen
->info
.chip_class
>= GFX9
||
339 base_level_info
->mode
== RADEON_SURF_MODE_2D
)
340 state
[0] |= tex
->surface
.tile_swizzle
;
342 if (sscreen
->info
.chip_class
>= GFX8
) {
343 state
[6] &= C_008F28_COMPRESSION_EN
;
345 if (vi_dcc_enabled(tex
, first_level
)) {
346 meta_va
= (!tex
->dcc_separate_buffer
? tex
->buffer
.gpu_address
: 0) +
349 if (sscreen
->info
.chip_class
== GFX8
) {
350 meta_va
+= base_level_info
->dcc_offset
;
351 assert(base_level_info
->mode
== RADEON_SURF_MODE_2D
);
354 unsigned dcc_tile_swizzle
= tex
->surface
.tile_swizzle
<< 8;
355 dcc_tile_swizzle
&= tex
->surface
.dcc_alignment
- 1;
356 meta_va
|= dcc_tile_swizzle
;
357 } else if (vi_tc_compat_htile_enabled(tex
, first_level
)) {
358 meta_va
= tex
->buffer
.gpu_address
+ tex
->htile_offset
;
362 state
[6] |= S_008F28_COMPRESSION_EN(1);
365 if (sscreen
->info
.chip_class
>= GFX8
&& sscreen
->info
.chip_class
<= GFX9
)
366 state
[7] = meta_va
>> 8;
368 if (sscreen
->info
.chip_class
>= GFX10
) {
369 state
[3] &= C_00A00C_SW_MODE
;
372 state
[3] |= S_00A00C_SW_MODE(tex
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
374 state
[3] |= S_00A00C_SW_MODE(tex
->surface
.u
.gfx9
.surf
.swizzle_mode
);
377 state
[6] &= C_00A018_META_DATA_ADDRESS_LO
&
378 C_00A018_META_PIPE_ALIGNED
;
381 struct gfx9_surf_meta_flags meta
;
384 meta
= tex
->surface
.u
.gfx9
.dcc
;
386 meta
= tex
->surface
.u
.gfx9
.htile
;
388 state
[6] |= S_00A018_META_PIPE_ALIGNED(meta
.pipe_aligned
) |
389 S_00A018_META_DATA_ADDRESS_LO(meta_va
>> 8);
392 state
[7] = meta_va
>> 16;
393 } else if (sscreen
->info
.chip_class
>= GFX9
) {
394 state
[3] &= C_008F1C_SW_MODE
;
395 state
[4] &= C_008F20_PITCH
;
398 state
[3] |= S_008F1C_SW_MODE(tex
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
399 state
[4] |= S_008F20_PITCH(tex
->surface
.u
.gfx9
.stencil
.epitch
);
401 state
[3] |= S_008F1C_SW_MODE(tex
->surface
.u
.gfx9
.surf
.swizzle_mode
);
402 state
[4] |= S_008F20_PITCH(tex
->surface
.u
.gfx9
.surf
.epitch
);
405 state
[5] &= C_008F24_META_DATA_ADDRESS
&
406 C_008F24_META_PIPE_ALIGNED
&
407 C_008F24_META_RB_ALIGNED
;
409 struct gfx9_surf_meta_flags meta
;
412 meta
= tex
->surface
.u
.gfx9
.dcc
;
414 meta
= tex
->surface
.u
.gfx9
.htile
;
416 state
[5] |= S_008F24_META_DATA_ADDRESS(meta_va
>> 40) |
417 S_008F24_META_PIPE_ALIGNED(meta
.pipe_aligned
) |
418 S_008F24_META_RB_ALIGNED(meta
.rb_aligned
);
422 unsigned pitch
= base_level_info
->nblk_x
* block_width
;
423 unsigned index
= si_tile_mode_index(tex
, base_level
, is_stencil
);
425 state
[3] &= C_008F1C_TILING_INDEX
;
426 state
[3] |= S_008F1C_TILING_INDEX(index
);
427 state
[4] &= C_008F20_PITCH
;
428 state
[4] |= S_008F20_PITCH(pitch
- 1);
432 static void si_set_sampler_state_desc(struct si_sampler_state
*sstate
,
433 struct si_sampler_view
*sview
,
434 struct si_texture
*tex
,
437 if (sview
&& sview
->is_integer
)
438 memcpy(desc
, sstate
->integer_val
, 4*4);
439 else if (tex
&& tex
->upgraded_depth
&&
440 (!sview
|| !sview
->is_stencil_sampler
))
441 memcpy(desc
, sstate
->upgraded_depth_val
, 4*4);
443 memcpy(desc
, sstate
->val
, 4*4);
446 static void si_set_sampler_view_desc(struct si_context
*sctx
,
447 struct si_sampler_view
*sview
,
448 struct si_sampler_state
*sstate
,
451 struct pipe_sampler_view
*view
= &sview
->base
;
452 struct si_texture
*tex
= (struct si_texture
*)view
->texture
;
453 bool is_buffer
= tex
->buffer
.b
.b
.target
== PIPE_BUFFER
;
455 if (unlikely(!is_buffer
&& sview
->dcc_incompatible
)) {
456 if (vi_dcc_enabled(tex
, view
->u
.tex
.first_level
))
457 if (!si_texture_disable_dcc(sctx
, tex
))
458 si_decompress_dcc(sctx
, tex
);
460 sview
->dcc_incompatible
= false;
463 assert(tex
); /* views with texture == NULL aren't supported */
464 memcpy(desc
, sview
->state
, 8*4);
467 si_set_buf_desc_address(&tex
->buffer
,
468 sview
->base
.u
.buf
.offset
,
471 bool is_separate_stencil
= tex
->db_compatible
&&
472 sview
->is_stencil_sampler
;
474 si_set_mutable_tex_desc_fields(sctx
->screen
, tex
,
475 sview
->base_level_info
,
477 sview
->base
.u
.tex
.first_level
,
483 if (!is_buffer
&& tex
->surface
.fmask_size
) {
484 memcpy(desc
+ 8, sview
->fmask_state
, 8*4);
486 /* Disable FMASK and bind sampler state in [12:15]. */
487 memcpy(desc
+ 8, null_texture_descriptor
, 4*4);
490 si_set_sampler_state_desc(sstate
, sview
,
491 is_buffer
? NULL
: tex
,
496 static bool color_needs_decompression(struct si_texture
*tex
)
498 return tex
->surface
.fmask_size
||
499 (tex
->dirty_level_mask
&&
500 (tex
->cmask_buffer
|| tex
->dcc_offset
));
503 static bool depth_needs_decompression(struct si_texture
*tex
)
505 /* If the depth/stencil texture is TC-compatible, no decompression
506 * will be done. The decompression function will only flush DB caches
507 * to make it coherent with shaders. That's necessary because the driver
508 * doesn't flush DB caches in any other case.
510 return tex
->db_compatible
;
513 static void si_set_sampler_view(struct si_context
*sctx
,
515 unsigned slot
, struct pipe_sampler_view
*view
,
516 bool disallow_early_out
)
518 struct si_samplers
*samplers
= &sctx
->samplers
[shader
];
519 struct si_sampler_view
*sview
= (struct si_sampler_view
*)view
;
520 struct si_descriptors
*descs
= si_sampler_and_image_descriptors(sctx
, shader
);
521 unsigned desc_slot
= si_get_sampler_slot(slot
);
522 uint32_t *desc
= descs
->list
+ desc_slot
* 16;
524 if (samplers
->views
[slot
] == view
&& !disallow_early_out
)
528 struct si_texture
*tex
= (struct si_texture
*)view
->texture
;
530 si_set_sampler_view_desc(sctx
, sview
,
531 samplers
->sampler_states
[slot
], desc
);
533 if (tex
->buffer
.b
.b
.target
== PIPE_BUFFER
) {
534 tex
->buffer
.bind_history
|= PIPE_BIND_SAMPLER_VIEW
;
535 samplers
->needs_depth_decompress_mask
&= ~(1u << slot
);
536 samplers
->needs_color_decompress_mask
&= ~(1u << slot
);
538 if (depth_needs_decompression(tex
)) {
539 samplers
->needs_depth_decompress_mask
|= 1u << slot
;
541 samplers
->needs_depth_decompress_mask
&= ~(1u << slot
);
543 if (color_needs_decompression(tex
)) {
544 samplers
->needs_color_decompress_mask
|= 1u << slot
;
546 samplers
->needs_color_decompress_mask
&= ~(1u << slot
);
549 if (tex
->dcc_offset
&&
550 p_atomic_read(&tex
->framebuffers_bound
))
551 sctx
->need_check_render_feedback
= true;
554 pipe_sampler_view_reference(&samplers
->views
[slot
], view
);
555 samplers
->enabled_mask
|= 1u << slot
;
557 /* Since this can flush, it must be done after enabled_mask is
559 si_sampler_view_add_buffer(sctx
, view
->texture
,
561 sview
->is_stencil_sampler
, true);
563 pipe_sampler_view_reference(&samplers
->views
[slot
], NULL
);
564 memcpy(desc
, null_texture_descriptor
, 8*4);
565 /* Only clear the lower dwords of FMASK. */
566 memcpy(desc
+ 8, null_texture_descriptor
, 4*4);
567 /* Re-set the sampler state if we are transitioning from FMASK. */
568 if (samplers
->sampler_states
[slot
])
569 si_set_sampler_state_desc(samplers
->sampler_states
[slot
], NULL
, NULL
,
572 samplers
->enabled_mask
&= ~(1u << slot
);
573 samplers
->needs_depth_decompress_mask
&= ~(1u << slot
);
574 samplers
->needs_color_decompress_mask
&= ~(1u << slot
);
577 sctx
->descriptors_dirty
|= 1u << si_sampler_and_image_descriptors_idx(shader
);
580 static void si_update_shader_needs_decompress_mask(struct si_context
*sctx
,
583 struct si_samplers
*samplers
= &sctx
->samplers
[shader
];
584 unsigned shader_bit
= 1 << shader
;
586 if (samplers
->needs_depth_decompress_mask
||
587 samplers
->needs_color_decompress_mask
||
588 sctx
->images
[shader
].needs_color_decompress_mask
)
589 sctx
->shader_needs_decompress_mask
|= shader_bit
;
591 sctx
->shader_needs_decompress_mask
&= ~shader_bit
;
594 static void si_set_sampler_views(struct pipe_context
*ctx
,
595 enum pipe_shader_type shader
, unsigned start
,
597 struct pipe_sampler_view
**views
)
599 struct si_context
*sctx
= (struct si_context
*)ctx
;
602 if (!count
|| shader
>= SI_NUM_SHADERS
)
606 for (i
= 0; i
< count
; i
++)
607 si_set_sampler_view(sctx
, shader
, start
+ i
, views
[i
], false);
609 for (i
= 0; i
< count
; i
++)
610 si_set_sampler_view(sctx
, shader
, start
+ i
, NULL
, false);
613 si_update_shader_needs_decompress_mask(sctx
, shader
);
617 si_samplers_update_needs_color_decompress_mask(struct si_samplers
*samplers
)
619 unsigned mask
= samplers
->enabled_mask
;
622 int i
= u_bit_scan(&mask
);
623 struct pipe_resource
*res
= samplers
->views
[i
]->texture
;
625 if (res
&& res
->target
!= PIPE_BUFFER
) {
626 struct si_texture
*tex
= (struct si_texture
*)res
;
628 if (color_needs_decompression(tex
)) {
629 samplers
->needs_color_decompress_mask
|= 1u << i
;
631 samplers
->needs_color_decompress_mask
&= ~(1u << i
);
640 si_release_image_views(struct si_images
*images
)
644 for (i
= 0; i
< SI_NUM_IMAGES
; ++i
) {
645 struct pipe_image_view
*view
= &images
->views
[i
];
647 pipe_resource_reference(&view
->resource
, NULL
);
652 si_image_views_begin_new_cs(struct si_context
*sctx
, struct si_images
*images
)
654 uint mask
= images
->enabled_mask
;
656 /* Add buffers to the CS. */
658 int i
= u_bit_scan(&mask
);
659 struct pipe_image_view
*view
= &images
->views
[i
];
661 assert(view
->resource
);
663 si_sampler_view_add_buffer(sctx
, view
->resource
,
664 RADEON_USAGE_READWRITE
, false, false);
669 si_disable_shader_image(struct si_context
*ctx
, unsigned shader
, unsigned slot
)
671 struct si_images
*images
= &ctx
->images
[shader
];
673 if (images
->enabled_mask
& (1u << slot
)) {
674 struct si_descriptors
*descs
= si_sampler_and_image_descriptors(ctx
, shader
);
675 unsigned desc_slot
= si_get_image_slot(slot
);
677 pipe_resource_reference(&images
->views
[slot
].resource
, NULL
);
678 images
->needs_color_decompress_mask
&= ~(1 << slot
);
680 memcpy(descs
->list
+ desc_slot
*8, null_image_descriptor
, 8*4);
681 images
->enabled_mask
&= ~(1u << slot
);
682 ctx
->descriptors_dirty
|= 1u << si_sampler_and_image_descriptors_idx(shader
);
687 si_mark_image_range_valid(const struct pipe_image_view
*view
)
689 struct si_resource
*res
= si_resource(view
->resource
);
691 if (res
->b
.b
.target
!= PIPE_BUFFER
)
694 util_range_add(&res
->valid_buffer_range
,
696 view
->u
.buf
.offset
+ view
->u
.buf
.size
);
699 static void si_set_shader_image_desc(struct si_context
*ctx
,
700 const struct pipe_image_view
*view
,
701 bool skip_decompress
,
702 uint32_t *desc
, uint32_t *fmask_desc
)
704 struct si_screen
*screen
= ctx
->screen
;
705 struct si_resource
*res
;
707 res
= si_resource(view
->resource
);
709 if (res
->b
.b
.target
== PIPE_BUFFER
||
710 view
->shader_access
& SI_IMAGE_ACCESS_AS_BUFFER
) {
711 if (view
->access
& PIPE_IMAGE_ACCESS_WRITE
)
712 si_mark_image_range_valid(view
);
714 si_make_buffer_descriptor(screen
, res
,
717 view
->u
.buf
.size
, desc
);
718 si_set_buf_desc_address(res
, view
->u
.buf
.offset
, desc
+ 4);
720 static const unsigned char swizzle
[4] = { 0, 1, 2, 3 };
721 struct si_texture
*tex
= (struct si_texture
*)res
;
722 unsigned level
= view
->u
.tex
.level
;
723 unsigned width
, height
, depth
, hw_level
;
724 bool uses_dcc
= vi_dcc_enabled(tex
, level
);
725 unsigned access
= view
->access
;
727 /* Clear the write flag when writes can't occur.
