2 * Copyright 2013 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 * Marek Olšák <marek.olsak@amd.com>
27 /* Resource binding slots and sampler states (each described with 8 or
28 * 4 dwords) are stored in lists in memory which is accessed by shaders
29 * using scalar load instructions.
31 * This file is responsible for managing such lists. It keeps a copy of all
32 * descriptors in CPU memory and re-uploads a whole list if some slots have
35 * This code is also reponsible for updating shader pointers to those lists.
37 * Note that CP DMA can't be used for updating the lists, because a GPU hang
38 * could leave the list in a mid-IB state and the next IB would get wrong
39 * descriptors and the whole context would be unusable at that point.
40 * (Note: The register shadowing can't be used due to the same reason)
42 * Also, uploading descriptors to newly allocated memory doesn't require
46 * Possible scenarios for one 16 dword image+sampler slot:
48 * | Image | w/ FMASK | Buffer | NULL
49 * [ 0: 3] Image[0:3] | Image[0:3] | Null[0:3] | Null[0:3]
50 * [ 4: 7] Image[4:7] | Image[4:7] | Buffer[0:3] | 0
51 * [ 8:11] Null[0:3] | Fmask[0:3] | Null[0:3] | Null[0:3]
52 * [12:15] Sampler[0:3] | Fmask[4:7] | Sampler[0:3] | Sampler[0:3]
54 * FMASK implies MSAA, therefore no sampler state.
55 * Sampler states are never unbound except when FMASK is bound.
58 #include "radeon/r600_cs.h"
63 #include "util/hash_table.h"
64 #include "util/u_format.h"
65 #include "util/u_memory.h"
66 #include "util/u_upload_mgr.h"
69 /* NULL image and buffer descriptor for textures (alpha = 1) and images
72 * For images, all fields must be zero except for the swizzle, which
73 * supports arbitrary combinations of 0s and 1s. The texture type must be
74 * any valid type (e.g. 1D). If the texture type isn't set, the hw hangs.
76 * For buffers, all fields must be zero. If they are not, the hw hangs.
78 * This is the only reason why the buffer descriptor must be in words [4:7].
80 static uint32_t null_texture_descriptor
[8] = {
84 S_008F1C_DST_SEL_W(V_008F1C_SQ_SEL_1
) |
85 S_008F1C_TYPE(V_008F1C_SQ_RSRC_IMG_1D
)
86 /* the rest must contain zeros, which is also used by the buffer
90 static uint32_t null_image_descriptor
[8] = {
94 S_008F1C_TYPE(V_008F1C_SQ_RSRC_IMG_1D
)
95 /* the rest must contain zeros, which is also used by the buffer
99 static uint16_t si_ce_ram_size(struct si_context
*sctx
)
101 return sctx
->b
.chip_class
>= GFX9
? 4096 : 32768;
104 static void si_init_descriptor_list(uint32_t *desc_list
,
105 unsigned element_dw_size
,
106 unsigned num_elements
,
107 const uint32_t *null_descriptor
)
111 /* Initialize the array to NULL descriptors if the element size is 8. */
112 if (null_descriptor
) {
113 assert(element_dw_size
% 8 == 0);
114 for (i
= 0; i
< num_elements
* element_dw_size
/ 8; i
++)
115 memcpy(desc_list
+ i
* 8, null_descriptor
, 8 * 4);
119 static void si_init_descriptors(struct si_context
*sctx
,
120 struct si_descriptors
*desc
,
121 unsigned shader_userdata_index
,
122 unsigned element_dw_size
,
123 unsigned num_elements
,
124 unsigned first_ce_slot
,
125 unsigned num_ce_slots
,
128 desc
->list
= CALLOC(num_elements
, element_dw_size
* 4);
129 desc
->element_dw_size
= element_dw_size
;
130 desc
->num_elements
= num_elements
;
131 desc
->first_ce_slot
= sctx
->ce_ib
? first_ce_slot
: 0;
132 desc
->num_ce_slots
= sctx
->ce_ib
? num_ce_slots
: 0;
133 desc
->dirty_mask
= 0;
134 desc
->shader_userdata_offset
= shader_userdata_index
* 4;
136 if (desc
->num_ce_slots
) {
137 assert(num_elements
<= sizeof(desc
->dirty_mask
)*8);
139 desc
->uses_ce
= true;
140 desc
->ce_offset
= *ce_offset
;
141 desc
->dirty_mask
= u_bit_consecutive64(0, num_elements
);
143 *ce_offset
+= element_dw_size
* desc
->num_ce_slots
* 4;
147 static void si_release_descriptors(struct si_descriptors
*desc
)
149 r600_resource_reference(&desc
->buffer
, NULL
);
153 static bool si_ce_upload(struct si_context
*sctx
, unsigned ce_offset
, unsigned size
,
154 unsigned *out_offset
, struct r600_resource
**out_buf
)
157 unsigned cache_line_size
= sctx
->screen
->b
.info
.tcc_cache_line_size
;
159 /* The base and size should be aligned to the L2 cache line size
160 * for optimal performance. (all dumps should rewrite whole lines)
162 size
= align(size
, cache_line_size
);
164 (void)si_ce_ram_size
; /* silence an "unused" warning */
165 assert(ce_offset
+ size
<= si_ce_ram_size(sctx
));
167 u_suballocator_alloc(sctx
->ce_suballocator
, size
, cache_line_size
,
168 out_offset
, (struct pipe_resource
**)out_buf
);
172 va
= (*out_buf
)->gpu_address
+ *out_offset
;
174 radeon_emit(sctx
->ce_ib
, PKT3(PKT3_DUMP_CONST_RAM
, 3, 0));
175 radeon_emit(sctx
->ce_ib
, ce_offset
);
176 radeon_emit(sctx
->ce_ib
, size
/ 4);
177 radeon_emit(sctx
->ce_ib
, va
);
178 radeon_emit(sctx
->ce_ib
, va
>> 32);
180 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, *out_buf
,
181 RADEON_USAGE_READWRITE
, RADEON_PRIO_DESCRIPTORS
);
183 sctx
->ce_need_synchronization
= true;
187 void si_ce_save_all_descriptors_at_ib_end(struct si_context
* sctx
)
189 bool success
= si_ce_upload(sctx
, 0, sctx
->total_ce_ram_allocated
,
190 &sctx
->ce_ram_saved_offset
,
191 &sctx
->ce_ram_saved_buffer
);
196 void si_ce_restore_all_descriptors_at_ib_start(struct si_context
*sctx
)
198 if (!sctx
->ce_ram_saved_buffer
)
201 struct radeon_winsys_cs
*ib
= sctx
->ce_preamble_ib
;
205 uint64_t va
= sctx
->ce_ram_saved_buffer
->gpu_address
+
206 sctx
->ce_ram_saved_offset
;
208 radeon_emit(ib
, PKT3(PKT3_LOAD_CONST_RAM
, 3, 0));
210 radeon_emit(ib
, va
>> 32);
211 radeon_emit(ib
, sctx
->total_ce_ram_allocated
/ 4);
214 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
215 sctx
->ce_ram_saved_buffer
,
216 RADEON_USAGE_READ
, RADEON_PRIO_DESCRIPTORS
);
219 void si_ce_enable_loads(struct radeon_winsys_cs
*ib
)
221 radeon_emit(ib
, PKT3(PKT3_CONTEXT_CONTROL
, 1, 0));
222 radeon_emit(ib
, CONTEXT_CONTROL_LOAD_ENABLE(1) |
223 CONTEXT_CONTROL_LOAD_CE_RAM(1));
224 radeon_emit(ib
, CONTEXT_CONTROL_SHADOW_ENABLE(1));
227 static bool si_upload_descriptors(struct si_context
*sctx
,
228 struct si_descriptors
*desc
,
229 struct r600_atom
* atom
)
231 unsigned slot_size
= desc
->element_dw_size
* 4;
232 unsigned first_slot_offset
= desc
->first_active_slot
* slot_size
;
233 unsigned upload_size
= desc
->num_active_slots
* slot_size
;
235 /* Skip the upload if no shader is using the descriptors. dirty_mask
236 * will stay dirty and the descriptors will be uploaded when there is
237 * a shader using them.
243 const uint32_t *list
= desc
->list
+
244 desc
->first_ce_slot
* desc
->element_dw_size
;
245 uint64_t mask
= (desc
->dirty_mask
>> desc
->first_ce_slot
) &
246 u_bit_consecutive64(0, desc
->num_ce_slots
);
251 u_bit_scan_consecutive_range64(&mask
, &begin
, &count
);
253 begin
*= desc
->element_dw_size
;
254 count
*= desc
->element_dw_size
;
256 radeon_emit(sctx
->ce_ib
,
257 PKT3(PKT3_WRITE_CONST_RAM
, count
, 0));
258 radeon_emit(sctx
->ce_ib
, desc
->ce_offset
+ begin
* 4);
259 radeon_emit_array(sctx
->ce_ib
, list
+ begin
, count
);
262 if (!si_ce_upload(sctx
,
264 (first_slot_offset
- desc
->first_ce_slot
* slot_size
),
265 upload_size
, (unsigned*)&desc
->buffer_offset
,
271 u_upload_alloc(sctx
->b
.b
.const_uploader
, 0, upload_size
,
272 si_optimal_tcc_alignment(sctx
, upload_size
),
273 (unsigned*)&desc
->buffer_offset
,
274 (struct pipe_resource
**)&desc
->buffer
,
277 return false; /* skip the draw call */
279 util_memcpy_cpu_to_le32(ptr
, (char*)desc
->list
+ first_slot_offset
,
281 desc
->gpu_list
= ptr
- first_slot_offset
/ 4;
283 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, desc
->buffer
,
284 RADEON_USAGE_READ
, RADEON_PRIO_DESCRIPTORS
);
287 /* The shader pointer should point to slot 0. */
288 desc
->buffer_offset
-= first_slot_offset
;
290 desc
->dirty_mask
= 0;
293 si_mark_atom_dirty(sctx
, atom
);
299 si_descriptors_begin_new_cs(struct si_context
*sctx
, struct si_descriptors
*desc
)
304 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, desc
->buffer
,
305 RADEON_USAGE_READ
, RADEON_PRIO_DESCRIPTORS
);
311 si_sampler_and_image_descriptors_idx(unsigned shader
)
313 return SI_DESCS_FIRST_SHADER
+ shader
* SI_NUM_SHADER_DESCS
+
314 SI_SHADER_DESCS_SAMPLERS_AND_IMAGES
;
317 static struct si_descriptors
*
318 si_sampler_and_image_descriptors(struct si_context
*sctx
, unsigned shader
)
320 return &sctx
->descriptors
[si_sampler_and_image_descriptors_idx(shader
)];
323 static void si_release_sampler_views(struct si_sampler_views
*views
)
327 for (i
= 0; i
< ARRAY_SIZE(views
->views
); i
++) {
328 pipe_sampler_view_reference(&views
->views
[i
], NULL
);
332 static void si_sampler_view_add_buffer(struct si_context
*sctx
,
333 struct pipe_resource
*resource
,
334 enum radeon_bo_usage usage
,
335 bool is_stencil_sampler
,
338 struct r600_resource
*rres
;
339 struct r600_texture
*rtex
;
340 enum radeon_bo_priority priority
;
345 if (resource
->target
!= PIPE_BUFFER
) {
346 struct r600_texture
*tex
= (struct r600_texture
*)resource
;
348 if (tex
->is_depth
&& !r600_can_sample_zs(tex
, is_stencil_sampler
))
349 resource
= &tex
->flushed_depth_texture
->resource
.b
.b
;
352 rres
= (struct r600_resource
*)resource
;
353 priority
= r600_get_sampler_view_priority(rres
);
355 radeon_add_to_buffer_list_check_mem(&sctx
->b
, &sctx
->b
.gfx
,
356 rres
, usage
, priority
,
359 if (resource
->target
== PIPE_BUFFER
)
362 /* Now add separate DCC or HTILE. */
363 rtex
= (struct r600_texture
*)resource
;
364 if (rtex
->dcc_separate_buffer
) {
365 radeon_add_to_buffer_list_check_mem(&sctx
->b
, &sctx
->b
.gfx
,
366 rtex
->dcc_separate_buffer
, usage
,
367 RADEON_PRIO_DCC
, check_mem
);
371 static void si_sampler_views_begin_new_cs(struct si_context
*sctx
,
372 struct si_sampler_views
*views
)
374 unsigned mask
= views
->enabled_mask
;
376 /* Add buffers to the CS. */
378 int i
= u_bit_scan(&mask
);
379 struct si_sampler_view
*sview
= (struct si_sampler_view
*)views
->views
[i
];
381 si_sampler_view_add_buffer(sctx
, sview
->base
.texture
,
383 sview
->is_stencil_sampler
, false);
387 /* Set buffer descriptor fields that can be changed by reallocations. */
388 static void si_set_buf_desc_address(struct r600_resource
*buf
,
389 uint64_t offset
, uint32_t *state
)
391 uint64_t va
= buf
->gpu_address
+ offset
;
394 state
[1] &= C_008F04_BASE_ADDRESS_HI
;
395 state
[1] |= S_008F04_BASE_ADDRESS_HI(va
>> 32);
398 /* Set texture descriptor fields that can be changed by reallocations.