728 * Note that DCC_DECOMPRESS for MSAA doesn't work in some cases,
729 * so we don't wanna trigger it.
732 (!fmask_desc
&& tex
->surface
.fmask_size
!= 0)) {
733 assert(!"Z/S and MSAA image stores are not supported");
734 access
&= ~PIPE_IMAGE_ACCESS_WRITE
;
737 assert(!tex
->is_depth
);
738 assert(fmask_desc
|| tex
->surface
.fmask_size
== 0);
740 if (uses_dcc
&& !skip_decompress
&&
741 (view
->access
& PIPE_IMAGE_ACCESS_WRITE
||
742 !vi_dcc_formats_compatible(res
->b
.b
.format
, view
->format
))) {
743 /* If DCC can't be disabled, at least decompress it.
744 * The decompression is relatively cheap if the surface
745 * has been decompressed already.
747 if (!si_texture_disable_dcc(ctx
, tex
))
748 si_decompress_dcc(ctx
, tex
);
751 if (ctx
->chip_class
>= GFX9
) {
752 /* Always set the base address. The swizzle modes don't
753 * allow setting mipmap level offsets as the base.
755 width
= res
->b
.b
.width0
;
756 height
= res
->b
.b
.height0
;
757 depth
= res
->b
.b
.depth0
;
760 /* Always force the base level to the selected level.
762 * This is required for 3D textures, where otherwise
763 * selecting a single slice for non-layered bindings
764 * fails. It doesn't hurt the other targets.
766 width
= u_minify(res
->b
.b
.width0
, level
);
767 height
= u_minify(res
->b
.b
.height0
, level
);
768 depth
= u_minify(res
->b
.b
.depth0
, level
);
772 screen
->make_texture_descriptor(screen
, tex
,
773 false, res
->b
.b
.target
,
774 view
->format
, swizzle
,
776 view
->u
.tex
.first_layer
,
777 view
->u
.tex
.last_layer
,
778 width
, height
, depth
,
780 si_set_mutable_tex_desc_fields(screen
, tex
,
781 &tex
->surface
.u
.legacy
.level
[level
],
783 util_format_get_blockwidth(view
->format
),
788 static void si_set_shader_image(struct si_context
*ctx
,
790 unsigned slot
, const struct pipe_image_view
*view
,
791 bool skip_decompress
)
793 struct si_images
*images
= &ctx
->images
[shader
];
794 struct si_descriptors
*descs
= si_sampler_and_image_descriptors(ctx
, shader
);
795 struct si_resource
*res
;
796 unsigned desc_slot
= si_get_image_slot(slot
);
797 uint32_t *desc
= descs
->list
+ desc_slot
* 8;
799 if (!view
|| !view
->resource
) {
800 si_disable_shader_image(ctx
, shader
, slot
);
804 res
= si_resource(view
->resource
);
806 if (&images
->views
[slot
] != view
)
807 util_copy_image_view(&images
->views
[slot
], view
);
809 si_set_shader_image_desc(ctx
, view
, skip_decompress
, desc
, NULL
);
811 if (res
->b
.b
.target
== PIPE_BUFFER
||
812 view
->shader_access
& SI_IMAGE_ACCESS_AS_BUFFER
) {
813 images
->needs_color_decompress_mask
&= ~(1 << slot
);
814 res
->bind_history
|= PIPE_BIND_SHADER_IMAGE
;
816 struct si_texture
*tex
= (struct si_texture
*)res
;
817 unsigned level
= view
->u
.tex
.level
;
819 if (color_needs_decompression(tex
)) {
820 images
->needs_color_decompress_mask
|= 1 << slot
;
822 images
->needs_color_decompress_mask
&= ~(1 << slot
);
825 if (vi_dcc_enabled(tex
, level
) &&
826 p_atomic_read(&tex
->framebuffers_bound
))
827 ctx
->need_check_render_feedback
= true;
830 images
->enabled_mask
|= 1u << slot
;
831 ctx
->descriptors_dirty
|= 1u << si_sampler_and_image_descriptors_idx(shader
);
833 /* Since this can flush, it must be done after enabled_mask is updated. */
834 si_sampler_view_add_buffer(ctx
, &res
->b
.b
,
835 (view
->access
& PIPE_IMAGE_ACCESS_WRITE
) ?
836 RADEON_USAGE_READWRITE
: RADEON_USAGE_READ
,
841 si_set_shader_images(struct pipe_context
*pipe
,
842 enum pipe_shader_type shader
,
843 unsigned start_slot
, unsigned count
,
844 const struct pipe_image_view
*views
)
846 struct si_context
*ctx
= (struct si_context
*)pipe
;
849 assert(shader
< SI_NUM_SHADERS
);
854 assert(start_slot
+ count
<= SI_NUM_IMAGES
);
857 for (i
= 0, slot
= start_slot
; i
< count
; ++i
, ++slot
)
858 si_set_shader_image(ctx
, shader
, slot
, &views
[i
], false);
860 for (i
= 0, slot
= start_slot
; i
< count
; ++i
, ++slot
)
861 si_set_shader_image(ctx
, shader
, slot
, NULL
, false);
864 si_update_shader_needs_decompress_mask(ctx
, shader
);
868 si_images_update_needs_color_decompress_mask(struct si_images
*images
)
870 unsigned mask
= images
->enabled_mask
;
873 int i
= u_bit_scan(&mask
);
874 struct pipe_resource
*res
= images
->views
[i
].resource
;
876 if (res
&& res
->target
!= PIPE_BUFFER
) {
877 struct si_texture
*tex
= (struct si_texture
*)res
;
879 if (color_needs_decompression(tex
)) {
880 images
->needs_color_decompress_mask
|= 1 << i
;
882 images
->needs_color_decompress_mask
&= ~(1 << i
);
888 void si_update_ps_colorbuf0_slot(struct si_context
*sctx
)
890 struct si_buffer_resources
*buffers
= &sctx
->rw_buffers
;
891 struct si_descriptors
*descs
= &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
];
892 unsigned slot
= SI_PS_IMAGE_COLORBUF0
;
893 struct pipe_surface
*surf
= NULL
;
895 /* si_texture_disable_dcc can get us here again. */
896 if (sctx
->blitter
->running
)
899 /* See whether FBFETCH is used and color buffer 0 is set. */
900 if (sctx
->ps_shader
.cso
&&
901 sctx
->ps_shader
.cso
->info
.opcode_count
[TGSI_OPCODE_FBFETCH
] &&
902 sctx
->framebuffer
.state
.nr_cbufs
&&
903 sctx
->framebuffer
.state
.cbufs
[0])
904 surf
= sctx
->framebuffer
.state
.cbufs
[0];
906 /* Return if FBFETCH transitions from disabled to disabled. */
907 if (!buffers
->buffers
[slot
] && !surf
)
910 sctx
->ps_uses_fbfetch
= surf
!= NULL
;
911 si_update_ps_iter_samples(sctx
);
914 struct si_texture
*tex
= (struct si_texture
*)surf
->texture
;
915 struct pipe_image_view view
;
918 assert(!tex
->is_depth
);
920 /* Disable DCC, because the texture is used as both a sampler
923 si_texture_disable_dcc(sctx
, tex
);
925 if (tex
->buffer
.b
.b
.nr_samples
<= 1 && tex
->cmask_buffer
) {
927 assert(tex
->cmask_buffer
!= &tex
->buffer
);
928 si_eliminate_fast_color_clear(sctx
, tex
);
929 si_texture_discard_cmask(sctx
->screen
, tex
);
932 view
.resource
= surf
->texture
;
933 view
.format
= surf
->format
;
934 view
.access
= PIPE_IMAGE_ACCESS_READ
;
935 view
.u
.tex
.first_layer
= surf
->u
.tex
.first_layer
;
936 view
.u
.tex
.last_layer
= surf
->u
.tex
.last_layer
;
937 view
.u
.tex
.level
= surf
->u
.tex
.level
;
939 /* Set the descriptor. */
940 uint32_t *desc
= descs
->list
+ slot
*4;
941 memset(desc
, 0, 16 * 4);
942 si_set_shader_image_desc(sctx
, &view
, true, desc
, desc
+ 8);
944 pipe_resource_reference(&buffers
->buffers
[slot
], &tex
->buffer
.b
.b
);
945 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
946 &tex
->buffer
, RADEON_USAGE_READ
,
947 RADEON_PRIO_SHADER_RW_IMAGE
);
948 buffers
->enabled_mask
|= 1u << slot
;
950 /* Clear the descriptor. */
951 memset(descs
->list
+ slot
*4, 0, 8*4);
952 pipe_resource_reference(&buffers
->buffers
[slot
], NULL
);
953 buffers
->enabled_mask
&= ~(1u << slot
);
956 sctx
->descriptors_dirty
|= 1u << SI_DESCS_RW_BUFFERS
;
961 static void si_bind_sampler_states(struct pipe_context
*ctx
,
962 enum pipe_shader_type shader
,
963 unsigned start
, unsigned count
, void **states
)
965 struct si_context
*sctx
= (struct si_context
*)ctx
;
966 struct si_samplers
*samplers
= &sctx
->samplers
[shader
];
967 struct si_descriptors
*desc
= si_sampler_and_image_descriptors(sctx
, shader
);
968 struct si_sampler_state
**sstates
= (struct si_sampler_state
**)states
;
971 if (!count
|| shader
>= SI_NUM_SHADERS
|| !sstates
)
974 for (i
= 0; i
< count
; i
++) {
975 unsigned slot
= start
+ i
;
976 unsigned desc_slot
= si_get_sampler_slot(slot
);
979 sstates
[i
] == samplers
->sampler_states
[slot
])
983 assert(sstates
[i
]->magic
== SI_SAMPLER_STATE_MAGIC
);
985 samplers
->sampler_states
[slot
] = sstates
[i
];
987 /* If FMASK is bound, don't overwrite it.
988 * The sampler state will be set after FMASK is unbound.
990 struct si_sampler_view
*sview
=
991 (struct si_sampler_view
*)samplers
->views
[slot
];
993 struct si_texture
*tex
= NULL
;
995 if (sview
&& sview
->base
.texture
&&
996 sview
->base
.texture
->target
!= PIPE_BUFFER
)
997 tex
= (struct si_texture
*)sview
->base
.texture
;
999 if (tex
&& tex
->surface
.fmask_size
)
1002 si_set_sampler_state_desc(sstates
[i
], sview
, tex
,
1003 desc
->list
+ desc_slot
* 16 + 12);
1005 sctx
->descriptors_dirty
|= 1u << si_sampler_and_image_descriptors_idx(shader
);
1009 /* BUFFER RESOURCES */
1011 static void si_init_buffer_resources(struct si_buffer_resources
*buffers
,
1012 struct si_descriptors
*descs
,
1013 unsigned num_buffers
,
1014 short shader_userdata_rel_index
,
1015 enum radeon_bo_priority priority
,
1016 enum radeon_bo_priority priority_constbuf
)
1018 buffers
->priority
= priority
;
1019 buffers
->priority_constbuf
= priority_constbuf
;
1020 buffers
->buffers
= CALLOC(num_buffers
, sizeof(struct pipe_resource
*));
1021 buffers
->offsets
= CALLOC(num_buffers
, sizeof(buffers
->offsets
[0]));
1023 si_init_descriptors(descs
, shader_userdata_rel_index
, 4, num_buffers
);
1026 static void si_release_buffer_resources(struct si_buffer_resources
*buffers
,
1027 struct si_descriptors
*descs
)
1031 for (i
= 0; i
< descs
->num_elements
; i
++) {
1032 pipe_resource_reference(&buffers
->buffers
[i
], NULL
);
1035 FREE(buffers
->buffers
);
1036 FREE(buffers
->offsets
);
1039 static void si_buffer_resources_begin_new_cs(struct si_context
*sctx
,
1040 struct si_buffer_resources
*buffers
)
1042 unsigned mask
= buffers
->enabled_mask
;
1044 /* Add buffers to the CS. */
1046 int i
= u_bit_scan(&mask
);
1048 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
1049 si_resource(buffers
->buffers
[i
]),
1050 buffers
->writable_mask
& (1u << i
) ? RADEON_USAGE_READWRITE
:
1052 i
< SI_NUM_SHADER_BUFFERS
? buffers
->priority
:
1053 buffers
->priority_constbuf
);
1057 static void si_get_buffer_from_descriptors(struct si_buffer_resources
*buffers
,
1058 struct si_descriptors
*descs
,
1059 unsigned idx
, struct pipe_resource
**buf
,
1060 unsigned *offset
, unsigned *size
)
1062 pipe_resource_reference(buf
, buffers
->buffers
[idx
]);
1064 struct si_resource
*res
= si_resource(*buf
);
1065 const uint32_t *desc
= descs
->list
+ idx
* 4;
1070 assert(G_008F04_STRIDE(desc
[1]) == 0);
1071 va
= si_desc_extract_buffer_address(desc
);
1073 assert(va
>= res
->gpu_address
&& va
+ *size
<= res
->gpu_address
+ res
->bo_size
);
1074 *offset
= va
- res
->gpu_address
;
1078 /* VERTEX BUFFERS */
1080 static void si_vertex_buffers_begin_new_cs(struct si_context
*sctx
)
1082 int count
= sctx
->vertex_elements
? sctx
->vertex_elements
->count
: 0;
1085 for (i
= 0; i
< count
; i
++) {
1086 int vb
= sctx
->vertex_elements
->vertex_buffer_index
[i
];
1088 if (vb
>= ARRAY_SIZE(sctx
->vertex_buffer
))
1090 if (!sctx
->vertex_buffer
[vb
].buffer
.resource
)
1093 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
1094 si_resource(sctx
->vertex_buffer
[vb
].buffer
.resource
),
1095 RADEON_USAGE_READ
, RADEON_PRIO_VERTEX_BUFFER
);
1098 if (!sctx
->vb_descriptors_buffer
)
1100 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
1101 sctx
->vb_descriptors_buffer
, RADEON_USAGE_READ
,
1102 RADEON_PRIO_DESCRIPTORS
);
1105 bool si_upload_vertex_buffer_descriptors(struct si_context
*sctx
)
1107 struct si_vertex_elements
*velems
= sctx
->vertex_elements
;
1109 unsigned desc_list_byte_size
;
1110 unsigned first_vb_use_mask
;
1113 if (!sctx
->vertex_buffers_dirty
|| !velems
)
1116 count
= velems
->count
;
1121 desc_list_byte_size
= velems
->desc_list_byte_size
;
1122 first_vb_use_mask
= velems
->first_vb_use_mask
;
1124 /* Vertex buffer descriptors are the only ones which are uploaded
1125 * directly through a staging buffer and don't go through
1126 * the fine-grained upload path.
1128 u_upload_alloc(sctx
->b
.const_uploader
, 0,
1129 desc_list_byte_size
,
1130 si_optimal_tcc_alignment(sctx
, desc_list_byte_size
),
1131 &sctx
->vb_descriptors_offset
,
1132 (struct pipe_resource
**)&sctx
->vb_descriptors_buffer
,
1134 if (!sctx
->vb_descriptors_buffer
) {
1135 sctx
->vb_descriptors_offset
= 0;
1136 sctx
->vb_descriptors_gpu_list
= NULL
;
1140 sctx
->vb_descriptors_gpu_list
= ptr
;
1141 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
1142 sctx
->vb_descriptors_buffer
, RADEON_USAGE_READ
,
1143 RADEON_PRIO_DESCRIPTORS
);
1145 assert(count
<= SI_MAX_ATTRIBS
);
1147 for (i
= 0; i
< count
; i
++) {
1148 struct pipe_vertex_buffer
*vb
;
1149 struct si_resource
*buf
;
1150 unsigned vbo_index
= velems
->vertex_buffer_index
[i
];
1151 uint32_t *desc
= &ptr
[i
*4];
1153 vb
= &sctx
->vertex_buffer
[vbo_index
];
1154 buf
= si_resource(vb
->buffer
.resource
);
1156 memset(desc
, 0, 16);
1160 int64_t offset
= (int64_t)((int)vb
->buffer_offset
) +
1161 velems
->src_offset
[i
];
1162 uint64_t va
= buf
->gpu_address
+ offset
;
1164 int64_t num_records
= (int64_t)buf
->b
.b
.width0
- offset
;
1165 if (sctx
->chip_class
!= GFX8
&& vb
->stride
) {
1166 /* Round up by rounding down and adding 1 */
1167 num_records
= (num_records
- velems
->format_size
[i
]) /
1170 assert(num_records
>= 0 && num_records
<= UINT_MAX
);
1172 uint32_t rsrc_word3
= velems
->rsrc_word3
[i
];
1174 /* OOB_SELECT chooses the out-of-bounds check:
1175 * - 1: index >= NUM_RECORDS (Structured)
1176 * - 3: offset >= NUM_RECORDS (Raw)
1178 if (sctx
->chip_class
>= GFX10
)
1179 rsrc_word3
|= S_008F0C_OOB_SELECT(vb
->stride
? 1 : 3);
1182 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
1183 S_008F04_STRIDE(vb
->stride
);
1184 desc
[2] = num_records
;
1185 desc
[3] = rsrc_word3
;
1187 if (first_vb_use_mask
& (1 << i
)) {
1188 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
1189 si_resource(vb
->buffer
.resource
),
1190 RADEON_USAGE_READ
, RADEON_PRIO_VERTEX_BUFFER
);
1194 /* Don't flush the const cache. It would have a very negative effect
1195 * on performance (confirmed by testing). New descriptors are always
1196 * uploaded to a fresh new buffer, so I don't think flushing the const
1197 * cache is needed. */
1198 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.shader_pointers
);
1199 sctx
->vertex_buffers_dirty
= false;
1200 sctx
->vertex_buffer_pointer_dirty
= true;
1201 sctx
->prefetch_L2_mask
|= SI_PREFETCH_VBO_DESCRIPTORS
;
1206 /* CONSTANT BUFFERS */
1208 static struct si_descriptors
*
1209 si_const_and_shader_buffer_descriptors(struct si_context
*sctx
, unsigned shader
)
1211 return &sctx
->descriptors
[si_const_and_shader_buffer_descriptors_idx(shader
)];
1214 void si_upload_const_buffer(struct si_context
*sctx
, struct si_resource
**buf
,
1215 const uint8_t *ptr
, unsigned size
, uint32_t *const_offset
)
1219 u_upload_alloc(sctx
->b
.const_uploader
, 0, size
,
1220 si_optimal_tcc_alignment(sctx
, size
),
1222 (struct pipe_resource
**)buf
, &tmp
);
1224 util_memcpy_cpu_to_le32(tmp
, ptr
, size
);
1227 static void si_set_constant_buffer(struct si_context
*sctx
,
1228 struct si_buffer_resources
*buffers
,
1229 unsigned descriptors_idx
,
1230 uint slot
, const struct pipe_constant_buffer
*input
)
1232 struct si_descriptors
*descs
= &sctx
->descriptors
[descriptors_idx
];
1233 assert(slot
< descs
->num_elements
);
1234 pipe_resource_reference(&buffers
->buffers
[slot
], NULL
);
1236 /* GFX7 cannot unbind a constant buffer (S_BUFFER_LOAD is buggy
1237 * with a NULL buffer). We need to use a dummy buffer instead. */
1238 if (sctx
->chip_class
== GFX7
&&
1239 (!input
|| (!input
->buffer
&& !input
->user_buffer
)))
1240 input
= &sctx
->null_const_buf
;
1242 if (input
&& (input
->buffer
|| input
->user_buffer
)) {
1243 struct pipe_resource
*buffer
= NULL
;
1245 unsigned buffer_offset
;
1247 /* Upload the user buffer if needed. */
1248 if (input
->user_buffer
) {
1249 si_upload_const_buffer(sctx
,
1250 (struct si_resource
**)&buffer
, input
->user_buffer
,
1251 input
->buffer_size
, &buffer_offset
);
1253 /* Just unbind on failure. */
1254 si_set_constant_buffer(sctx
, buffers
, descriptors_idx
, slot
, NULL
);
1258 pipe_resource_reference(&buffer
, input
->buffer
);
1259 buffer_offset
= input
->buffer_offset
;
1262 va
= si_resource(buffer
)->gpu_address
+ buffer_offset
;
1264 /* Set the descriptor. */
1265 uint32_t *desc
= descs
->list
+ slot
*4;
1267 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
1269 desc
[2] = input
->buffer_size
;
1270 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1271 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1272 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1273 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
1275 if (sctx
->chip_class
>= GFX10
) {
1276 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
1277 S_008F0C_OOB_SELECT(3) |
1278 S_008F0C_RESOURCE_LEVEL(1);
1280 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1281 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
1284 buffers
->buffers
[slot
] = buffer
;
1285 buffers
->offsets
[slot
] = buffer_offset
;
1286 radeon_add_to_gfx_buffer_list_check_mem(sctx
,
1287 si_resource(buffer
),
1289 buffers
->priority_constbuf
, true);
1290 buffers
->enabled_mask
|= 1u << slot
;
1292 /* Clear the descriptor. */
1293 memset(descs
->list
+ slot
*4, 0, sizeof(uint32_t) * 4);
1294 buffers
->enabled_mask
&= ~(1u << slot
);
1297 sctx
->descriptors_dirty
|= 1u << descriptors_idx
;
1300 static void si_pipe_set_constant_buffer(struct pipe_context
*ctx
,
1301 enum pipe_shader_type shader
, uint slot
,
1302 const struct pipe_constant_buffer
*input
)
1304 struct si_context
*sctx
= (struct si_context
*)ctx
;
1306 if (shader
>= SI_NUM_SHADERS
)
1309 if (slot
== 0 && input
&& input
->buffer
&&
1310 !(si_resource(input
->buffer
)->flags
& RADEON_FLAG_32BIT
)) {
1311 assert(!"constant buffer 0 must have a 32-bit VM address, use const_uploader");
1315 if (input
&& input
->buffer
)
1316 si_resource(input
->buffer
)->bind_history
|= PIPE_BIND_CONSTANT_BUFFER
;
1318 slot
= si_get_constbuf_slot(slot
);
1319 si_set_constant_buffer(sctx
, &sctx
->const_and_shader_buffers
[shader
],
1320 si_const_and_shader_buffer_descriptors_idx(shader
),
1324 void si_get_pipe_constant_buffer(struct si_context
*sctx
, uint shader
,
1325 uint slot
, struct pipe_constant_buffer
*cbuf
)
1327 cbuf
->user_buffer
= NULL
;
1328 si_get_buffer_from_descriptors(
1329 &sctx
->const_and_shader_buffers
[shader
],
1330 si_const_and_shader_buffer_descriptors(sctx