401 * \param base_level_info information of the level of BASE_ADDRESS
402 * \param base_level the level of BASE_ADDRESS
403 * \param first_level pipe_sampler_view.u.tex.first_level
404 * \param block_width util_format_get_blockwidth()
405 * \param is_stencil select between separate Z & Stencil
406 * \param state descriptor to update
408 void si_set_mutable_tex_desc_fields(struct si_screen
*sscreen
,
409 struct r600_texture
*tex
,
410 const struct legacy_surf_level
*base_level_info
,
411 unsigned base_level
, unsigned first_level
,
412 unsigned block_width
, bool is_stencil
,
415 uint64_t va
, meta_va
= 0;
417 if (tex
->is_depth
&& !r600_can_sample_zs(tex
, is_stencil
)) {
418 tex
= tex
->flushed_depth_texture
;
422 va
= tex
->resource
.gpu_address
;
424 if (sscreen
->b
.chip_class
>= GFX9
) {
425 /* Only stencil_offset needs to be added here. */
427 va
+= tex
->surface
.u
.gfx9
.stencil_offset
;
429 va
+= tex
->surface
.u
.gfx9
.surf_offset
;
431 va
+= base_level_info
->offset
;
435 state
[1] &= C_008F14_BASE_ADDRESS_HI
;
436 state
[1] |= S_008F14_BASE_ADDRESS_HI(va
>> 40);
438 /* Only macrotiled modes can set tile swizzle.
439 * GFX9 doesn't use (legacy) base_level_info.
441 if (sscreen
->b
.chip_class
>= GFX9
||
442 base_level_info
->mode
== RADEON_SURF_MODE_2D
)
443 state
[0] |= tex
->surface
.tile_swizzle
;
445 if (sscreen
->b
.chip_class
>= VI
) {
446 state
[6] &= C_008F28_COMPRESSION_EN
;
449 if (vi_dcc_enabled(tex
, first_level
)) {
450 meta_va
= (!tex
->dcc_separate_buffer
? tex
->resource
.gpu_address
: 0) +
453 if (sscreen
->b
.chip_class
== VI
) {
454 meta_va
+= base_level_info
->dcc_offset
;
455 assert(base_level_info
->mode
== RADEON_SURF_MODE_2D
);
458 meta_va
|= (uint32_t)tex
->surface
.tile_swizzle
<< 8;
459 } else if (tex
->tc_compatible_htile
&& first_level
== 0) {
460 meta_va
= tex
->resource
.gpu_address
+ tex
->htile_offset
;
464 state
[6] |= S_008F28_COMPRESSION_EN(1);
465 state
[7] = meta_va
>> 8;
469 if (sscreen
->b
.chip_class
>= GFX9
) {
470 state
[3] &= C_008F1C_SW_MODE
;
471 state
[4] &= C_008F20_PITCH_GFX9
;
474 state
[3] |= S_008F1C_SW_MODE(tex
->surface
.u
.gfx9
.stencil
.swizzle_mode
);
475 state
[4] |= S_008F20_PITCH_GFX9(tex
->surface
.u
.gfx9
.stencil
.epitch
);
477 state
[3] |= S_008F1C_SW_MODE(tex
->surface
.u
.gfx9
.surf
.swizzle_mode
);
478 state
[4] |= S_008F20_PITCH_GFX9(tex
->surface
.u
.gfx9
.surf
.epitch
);
481 state
[5] &= C_008F24_META_DATA_ADDRESS
&
482 C_008F24_META_PIPE_ALIGNED
&
483 C_008F24_META_RB_ALIGNED
;
485 struct gfx9_surf_meta_flags meta
;
488 meta
= tex
->surface
.u
.gfx9
.dcc
;
490 meta
= tex
->surface
.u
.gfx9
.htile
;
492 state
[5] |= S_008F24_META_DATA_ADDRESS(meta_va
>> 40) |
493 S_008F24_META_PIPE_ALIGNED(meta
.pipe_aligned
) |
494 S_008F24_META_RB_ALIGNED(meta
.rb_aligned
);
498 unsigned pitch
= base_level_info
->nblk_x
* block_width
;
499 unsigned index
= si_tile_mode_index(tex
, base_level
, is_stencil
);
501 state
[3] &= C_008F1C_TILING_INDEX
;
502 state
[3] |= S_008F1C_TILING_INDEX(index
);
503 state
[4] &= C_008F20_PITCH_GFX6
;
504 state
[4] |= S_008F20_PITCH_GFX6(pitch
- 1);
508 static void si_set_sampler_view_desc(struct si_context
*sctx
,
509 struct si_sampler_view
*sview
,
510 struct si_sampler_state
*sstate
,
513 struct pipe_sampler_view
*view
= &sview
->base
;
514 struct r600_texture
*rtex
= (struct r600_texture
*)view
->texture
;
515 bool is_buffer
= rtex
->resource
.b
.b
.target
== PIPE_BUFFER
;
517 if (unlikely(!is_buffer
&& sview
->dcc_incompatible
)) {
518 if (vi_dcc_enabled(rtex
, view
->u
.tex
.first_level
))
519 if (!r600_texture_disable_dcc(&sctx
->b
, rtex
))
520 sctx
->b
.decompress_dcc(&sctx
->b
.b
, rtex
);
522 sview
->dcc_incompatible
= false;
525 assert(rtex
); /* views with texture == NULL aren't supported */
526 memcpy(desc
, sview
->state
, 8*4);
529 si_set_buf_desc_address(&rtex
->resource
,
530 sview
->base
.u
.buf
.offset
,
533 bool is_separate_stencil
= rtex
->db_compatible
&&
534 sview
->is_stencil_sampler
;
536 si_set_mutable_tex_desc_fields(sctx
->screen
, rtex
,
537 sview
->base_level_info
,
539 sview
->base
.u
.tex
.first_level
,
545 if (!is_buffer
&& rtex
->fmask
.size
) {
546 memcpy(desc
+ 8, sview
->fmask_state
, 8*4);
548 /* Disable FMASK and bind sampler state in [12:15]. */
549 memcpy(desc
+ 8, null_texture_descriptor
, 4*4);
552 memcpy(desc
+ 12, sstate
->val
, 4*4);
556 static void si_set_sampler_view(struct si_context
*sctx
,
558 unsigned slot
, struct pipe_sampler_view
*view
,
559 bool disallow_early_out
)
561 struct si_sampler_views
*views
= &sctx
->samplers
[shader
].views
;
562 struct si_sampler_view
*rview
= (struct si_sampler_view
*)view
;
563 struct si_descriptors
*descs
= si_sampler_and_image_descriptors(sctx
, shader
);
564 unsigned desc_slot
= si_get_sampler_slot(slot
);
565 uint32_t *desc
= descs
->list
+ desc_slot
* 16;
567 if (views
->views
[slot
] == view
&& !disallow_early_out
)
571 struct r600_texture
*rtex
= (struct r600_texture
*)view
->texture
;
573 si_set_sampler_view_desc(sctx
, rview
,
574 views
->sampler_states
[slot
], desc
);
576 if (rtex
->resource
.b
.b
.target
== PIPE_BUFFER
)
577 rtex
->resource
.bind_history
|= PIPE_BIND_SAMPLER_VIEW
;
579 pipe_sampler_view_reference(&views
->views
[slot
], view
);
580 views
->enabled_mask
|= 1u << slot
;
582 /* Since this can flush, it must be done after enabled_mask is
584 si_sampler_view_add_buffer(sctx
, view
->texture
,
586 rview
->is_stencil_sampler
, true);
588 pipe_sampler_view_reference(&views
->views
[slot
], NULL
);
589 memcpy(desc
, null_texture_descriptor
, 8*4);
590 /* Only clear the lower dwords of FMASK. */
591 memcpy(desc
+ 8, null_texture_descriptor
, 4*4);
592 /* Re-set the sampler state if we are transitioning from FMASK. */
593 if (views
->sampler_states
[slot
])
595 views
->sampler_states
[slot
]->val
, 4*4);
597 views
->enabled_mask
&= ~(1u << slot
);
600 descs
->dirty_mask
|= 1ull << desc_slot
;
601 sctx
->descriptors_dirty
|= 1u << si_sampler_and_image_descriptors_idx(shader
);
604 static bool color_needs_decompression(struct r600_texture
*rtex
)
606 return rtex
->fmask
.size
||
607 (rtex
->dirty_level_mask
&&
608 (rtex
->cmask
.size
|| rtex
->dcc_offset
));
611 static bool depth_needs_decompression(struct r600_texture
*rtex
)
613 /* If the depth/stencil texture is TC-compatible, no decompression
614 * will be done. The decompression function will only flush DB caches
615 * to make it coherent with shaders. That's necessary because the driver
616 * doesn't flush DB caches in any other case.
618 return rtex
->db_compatible
;
621 static void si_update_shader_needs_decompress_mask(struct si_context
*sctx
,
624 struct si_textures_info
*samplers
= &sctx
->samplers
[shader
];
625 unsigned shader_bit
= 1 << shader
;
627 if (samplers
->needs_depth_decompress_mask
||
628 samplers
->needs_color_decompress_mask
||
629 sctx
->images
[shader
].needs_color_decompress_mask
)
630 sctx
->shader_needs_decompress_mask
|= shader_bit
;
632 sctx
->shader_needs_decompress_mask
&= ~shader_bit
;
635 static void si_set_sampler_views(struct pipe_context
*ctx
,
636 enum pipe_shader_type shader
, unsigned start
,
638 struct pipe_sampler_view
**views
)
640 struct si_context
*sctx
= (struct si_context
*)ctx
;
641 struct si_textures_info
*samplers
= &sctx
->samplers
[shader
];
644 if (!count
|| shader
>= SI_NUM_SHADERS
)
647 for (i
= 0; i
< count
; i
++) {
648 unsigned slot
= start
+ i
;
650 if (!views
|| !views
[i
]) {
651 samplers
->needs_depth_decompress_mask
&= ~(1u << slot
);
652 samplers
->needs_color_decompress_mask
&= ~(1u << slot
);
653 si_set_sampler_view(sctx
, shader
, slot
, NULL
, false);
657 si_set_sampler_view(sctx
, shader
, slot
, views
[i
], false);
659 if (views
[i
]->texture
&& views
[i
]->texture
->target
!= PIPE_BUFFER
) {
660 struct r600_texture
*rtex
=
661 (struct r600_texture
*)views
[i
]->texture
;
663 if (depth_needs_decompression(rtex
)) {
664 samplers
->needs_depth_decompress_mask
|= 1u << slot
;
666 samplers
->needs_depth_decompress_mask
&= ~(1u << slot
);
668 if (color_needs_decompression(rtex
)) {
669 samplers
->needs_color_decompress_mask
|= 1u << slot
;
671 samplers
->needs_color_decompress_mask
&= ~(1u << slot
);
674 if (rtex
->dcc_offset
&&
675 p_atomic_read(&rtex
->framebuffers_bound
))
676 sctx
->need_check_render_feedback
= true;
678 samplers
->needs_depth_decompress_mask
&= ~(1u << slot
);
679 samplers
->needs_color_decompress_mask
&= ~(1u << slot
);
683 si_update_shader_needs_decompress_mask(sctx
, shader
);
687 si_samplers_update_needs_color_decompress_mask(struct si_textures_info
*samplers
)
689 unsigned mask
= samplers
->views
.enabled_mask
;
692 int i
= u_bit_scan(&mask
);
693 struct pipe_resource
*res
= samplers
->views
.views
[i
]->texture
;
695 if (res
&& res
->target
!= PIPE_BUFFER
) {
696 struct r600_texture
*rtex
= (struct r600_texture
*)res
;
698 if (color_needs_decompression(rtex
)) {
699 samplers
->needs_color_decompress_mask
|= 1u << i
;
701 samplers
->needs_color_decompress_mask
&= ~(1u << i
);
710 si_release_image_views(struct si_images_info
*images
)
714 for (i
= 0; i
< SI_NUM_IMAGES
; ++i
) {
715 struct pipe_image_view
*view
= &images
->views
[i
];
717 pipe_resource_reference(&view
->resource
, NULL
);
722 si_image_views_begin_new_cs(struct si_context
*sctx
, struct si_images_info
*images
)
724 uint mask
= images
->enabled_mask
;
726 /* Add buffers to the CS. */
728 int i
= u_bit_scan(&mask
);
729 struct pipe_image_view
*view
= &images
->views
[i
];
731 assert(view
->resource
);
733 si_sampler_view_add_buffer(sctx
, view
->resource
,
734 RADEON_USAGE_READWRITE
, false, false);
739 si_disable_shader_image(struct si_context
*ctx
, unsigned shader
, unsigned slot
)
741 struct si_images_info
*images
= &ctx
->images
[shader
];
743 if (images
->enabled_mask
& (1u << slot
)) {
744 struct si_descriptors
*descs
= si_sampler_and_image_descriptors(ctx
, shader
);
745 unsigned desc_slot
= si_get_image_slot(slot
);
747 pipe_resource_reference(&images
->views
[slot
].resource
, NULL
);
748 images
->needs_color_decompress_mask
&= ~(1 << slot
);
750 memcpy(descs
->list
+ desc_slot
*8, null_image_descriptor
, 8*4);
751 images
->enabled_mask
&= ~(1u << slot
);
752 /* two 8-byte images share one 16-byte slot */
753 descs
->dirty_mask
|= 1u << (desc_slot
/ 2);
754 ctx
->descriptors_dirty
|= 1u << si_sampler_and_image_descriptors_idx(shader
);
759 si_mark_image_range_valid(const struct pipe_image_view
*view
)
761 struct r600_resource
*res
= (struct r600_resource
*)view
->resource
;
763 assert(res
&& res
->b
.b
.target
== PIPE_BUFFER
);
765 util_range_add(&res
->valid_buffer_range
,
767 view
->u
.buf
.offset
+ view
->u
.buf
.size
);
770 static void si_set_shader_image_desc(struct si_context
*ctx
,
771 const struct pipe_image_view
*view
,
772 bool skip_decompress
,
775 struct si_screen
*screen
= ctx
->screen
;
776 struct r600_resource
*res
;
778 res
= (struct r600_resource
*)view
->resource
;
780 if (res
->b
.b
.target
== PIPE_BUFFER
) {
781 if (view
->access
& PIPE_IMAGE_ACCESS_WRITE
)
782 si_mark_image_range_valid(view
);
784 si_make_buffer_descriptor(screen
, res
,
787 view
->u
.buf
.size
, desc
);
788 si_set_buf_desc_address(res
, view
->u
.buf
.offset
, desc
+ 4);
790 static const unsigned char swizzle
[4] = { 0, 1, 2, 3 };
791 struct r600_texture
*tex
= (struct r600_texture
*)res
;
792 unsigned level
= view
->u
.tex
.level
;
793 unsigned width
, height
, depth
, hw_level
;
794 bool uses_dcc
= vi_dcc_enabled(tex
, level
);
796 assert(!tex
->is_depth
);
797 assert(tex
->fmask
.size
== 0);
799 if (uses_dcc
&& !skip_decompress
&&
800 (view
->access
& PIPE_IMAGE_ACCESS_WRITE
||
801 !vi_dcc_formats_compatible(res
->b
.b
.format
, view
->format
))) {
802 /* If DCC can't be disabled, at least decompress it.
803 * The decompression is relatively cheap if the surface
804 * has been decompressed already.
806 if (!r600_texture_disable_dcc(&ctx
->b
, tex
))
807 ctx
->b
.decompress_dcc(&ctx
->b
.b
, tex
);
810 if (ctx
->b
.chip_class
>= GFX9
) {
811 /* Always set the base address. The swizzle modes don't
812 * allow setting mipmap level offsets as the base.
814 width
= res
->b
.b
.width0
;
815 height
= res
->b
.b
.height0
;
816 depth
= res
->b
.b
.depth0
;
819 /* Always force the base level to the selected level.
821 * This is required for 3D textures, where otherwise
822 * selecting a single slice for non-layered bindings
823 * fails. It doesn't hurt the other targets.
825 width
= u_minify(res
->b
.b
.width0
, level
);
826 height
= u_minify(res
->b
.b
.height0
, level
);
827 depth
= u_minify(res
->b
.b
.depth0
, level
);
831 si_make_texture_descriptor(screen
, tex
,
832 false, res
->b
.b
.target
,
833 view
->format
, swizzle
,
835 view
->u
.tex
.first_layer
,
836 view
->u
.tex
.last_layer
,
837 width
, height
, depth
,
839 si_set_mutable_tex_desc_fields(screen
, tex
,
840 &tex
->surface
.u
.legacy
.level
[level
],
842 util_format_get_blockwidth(view
->format
),
847 static void si_set_shader_image(struct si_context
*ctx
,
849 unsigned slot
, const struct pipe_image_view
*view
,
850 bool skip_decompress
)
852 struct si_images_info
*images
= &ctx
->images
[shader
];
853 struct si_descriptors
*descs
= si_sampler_and_image_descriptors(ctx
, shader
);
854 struct r600_resource
*res
;
855 unsigned desc_slot
= si_get_image_slot(slot
);
856 uint32_t *desc
= descs
->list
+ desc_slot
* 8;
858 if (!view
|| !view
->resource
) {
859 si_disable_shader_image(ctx
, shader
, slot
);
863 res
= (struct r600_resource
*)view
->resource
;
865 if (&images
->views
[slot
] != view
)
866 util_copy_image_view(&images
->views
[slot
], view
);
868 si_set_shader_image_desc(ctx
, view
, skip_decompress
, desc
);
870 if (res
->b
.b
.target
== PIPE_BUFFER
) {
871 images
->needs_color_decompress_mask
&= ~(1 << slot
);
872 res
->bind_history
|= PIPE_BIND_SHADER_IMAGE
;
874 struct r600_texture
*tex
= (struct r600_texture
*)res
;
875 unsigned level
= view
->u
.tex
.level
;
877 if (color_needs_decompression(tex
)) {
878 images
->needs_color_decompress_mask
|= 1 << slot
;
880 images
->needs_color_decompress_mask
&= ~(1 << slot
);
883 if (vi_dcc_enabled(tex
, level
) &&
884 p_atomic_read(&tex
->framebuffers_bound
))
885 ctx
->need_check_render_feedback
= true;
888 images
->enabled_mask
|= 1u << slot
;
889 /* two 8-byte images share one 16-byte slot */
890 descs
->dirty_mask
|= 1u << (desc_slot
/ 2);
891 ctx
->descriptors_dirty
|= 1u << si_sampler_and_image_descriptors_idx(shader
);
893 /* Since this can flush, it must be done after enabled_mask is updated. */
894 si_sampler_view_add_buffer(ctx
, &res
->b
.b
,
895 (view
->access
& PIPE_IMAGE_ACCESS_WRITE
) ?
896 RADEON_USAGE_READWRITE
: RADEON_USAGE_READ
,
901 si_set_shader_images(struct pipe_context
*pipe
,
902 enum pipe_shader_type shader
,
903 unsigned start_slot
, unsigned count
,
904 const struct pipe_image_view
*views
)
906 struct si_context
*ctx
= (struct si_context
*)pipe
;
909 assert(shader
< SI_NUM_SHADERS
);
914 assert(start_slot
+ count
<= SI_NUM_IMAGES
);
917 for (i
= 0, slot
= start_slot
; i
< count
; ++i
, ++slot
)
918 si_set_shader_image(ctx
, shader
, slot
, &views
[i
], false);
920 for (i
= 0, slot
= start_slot
; i
< count
; ++i
, ++slot
)
921 si_set_shader_image(ctx
, shader
, slot
, NULL
, false);
924 si_update_shader_needs_decompress_mask(ctx
, shader
);
928 si_images_update_needs_color_decompress_mask(struct si_images_info
*images
)
930 unsigned mask
= images
->enabled_mask
;
933 int i
= u_bit_scan(&mask
);
934 struct pipe_resource
*res
= images
->views
[i
].resource
;
936 if (res
&& res
->target
!= PIPE_BUFFER
) {
937 struct r600_texture
*rtex
= (struct r600_texture
*)res
;
939 if (color_needs_decompression(rtex
)) {
940 images
->needs_color_decompress_mask
|= 1 << i
;
942 images
->needs_color_decompress_mask
&= ~(1 << i
);
950 static void si_bind_sampler_states(struct pipe_context
*ctx
,
951 enum pipe_shader_type shader
,
952 unsigned start
, unsigned count
, void **states
)
954 struct si_context
*sctx
= (struct si_context
*)ctx
;
955 struct si_textures_info
*samplers
= &sctx
->samplers
[shader
];
956 struct si_descriptors
*desc
= si_sampler_and_image_descriptors(sctx
, shader
);
957 struct si_sampler_state
**sstates
= (struct si_sampler_state
**)states
;
960 if (!count
|| shader
>= SI_NUM_SHADERS
)
963 for (i
= 0; i
< count
; i
++) {
964 unsigned slot
= start
+ i
;
965 unsigned desc_slot
= si_get_sampler_slot(slot
);
968 sstates
[i
] == samplers
->views
.sampler_states
[slot
])
972 assert(sstates
[i
]->magic
== SI_SAMPLER_STATE_MAGIC
);
974 samplers
->views
.sampler_states
[slot
] = sstates
[i
];
976 /* If FMASK is bound, don't overwrite it.
977 * The sampler state will be set after FMASK is unbound.