, shader
),
1331 si_get_constbuf_slot(slot
),
1332 &cbuf
->buffer
, &cbuf
->buffer_offset
, &cbuf
->buffer_size
);
1335 /* SHADER BUFFERS */
1337 static void si_set_shader_buffer(struct si_context
*sctx
,
1338 struct si_buffer_resources
*buffers
,
1339 unsigned descriptors_idx
,
1340 uint slot
, const struct pipe_shader_buffer
*sbuffer
,
1341 bool writable
, enum radeon_bo_priority priority
)
1343 struct si_descriptors
*descs
= &sctx
->descriptors
[descriptors_idx
];
1344 uint32_t *desc
= descs
->list
+ slot
* 4;
1346 if (!sbuffer
|| !sbuffer
->buffer
) {
1347 pipe_resource_reference(&buffers
->buffers
[slot
], NULL
);
1348 memset(desc
, 0, sizeof(uint32_t) * 4);
1349 buffers
->enabled_mask
&= ~(1u << slot
);
1350 buffers
->writable_mask
&= ~(1u << slot
);
1351 sctx
->descriptors_dirty
|= 1u << descriptors_idx
;
1355 struct si_resource
*buf
= si_resource(sbuffer
->buffer
);
1356 uint64_t va
= buf
->gpu_address
+ sbuffer
->buffer_offset
;
1359 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
1361 desc
[2] = sbuffer
->buffer_size
;
1362 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1363 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1364 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1365 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
);
1367 if (sctx
->chip_class
>= GFX10
) {
1368 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
1369 S_008F0C_OOB_SELECT(3) |
1370 S_008F0C_RESOURCE_LEVEL(1);
1372 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1373 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
1376 pipe_resource_reference(&buffers
->buffers
[slot
], &buf
->b
.b
);
1377 buffers
->offsets
[slot
] = sbuffer
->buffer_offset
;
1378 radeon_add_to_gfx_buffer_list_check_mem(sctx
, buf
,
1379 writable
? RADEON_USAGE_READWRITE
:
1383 buffers
->writable_mask
|= 1u << slot
;
1385 buffers
->writable_mask
&= ~(1u << slot
);
1387 buffers
->enabled_mask
|= 1u << slot
;
1388 sctx
->descriptors_dirty
|= 1u << descriptors_idx
;
1390 util_range_add(&buf
->valid_buffer_range
, sbuffer
->buffer_offset
,
1391 sbuffer
->buffer_offset
+ sbuffer
->buffer_size
);
1394 static void si_set_shader_buffers(struct pipe_context
*ctx
,
1395 enum pipe_shader_type shader
,
1396 unsigned start_slot
, unsigned count
,
1397 const struct pipe_shader_buffer
*sbuffers
,
1398 unsigned writable_bitmask
)
1400 struct si_context
*sctx
= (struct si_context
*)ctx
;
1401 struct si_buffer_resources
*buffers
= &sctx
->const_and_shader_buffers
[shader
];
1402 unsigned descriptors_idx
= si_const_and_shader_buffer_descriptors_idx(shader
);
1405 assert(start_slot
+ count
<= SI_NUM_SHADER_BUFFERS
);
1407 for (i
= 0; i
< count
; ++i
) {
1408 const struct pipe_shader_buffer
*sbuffer
= sbuffers
? &sbuffers
[i
] : NULL
;
1409 unsigned slot
= si_get_shaderbuf_slot(start_slot
+ i
);
1411 if (sbuffer
&& sbuffer
->buffer
)
1412 si_resource(sbuffer
->buffer
)->bind_history
|= PIPE_BIND_SHADER_BUFFER
;
1414 si_set_shader_buffer(sctx
, buffers
, descriptors_idx
, slot
, sbuffer
,
1415 !!(writable_bitmask
& (1u << i
)),
1420 void si_get_shader_buffers(struct si_context
*sctx
,
1421 enum pipe_shader_type shader
,
1422 uint start_slot
, uint count
,
1423 struct pipe_shader_buffer
*sbuf
)
1425 struct si_buffer_resources
*buffers
= &sctx
->const_and_shader_buffers
[shader
];
1426 struct si_descriptors
*descs
= si_const_and_shader_buffer_descriptors(sctx
, shader
);
1428 for (unsigned i
= 0; i
< count
; ++i
) {
1429 si_get_buffer_from_descriptors(
1431 si_get_shaderbuf_slot(start_slot
+ i
),
1432 &sbuf
[i
].buffer
, &sbuf
[i
].buffer_offset
,
1433 &sbuf
[i
].buffer_size
);
1439 void si_set_rw_buffer(struct si_context
*sctx
,
1440 uint slot
, const struct pipe_constant_buffer
*input
)
1442 si_set_constant_buffer(sctx
, &sctx
->rw_buffers
, SI_DESCS_RW_BUFFERS
,
1446 void si_set_rw_shader_buffer(struct si_context
*sctx
, uint slot
,
1447 const struct pipe_shader_buffer
*sbuffer
)
1449 si_set_shader_buffer(sctx
, &sctx
->rw_buffers
, SI_DESCS_RW_BUFFERS
,
1450 slot
, sbuffer
, true, RADEON_PRIO_SHADER_RW_BUFFER
);
1453 void si_set_ring_buffer(struct si_context
*sctx
, uint slot
,
1454 struct pipe_resource
*buffer
,
1455 unsigned stride
, unsigned num_records
,
1456 bool add_tid
, bool swizzle
,
1457 unsigned element_size
, unsigned index_stride
, uint64_t offset
)
1459 struct si_buffer_resources
*buffers
= &sctx
->rw_buffers
;
1460 struct si_descriptors
*descs
= &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
];
1462 /* The stride field in the resource descriptor has 14 bits */
1463 assert(stride
< (1 << 14));
1465 assert(slot
< descs
->num_elements
);
1466 pipe_resource_reference(&buffers
->buffers
[slot
], NULL
);
1471 va
= si_resource(buffer
)->gpu_address
+ offset
;
1473 switch (element_size
) {
1475 assert(!"Unsupported ring buffer element size");
1491 switch (index_stride
) {
1493 assert(!"Unsupported ring buffer index stride");
1509 if (sctx
->chip_class
>= GFX8
&& stride
)
1510 num_records
*= stride
;
1512 /* Set the descriptor. */
1513 uint32_t *desc
= descs
->list
+ slot
*4;
1515 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
1516 S_008F04_STRIDE(stride
) |
1517 S_008F04_SWIZZLE_ENABLE(swizzle
);
1518 desc
[2] = num_records
;
1519 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1520 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1521 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1522 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1523 S_008F0C_INDEX_STRIDE(index_stride
) |
1524 S_008F0C_ADD_TID_ENABLE(add_tid
);
1526 if (sctx
->chip_class
>= GFX9
)
1527 assert(!swizzle
|| element_size
== 1); /* always 4 bytes on GFX9 */
1529 desc
[3] |= S_008F0C_ELEMENT_SIZE(element_size
);
1531 if (sctx
->chip_class
>= GFX10
) {
1532 desc
[3] |= S_008F0C_FORMAT(V_008F0C_IMG_FORMAT_32_FLOAT
) |
1533 S_008F0C_OOB_SELECT(2) |
1534 S_008F0C_RESOURCE_LEVEL(1);
1536 desc
[3] |= S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1537 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
1540 pipe_resource_reference(&buffers
->buffers
[slot
], buffer
);
1541 radeon_add_to_buffer_list(sctx
, sctx
->gfx_cs
,
1542 si_resource(buffer
),
1543 RADEON_USAGE_READWRITE
, buffers
->priority
);
1544 buffers
->enabled_mask
|= 1u << slot
;
1546 /* Clear the descriptor. */
1547 memset(descs
->list
+ slot
*4, 0, sizeof(uint32_t) * 4);
1548 buffers
->enabled_mask
&= ~(1u << slot
);
1551 sctx
->descriptors_dirty
|= 1u << SI_DESCS_RW_BUFFERS
;
1554 /* INTERNAL CONST BUFFERS */
1556 static void si_set_polygon_stipple(struct pipe_context
*ctx
,
1557 const struct pipe_poly_stipple
*state
)
1559 struct si_context
*sctx
= (struct si_context
*)ctx
;
1560 struct pipe_constant_buffer cb
= {};
1561 unsigned stipple
[32];
1564 for (i
= 0; i
< 32; i
++)
1565 stipple
[i
] = util_bitreverse(state
->stipple
[i
]);
1567 cb
.user_buffer
= stipple
;
1568 cb
.buffer_size
= sizeof(stipple
);
1570 si_set_rw_buffer(sctx
, SI_PS_CONST_POLY_STIPPLE
, &cb
);
1573 /* TEXTURE METADATA ENABLE/DISABLE */
1576 si_resident_handles_update_needs_color_decompress(struct si_context
*sctx
)
1578 util_dynarray_clear(&sctx
->resident_tex_needs_color_decompress
);
1579 util_dynarray_clear(&sctx
->resident_img_needs_color_decompress
);
1581 util_dynarray_foreach(&sctx
->resident_tex_handles
,
1582 struct si_texture_handle
*, tex_handle
) {
1583 struct pipe_resource
*res
= (*tex_handle
)->view
->texture
;
1584 struct si_texture
*tex
;
1586 if (!res
|| res
->target
== PIPE_BUFFER
)
1589 tex
= (struct si_texture
*)res
;
1590 if (!color_needs_decompression(tex
))
1593 util_dynarray_append(&sctx
->resident_tex_needs_color_decompress
,
1594 struct si_texture_handle
*, *tex_handle
);
1597 util_dynarray_foreach(&sctx
->resident_img_handles
,
1598 struct si_image_handle
*, img_handle
) {
1599 struct pipe_image_view
*view
= &(*img_handle
)->view
;
1600 struct pipe_resource
*res
= view
->resource
;
1601 struct si_texture
*tex
;
1603 if (!res
|| res
->target
== PIPE_BUFFER
)
1606 tex
= (struct si_texture
*)res
;
1607 if (!color_needs_decompression(tex
))
1610 util_dynarray_append(&sctx
->resident_img_needs_color_decompress
,
1611 struct si_image_handle
*, *img_handle
);
1615 /* CMASK can be enabled (for fast clear) and disabled (for texture export)
1616 * while the texture is bound, possibly by a different context. In that case,
1617 * call this function to update needs_*_decompress_masks.
1619 void si_update_needs_color_decompress_masks(struct si_context
*sctx
)
1621 for (int i
= 0; i
< SI_NUM_SHADERS
; ++i
) {
1622 si_samplers_update_needs_color_decompress_mask(&sctx
->samplers
[i
]);
1623 si_images_update_needs_color_decompress_mask(&sctx
->images
[i
]);
1624 si_update_shader_needs_decompress_mask(sctx
, i
);
1627 si_resident_handles_update_needs_color_decompress(sctx
);
1630 /* BUFFER DISCARD/INVALIDATION */
1632 /* Reset descriptors of buffer resources after \p buf has been invalidated.
1633 * If buf == NULL, reset all descriptors.
1635 static void si_reset_buffer_resources(struct si_context
*sctx
,
1636 struct si_buffer_resources
*buffers
,
1637 unsigned descriptors_idx
,
1639 struct pipe_resource
*buf
,
1640 enum radeon_bo_priority priority
)
1642 struct si_descriptors
*descs
= &sctx
->descriptors
[descriptors_idx
];
1643 unsigned mask
= buffers
->enabled_mask
& slot_mask
;
1646 unsigned i
= u_bit_scan(&mask
);
1647 struct pipe_resource
*buffer
= buffers
->buffers
[i
];
1649 if (buffer
&& (!buf
|| buffer
== buf
)) {
1650 si_set_buf_desc_address(si_resource(buffer
), buffers
->offsets
[i
],
1652 sctx
->descriptors_dirty
|= 1u << descriptors_idx
;
1654 radeon_add_to_gfx_buffer_list_check_mem(sctx
,
1655 si_resource(buffer
),
1656 buffers
->writable_mask
& (1u << i
) ?
1657 RADEON_USAGE_READWRITE
:
1664 /* Update all buffer bindings where the buffer is bound, including
1665 * all resource descriptors. This is invalidate_buffer without
1668 * If buf == NULL, update all buffer bindings.
1670 void si_rebind_buffer(struct si_context
*sctx
, struct pipe_resource
*buf
)
1672 struct si_resource
*buffer
= si_resource(buf
);
1674 unsigned num_elems
= sctx
->vertex_elements
?
1675 sctx
->vertex_elements
->count
: 0;
1677 /* We changed the buffer, now we need to bind it where the old one
1678 * was bound. This consists of 2 things:
1679 * 1) Updating the resource descriptor and dirtying it.
1680 * 2) Adding a relocation to the CS, so that it's usable.