979 if (samplers
->views
.views
[slot
] &&
980 samplers
->views
.views
[slot
]->texture
&&
981 samplers
->views
.views
[slot
]->texture
->target
!= PIPE_BUFFER
&&
982 ((struct r600_texture
*)samplers
->views
.views
[slot
]->texture
)->fmask
.size
)
985 memcpy(desc
->list
+ desc_slot
* 16 + 12, sstates
[i
]->val
, 4*4);
986 desc
->dirty_mask
|= 1ull << desc_slot
;
987 sctx
->descriptors_dirty
|= 1u << si_sampler_and_image_descriptors_idx(shader
);
991 /* BUFFER RESOURCES */
993 static void si_init_buffer_resources(struct si_context
*sctx
,
994 struct si_buffer_resources
*buffers
,
995 struct si_descriptors
*descs
,
996 unsigned num_buffers
,
997 unsigned first_ce_slot
,
998 unsigned num_ce_slots
,
999 unsigned shader_userdata_index
,
1000 enum radeon_bo_usage shader_usage
,
1001 enum radeon_bo_usage shader_usage_constbuf
,
1002 enum radeon_bo_priority priority
,
1003 enum radeon_bo_priority priority_constbuf
,
1004 unsigned *ce_offset
)
1006 buffers
->shader_usage
= shader_usage
;
1007 buffers
->shader_usage_constbuf
= shader_usage_constbuf
;
1008 buffers
->priority
= priority
;
1009 buffers
->priority_constbuf
= priority_constbuf
;
1010 buffers
->buffers
= CALLOC(num_buffers
, sizeof(struct pipe_resource
*));
1012 si_init_descriptors(sctx
, descs
, shader_userdata_index
, 4, num_buffers
,
1013 first_ce_slot
, num_ce_slots
, ce_offset
);
1016 static void si_release_buffer_resources(struct si_buffer_resources
*buffers
,
1017 struct si_descriptors
*descs
)
1021 for (i
= 0; i
< descs
->num_elements
; i
++) {
1022 pipe_resource_reference(&buffers
->buffers
[i
], NULL
);
1025 FREE(buffers
->buffers
);
1028 static void si_buffer_resources_begin_new_cs(struct si_context
*sctx
,
1029 struct si_buffer_resources
*buffers
)
1031 unsigned mask
= buffers
->enabled_mask
;
1033 /* Add buffers to the CS. */
1035 int i
= u_bit_scan(&mask
);
1037 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
1038 r600_resource(buffers
->buffers
[i
]),
1039 i
< SI_NUM_SHADER_BUFFERS
? buffers
->shader_usage
:
1040 buffers
->shader_usage_constbuf
,
1041 i
< SI_NUM_SHADER_BUFFERS
? buffers
->priority
:
1042 buffers
->priority_constbuf
);
1046 static void si_get_buffer_from_descriptors(struct si_buffer_resources
*buffers
,
1047 struct si_descriptors
*descs
,
1048 unsigned idx
, struct pipe_resource
**buf
,
1049 unsigned *offset
, unsigned *size
)
1051 pipe_resource_reference(buf
, buffers
->buffers
[idx
]);
1053 struct r600_resource
*res
= r600_resource(*buf
);
1054 const uint32_t *desc
= descs
->list
+ idx
* 4;
1059 assert(G_008F04_STRIDE(desc
[1]) == 0);
1060 va
= ((uint64_t)desc
[1] << 32) | desc
[0];
1062 assert(va
>= res
->gpu_address
&& va
+ *size
<= res
->gpu_address
+ res
->bo_size
);
1063 *offset
= va
- res
->gpu_address
;
1067 /* VERTEX BUFFERS */
1069 static void si_vertex_buffers_begin_new_cs(struct si_context
*sctx
)
1071 struct si_descriptors
*desc
= &sctx
->vertex_buffers
;
1072 int count
= sctx
->vertex_elements
? sctx
->vertex_elements
->count
: 0;
1075 for (i
= 0; i
< count
; i
++) {
1076 int vb
= sctx
->vertex_elements
->vertex_buffer_index
[i
];
1078 if (vb
>= ARRAY_SIZE(sctx
->vertex_buffer
))
1080 if (!sctx
->vertex_buffer
[vb
].buffer
.resource
)
1083 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
1084 (struct r600_resource
*)sctx
->vertex_buffer
[vb
].buffer
.resource
,
1085 RADEON_USAGE_READ
, RADEON_PRIO_VERTEX_BUFFER
);
1090 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
1091 desc
->buffer
, RADEON_USAGE_READ
,
1092 RADEON_PRIO_DESCRIPTORS
);
1095 bool si_upload_vertex_buffer_descriptors(struct si_context
*sctx
)
1097 struct si_vertex_elements
*velems
= sctx
->vertex_elements
;
1098 struct si_descriptors
*desc
= &sctx
->vertex_buffers
;
1100 unsigned desc_list_byte_size
;
1101 unsigned first_vb_use_mask
;
1105 if (!sctx
->vertex_buffers_dirty
|| !velems
)
1108 count
= velems
->count
;
1113 desc_list_byte_size
= velems
->desc_list_byte_size
;
1114 first_vb_use_mask
= velems
->first_vb_use_mask
;
1116 /* Vertex buffer descriptors are the only ones which are uploaded
1117 * directly through a staging buffer and don't go through
1118 * the fine-grained upload path.
1120 u_upload_alloc(sctx
->b
.b
.const_uploader
, 0,
1121 desc_list_byte_size
,
1122 si_optimal_tcc_alignment(sctx
, desc_list_byte_size
),
1123 (unsigned*)&desc
->buffer_offset
,
1124 (struct pipe_resource
**)&desc
->buffer
, (void**)&ptr
);
1129 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
1130 desc
->buffer
, RADEON_USAGE_READ
,
1131 RADEON_PRIO_DESCRIPTORS
);
1133 assert(count
<= SI_MAX_ATTRIBS
);
1135 for (i
= 0; i
< count
; i
++) {
1136 struct pipe_vertex_buffer
*vb
;
1137 struct r600_resource
*rbuffer
;
1139 unsigned vbo_index
= velems
->vertex_buffer_index
[i
];
1140 uint32_t *desc
= &ptr
[i
*4];
1142 vb
= &sctx
->vertex_buffer
[vbo_index
];
1143 rbuffer
= (struct r600_resource
*)vb
->buffer
.resource
;
1145 memset(desc
, 0, 16);
1149 offset
= vb
->buffer_offset
+ velems
->src_offset
[i
];
1150 va
= rbuffer
->gpu_address
+ offset
;
1152 /* Fill in T# buffer resource description */
1154 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
1155 S_008F04_STRIDE(vb
->stride
);
1157 if (sctx
->b
.chip_class
!= VI
&& vb
->stride
) {
1158 /* Round up by rounding down and adding 1 */
1159 desc
[2] = (vb
->buffer
.resource
->width0
- offset
-
1160 velems
->format_size
[i
]) /
1163 desc
[2] = vb
->buffer
.resource
->width0
- offset
;
1166 desc
[3] = velems
->rsrc_word3
[i
];
1168 if (first_vb_use_mask
& (1 << i
)) {
1169 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
1170 (struct r600_resource
*)vb
->buffer
.resource
,
1171 RADEON_USAGE_READ
, RADEON_PRIO_VERTEX_BUFFER
);
1175 /* Don't flush the const cache. It would have a very negative effect
1176 * on performance (confirmed by testing). New descriptors are always
1177 * uploaded to a fresh new buffer, so I don't think flushing the const
1178 * cache is needed. */
1179 si_mark_atom_dirty(sctx
, &sctx
->shader_pointers
.atom
);
1180 sctx
->vertex_buffers_dirty
= false;
1181 sctx
->vertex_buffer_pointer_dirty
= true;
1182 sctx
->prefetch_L2_mask
|= SI_PREFETCH_VBO_DESCRIPTORS
;
1187 /* CONSTANT BUFFERS */
1190 si_const_and_shader_buffer_descriptors_idx(unsigned shader
)
1192 return SI_DESCS_FIRST_SHADER
+ shader
* SI_NUM_SHADER_DESCS
+
1193 SI_SHADER_DESCS_CONST_AND_SHADER_BUFFERS
;
1196 static struct si_descriptors
*
1197 si_const_and_shader_buffer_descriptors(struct si_context
*sctx
, unsigned shader
)
1199 return &sctx
->descriptors
[si_const_and_shader_buffer_descriptors_idx(shader
)];
1202 void si_upload_const_buffer(struct si_context
*sctx
, struct r600_resource
**rbuffer
,
1203 const uint8_t *ptr
, unsigned size
, uint32_t *const_offset
)
1207 u_upload_alloc(sctx
->b
.b
.const_uploader
, 0, size
,
1208 si_optimal_tcc_alignment(sctx
, size
),
1210 (struct pipe_resource
**)rbuffer
, &tmp
);
1212 util_memcpy_cpu_to_le32(tmp
, ptr
, size
);
1215 static void si_set_constant_buffer(struct si_context
*sctx
,
1216 struct si_buffer_resources
*buffers
,
1217 unsigned descriptors_idx
,
1218 uint slot
, const struct pipe_constant_buffer
*input
)
1220 struct si_descriptors
*descs
= &sctx
->descriptors
[descriptors_idx
];
1221 assert(slot
< descs
->num_elements
);
1222 pipe_resource_reference(&buffers
->buffers
[slot
], NULL
);
1224 /* CIK cannot unbind a constant buffer (S_BUFFER_LOAD is buggy
1225 * with a NULL buffer). We need to use a dummy buffer instead. */
1226 if (sctx
->b
.chip_class
== CIK
&&
1227 (!input
|| (!input
->buffer
&& !input
->user_buffer
)))
1228 input
= &sctx
->null_const_buf
;
1230 if (input
&& (input
->buffer
|| input
->user_buffer
)) {
1231 struct pipe_resource
*buffer
= NULL
;
1234 /* Upload the user buffer if needed. */
1235 if (input
->user_buffer
) {
1236 unsigned buffer_offset
;
1238 si_upload_const_buffer(sctx
,
1239 (struct r600_resource
**)&buffer
, input
->user_buffer
,
1240 input
->buffer_size
, &buffer_offset
);
1242 /* Just unbind on failure. */
1243 si_set_constant_buffer(sctx
, buffers
, descriptors_idx
, slot
, NULL
);
1246 va
= r600_resource(buffer
)->gpu_address
+ buffer_offset
;
1248 pipe_resource_reference(&buffer
, input
->buffer
);
1249 va
= r600_resource(buffer
)->gpu_address
+ input
->buffer_offset
;
1250 /* Only track usage for non-user buffers. */
1251 r600_resource(buffer
)->bind_history
|= PIPE_BIND_CONSTANT_BUFFER
;
1254 /* Set the descriptor. */
1255 uint32_t *desc
= descs
->list
+ slot
*4;
1257 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
1259 desc
[2] = input
->buffer_size
;
1260 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1261 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1262 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1263 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1264 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1265 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
1267 buffers
->buffers
[slot
] = buffer
;
1268 radeon_add_to_buffer_list_check_mem(&sctx
->b
, &sctx
->b
.gfx
,
1269 (struct r600_resource
*)buffer
,
1270 buffers
->shader_usage_constbuf
,
1271 buffers
->priority_constbuf
, true);
1272 buffers
->enabled_mask
|= 1u << slot
;
1274 /* Clear the descriptor. */
1275 memset(descs
->list
+ slot
*4, 0, sizeof(uint32_t) * 4);
1276 buffers
->enabled_mask
&= ~(1u << slot
);
1279 descs
->dirty_mask
|= 1u << slot
;
1280 sctx
->descriptors_dirty
|= 1u << descriptors_idx
;
1283 void si_set_rw_buffer(struct si_context
*sctx
,
1284 uint slot
, const struct pipe_constant_buffer
*input
)
1286 si_set_constant_buffer(sctx
, &sctx
->rw_buffers
,
1287 SI_DESCS_RW_BUFFERS
, slot
, input
);
1290 static void si_pipe_set_constant_buffer(struct pipe_context
*ctx
,
1291 enum pipe_shader_type shader
, uint slot
,
1292 const struct pipe_constant_buffer
*input
)
1294 struct si_context
*sctx
= (struct si_context
*)ctx
;
1296 if (shader
>= SI_NUM_SHADERS
)
1299 slot
= si_get_constbuf_slot(slot
);
1300 si_set_constant_buffer(sctx
, &sctx
->const_and_shader_buffers
[shader
],
1301 si_const_and_shader_buffer_descriptors_idx(shader
),
1305 void si_get_pipe_constant_buffer(struct si_context
*sctx
, uint shader
,
1306 uint slot
, struct pipe_constant_buffer
*cbuf
)
1308 cbuf
->user_buffer
= NULL
;
1309 si_get_buffer_from_descriptors(
1310 &sctx
->const_and_shader_buffers
[shader
],
1311 si_const_and_shader_buffer_descriptors(sctx
, shader
),
1312 si_get_constbuf_slot(slot
),
1313 &cbuf
->buffer
, &cbuf
->buffer_offset
, &cbuf
->buffer_size
);
1316 /* SHADER BUFFERS */
1318 static void si_set_shader_buffers(struct pipe_context
*ctx
,
1319 enum pipe_shader_type shader
,
1320 unsigned start_slot
, unsigned count
,
1321 const struct pipe_shader_buffer
*sbuffers
)
1323 struct si_context
*sctx
= (struct si_context
*)ctx
;
1324 struct si_buffer_resources
*buffers
= &sctx
->const_and_shader_buffers
[shader
];
1325 struct si_descriptors
*descs
= si_const_and_shader_buffer_descriptors(sctx
, shader
);
1328 assert(start_slot
+ count
<= SI_NUM_SHADER_BUFFERS
);
1330 for (i
= 0; i
< count
; ++i
) {
1331 const struct pipe_shader_buffer
*sbuffer
= sbuffers
? &sbuffers
[i
] : NULL
;
1332 struct r600_resource
*buf
;
1333 unsigned slot
= si_get_shaderbuf_slot(start_slot
+ i
);
1334 uint32_t *desc
= descs
->list
+ slot
* 4;
1337 if (!sbuffer
|| !sbuffer
->buffer
) {
1338 pipe_resource_reference(&buffers
->buffers
[slot
], NULL
);
1339 memset(desc
, 0, sizeof(uint32_t) * 4);
1340 buffers
->enabled_mask
&= ~(1u << slot
);
1341 descs
->dirty_mask
|= 1u << slot
;
1342 sctx
->descriptors_dirty
|=
1343 1u << si_const_and_shader_buffer_descriptors_idx(shader
);
1347 buf
= (struct r600_resource
*)sbuffer
->buffer
;
1348 va
= buf
->gpu_address
+ sbuffer
->buffer_offset
;
1351 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
1353 desc
[2] = sbuffer
->buffer_size
;
1354 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1355 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1356 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1357 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1358 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1359 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
1361 pipe_resource_reference(&buffers
->buffers
[slot
], &buf
->b
.b
);
1362 radeon_add_to_buffer_list_check_mem(&sctx
->b
, &sctx
->b
.gfx
, buf
,
1363 buffers
->shader_usage
,
1364 buffers
->priority
, true);
1365 buf
->bind_history
|= PIPE_BIND_SHADER_BUFFER
;
1367 buffers
->enabled_mask
|= 1u << slot
;
1368 descs
->dirty_mask
|= 1u << slot
;
1369 sctx
->descriptors_dirty
|=
1370 1u << si_const_and_shader_buffer_descriptors_idx(shader
);
1372 util_range_add(&buf
->valid_buffer_range
, sbuffer
->buffer_offset
,
1373 sbuffer
->buffer_offset
+ sbuffer
->buffer_size
);
1377 void si_get_shader_buffers(struct si_context
*sctx
,
1378 enum pipe_shader_type shader
,
1379 uint start_slot
, uint count
,
1380 struct pipe_shader_buffer
*sbuf
)
1382 struct si_buffer_resources
*buffers
= &sctx
->const_and_shader_buffers
[shader
];
1383 struct si_descriptors
*descs
= si_const_and_shader_buffer_descriptors(sctx
, shader
);
1385 for (unsigned i
= 0; i
< count
; ++i
) {
1386 si_get_buffer_from_descriptors(
1388 si_get_shaderbuf_slot(start_slot
+ i
),
1389 &sbuf
[i
].buffer
, &sbuf
[i
].buffer_offset
,
1390 &sbuf
[i
].buffer_size
);
1396 void si_set_ring_buffer(struct pipe_context
*ctx
, uint slot
,
1397 struct pipe_resource
*buffer
,
1398 unsigned stride
, unsigned num_records
,
1399 bool add_tid
, bool swizzle
,
1400 unsigned element_size
, unsigned index_stride
, uint64_t offset
)
1402 struct si_context
*sctx
= (struct si_context
*)ctx
;
1403 struct si_buffer_resources
*buffers
= &sctx
->rw_buffers
;
1404 struct si_descriptors
*descs
= &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
];
1406 /* The stride field in the resource descriptor has 14 bits */
1407 assert(stride
< (1 << 14));
1409 assert(slot
< descs
->num_elements
);
1410 pipe_resource_reference(&buffers
->buffers
[slot
], NULL
);
1415 va
= r600_resource(buffer
)->gpu_address
+ offset
;
1417 switch (element_size
) {
1419 assert(!"Unsupported ring buffer element size");
1435 switch (index_stride
) {
1437 assert(!"Unsupported ring buffer index stride");
1453 if (sctx
->b
.chip_class
>= VI
&& stride
)
1454 num_records
*= stride
;
1456 /* Set the descriptor. */
1457 uint32_t *desc
= descs
->list
+ slot
*4;
1459 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32) |
1460 S_008F04_STRIDE(stride
) |
1461 S_008F04_SWIZZLE_ENABLE(swizzle
);
1462 desc
[2] = num_records
;
1463 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1464 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1465 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1466 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1467 S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT
) |
1468 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
) |
1469 S_008F0C_INDEX_STRIDE(index_stride
) |
1470 S_008F0C_ADD_TID_ENABLE(add_tid
);
1472 if (sctx
->b
.chip_class
>= GFX9
)
1473 assert(!swizzle
|| element_size
== 1); /* always 4 bytes on GFX9 */
1475 desc
[3] |= S_008F0C_ELEMENT_SIZE(element_size
);
1477 pipe_resource_reference(&buffers
->buffers
[slot
], buffer
);
1478 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
1479 (struct r600_resource
*)buffer
,
1480 buffers
->shader_usage
, buffers
->priority
);
1481 buffers
->enabled_mask
|= 1u << slot
;
1483 /* Clear the descriptor. */
1484 memset(descs
->list
+ slot
*4, 0, sizeof(uint32_t) * 4);
1485 buffers
->enabled_mask
&= ~(1u << slot
);
1488 descs
->dirty_mask
|= 1u << slot
;
1489 sctx
->descriptors_dirty
|= 1u << SI_DESCS_RW_BUFFERS
;
1492 /* STREAMOUT BUFFERS */
1494 static void si_set_streamout_targets(struct pipe_context
*ctx
,
1495 unsigned num_targets
,
1496 struct pipe_stream_output_target
**targets
,
1497 const unsigned *offsets
)
1499 struct si_context
*sctx
= (struct si_context
*)ctx
;
1500 struct si_buffer_resources
*buffers
= &sctx
->rw_buffers
;
1501 struct si_descriptors
*descs
= &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
];
1502 unsigned old_num_targets
= sctx
->b
.streamout
.num_targets
;
1505 /* We are going to unbind the buffers. Mark which caches need to be flushed. */
1506 if (sctx
->b
.streamout
.num_targets
&& sctx
->b
.streamout
.begin_emitted
) {
1507 /* Since streamout uses vector writes which go through TC L2
1508 * and most other clients can use TC L2 as well, we don't need
1511 * The only cases which requires flushing it is VGT DMA index
1512 * fetching (on <= CIK) and indirect draw data, which are rare
1513 * cases. Thus, flag the TC L2 dirtiness in the resource and
1514 * handle it at draw call time.
1516 for (i
= 0; i
< sctx
->b
.streamout
.num_targets
; i
++)
1517 if (sctx
->b
.streamout
.targets
[i
])
1518 r600_resource(sctx
->b
.streamout
.targets
[i
]->b
.buffer
)->TC_L2_dirty
= true;
1520 /* Invalidate the scalar cache in case a streamout buffer is
1521 * going to be used as a constant buffer.
1523 * Invalidate TC L1, because streamout bypasses it (done by
1524 * setting GLC=1 in the store instruction), but it can contain
1525 * outdated data of streamout buffers.
1527 * VS_PARTIAL_FLUSH is required if the buffers are going to be
1528 * used as an input immediately.
1530 sctx
->b
.flags
|= SI_CONTEXT_INV_SMEM_L1
|
1531 SI_CONTEXT_INV_VMEM_L1
|
1532 SI_CONTEXT_VS_PARTIAL_FLUSH
;
1535 /* All readers of the streamout targets need to be finished before we can
1536 * start writing to the targets.
1539 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
1540 SI_CONTEXT_CS_PARTIAL_FLUSH
;
1542 /* Streamout buffers must be bound in 2 places:
1543 * 1) in VGT by setting the VGT_STRMOUT registers
1544 * 2) as shader resources
1547 /* Set the VGT regs. */
1548 r600_set_streamout_targets(ctx
, num_targets
, targets
, offsets
);
1550 /* Set the shader resources.*/
1551 for (i
= 0; i
< num_targets
; i
++) {
1552 bufidx
= SI_VS_STREAMOUT_BUF0
+ i
;
1555 struct pipe_resource
*buffer
= targets
[i
]->buffer
;
1556 uint64_t va
= r600_resource(buffer
)->gpu_address
;
1558 /* Set the descriptor.
1560 * On VI, the format must be non-INVALID, otherwise
1561 * the buffer will be considered not bound and store
1562 * instructions will be no-ops.