1683 /* Vertex buffers. */
1686 sctx
->vertex_buffers_dirty
= true;
1687 } else if (buffer
->bind_history
& PIPE_BIND_VERTEX_BUFFER
) {
1688 for (i
= 0; i
< num_elems
; i
++) {
1689 int vb
= sctx
->vertex_elements
->vertex_buffer_index
[i
];
1691 if (vb
>= ARRAY_SIZE(sctx
->vertex_buffer
))
1693 if (!sctx
->vertex_buffer
[vb
].buffer
.resource
)
1696 if (sctx
->vertex_buffer
[vb
].buffer
.resource
== buf
) {
1697 sctx
->vertex_buffers_dirty
= true;
1703 /* Streamout buffers. (other internal buffers can't be invalidated) */
1704 if (!buffer
|| buffer
->bind_history
& PIPE_BIND_STREAM_OUTPUT
) {
1705 for (i
= SI_VS_STREAMOUT_BUF0
; i
<= SI_VS_STREAMOUT_BUF3
; i
++) {
1706 struct si_buffer_resources
*buffers
= &sctx
->rw_buffers
;
1707 struct si_descriptors
*descs
=
1708 &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
];
1709 struct pipe_resource
*buffer
= buffers
->buffers
[i
];
1711 if (!buffer
|| (buf
&& buffer
!= buf
))
1714 si_set_buf_desc_address(si_resource(buffer
), buffers
->offsets
[i
],
1716 sctx
->descriptors_dirty
|= 1u << SI_DESCS_RW_BUFFERS
;
1718 radeon_add_to_gfx_buffer_list_check_mem(sctx
,
1719 si_resource(buffer
),
1721 RADEON_PRIO_SHADER_RW_BUFFER
,
1724 /* Update the streamout state. */
1725 if (sctx
->streamout
.begin_emitted
)
1726 si_emit_streamout_end(sctx
);
1727 sctx
->streamout
.append_bitmask
=
1728 sctx
->streamout
.enabled_mask
;
1729 si_streamout_buffers_dirty(sctx
);
1733 /* Constant and shader buffers. */
1734 if (!buffer
|| buffer
->bind_history
& PIPE_BIND_CONSTANT_BUFFER
) {
1735 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++)
1736 si_reset_buffer_resources(sctx
, &sctx
->const_and_shader_buffers
[shader
],
1737 si_const_and_shader_buffer_descriptors_idx(shader
),
1738 u_bit_consecutive(SI_NUM_SHADER_BUFFERS
, SI_NUM_CONST_BUFFERS
),
1740 sctx
->const_and_shader_buffers
[shader
].priority_constbuf
);
1743 if (!buffer
|| buffer
->bind_history
& PIPE_BIND_SHADER_BUFFER
) {
1744 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++)
1745 si_reset_buffer_resources(sctx
, &sctx
->const_and_shader_buffers
[shader
],
1746 si_const_and_shader_buffer_descriptors_idx(shader
),
1747 u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS
),
1749 sctx
->const_and_shader_buffers
[shader
].priority
);
1752 if (!buffer
|| buffer
->bind_history
& PIPE_BIND_SAMPLER_VIEW
) {
1753 /* Texture buffers - update bindings. */
1754 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++) {
1755 struct si_samplers
*samplers
= &sctx
->samplers
[shader
];
1756 struct si_descriptors
*descs
=
1757 si_sampler_and_image_descriptors(sctx
, shader
);
1758 unsigned mask
= samplers
->enabled_mask
;
1761 unsigned i
= u_bit_scan(&mask
);
1762 struct pipe_resource
*buffer
= samplers
->views
[i
]->texture
;
1764 if (buffer
&& buffer
->target
== PIPE_BUFFER
&&
1765 (!buf
|| buffer
== buf
)) {
1766 unsigned desc_slot
= si_get_sampler_slot(i
);
1768 si_set_buf_desc_address(si_resource(buffer
),
1769 samplers
->views
[i
]->u
.buf
.offset
,
1770 descs
->list
+ desc_slot
* 16 + 4);
1771 sctx
->descriptors_dirty
|=
1772 1u << si_sampler_and_image_descriptors_idx(shader
);
1774 radeon_add_to_gfx_buffer_list_check_mem(
1775 sctx
, si_resource(buffer
),
1777 RADEON_PRIO_SAMPLER_BUFFER
, true);
1784 if (!buffer
|| buffer
->bind_history
& PIPE_BIND_SHADER_IMAGE
) {
1785 for (shader
= 0; shader
< SI_NUM_SHADERS
; ++shader
) {
1786 struct si_images
*images
= &sctx
->images
[shader
];
1787 struct si_descriptors
*descs
=
1788 si_sampler_and_image_descriptors(sctx
, shader
);
1789 unsigned mask
= images
->enabled_mask
;
1792 unsigned i
= u_bit_scan(&mask
);
1793 struct pipe_resource
*buffer
= images
->views
[i
].resource
;
1795 if (buffer
&& buffer
->target
== PIPE_BUFFER
&&
1796 (!buf
|| buffer
== buf
)) {
1797 unsigned desc_slot
= si_get_image_slot(i
);
1799 if (images
->views
[i
].access
& PIPE_IMAGE_ACCESS_WRITE
)
1800 si_mark_image_range_valid(&images
->views
[i
]);
1802 si_set_buf_desc_address(si_resource(buffer
),
1803 images
->views
[i
].u
.buf
.offset
,
1804 descs
->list
+ desc_slot
* 8 + 4);
1805 sctx
->descriptors_dirty
|=
1806 1u << si_sampler_and_image_descriptors_idx(shader
);
1808 radeon_add_to_gfx_buffer_list_check_mem(
1809 sctx
, si_resource(buffer
),
1810 RADEON_USAGE_READWRITE
,
1811 RADEON_PRIO_SAMPLER_BUFFER
, true);
1817 /* Bindless texture handles */
1818 if (!buffer
|| buffer
->texture_handle_allocated
) {
1819 struct si_descriptors
*descs
= &sctx
->bindless_descriptors
;
1821 util_dynarray_foreach(&sctx
->resident_tex_handles
,
1822 struct si_texture_handle
*, tex_handle
) {
1823 struct pipe_sampler_view
*view
= (*tex_handle
)->view
;
1824 unsigned desc_slot
= (*tex_handle
)->desc_slot
;
1825 struct pipe_resource
*buffer
= view
->texture
;
1827 if (buffer
&& buffer
->target
== PIPE_BUFFER
&&
1828 (!buf
|| buffer
== buf
)) {
1829 si_set_buf_desc_address(si_resource(buffer
),
1832 desc_slot
* 16 + 4);
1834 (*tex_handle
)->desc_dirty
= true;
1835 sctx
->bindless_descriptors_dirty
= true;
1837 radeon_add_to_gfx_buffer_list_check_mem(
1838 sctx
, si_resource(buffer
),
1840 RADEON_PRIO_SAMPLER_BUFFER
, true);
1845 /* Bindless image handles */
1846 if (!buffer
|| buffer
->image_handle_allocated
) {
1847 struct si_descriptors
*descs
= &sctx
->bindless_descriptors
;
1849 util_dynarray_foreach(&sctx
->resident_img_handles
,
1850 struct si_image_handle
*, img_handle
) {
1851 struct pipe_image_view
*view
= &(*img_handle
)->view
;
1852 unsigned desc_slot
= (*img_handle
)->desc_slot
;
1853 struct pipe_resource
*buffer
= view
->resource
;
1855 if (buffer
&& buffer
->target
== PIPE_BUFFER
&&
1856 (!buf
|| buffer
== buf
)) {
1857 if (view
->access
& PIPE_IMAGE_ACCESS_WRITE
)
1858 si_mark_image_range_valid(view
);
1860 si_set_buf_desc_address(si_resource(buffer
),
1863 desc_slot
* 16 + 4);
1865 (*img_handle
)->desc_dirty
= true;
1866 sctx
->bindless_descriptors_dirty
= true;
1868 radeon_add_to_gfx_buffer_list_check_mem(
1869 sctx
, si_resource(buffer
),
1870 RADEON_USAGE_READWRITE
,
1871 RADEON_PRIO_SAMPLER_BUFFER
, true);
1877 /* Do the same for other contexts. They will invoke this function
1878 * with buffer == NULL.
1880 unsigned new_counter
= p_atomic_inc_return(&sctx
->screen
->dirty_buf_counter
);
1882 /* Skip the update for the current context, because we have already updated
1883 * the buffer bindings.
1885 if (new_counter
== sctx
->last_dirty_buf_counter
+ 1)
1886 sctx
->last_dirty_buf_counter
= new_counter
;
1890 static void si_upload_bindless_descriptor(struct si_context
*sctx
,
1892 unsigned num_dwords
)
1894 struct si_descriptors
*desc
= &sctx
->bindless_descriptors
;
1895 unsigned desc_slot_offset
= desc_slot
* 16;
1899 data
= desc
->list
+ desc_slot_offset
;
1900 va
= desc
->gpu_address
+ desc_slot_offset
* 4;
1902 si_cp_write_data(sctx
, desc
->buffer
, va
- desc
->buffer
->gpu_address
,
1903 num_dwords
* 4, V_370_TC_L2
, V_370_ME
, data
);
1906 static void si_upload_bindless_descriptors(struct si_context
*sctx
)
1908 if (!sctx
->bindless_descriptors_dirty
)
1911 /* Wait for graphics/compute to be idle before updating the resident
1912 * descriptors directly in memory, in case the GPU is using them.
1914 sctx
->flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
1915 SI_CONTEXT_CS_PARTIAL_FLUSH
;
1916 sctx
->emit_cache_flush(sctx
);
1918 util_dynarray_foreach(&sctx
->resident_tex_handles
,
1919 struct si_texture_handle
*, tex_handle
) {
1920 unsigned desc_slot
= (*tex_handle
)->desc_slot
;
1922 if (!(*tex_handle
)->desc_dirty
)
1925 si_upload_bindless_descriptor(sctx
, desc_slot
, 16);
1926 (*tex_handle
)->desc_dirty
= false;
1929 util_dynarray_foreach(&sctx
->resident_img_handles
,
1930 struct si_image_handle
*, img_handle
) {
1931 unsigned desc_slot
= (*img_handle
)->desc_slot
;
1933 if (!(*img_handle
)->desc_dirty
)
1936 si_upload_bindless_descriptor(sctx
, desc_slot
, 8);
1937 (*img_handle
)->desc_dirty
= false;
1940 /* Invalidate L1 because it doesn't know that L2 changed. */
1941 sctx
->flags
|= SI_CONTEXT_INV_SCACHE
;
1942 sctx
->emit_cache_flush(sctx
);
1944 sctx
->bindless_descriptors_dirty
= false;
1947 /* Update mutable image descriptor fields of all resident textures. */
1948 static void si_update_bindless_texture_descriptor(struct si_context
*sctx
,
1949 struct si_texture_handle
*tex_handle
)
1951 struct si_sampler_view
*sview
= (struct si_sampler_view
*)tex_handle
->view
;
1952 struct si_descriptors
*desc
= &sctx
->bindless_descriptors
;
1953 unsigned desc_slot_offset
= tex_handle
->desc_slot
* 16;
1954 uint32_t desc_list
[16];
1956 if (sview
->base
.texture
->target
== PIPE_BUFFER
)
1959 memcpy(desc_list
, desc
->list
+ desc_slot_offset
, sizeof(desc_list
));
1960 si_set_sampler_view_desc(sctx
, sview
, &tex_handle
->sstate
,
1961 desc
->list
+ desc_slot_offset
);
1963 if (memcmp(desc_list
, desc
->list
+ desc_slot_offset
,
1964 sizeof(desc_list
))) {
1965 tex_handle
->desc_dirty
= true;
1966 sctx
->bindless_descriptors_dirty
= true;
1970 static void si_update_bindless_image_descriptor(struct si_context
*sctx
,
1971 struct si_image_handle
*img_handle
)
1973 struct si_descriptors
*desc
= &sctx
->bindless_descriptors
;
1974 unsigned desc_slot_offset
= img_handle
->desc_slot
* 16;
1975 struct pipe_image_view
*view
= &img_handle
->view
;
1976 uint32_t desc_list
[8];
1978 if (view
->resource
->target
== PIPE_BUFFER
)
1981 memcpy(desc_list
, desc
->list
+ desc_slot_offset
,
1983 si_set_shader_image_desc(sctx
, view
, true,
1984 desc
->list
+ desc_slot_offset
, NULL
);
1986 if (memcmp(desc_list
, desc
->list
+ desc_slot_offset
,
1987 sizeof(desc_list
))) {
1988 img_handle
->desc_dirty
= true;
1989 sctx
->bindless_descriptors_dirty
= true;
1993 static void si_update_all_resident_texture_descriptors(struct si_context
*sctx
)
1995 util_dynarray_foreach(&sctx
->resident_tex_handles
,
1996 struct si_texture_handle
*, tex_handle
) {
1997 si_update_bindless_texture_descriptor(sctx
, *tex_handle
);
2000 util_dynarray_foreach(&sctx
->resident_img_handles
,
2001 struct si_image_handle
*, img_handle
) {
2002 si_update_bindless_image_descriptor(sctx
, *img_handle
);
2005 si_upload_bindless_descriptors(sctx
);
2008 /* Update mutable image descriptor fields of all bound textures. */
2009 void si_update_all_texture_descriptors(struct si_context
*sctx
)
2013 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++) {
2014 struct si_samplers
*samplers
= &sctx
->samplers
[shader
];
2015 struct si_images
*images
= &sctx
->images
[shader
];
2019 mask
= images
->enabled_mask
;
2021 unsigned i
= u_bit_scan(&mask
);
2022 struct pipe_image_view
*view
= &images
->views
[i
];
2024 if (!view
->resource
||
2025 view
->resource
->target
== PIPE_BUFFER
)
2028 si_set_shader_image(sctx
, shader
, i
, view
, true);
2031 /* Sampler views. */
2032 mask
= samplers
->enabled_mask
;
2034 unsigned i
= u_bit_scan(&mask
);
2035 struct pipe_sampler_view
*view
= samplers
->views
[i
];
2039 view
->texture
->target
== PIPE_BUFFER
)
2042 si_set_sampler_view(sctx
, shader
, i
,
2043 samplers
->views
[i
], true);
2046 si_update_shader_needs_decompress_mask(sctx
, shader
);
2049 si_update_all_resident_texture_descriptors(sctx
);
2050 si_update_ps_colorbuf0_slot(sctx
);
2053 /* SHADER USER DATA */
2055 static void si_mark_shader_pointers_dirty(struct si_context
*sctx
,
2058 sctx
->shader_pointers_dirty
|=
2059 u_bit_consecutive(SI_DESCS_FIRST_SHADER
+ shader
* SI_NUM_SHADER_DESCS
,
2060 SI_NUM_SHADER_DESCS
);
2062 if (shader
== PIPE_SHADER_VERTEX
)
2063 sctx
->vertex_buffer_pointer_dirty
= sctx
->vb_descriptors_buffer
!= NULL
;
2065 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.shader_pointers
);
2068 static void si_shader_pointers_begin_new_cs(struct si_context
*sctx
)
2070 sctx
->shader_pointers_dirty
= u_bit_consecutive(0, SI_NUM_DESCS
);
2071 sctx
->vertex_buffer_pointer_dirty
= sctx
->vb_descriptors_buffer
!= NULL
;
2072 si_mark_atom_dirty(sctx
, &sctx
->atoms
.s
.shader_pointers
);
2073 sctx
->graphics_bindless_pointer_dirty
= sctx
->bindless_descriptors
.buffer
!= NULL
;
2074 sctx
->compute_bindless_pointer_dirty
= sctx
->bindless_descriptors
.buffer
!= NULL
;
2077 /* Set a base register address for user data constants in the given shader.
2078 * This assigns a mapping from PIPE_SHADER_* to SPI_SHADER_USER_DATA_*.
2080 static void si_set_user_data_base(struct si_context
*sctx
,
2081 unsigned shader
, uint32_t new_base
)
2083 uint32_t *base
= &sctx
->shader_pointers
.sh_base
[shader
];
2085 if (*base
!= new_base
) {
2089 si_mark_shader_pointers_dirty(sctx
, shader
);
2091 /* Any change in enabled shader stages requires re-emitting
2092 * the VS state SGPR, because it contains the clamp_vertex_color
2093 * state, which can be done in VS, TES, and GS.
2095 sctx
->last_vs_state
= ~0;
2099 /* This must be called when these are changed between enabled and disabled
2101 * - tessellation evaluation shader
2104 void si_shader_change_notify(struct si_context
*sctx
)
2106 /* VS can be bound as VS, ES, or LS. */
2107 if (sctx
->tes_shader
.cso
) {
2108 if (sctx
->chip_class
>= GFX10
) {
2109 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2110 R_00B430_SPI_SHADER_USER_DATA_HS_0
);
2111 } else if (sctx
->chip_class
>= GFX9
) {
2112 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2113 R_00B430_SPI_SHADER_USER_DATA_LS_0
);
2115 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2116 R_00B530_SPI_SHADER_USER_DATA_LS_0
);
2118 } else if (sctx
->chip_class
>= GFX10
) {
2119 if (sctx
->ngg
|| sctx
->gs_shader
.cso
) {
2120 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2121 R_00B230_SPI_SHADER_USER_DATA_GS_0
);
2123 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2124 R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2126 } else if (sctx
->gs_shader
.cso
) {
2127 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2128 R_00B330_SPI_SHADER_USER_DATA_ES_0
);
2130 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2131 R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2134 /* TES can be bound as ES, VS, or not bound. */
2135 if (sctx
->tes_shader
.cso
) {
2136 if (sctx
->chip_class
>= GFX10
) {
2137 if (sctx
->ngg
|| sctx
->gs_shader
.cso
) {
2138 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_EVAL
,
2139 R_00B230_SPI_SHADER_USER_DATA_GS_0
);
2141 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_EVAL
,
2142 R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2144 } else if (sctx
->gs_shader
.cso
) {
2145 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_EVAL
,
2146 R_00B330_SPI_SHADER_USER_DATA_ES_0
);
2148 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_EVAL
,
2149 R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2152 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_EVAL
, 0);
2156 static void si_emit_shader_pointer_head(struct radeon_cmdbuf
*cs
,
2158 unsigned pointer_count
)
2160 radeon_emit(cs
, PKT3(PKT3_SET_SH_REG
, pointer_count
, 0));
2161 radeon_emit(cs
, (sh_offset
- SI_SH_REG_OFFSET
) >> 2);
2164 static void si_emit_shader_pointer_body(struct si_screen
*sscreen
,
2165 struct radeon_cmdbuf
*cs
,
2168 radeon_emit(cs
, va
);
2170 assert(va
== 0 || (va
>> 32) == sscreen
->info
.address32_hi
);
2173 static void si_emit_shader_pointer(struct si_context
*sctx
,
2174 struct si_descriptors
*desc
,
2177 struct radeon_cmdbuf
*cs
= sctx
->gfx_cs
;
2178 unsigned sh_offset
= sh_base
+ desc
->shader_userdata_offset
;
2180 si_emit_shader_pointer_head(cs
, sh_offset
, 1);
2181 si_emit_shader_pointer_body(sctx
->screen
, cs
, desc
->gpu_address
);
2184 static void si_emit_consecutive_shader_pointers(struct si_context
*sctx
,
2185 unsigned pointer_mask
,
2191 struct radeon_cmdbuf
*cs
= sctx
->gfx_cs
;
2192 unsigned mask
= sctx
->shader_pointers_dirty
& pointer_mask
;
2196 u_bit_scan_consecutive_range(&mask
, &start
, &count
);
2198 struct si_descriptors
*descs
= &sctx
->descriptors
[start
];
2199 unsigned sh_offset
= sh_base
+ descs
->shader_userdata_offset
;
2201 si_emit_shader_pointer_head(cs
, sh_offset
, count
);
2202 for (int i
= 0; i
< count
; i
++)
2203 si_emit_shader_pointer_body(sctx
->screen
, cs
,
2204 descs
[i
].gpu_address
);
2208 static void si_emit_global_shader_pointers(struct si_context
*sctx
,
2209 struct si_descriptors
*descs
)
2211 if (sctx
->chip_class
== GFX9
) {
2212 /* Broadcast it to all shader stages. */
2213 si_emit_shader_pointer(sctx
, descs
,
2214 R_00B530_SPI_SHADER_USER_DATA_COMMON_0
);
2218 si_emit_shader_pointer(sctx
, descs
,
2219 R_00B030_SPI_SHADER_USER_DATA_PS_0
);
2220 si_emit_shader_pointer(sctx
, descs
,
2221 R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2222 si_emit_shader_pointer(sctx
, descs
,
2223 R_00B330_SPI_SHADER_USER_DATA_ES_0
);
2224 si_emit_shader_pointer(sctx
, descs
,
2225 R_00B230_SPI_SHADER_USER_DATA_GS_0
);
2226 si_emit_shader_pointer(sctx
, descs
,
2227 R_00B430_SPI_SHADER_USER_DATA_HS_0
);
2228 si_emit_shader_pointer(sctx
, descs
,
2229 R_00B530_SPI_SHADER_USER_DATA_LS_0
);
2232 void si_emit_graphics_shader_pointers(struct si_context
*sctx
)
2234 uint32_t *sh_base
= sctx
->shader_pointers
.sh_base
;
2236 if (sctx
->shader_pointers_dirty
& (1 << SI_DESCS_RW_BUFFERS
)) {
2237 si_emit_global_shader_pointers(sctx
,
2238 &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
]);
2241 si_emit_consecutive_shader_pointers(sctx
, SI_DESCS_SHADER_MASK(VERTEX
),
2242 sh_base
[PIPE_SHADER_VERTEX
]);
2243 si_emit_consecutive_shader_pointers(sctx
, SI_DESCS_SHADER_MASK(TESS_EVAL
),
2244 sh_base
[PIPE_SHADER_TESS_EVAL
]);
2245 si_emit_consecutive_shader_pointers(sctx
, SI_DESCS_SHADER_MASK(FRAGMENT
),
2246 sh_base
[PIPE_SHADER_FRAGMENT
]);
2247 si_emit_consecutive_shader_pointers(sctx
, SI_DESCS_SHADER_MASK(TESS_CTRL
),
2248 sh_base
[PIPE_SHADER_TESS_CTRL
]);
2249 si_emit_consecutive_shader_pointers(sctx
, SI_DESCS_SHADER_MASK(GEOMETRY
),
2250 sh_base
[PIPE_SHADER_GEOMETRY
]);
2252 sctx
->shader_pointers_dirty
&=
2253 ~u_bit_consecutive(SI_DESCS_RW_BUFFERS
, SI_DESCS_FIRST_COMPUTE
);
2255 if (sctx
->vertex_buffer_pointer_dirty
) {
2256 struct radeon_cmdbuf
*cs
= sctx
->gfx_cs
;
2258 /* Find the location of the VB descriptor pointer. */
2259 /* TODO: In the future, the pointer will be packed in unused
2260 * bits of the first 2 VB descriptors. */
2261 unsigned sh_dw_offset
= SI_VS_NUM_USER_SGPR
;
2262 if (sctx
->chip_class
>= GFX9
) {
2263 if (sctx
->tes_shader
.cso
)
2264 sh_dw_offset
= GFX9_TCS_NUM_USER_SGPR
;
2265 else if (sctx
->gs_shader
.cso
)
2266 sh_dw_offset
= GFX9_VSGS_NUM_USER_SGPR
;
2269 unsigned sh_offset
= sh_base
[PIPE_SHADER_VERTEX
] + sh_dw_offset
* 4;
2270 si_emit_shader_pointer_head(cs
, sh_offset
, 1);
2271 si_emit_shader_pointer_body(sctx
->screen
, cs
,
2272 sctx
->vb_descriptors_buffer
->gpu_address
+
2273 sctx
->vb_descriptors_offset
);
2274 sctx
->vertex_buffer_pointer_dirty
= false;
2277 if (sctx
->graphics_bindless_pointer_dirty
) {
2278 si_emit_global_shader_pointers(sctx
,
2279 &sctx
->bindless_descriptors
);
2280 sctx
->graphics_bindless_pointer_dirty
= false;
2284 void si_emit_compute_shader_pointers(struct si_context
*sctx
)
2286 unsigned base
= R_00B900_COMPUTE_USER_DATA_0
;
2288 si_emit_consecutive_shader_pointers(sctx
, SI_DESCS_SHADER_MASK(COMPUTE
),
2289 R_00B900_COMPUTE_USER_DATA_0
);
2290 sctx
->shader_pointers_dirty
&= ~SI_DESCS_SHADER_MASK(COMPUTE
);
2292 if (sctx
->compute_bindless_pointer_dirty
) {
2293 si_emit_shader_pointer(sctx
, &sctx
->bindless_descriptors
, base
);
2294 sctx
->compute_bindless_pointer_dirty
= false;
2300 static void si_init_bindless_descriptors(struct si_context
*sctx
,
2301 struct si_descriptors
*desc
,
2302 short shader_userdata_rel_index
,
2303 unsigned num_elements
)
2305 MAYBE_UNUSED
unsigned desc_slot
;
2307 si_init_descriptors(desc
, shader_userdata_rel_index
, 16, num_elements
);
2308 sctx
->bindless_descriptors
.num_active_slots
= num_elements
;
2310 /* The first bindless descriptor is stored at slot 1, because 0 is not
2311 * considered to be a valid handle.
2313 sctx
->num_bindless_descriptors
= 1;
2315 /* Track which bindless slots are used (or not). */
2316 util_idalloc_init(&sctx
->bindless_used_slots
);
2317 util_idalloc_resize(&sctx
->bindless_used_slots
, num_elements
);
2319 /* Reserve slot 0 because it's an invalid handle for bindless. */
2320 desc_slot
= util_idalloc_alloc(&sctx
->bindless_used_slots
);
2321 assert(desc_slot
== 0);
2324 static void si_release_bindless_descriptors(struct si_context
*sctx
)
2326 si_release_descriptors(&sctx
->bindless_descriptors
);
2327 util_idalloc_fini(&sctx
->bindless_used_slots
);
2330 static unsigned si_get_first_free_bindless_slot(struct si_context
*sctx
)
2332 struct si_descriptors
*desc
= &sctx
->bindless_descriptors
;
2335 desc_slot
= util_idalloc_alloc(&sctx
->bindless_used_slots
);
2336 if (desc_slot
>= desc
->num_elements
) {
2337 /* The array of bindless descriptors is full, resize it. */
2338 unsigned slot_size
= desc
->element_dw_size
* 4;
2339 unsigned new_num_elements
= desc
->num_elements
* 2;
2341 desc
->list
= REALLOC(desc
->list
, desc
->num_elements
* slot_size
,
2342 new_num_elements
* slot_size
);
2343 desc
->num_elements
= new_num_elements
;
2344 desc
->num_active_slots
= new_num_elements
;
2352 si_create_bindless_descriptor(struct si_context
*sctx
, uint32_t *desc_list
,
2355 struct si_descriptors
*desc
= &sctx
->bindless_descriptors
;
2356 unsigned desc_slot
, desc_slot_offset
;
2358 /* Find a free slot. */
2359 desc_slot
= si_get_first_free_bindless_slot(sctx
);
2361 /* For simplicity, sampler and image bindless descriptors use fixed
2362 * 16-dword slots for now. Image descriptors only need 8-dword but this
2363 * doesn't really matter because no real apps use image handles.