1564 uint32_t *desc
= descs
->list
+ bufidx
*4;
1566 desc
[1] = S_008F04_BASE_ADDRESS_HI(va
>> 32);
1567 desc
[2] = 0xffffffff;
1568 desc
[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X
) |
1569 S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y
) |
1570 S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z
) |
1571 S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W
) |
1572 S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32
);
1574 /* Set the resource. */
1575 pipe_resource_reference(&buffers
->buffers
[bufidx
],
1577 radeon_add_to_buffer_list_check_mem(&sctx
->b
, &sctx
->b
.gfx
,
1578 (struct r600_resource
*)buffer
,
1579 buffers
->shader_usage
,
1580 RADEON_PRIO_SHADER_RW_BUFFER
,
1582 r600_resource(buffer
)->bind_history
|= PIPE_BIND_STREAM_OUTPUT
;
1584 buffers
->enabled_mask
|= 1u << bufidx
;
1586 /* Clear the descriptor and unset the resource. */
1587 memset(descs
->list
+ bufidx
*4, 0,
1588 sizeof(uint32_t) * 4);
1589 pipe_resource_reference(&buffers
->buffers
[bufidx
],
1591 buffers
->enabled_mask
&= ~(1u << bufidx
);
1593 descs
->dirty_mask
|= 1u << bufidx
;
1595 for (; i
< old_num_targets
; i
++) {
1596 bufidx
= SI_VS_STREAMOUT_BUF0
+ i
;
1597 /* Clear the descriptor and unset the resource. */
1598 memset(descs
->list
+ bufidx
*4, 0, sizeof(uint32_t) * 4);
1599 pipe_resource_reference(&buffers
->buffers
[bufidx
], NULL
);
1600 buffers
->enabled_mask
&= ~(1u << bufidx
);
1601 descs
->dirty_mask
|= 1u << bufidx
;
1604 sctx
->descriptors_dirty
|= 1u << SI_DESCS_RW_BUFFERS
;
1607 static void si_desc_reset_buffer_offset(struct pipe_context
*ctx
,
1608 uint32_t *desc
, uint64_t old_buf_va
,
1609 struct pipe_resource
*new_buf
)
1611 /* Retrieve the buffer offset from the descriptor. */
1612 uint64_t old_desc_va
=
1613 desc
[0] | ((uint64_t)G_008F04_BASE_ADDRESS_HI(desc
[1]) << 32);
1615 assert(old_buf_va
<= old_desc_va
);
1616 uint64_t offset_within_buffer
= old_desc_va
- old_buf_va
;
1618 /* Update the descriptor. */
1619 si_set_buf_desc_address(r600_resource(new_buf
), offset_within_buffer
,
1623 /* INTERNAL CONST BUFFERS */
1625 static void si_set_polygon_stipple(struct pipe_context
*ctx
,
1626 const struct pipe_poly_stipple
*state
)
1628 struct si_context
*sctx
= (struct si_context
*)ctx
;
1629 struct pipe_constant_buffer cb
= {};
1630 unsigned stipple
[32];
1633 for (i
= 0; i
< 32; i
++)
1634 stipple
[i
] = util_bitreverse(state
->stipple
[i
]);
1636 cb
.user_buffer
= stipple
;
1637 cb
.buffer_size
= sizeof(stipple
);
1639 si_set_rw_buffer(sctx
, SI_PS_CONST_POLY_STIPPLE
, &cb
);
1642 /* TEXTURE METADATA ENABLE/DISABLE */
1645 si_resident_handles_update_needs_color_decompress(struct si_context
*sctx
)
1647 util_dynarray_clear(&sctx
->resident_tex_needs_color_decompress
);
1648 util_dynarray_clear(&sctx
->resident_img_needs_color_decompress
);
1650 util_dynarray_foreach(&sctx
->resident_tex_handles
,
1651 struct si_texture_handle
*, tex_handle
) {
1652 struct pipe_resource
*res
= (*tex_handle
)->view
->texture
;
1653 struct r600_texture
*rtex
;
1655 if (!res
|| res
->target
== PIPE_BUFFER
)
1658 rtex
= (struct r600_texture
*)res
;
1659 if (!color_needs_decompression(rtex
))
1662 util_dynarray_append(&sctx
->resident_tex_needs_color_decompress
,
1663 struct si_texture_handle
*, *tex_handle
);
1666 util_dynarray_foreach(&sctx
->resident_img_handles
,
1667 struct si_image_handle
*, img_handle
) {
1668 struct pipe_image_view
*view
= &(*img_handle
)->view
;
1669 struct pipe_resource
*res
= view
->resource
;
1670 struct r600_texture
*rtex
;
1672 if (!res
|| res
->target
== PIPE_BUFFER
)
1675 rtex
= (struct r600_texture
*)res
;
1676 if (!color_needs_decompression(rtex
))
1679 util_dynarray_append(&sctx
->resident_img_needs_color_decompress
,
1680 struct si_image_handle
*, *img_handle
);
1684 /* CMASK can be enabled (for fast clear) and disabled (for texture export)
1685 * while the texture is bound, possibly by a different context. In that case,
1686 * call this function to update needs_*_decompress_masks.
1688 void si_update_needs_color_decompress_masks(struct si_context
*sctx
)
1690 for (int i
= 0; i
< SI_NUM_SHADERS
; ++i
) {
1691 si_samplers_update_needs_color_decompress_mask(&sctx
->samplers
[i
]);
1692 si_images_update_needs_color_decompress_mask(&sctx
->images
[i
]);
1693 si_update_shader_needs_decompress_mask(sctx
, i
);
1696 si_resident_handles_update_needs_color_decompress(sctx
);
1699 /* BUFFER DISCARD/INVALIDATION */
1701 /** Reset descriptors of buffer resources after \p buf has been invalidated. */
1702 static void si_reset_buffer_resources(struct si_context
*sctx
,
1703 struct si_buffer_resources
*buffers
,
1704 unsigned descriptors_idx
,
1706 struct pipe_resource
*buf
,
1708 enum radeon_bo_usage usage
,
1709 enum radeon_bo_priority priority
)
1711 struct si_descriptors
*descs
= &sctx
->descriptors
[descriptors_idx
];
1712 unsigned mask
= buffers
->enabled_mask
& slot_mask
;
1715 unsigned i
= u_bit_scan(&mask
);
1716 if (buffers
->buffers
[i
] == buf
) {
1717 si_desc_reset_buffer_offset(&sctx
->b
.b
,
1720 descs
->dirty_mask
|= 1u << i
;
1721 sctx
->descriptors_dirty
|= 1u << descriptors_idx
;
1723 radeon_add_to_buffer_list_check_mem(&sctx
->b
, &sctx
->b
.gfx
,
1724 (struct r600_resource
*)buf
,
1725 usage
, priority
, true);
1730 static void si_rebind_buffer(struct pipe_context
*ctx
, struct pipe_resource
*buf
,
1733 struct si_context
*sctx
= (struct si_context
*)ctx
;
1734 struct r600_resource
*rbuffer
= r600_resource(buf
);
1736 unsigned num_elems
= sctx
->vertex_elements
?
1737 sctx
->vertex_elements
->count
: 0;
1739 /* We changed the buffer, now we need to bind it where the old one
1740 * was bound. This consists of 2 things:
1741 * 1) Updating the resource descriptor and dirtying it.
1742 * 2) Adding a relocation to the CS, so that it's usable.
1745 /* Vertex buffers. */
1746 if (rbuffer
->bind_history
& PIPE_BIND_VERTEX_BUFFER
) {
1747 for (i
= 0; i
< num_elems
; i
++) {
1748 int vb
= sctx
->vertex_elements
->vertex_buffer_index
[i
];
1750 if (vb
>= ARRAY_SIZE(sctx
->vertex_buffer
))
1752 if (!sctx
->vertex_buffer
[vb
].buffer
.resource
)
1755 if (sctx
->vertex_buffer
[vb
].buffer
.resource
== buf
) {
1756 sctx
->vertex_buffers_dirty
= true;
1762 /* Streamout buffers. (other internal buffers can't be invalidated) */
1763 if (rbuffer
->bind_history
& PIPE_BIND_STREAM_OUTPUT
) {
1764 for (i
= SI_VS_STREAMOUT_BUF0
; i
<= SI_VS_STREAMOUT_BUF3
; i
++) {
1765 struct si_buffer_resources
*buffers
= &sctx
->rw_buffers
;
1766 struct si_descriptors
*descs
=
1767 &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
];
1769 if (buffers
->buffers
[i
] != buf
)
1772 si_desc_reset_buffer_offset(ctx
, descs
->list
+ i
*4,
1774 descs
->dirty_mask
|= 1u << i
;
1775 sctx
->descriptors_dirty
|= 1u << SI_DESCS_RW_BUFFERS
;
1777 radeon_add_to_buffer_list_check_mem(&sctx
->b
, &sctx
->b
.gfx
,
1778 rbuffer
, buffers
->shader_usage
,
1779 RADEON_PRIO_SHADER_RW_BUFFER
,
1782 /* Update the streamout state. */
1783 if (sctx
->b
.streamout
.begin_emitted
)
1784 r600_emit_streamout_end(&sctx
->b
);
1785 sctx
->b
.streamout
.append_bitmask
=
1786 sctx
->b
.streamout
.enabled_mask
;
1787 r600_streamout_buffers_dirty(&sctx
->b
);
1791 /* Constant and shader buffers. */
1792 if (rbuffer
->bind_history
& PIPE_BIND_CONSTANT_BUFFER
) {
1793 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++)
1794 si_reset_buffer_resources(sctx
, &sctx
->const_and_shader_buffers
[shader
],
1795 si_const_and_shader_buffer_descriptors_idx(shader
),
1796 u_bit_consecutive(SI_NUM_SHADER_BUFFERS
, SI_NUM_CONST_BUFFERS
),
1798 sctx
->const_and_shader_buffers
[shader
].shader_usage_constbuf
,
1799 sctx
->const_and_shader_buffers
[shader
].priority_constbuf
);
1802 if (rbuffer
->bind_history
& PIPE_BIND_SHADER_BUFFER
) {
1803 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++)
1804 si_reset_buffer_resources(sctx
, &sctx
->const_and_shader_buffers
[shader
],
1805 si_const_and_shader_buffer_descriptors_idx(shader
),
1806 u_bit_consecutive(0, SI_NUM_SHADER_BUFFERS
),
1808 sctx
->const_and_shader_buffers
[shader
].shader_usage
,
1809 sctx
->const_and_shader_buffers
[shader
].priority
);
1812 if (rbuffer
->bind_history
& PIPE_BIND_SAMPLER_VIEW
) {
1813 /* Texture buffers - update bindings. */
1814 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++) {
1815 struct si_sampler_views
*views
= &sctx
->samplers
[shader
].views
;
1816 struct si_descriptors
*descs
=
1817 si_sampler_and_image_descriptors(sctx
, shader
);
1818 unsigned mask
= views
->enabled_mask
;
1821 unsigned i
= u_bit_scan(&mask
);
1822 if (views
->views
[i
]->texture
== buf
) {
1823 unsigned desc_slot
= si_get_sampler_slot(i
);
1825 si_desc_reset_buffer_offset(ctx
,
1829 descs
->dirty_mask
|= 1ull << desc_slot
;
1830 sctx
->descriptors_dirty
|=
1831 1u << si_sampler_and_image_descriptors_idx(shader
);
1833 radeon_add_to_buffer_list_check_mem(&sctx
->b
, &sctx
->b
.gfx
,
1834 rbuffer
, RADEON_USAGE_READ
,
1835 RADEON_PRIO_SAMPLER_BUFFER
,
1843 if (rbuffer
->bind_history
& PIPE_BIND_SHADER_IMAGE
) {
1844 for (shader
= 0; shader
< SI_NUM_SHADERS
; ++shader
) {
1845 struct si_images_info
*images
= &sctx
->images
[shader
];
1846 struct si_descriptors
*descs
=
1847 si_sampler_and_image_descriptors(sctx
, shader
);
1848 unsigned mask
= images
->enabled_mask
;
1851 unsigned i
= u_bit_scan(&mask
);
1853 if (images
->views
[i
].resource
== buf
) {
1854 unsigned desc_slot
= si_get_image_slot(i
);
1856 if (images
->views
[i
].access
& PIPE_IMAGE_ACCESS_WRITE
)
1857 si_mark_image_range_valid(&images
->views
[i
]);
1859 si_desc_reset_buffer_offset(
1860 ctx
, descs
->list
+ desc_slot
* 8 + 4,
1862 /* two 8-byte images share one 16-byte slot */
1863 descs
->dirty_mask
|= 1u << (desc_slot
/ 2);
1864 sctx
->descriptors_dirty
|=
1865 1u << si_sampler_and_image_descriptors_idx(shader
);
1867 radeon_add_to_buffer_list_check_mem(
1868 &sctx
->b
, &sctx
->b
.gfx
, rbuffer
,
1869 RADEON_USAGE_READWRITE
,
1870 RADEON_PRIO_SAMPLER_BUFFER
, true);
1876 /* Bindless texture handles */
1877 if (rbuffer
->texture_handle_allocated
) {
1878 util_dynarray_foreach(&sctx
->resident_tex_handles
,
1879 struct si_texture_handle
*, tex_handle
) {
1880 struct pipe_sampler_view
*view
= (*tex_handle
)->view
;
1881 struct si_bindless_descriptor
*desc
= (*tex_handle
)->desc
;
1883 if (view
->texture
== buf
) {
1884 si_set_buf_desc_address(rbuffer
,
1886 &desc
->desc_list
[4]);
1888 sctx
->bindless_descriptors_dirty
= true;
1890 radeon_add_to_buffer_list_check_mem(
1891 &sctx
->b
, &sctx
->b
.gfx
, rbuffer
,
1893 RADEON_PRIO_SAMPLER_BUFFER
, true);
1898 /* Bindless image handles */
1899 if (rbuffer
->image_handle_allocated
) {
1900 util_dynarray_foreach(&sctx
->resident_img_handles
,
1901 struct si_image_handle
*, img_handle
) {
1902 struct pipe_image_view
*view
= &(*img_handle
)->view
;
1903 struct si_bindless_descriptor
*desc
= (*img_handle
)->desc
;
1905 if (view
->resource
== buf
) {
1906 if (view
->access
& PIPE_IMAGE_ACCESS_WRITE
)
1907 si_mark_image_range_valid(view
);
1909 si_set_buf_desc_address(rbuffer
,
1911 &desc
->desc_list
[4]);
1913 sctx
->bindless_descriptors_dirty
= true;
1915 radeon_add_to_buffer_list_check_mem(
1916 &sctx
->b
, &sctx
->b
.gfx
, rbuffer
,
1917 RADEON_USAGE_READWRITE
,
1918 RADEON_PRIO_SAMPLER_BUFFER
, true);
1924 /* Reallocate a buffer a update all resource bindings where the buffer is
1927 * This is used to avoid CPU-GPU synchronizations, because it makes the buffer
1928 * idle by discarding its contents. Apps usually tell us when to do this using
1929 * map_buffer flags, for example.