2365 desc_slot_offset
= desc_slot
* 16;
2367 /* Copy the descriptor into the array. */
2368 memcpy(desc
->list
+ desc_slot_offset
, desc_list
, size
);
2370 /* Re-upload the whole array of bindless descriptors into a new buffer.
2372 if (!si_upload_descriptors(sctx
, desc
))
2375 /* Make sure to re-emit the shader pointers for all stages. */
2376 sctx
->graphics_bindless_pointer_dirty
= true;
2377 sctx
->compute_bindless_pointer_dirty
= true;
2382 static void si_update_bindless_buffer_descriptor(struct si_context
*sctx
,
2384 struct pipe_resource
*resource
,
2388 struct si_descriptors
*desc
= &sctx
->bindless_descriptors
;
2389 struct si_resource
*buf
= si_resource(resource
);
2390 unsigned desc_slot_offset
= desc_slot
* 16;
2391 uint32_t *desc_list
= desc
->list
+ desc_slot_offset
+ 4;
2392 uint64_t old_desc_va
;
2394 assert(resource
->target
== PIPE_BUFFER
);
2396 /* Retrieve the old buffer addr from the descriptor. */
2397 old_desc_va
= si_desc_extract_buffer_address(desc_list
);
2399 if (old_desc_va
!= buf
->gpu_address
+ offset
) {
2400 /* The buffer has been invalidated when the handle wasn't
2401 * resident, update the descriptor and the dirty flag.
2403 si_set_buf_desc_address(buf
, offset
, &desc_list
[0]);
2409 static uint64_t si_create_texture_handle(struct pipe_context
*ctx
,
2410 struct pipe_sampler_view
*view
,
2411 const struct pipe_sampler_state
*state
)
2413 struct si_sampler_view
*sview
= (struct si_sampler_view
*)view
;
2414 struct si_context
*sctx
= (struct si_context
*)ctx
;
2415 struct si_texture_handle
*tex_handle
;
2416 struct si_sampler_state
*sstate
;
2417 uint32_t desc_list
[16];
2420 tex_handle
= CALLOC_STRUCT(si_texture_handle
);
2424 memset(desc_list
, 0, sizeof(desc_list
));
2425 si_init_descriptor_list(&desc_list
[0], 16, 1, null_texture_descriptor
);
2427 sstate
= ctx
->create_sampler_state(ctx
, state
);
2433 si_set_sampler_view_desc(sctx
, sview
, sstate
, &desc_list
[0]);
2434 memcpy(&tex_handle
->sstate
, sstate
, sizeof(*sstate
));
2435 ctx
->delete_sampler_state(ctx
, sstate
);
2437 tex_handle
->desc_slot
= si_create_bindless_descriptor(sctx
, desc_list
,
2439 if (!tex_handle
->desc_slot
) {
2444 handle
= tex_handle
->desc_slot
;
2446 if (!_mesa_hash_table_insert(sctx
->tex_handles
,
2447 (void *)(uintptr_t)handle
,
2453 pipe_sampler_view_reference(&tex_handle
->view
, view
);
2455 si_resource(sview
->base
.texture
)->texture_handle_allocated
= true;
2460 static void si_delete_texture_handle(struct pipe_context
*ctx
, uint64_t handle
)
2462 struct si_context
*sctx
= (struct si_context
*)ctx
;
2463 struct si_texture_handle
*tex_handle
;
2464 struct hash_entry
*entry
;
2466 entry
= _mesa_hash_table_search(sctx
->tex_handles
,
2467 (void *)(uintptr_t)handle
);
2471 tex_handle
= (struct si_texture_handle
*)entry
->data
;
2473 /* Allow this descriptor slot to be re-used. */
2474 util_idalloc_free(&sctx
->bindless_used_slots
, tex_handle
->desc_slot
);
2476 pipe_sampler_view_reference(&tex_handle
->view
, NULL
);
2477 _mesa_hash_table_remove(sctx
->tex_handles
, entry
);
2481 static void si_make_texture_handle_resident(struct pipe_context
*ctx
,
2482 uint64_t handle
, bool resident
)
2484 struct si_context
*sctx
= (struct si_context
*)ctx
;
2485 struct si_texture_handle
*tex_handle
;
2486 struct si_sampler_view
*sview
;
2487 struct hash_entry
*entry
;
2489 entry
= _mesa_hash_table_search(sctx
->tex_handles
,
2490 (void *)(uintptr_t)handle
);
2494 tex_handle
= (struct si_texture_handle
*)entry
->data
;
2495 sview
= (struct si_sampler_view
*)tex_handle
->view
;
2498 if (sview
->base
.texture
->target
!= PIPE_BUFFER
) {
2499 struct si_texture
*tex
=
2500 (struct si_texture
*)sview
->base
.texture
;
2502 if (depth_needs_decompression(tex
)) {
2503 util_dynarray_append(
2504 &sctx
->resident_tex_needs_depth_decompress
,
2505 struct si_texture_handle
*,
2509 if (color_needs_decompression(tex
)) {
2510 util_dynarray_append(
2511 &sctx
->resident_tex_needs_color_decompress
,
2512 struct si_texture_handle
*,
2516 if (tex
->dcc_offset
&&
2517 p_atomic_read(&tex
->framebuffers_bound
))
2518 sctx
->need_check_render_feedback
= true;
2520 si_update_bindless_texture_descriptor(sctx
, tex_handle
);
2522 si_update_bindless_buffer_descriptor(sctx
,
2523 tex_handle
->desc_slot
,
2524 sview
->base
.texture
,
2525 sview
->base
.u
.buf
.offset
,
2526 &tex_handle
->desc_dirty
);
2529 /* Re-upload the descriptor if it has been updated while it
2532 if (tex_handle
->desc_dirty
)
2533 sctx
->bindless_descriptors_dirty
= true;
2535 /* Add the texture handle to the per-context list. */
2536 util_dynarray_append(&sctx
->resident_tex_handles
,
2537 struct si_texture_handle
*, tex_handle
);
2539 /* Add the buffers to the current CS in case si_begin_new_cs()
2540 * is not going to be called.
2542 si_sampler_view_add_buffer(sctx
, sview
->base
.texture
,
2544 sview
->is_stencil_sampler
, false);
2546 /* Remove the texture handle from the per-context list. */
2547 util_dynarray_delete_unordered(&sctx
->resident_tex_handles
,
2548 struct si_texture_handle
*,
2551 if (sview
->base
.texture
->target
!= PIPE_BUFFER
) {
2552 util_dynarray_delete_unordered(
2553 &sctx
->resident_tex_needs_depth_decompress
,
2554 struct si_texture_handle
*, tex_handle
);
2556 util_dynarray_delete_unordered(
2557 &sctx
->resident_tex_needs_color_decompress
,
2558 struct si_texture_handle
*, tex_handle
);
2563 static uint64_t si_create_image_handle(struct pipe_context
*ctx
,
2564 const struct pipe_image_view
*view
)
2566 struct si_context
*sctx
= (struct si_context
*)ctx
;
2567 struct si_image_handle
*img_handle
;
2568 uint32_t desc_list
[8];
2571 if (!view
|| !view
->resource
)
2574 img_handle
= CALLOC_STRUCT(si_image_handle
);
2578 memset(desc_list
, 0, sizeof(desc_list
));
2579 si_init_descriptor_list(&desc_list
[0], 8, 1, null_image_descriptor
);
2581 si_set_shader_image_desc(sctx
, view
, false, &desc_list
[0], NULL
);
2583 img_handle
->desc_slot
= si_create_bindless_descriptor(sctx
, desc_list
,
2585 if (!img_handle
->desc_slot
) {
2590 handle
= img_handle
->desc_slot
;
2592 if (!_mesa_hash_table_insert(sctx
->img_handles
,
2593 (void *)(uintptr_t)handle
,
2599 util_copy_image_view(&img_handle
->view
, view
);
2601 si_resource(view
->resource
)->image_handle_allocated
= true;
2606 static void si_delete_image_handle(struct pipe_context
*ctx
, uint64_t handle
)
2608 struct si_context
*sctx
= (struct si_context
*)ctx
;
2609 struct si_image_handle
*img_handle
;
2610 struct hash_entry
*entry
;
2612 entry
= _mesa_hash_table_search(sctx
->img_handles
,
2613 (void *)(uintptr_t)handle
);
2617 img_handle
= (struct si_image_handle
*)entry
->data
;
2619 util_copy_image_view(&img_handle
->view
, NULL
);
2620 _mesa_hash_table_remove(sctx
->img_handles
, entry
);
2624 static void si_make_image_handle_resident(struct pipe_context
*ctx
,
2625 uint64_t handle
, unsigned access
,
2628 struct si_context
*sctx
= (struct si_context
*)ctx
;
2629 struct si_image_handle
*img_handle
;
2630 struct pipe_image_view
*view
;
2631 struct si_resource
*res
;
2632 struct hash_entry
*entry
;
2634 entry
= _mesa_hash_table_search(sctx
->img_handles
,
2635 (void *)(uintptr_t)handle
);
2639 img_handle
= (struct si_image_handle
*)entry
->data
;
2640 view
= &img_handle
->view
;
2641 res
= si_resource(view
->resource
);
2644 if (res
->b
.b
.target
!= PIPE_BUFFER
) {
2645 struct si_texture
*tex
= (struct si_texture
*)res
;
2646 unsigned level
= view
->u
.tex
.level
;
2648 if (color_needs_decompression(tex
)) {
2649 util_dynarray_append(
2650 &sctx
->resident_img_needs_color_decompress
,
2651 struct si_image_handle
*,
2655 if (vi_dcc_enabled(tex
, level
) &&
2656 p_atomic_read(&tex
->framebuffers_bound
))
2657 sctx
->need_check_render_feedback
= true;
2659 si_update_bindless_image_descriptor(sctx
, img_handle
);
2661 si_update_bindless_buffer_descriptor(sctx
,
2662 img_handle
->desc_slot
,
2665 &img_handle
->desc_dirty
);
2668 /* Re-upload the descriptor if it has been updated while it
2671 if (img_handle
->desc_dirty
)
2672 sctx
->bindless_descriptors_dirty
= true;
2674 /* Add the image handle to the per-context list. */
2675 util_dynarray_append(&sctx
->resident_img_handles
,
2676 struct si_image_handle
*, img_handle
);
2678 /* Add the buffers to the current CS in case si_begin_new_cs()
2679 * is not going to be called.
2681 si_sampler_view_add_buffer(sctx
, view
->resource
,
2682 (access
& PIPE_IMAGE_ACCESS_WRITE
) ?