1931 static void si_invalidate_buffer(struct pipe_context
*ctx
, struct pipe_resource
*buf
)
1933 struct si_context
*sctx
= (struct si_context
*)ctx
;
1934 struct r600_resource
*rbuffer
= r600_resource(buf
);
1935 uint64_t old_va
= rbuffer
->gpu_address
;
1937 /* Reallocate the buffer in the same pipe_resource. */
1938 r600_alloc_resource(&sctx
->screen
->b
, rbuffer
);
1940 si_rebind_buffer(ctx
, buf
, old_va
);
1943 static void si_upload_bindless_descriptor(struct si_context
*sctx
,
1944 struct si_bindless_descriptor
*desc
)
1946 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
1947 uint64_t va
= desc
->buffer
->gpu_address
+ desc
->offset
;
1948 unsigned num_dwords
= sizeof(desc
->desc_list
) / 4;
1950 radeon_emit(cs
, PKT3(PKT3_WRITE_DATA
, 2 + num_dwords
, 0));
1951 radeon_emit(cs
, S_370_DST_SEL(V_370_TC_L2
) |
1952 S_370_WR_CONFIRM(1) |
1953 S_370_ENGINE_SEL(V_370_ME
));
1954 radeon_emit(cs
, va
);
1955 radeon_emit(cs
, va
>> 32);
1956 radeon_emit_array(cs
, desc
->desc_list
, num_dwords
);
1959 static void si_upload_bindless_descriptors(struct si_context
*sctx
)
1961 if (!sctx
->bindless_descriptors_dirty
)
1964 /* Wait for graphics/compute to be idle before updating the resident
1965 * descriptors directly in memory, in case the GPU is using them.
1967 sctx
->b
.flags
|= SI_CONTEXT_PS_PARTIAL_FLUSH
|
1968 SI_CONTEXT_CS_PARTIAL_FLUSH
;
1969 si_emit_cache_flush(sctx
);
1971 util_dynarray_foreach(&sctx
->resident_tex_handles
,
1972 struct si_texture_handle
*, tex_handle
) {
1973 struct si_bindless_descriptor
*desc
= (*tex_handle
)->desc
;
1978 si_upload_bindless_descriptor(sctx
, desc
);
1979 desc
->dirty
= false;
1982 util_dynarray_foreach(&sctx
->resident_img_handles
,
1983 struct si_image_handle
*, img_handle
) {
1984 struct si_bindless_descriptor
*desc
= (*img_handle
)->desc
;
1989 si_upload_bindless_descriptor(sctx
, desc
);
1990 desc
->dirty
= false;
1993 /* Invalidate L1 because it doesn't know that L2 changed. */
1994 sctx
->b
.flags
|= SI_CONTEXT_INV_SMEM_L1
;
1995 si_emit_cache_flush(sctx
);
1997 sctx
->bindless_descriptors_dirty
= false;
2000 /* Update mutable image descriptor fields of all resident textures. */
2001 static void si_update_resident_texture_descriptor(struct si_context
*sctx
,
2002 struct si_texture_handle
*tex_handle
)
2004 struct si_sampler_view
*sview
= (struct si_sampler_view
*)tex_handle
->view
;
2005 struct si_bindless_descriptor
*desc
= tex_handle
->desc
;
2006 uint32_t desc_list
[16];
2008 if (sview
->base
.texture
->target
== PIPE_BUFFER
)
2011 memcpy(desc_list
, desc
->desc_list
, sizeof(desc_list
));
2012 si_set_sampler_view_desc(sctx
, sview
, &tex_handle
->sstate
,
2013 &desc
->desc_list
[0]);
2015 if (memcmp(desc_list
, desc
->desc_list
, sizeof(desc_list
))) {
2017 sctx
->bindless_descriptors_dirty
= true;
2021 static void si_update_resident_image_descriptor(struct si_context
*sctx
,
2022 struct si_image_handle
*img_handle
)
2024 struct si_bindless_descriptor
*desc
= img_handle
->desc
;
2025 struct pipe_image_view
*view
= &img_handle
->view
;
2026 uint32_t desc_list
[16];
2028 if (view
->resource
->target
== PIPE_BUFFER
)
2031 memcpy(desc_list
, desc
->desc_list
, sizeof(desc_list
));
2032 si_set_shader_image_desc(sctx
, view
, true, &desc
->desc_list
[0]);
2034 if (memcmp(desc_list
, desc
->desc_list
, sizeof(desc_list
))) {
2036 sctx
->bindless_descriptors_dirty
= true;
2040 static void si_update_all_resident_texture_descriptors(struct si_context
*sctx
)
2042 util_dynarray_foreach(&sctx
->resident_tex_handles
,
2043 struct si_texture_handle
*, tex_handle
) {
2044 si_update_resident_texture_descriptor(sctx
, *tex_handle
);
2047 util_dynarray_foreach(&sctx
->resident_img_handles
,
2048 struct si_image_handle
*, img_handle
) {
2049 si_update_resident_image_descriptor(sctx
, *img_handle
);
2052 si_upload_bindless_descriptors(sctx
);
2055 /* Update mutable image descriptor fields of all bound textures. */
2056 void si_update_all_texture_descriptors(struct si_context
*sctx
)
2060 for (shader
= 0; shader
< SI_NUM_SHADERS
; shader
++) {
2061 struct si_sampler_views
*samplers
= &sctx
->samplers
[shader
].views
;
2062 struct si_images_info
*images
= &sctx
->images
[shader
];
2066 mask
= images
->enabled_mask
;
2068 unsigned i
= u_bit_scan(&mask
);
2069 struct pipe_image_view
*view
= &images
->views
[i
];
2071 if (!view
->resource
||
2072 view
->resource
->target
== PIPE_BUFFER
)
2075 si_set_shader_image(sctx
, shader
, i
, view
, true);
2078 /* Sampler views. */
2079 mask
= samplers
->enabled_mask
;
2081 unsigned i
= u_bit_scan(&mask
);
2082 struct pipe_sampler_view
*view
= samplers
->views
[i
];
2086 view
->texture
->target
== PIPE_BUFFER
)
2089 si_set_sampler_view(sctx
, shader
, i
,
2090 samplers
->views
[i
], true);
2093 si_update_shader_needs_decompress_mask(sctx
, shader
);
2096 si_update_all_resident_texture_descriptors(sctx
);
2099 /* SHADER USER DATA */
2101 static void si_mark_shader_pointers_dirty(struct si_context
*sctx
,
2104 sctx
->shader_pointers_dirty
|=
2105 u_bit_consecutive(SI_DESCS_FIRST_SHADER
+ shader
* SI_NUM_SHADER_DESCS
,
2106 SI_NUM_SHADER_DESCS
);
2108 if (shader
== PIPE_SHADER_VERTEX
)
2109 sctx
->vertex_buffer_pointer_dirty
= sctx
->vertex_buffers
.buffer
!= NULL
;
2111 si_mark_atom_dirty(sctx
, &sctx
->shader_pointers
.atom
);
2114 static void si_shader_pointers_begin_new_cs(struct si_context
*sctx
)
2116 sctx
->shader_pointers_dirty
= u_bit_consecutive(0, SI_NUM_DESCS
);
2117 sctx
->vertex_buffer_pointer_dirty
= sctx
->vertex_buffers
.buffer
!= NULL
;
2118 si_mark_atom_dirty(sctx
, &sctx
->shader_pointers
.atom
);
2121 /* Set a base register address for user data constants in the given shader.
2122 * This assigns a mapping from PIPE_SHADER_* to SPI_SHADER_USER_DATA_*.
2124 static void si_set_user_data_base(struct si_context
*sctx
,
2125 unsigned shader
, uint32_t new_base
)
2127 uint32_t *base
= &sctx
->shader_pointers
.sh_base
[shader
];
2129 if (*base
!= new_base
) {
2133 si_mark_shader_pointers_dirty(sctx
, shader
);
2135 if (shader
== PIPE_SHADER_VERTEX
)
2136 sctx
->last_vs_state
= ~0;
2141 /* This must be called when these shaders are changed from non-NULL to NULL
2144 * - tessellation control shader
2145 * - tessellation evaluation shader
2147 void si_shader_change_notify(struct si_context
*sctx
)
2149 /* VS can be bound as VS, ES, or LS. */
2150 if (sctx
->tes_shader
.cso
) {
2151 if (sctx
->b
.chip_class
>= GFX9
) {
2152 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2153 R_00B430_SPI_SHADER_USER_DATA_LS_0
);
2155 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2156 R_00B530_SPI_SHADER_USER_DATA_LS_0
);
2158 } else if (sctx
->gs_shader
.cso
) {
2159 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2160 R_00B330_SPI_SHADER_USER_DATA_ES_0
);
2162 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
,
2163 R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2166 /* TES can be bound as ES, VS, or not bound. */
2167 if (sctx
->tes_shader
.cso
) {
2168 if (sctx
->gs_shader
.cso
)
2169 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_EVAL
,
2170 R_00B330_SPI_SHADER_USER_DATA_ES_0
);
2172 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_EVAL
,
2173 R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2175 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_EVAL
, 0);
2179 static void si_emit_shader_pointer(struct si_context
*sctx
,
2180 struct si_descriptors
*desc
,
2183 struct radeon_winsys_cs
*cs
= sctx
->b
.gfx
.cs
;
2187 return; /* the pointer is not used by current shaders */
2189 va
= desc
->buffer
->gpu_address
+
2190 desc
->buffer_offset
;
2192 radeon_emit(cs
, PKT3(PKT3_SET_SH_REG
, 2, 0));
2193 radeon_emit(cs
, (sh_base
+ desc
->shader_userdata_offset
- SI_SH_REG_OFFSET
) >> 2);
2194 radeon_emit(cs
, va
);
2195 radeon_emit(cs
, va
>> 32);
2198 static void si_emit_global_shader_pointers(struct si_context
*sctx
,
2199 struct si_descriptors
*descs
)
2201 si_emit_shader_pointer(sctx
, descs
,
2202 R_00B030_SPI_SHADER_USER_DATA_PS_0
);
2203 si_emit_shader_pointer(sctx
, descs
,
2204 R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2206 if (sctx
->b
.chip_class
>= GFX9
) {
2207 /* GFX9 merged LS-HS and ES-GS.