2683 RADEON_USAGE_READWRITE
:
2684 RADEON_USAGE_READ
, false, false);
2686 /* Remove the image handle from the per-context list. */
2687 util_dynarray_delete_unordered(&sctx
->resident_img_handles
,
2688 struct si_image_handle
*,
2691 if (res
->b
.b
.target
!= PIPE_BUFFER
) {
2692 util_dynarray_delete_unordered(
2693 &sctx
->resident_img_needs_color_decompress
,
2694 struct si_image_handle
*,
2700 static void si_resident_buffers_add_all_to_bo_list(struct si_context
*sctx
)
2702 unsigned num_resident_tex_handles
, num_resident_img_handles
;
2704 num_resident_tex_handles
= sctx
->resident_tex_handles
.size
/
2705 sizeof(struct si_texture_handle
*);
2706 num_resident_img_handles
= sctx
->resident_img_handles
.size
/
2707 sizeof(struct si_image_handle
*);
2709 /* Add all resident texture handles. */
2710 util_dynarray_foreach(&sctx
->resident_tex_handles
,
2711 struct si_texture_handle
*, tex_handle
) {
2712 struct si_sampler_view
*sview
=
2713 (struct si_sampler_view
*)(*tex_handle
)->view
;
2715 si_sampler_view_add_buffer(sctx
, sview
->base
.texture
,
2717 sview
->is_stencil_sampler
, false);
2720 /* Add all resident image handles. */
2721 util_dynarray_foreach(&sctx
->resident_img_handles
,
2722 struct si_image_handle
*, img_handle
) {
2723 struct pipe_image_view
*view
= &(*img_handle
)->view
;
2725 si_sampler_view_add_buffer(sctx
, view
->resource
,
2726 RADEON_USAGE_READWRITE
,
2730 sctx
->num_resident_handles
+= num_resident_tex_handles
+
2731 num_resident_img_handles
;
2732 assert(sctx
->bo_list_add_all_resident_resources
);
2733 sctx
->bo_list_add_all_resident_resources
= false;
2736 /* INIT/DEINIT/UPLOAD */
2738 void si_init_all_descriptors(struct si_context
*sctx
)
2741 unsigned first_shader
=
2742 sctx
->has_graphics
? 0 : PIPE_SHADER_COMPUTE
;
2744 for (i
= first_shader
; i
< SI_NUM_SHADERS
; i
++) {
2745 bool is_2nd
= sctx
->chip_class
>= GFX9
&&
2746 (i
== PIPE_SHADER_TESS_CTRL
||
2747 i
== PIPE_SHADER_GEOMETRY
);
2748 unsigned num_sampler_slots
= SI_NUM_IMAGES
/ 2 + SI_NUM_SAMPLERS
;
2749 unsigned num_buffer_slots
= SI_NUM_SHADER_BUFFERS
+ SI_NUM_CONST_BUFFERS
;
2751 struct si_descriptors
*desc
;
2754 if (i
== PIPE_SHADER_TESS_CTRL
) {
2755 rel_dw_offset
= (R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
-
2756 R_00B430_SPI_SHADER_USER_DATA_LS_0
) / 4;
2757 } else if (sctx
->chip_class
>= GFX10
) { /* PIPE_SHADER_GEOMETRY */
2758 rel_dw_offset
= (R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
-
2759 R_00B230_SPI_SHADER_USER_DATA_GS_0
) / 4;
2761 rel_dw_offset
= (R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
-
2762 R_00B330_SPI_SHADER_USER_DATA_ES_0
) / 4;
2765 rel_dw_offset
= SI_SGPR_CONST_AND_SHADER_BUFFERS
;
2767 desc
= si_const_and_shader_buffer_descriptors(sctx
, i
);
2768 si_init_buffer_resources(&sctx
->const_and_shader_buffers
[i
], desc
,
2769 num_buffer_slots
, rel_dw_offset
,
2770 RADEON_PRIO_SHADER_RW_BUFFER
,
2771 RADEON_PRIO_CONST_BUFFER
);
2772 desc
->slot_index_to_bind_directly
= si_get_constbuf_slot(0);
2775 if (i
== PIPE_SHADER_TESS_CTRL
) {
2776 rel_dw_offset
= (R_00B40C_SPI_SHADER_USER_DATA_ADDR_HI_HS
-
2777 R_00B430_SPI_SHADER_USER_DATA_LS_0
) / 4;
2778 } else if (sctx
->chip_class
>= GFX10
) { /* PIPE_SHADER_GEOMETRY */
2779 rel_dw_offset
= (R_00B20C_SPI_SHADER_USER_DATA_ADDR_HI_GS
-
2780 R_00B230_SPI_SHADER_USER_DATA_GS_0
) / 4;
2782 rel_dw_offset
= (R_00B20C_SPI_SHADER_USER_DATA_ADDR_HI_GS
-
2783 R_00B330_SPI_SHADER_USER_DATA_ES_0
) / 4;
2786 rel_dw_offset
= SI_SGPR_SAMPLERS_AND_IMAGES
;
2789 desc
= si_sampler_and_image_descriptors(sctx
, i
);
2790 si_init_descriptors(desc
, rel_dw_offset
, 16, num_sampler_slots
);
2793 for (j
= 0; j
< SI_NUM_IMAGES
; j
++)
2794 memcpy(desc
->list
+ j
* 8, null_image_descriptor
, 8 * 4);
2795 for (; j
< SI_NUM_IMAGES
+ SI_NUM_SAMPLERS
* 2; j
++)
2796 memcpy(desc
->list
+ j
* 8, null_texture_descriptor
, 8 * 4);
2799 si_init_buffer_resources(&sctx
->rw_buffers
,
2800 &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
],
2801 SI_NUM_RW_BUFFERS
, SI_SGPR_RW_BUFFERS
,
2802 /* The second priority is used by
2803 * const buffers in RW buffer slots. */
2804 RADEON_PRIO_SHADER_RINGS
, RADEON_PRIO_CONST_BUFFER
);
2805 sctx
->descriptors
[SI_DESCS_RW_BUFFERS
].num_active_slots
= SI_NUM_RW_BUFFERS
;
2807 /* Initialize an array of 1024 bindless descriptors, when the limit is
2808 * reached, just make it larger and re-upload the whole array.
2810 si_init_bindless_descriptors(sctx
, &sctx
->bindless_descriptors
,
2811 SI_SGPR_BINDLESS_SAMPLERS_AND_IMAGES
,
2814 sctx
->descriptors_dirty
= u_bit_consecutive(0, SI_NUM_DESCS
);
2816 /* Set pipe_context functions. */
2817 sctx
->b
.bind_sampler_states
= si_bind_sampler_states
;
2818 sctx
->b
.set_shader_images
= si_set_shader_images
;
2819 sctx
->b
.set_constant_buffer
= si_pipe_set_constant_buffer
;
2820 sctx
->b
.set_shader_buffers
= si_set_shader_buffers
;
2821 sctx
->b
.set_sampler_views
= si_set_sampler_views
;
2822 sctx
->b
.create_texture_handle
= si_create_texture_handle
;
2823 sctx
->b
.delete_texture_handle
= si_delete_texture_handle
;
2824 sctx
->b
.make_texture_handle_resident
= si_make_texture_handle_resident
;
2825 sctx
->b
.create_image_handle
= si_create_image_handle
;
2826 sctx
->b
.delete_image_handle
= si_delete_image_handle
;
2827 sctx
->b
.make_image_handle_resident
= si_make_image_handle_resident
;
2829 if (!sctx
->has_graphics
)
2832 sctx
->b
.set_polygon_stipple
= si_set_polygon_stipple
;
2834 /* Shader user data. */
2835 sctx
->atoms
.s
.shader_pointers
.emit
= si_emit_graphics_shader_pointers
;
2837 /* Set default and immutable mappings. */
2839 assert(sctx
->chip_class
>= GFX10
);
2840 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
, R_00B230_SPI_SHADER_USER_DATA_GS_0
);
2842 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
, R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2845 if (sctx
->chip_class
== GFX9
) {
2846 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_CTRL
,
2847 R_00B430_SPI_SHADER_USER_DATA_LS_0
);
2848 si_set_user_data_base(sctx
, PIPE_SHADER_GEOMETRY
,
2849 R_00B330_SPI_SHADER_USER_DATA_ES_0
);
2851 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_CTRL
,
2852 R_00B430_SPI_SHADER_USER_DATA_HS_0
);
2853 si_set_user_data_base(sctx
, PIPE_SHADER_GEOMETRY
,
2854 R_00B230_SPI_SHADER_USER_DATA_GS_0
);
2856 si_set_user_data_base(sctx
, PIPE_SHADER_FRAGMENT
, R_00B030_SPI_SHADER_USER_DATA_PS_0
);
2859 static bool si_upload_shader_descriptors(struct si_context
*sctx
, unsigned mask
)
2861 unsigned dirty
= sctx
->descriptors_dirty
& mask
;
2863 /* Assume nothing will go wrong: */
2864 sctx
->shader_pointers_dirty
|= dirty
;
2867 unsigned i
= u_bit_scan(&dirty
);
2869 if (!si_upload_descriptors(sctx
, &sctx
->descriptors
[i
]))
2873 sctx
->descriptors_dirty
&= ~mask
;
2875 si_upload_bindless_descriptors(sctx
);
2880 bool si_upload_graphics_shader_descriptors(struct si_context
*sctx
)
2882 const unsigned mask
= u_bit_consecutive(0, SI_DESCS_FIRST_COMPUTE
);
2883 return si_upload_shader_descriptors(sctx
, mask
);
2886 bool si_upload_compute_shader_descriptors(struct si_context
*sctx
)
2888 /* Does not update rw_buffers as that is not needed for compute shaders
2889 * and the input buffer is using the same SGPR's anyway.
2891 const unsigned mask
= u_bit_consecutive(SI_DESCS_FIRST_COMPUTE
,
2892 SI_NUM_DESCS
- SI_DESCS_FIRST_COMPUTE
);
2893 return si_upload_shader_descriptors(sctx
, mask
);
2896 void si_release_all_descriptors(struct si_context
*sctx
)
2900 for (i
= 0; i
< SI_NUM_SHADERS
; i
++) {
2901 si_release_buffer_resources(&sctx
->const_and_shader_buffers
[i
],
2902 si_const_and_shader_buffer_descriptors(sctx
, i
));
2903 si_release_sampler_views(&sctx
->samplers
[i
]);
2904 si_release_image_views(&sctx
->images
[i
]);
2906 si_release_buffer_resources(&sctx
->rw_buffers
,
2907 &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
]);
2908 for (i
= 0; i
< SI_NUM_VERTEX_BUFFERS
; i
++)
2909 pipe_vertex_buffer_unreference(&sctx
->vertex_buffer
[i
]);
2911 for (i
= 0; i
< SI_NUM_DESCS
; ++i
)
2912 si_release_descriptors(&sctx
->descriptors
[i
]);
2914 si_resource_reference(&sctx
->vb_descriptors_buffer
, NULL
);
2915 sctx
->vb_descriptors_gpu_list
= NULL
; /* points into a mapped buffer */
2917 si_release_bindless_descriptors(sctx
);
2920 void si_gfx_resources_add_all_to_bo_list(struct si_context
*sctx
)
2922 for (unsigned i
= 0; i
< SI_NUM_GRAPHICS_SHADERS
; i
++) {
2923 si_buffer_resources_begin_new_cs(sctx
, &sctx
->const_and_shader_buffers
[i
]);
2924 si_sampler_views_begin_new_cs(sctx
, &sctx
->samplers
[i
]);
2925 si_image_views_begin_new_cs(sctx
, &sctx
->images
[i
]);
2927 si_buffer_resources_begin_new_cs(sctx
, &sctx
->rw_buffers
);
2928 si_vertex_buffers_begin_new_cs(sctx
);
2930 if (sctx
->bo_list_add_all_resident_resources
)
2931 si_resident_buffers_add_all_to_bo_list(sctx
);
2933 assert(sctx
->bo_list_add_all_gfx_resources
);
2934 sctx
->bo_list_add_all_gfx_resources
= false;
2937 void si_compute_resources_add_all_to_bo_list(struct si_context
*sctx
)
2939 unsigned sh
= PIPE_SHADER_COMPUTE
;
2941 si_buffer_resources_begin_new_cs(sctx
, &sctx
->const_and_shader_buffers
[sh
]);
2942 si_sampler_views_begin_new_cs(sctx
, &sctx
->samplers
[sh
]);
2943 si_image_views_begin_new_cs(sctx
, &sctx
->images
[sh
]);
2944 si_buffer_resources_begin_new_cs(sctx
, &sctx
->rw_buffers
);
2946 if (sctx
->bo_list_add_all_resident_resources
)
2947 si_resident_buffers_add_all_to_bo_list(sctx
);
2949 assert(sctx
->bo_list_add_all_compute_resources
);
2950 sctx
->bo_list_add_all_compute_resources
= false;
2953 void si_all_descriptors_begin_new_cs(struct si_context
*sctx
)
2955 for (unsigned i
= 0; i
< SI_NUM_DESCS
; ++i
)
2956 si_descriptors_begin_new_cs(sctx
, &sctx
->descriptors
[i
]);
2957 si_descriptors_begin_new_cs(sctx
, &sctx
->bindless_descriptors
);
2959 si_shader_pointers_begin_new_cs(sctx
);
2961 sctx
->bo_list_add_all_resident_resources
= true;
2962 sctx
->bo_list_add_all_gfx_resources
= true;
2963 sctx
->bo_list_add_all_compute_resources
= true;
2966 void si_set_active_descriptors(struct si_context
*sctx
, unsigned desc_idx
,
2967 uint64_t new_active_mask
)
2969 struct si_descriptors
*desc
= &sctx
->descriptors
[desc_idx
];
2971 /* Ignore no-op updates and updates that disable all slots. */
2972 if (!new_active_mask
||
2973 new_active_mask
== u_bit_consecutive64(desc
->first_active_slot
,
2974 desc
->num_active_slots
))
2978 u_bit_scan_consecutive_range64(&new_active_mask
, &first
, &count
);
2979 assert(new_active_mask
== 0);
2981 /* Upload/dump descriptors if slots are being enabled. */
2982 if (first
< desc
->first_active_slot
||
2983 first
+ count
> desc
->first_active_slot
+ desc
->num_active_slots
)
2984 sctx
->descriptors_dirty
|= 1u << desc_idx
;
2986 desc
->first_active_slot
= first
;
2987 desc
->num_active_slots
= count
;
2990 void si_set_active_descriptors_for_shader(struct si_context
*sctx
,
2991 struct si_shader_selector
*sel
)
2996 si_set_active_descriptors(sctx
,
2997 si_const_and_shader_buffer_descriptors_idx(sel
->type
),
2998 sel
->active_const_and_shader_buffers
);
2999 si_set_active_descriptors(sctx
,
3000 si_sampler_and_image_descriptors_idx(sel
->type
),
3001 sel
->active_samplers_and_images
);