2208 * Set RW_BUFFERS in the special registers, so that
2209 * it's preloaded into s[0:1] instead of s[8:9].
2211 si_emit_shader_pointer(sctx
, descs
,
2212 R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS
);
2213 si_emit_shader_pointer(sctx
, descs
,
2214 R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS
);
2216 si_emit_shader_pointer(sctx
, descs
,
2217 R_00B230_SPI_SHADER_USER_DATA_GS_0
);
2218 si_emit_shader_pointer(sctx
, descs
,
2219 R_00B330_SPI_SHADER_USER_DATA_ES_0
);
2220 si_emit_shader_pointer(sctx
, descs
,
2221 R_00B430_SPI_SHADER_USER_DATA_HS_0
);
2222 si_emit_shader_pointer(sctx
, descs
,
2223 R_00B530_SPI_SHADER_USER_DATA_LS_0
);
2227 void si_emit_graphics_shader_pointers(struct si_context
*sctx
,
2228 struct r600_atom
*atom
)
2231 uint32_t *sh_base
= sctx
->shader_pointers
.sh_base
;
2232 struct si_descriptors
*descs
;
2234 descs
= &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
];
2236 if (sctx
->shader_pointers_dirty
& (1 << SI_DESCS_RW_BUFFERS
))
2237 si_emit_global_shader_pointers(sctx
, descs
);
2239 mask
= sctx
->shader_pointers_dirty
&
2240 u_bit_consecutive(SI_DESCS_FIRST_SHADER
,
2241 SI_DESCS_FIRST_COMPUTE
- SI_DESCS_FIRST_SHADER
);
2244 unsigned i
= u_bit_scan(&mask
);
2245 unsigned shader
= (i
- SI_DESCS_FIRST_SHADER
) / SI_NUM_SHADER_DESCS
;
2246 unsigned base
= sh_base
[shader
];
2249 si_emit_shader_pointer(sctx
, descs
+ i
, base
);
2251 sctx
->shader_pointers_dirty
&=
2252 ~u_bit_consecutive(SI_DESCS_RW_BUFFERS
, SI_DESCS_FIRST_COMPUTE
);
2254 if (sctx
->vertex_buffer_pointer_dirty
) {
2255 si_emit_shader_pointer(sctx
, &sctx
->vertex_buffers
,
2256 sh_base
[PIPE_SHADER_VERTEX
]);
2257 sctx
->vertex_buffer_pointer_dirty
= false;
2261 void si_emit_compute_shader_pointers(struct si_context
*sctx
)
2263 unsigned base
= R_00B900_COMPUTE_USER_DATA_0
;
2264 struct si_descriptors
*descs
= sctx
->descriptors
;
2265 unsigned compute_mask
=
2266 u_bit_consecutive(SI_DESCS_FIRST_COMPUTE
, SI_NUM_SHADER_DESCS
);
2267 unsigned mask
= sctx
->shader_pointers_dirty
& compute_mask
;
2270 unsigned i
= u_bit_scan(&mask
);
2272 si_emit_shader_pointer(sctx
, descs
+ i
, base
);
2274 sctx
->shader_pointers_dirty
&= ~compute_mask
;
2279 struct si_bindless_descriptor_slab
2281 struct pb_slab base
;
2282 struct r600_resource
*buffer
;
2283 struct si_bindless_descriptor
*entries
;
2286 bool si_bindless_descriptor_can_reclaim_slab(void *priv
,
2287 struct pb_slab_entry
*entry
)
2289 /* Do not allow to reclaim any bindless descriptors for now because the
2290 * GPU might be using them. This should be improved later on.
2295 struct pb_slab
*si_bindless_descriptor_slab_alloc(void *priv
, unsigned heap
,
2296 unsigned entry_size
,
2297 unsigned group_index
)
2299 struct si_context
*sctx
= priv
;
2300 struct si_screen
*sscreen
= sctx
->screen
;
2301 struct si_bindless_descriptor_slab
*slab
;
2303 slab
= CALLOC_STRUCT(si_bindless_descriptor_slab
);
2307 /* Create a buffer in VRAM for 1024 bindless descriptors. */
2308 slab
->buffer
= (struct r600_resource
*)
2309 pipe_buffer_create(&sscreen
->b
.b
, 0,
2310 PIPE_USAGE_DEFAULT
, 64 * 1024);
2314 slab
->base
.num_entries
= slab
->buffer
->bo_size
/ entry_size
;
2315 slab
->base
.num_free
= slab
->base
.num_entries
;
2316 slab
->entries
= CALLOC(slab
->base
.num_entries
, sizeof(*slab
->entries
));
2320 LIST_INITHEAD(&slab
->base
.free
);
2322 for (unsigned i
= 0; i
< slab
->base
.num_entries
; ++i
) {
2323 struct si_bindless_descriptor
*desc
= &slab
->entries
[i
];
2325 desc
->entry
.slab
= &slab
->base
;
2326 desc
->entry
.group_index
= group_index
;
2327 desc
->buffer
= slab
->buffer
;
2328 desc
->offset
= i
* entry_size
;
2330 LIST_ADDTAIL(&desc
->entry
.head
, &slab
->base
.free
);
2333 /* Add the descriptor to the per-context list. */
2334 util_dynarray_append(&sctx
->bindless_descriptors
,
2335 struct r600_resource
*, slab
->buffer
);
2340 r600_resource_reference(&slab
->buffer
, NULL
);
2346 void si_bindless_descriptor_slab_free(void *priv
, struct pb_slab
*pslab
)
2348 struct si_context
*sctx
= priv
;
2349 struct si_bindless_descriptor_slab
*slab
=
2350 (struct si_bindless_descriptor_slab
*)pslab
;
2352 /* Remove the descriptor from the per-context list. */
2353 util_dynarray_delete_unordered(&sctx
->bindless_descriptors
,
2354 struct r600_resource
*, slab
->buffer
);
2356 r600_resource_reference(&slab
->buffer
, NULL
);
2357 FREE(slab
->entries
);
2361 static struct si_bindless_descriptor
*
2362 si_create_bindless_descriptor(struct si_context
*sctx
, uint32_t *desc_list
,
2365 struct si_screen
*sscreen
= sctx
->screen
;
2366 struct si_bindless_descriptor
*desc
;
2367 struct pb_slab_entry
*entry
;
2370 /* Sub-allocate the bindless descriptor from a slab to avoid dealing
2371 * with a ton of buffers and for reducing the winsys overhead.
2373 entry
= pb_slab_alloc(&sctx
->bindless_descriptor_slabs
, 64, 0);
2378 desc
= container_of(entry
, desc
, entry
);
2380 /* Upload the descriptor directly in VRAM. Because the slabs are
2381 * currently never reclaimed, we don't need to synchronize the
2384 ptr
= sscreen
->b
.ws
->buffer_map(desc
->buffer
->buf
, NULL
,
2385 PIPE_TRANSFER_WRITE
|
2386 PIPE_TRANSFER_UNSYNCHRONIZED
);
2387 util_memcpy_cpu_to_le32(ptr
+ desc
->offset
, desc_list
, size
);
2389 /* Keep track of the initial descriptor especially for buffers
2390 * invalidation because we might need to know the previous address.
2392 memcpy(desc
->desc_list
, desc_list
, sizeof(desc
->desc_list
));
2397 static void si_invalidate_bindless_buf_desc(struct si_context
*sctx
,
2398 struct si_bindless_descriptor
*desc
,
2399 struct pipe_resource
*resource
,
2402 struct r600_resource
*buf
= r600_resource(resource
);
2403 uint32_t *desc_list
= desc
->desc_list
+ 4;
2404 uint64_t old_desc_va
;
2406 assert(resource
->target
== PIPE_BUFFER
);
2408 /* Retrieve the old buffer addr from the descriptor. */
2409 old_desc_va
= desc_list
[0];
2410 old_desc_va
|= ((uint64_t)G_008F04_BASE_ADDRESS_HI(desc_list
[1]) << 32);
2412 if (old_desc_va
!= buf
->gpu_address
+ offset
) {
2413 /* The buffer has been invalidated when the handle wasn't
2414 * resident, update the descriptor and the dirty flag.
2416 si_set_buf_desc_address(buf
, offset
, &desc_list
[0]);
2419 sctx
->bindless_descriptors_dirty
= true;
2423 static uint64_t si_create_texture_handle(struct pipe_context
*ctx
,
2424 struct pipe_sampler_view
*view
,
2425 const struct pipe_sampler_state
*state
)
2427 struct si_sampler_view
*sview
= (struct si_sampler_view
*)view
;
2428 struct si_context
*sctx
= (struct si_context
*)ctx
;
2429 struct si_texture_handle
*tex_handle
;
2430 struct si_sampler_state
*sstate
;
2431 uint32_t desc_list
[16];
2434 tex_handle
= CALLOC_STRUCT(si_texture_handle
);
2438 memset(desc_list
, 0, sizeof(desc_list
));
2439 si_init_descriptor_list(&desc_list
[0], 16, 1, null_texture_descriptor
);
2441 sstate
= ctx
->create_sampler_state(ctx
, state
);
2447 si_set_sampler_view_desc(sctx
, sview
, sstate
, &desc_list
[0]);
2448 memcpy(&tex_handle
->sstate
, sstate
, sizeof(*sstate
));
2449 ctx
->delete_sampler_state(ctx
, sstate
);
2451 tex_handle
->desc
= si_create_bindless_descriptor(sctx
, desc_list
,
2453 if (!tex_handle
->desc
) {
2458 handle
= tex_handle
->desc
->buffer
->gpu_address
+
2459 tex_handle
->desc
->offset
;
2461 if (!_mesa_hash_table_insert(sctx
->tex_handles
, (void *)handle
,
2463 pb_slab_free(&sctx
->bindless_descriptor_slabs
,
2464 &tex_handle
->desc
->entry
);
2469 pipe_sampler_view_reference(&tex_handle
->view
, view
);
2471 r600_resource(sview
->base
.texture
)->texture_handle_allocated
= true;
2476 static void si_delete_texture_handle(struct pipe_context
*ctx
, uint64_t handle
)
2478 struct si_context
*sctx
= (struct si_context
*)ctx
;
2479 struct si_texture_handle
*tex_handle
;
2480 struct hash_entry
*entry
;
2482 entry
= _mesa_hash_table_search(sctx
->tex_handles
, (void *)handle
);
2486 tex_handle
= (struct si_texture_handle
*)entry
->data
;
2488 pipe_sampler_view_reference(&tex_handle
->view
, NULL
);
2489 _mesa_hash_table_remove(sctx
->tex_handles
, entry
);
2490 pb_slab_free(&sctx
->bindless_descriptor_slabs
,
2491 &tex_handle
->desc
->entry
);
2495 static void si_make_texture_handle_resident(struct pipe_context
*ctx
,
2496 uint64_t handle
, bool resident
)
2498 struct si_context
*sctx
= (struct si_context
*)ctx
;
2499 struct si_texture_handle
*tex_handle
;
2500 struct si_sampler_view
*sview
;
2501 struct hash_entry
*entry
;
2503 entry
= _mesa_hash_table_search(sctx
->tex_handles
, (void *)handle
);
2507 tex_handle
= (struct si_texture_handle
*)entry
->data
;
2508 sview
= (struct si_sampler_view
*)tex_handle
->view
;
2511 if (sview
->base
.texture
->target
!= PIPE_BUFFER
) {
2512 struct r600_texture
*rtex
=
2513 (struct r600_texture
*)sview
->base
.texture
;
2515 if (depth_needs_decompression(rtex
)) {
2516 util_dynarray_append(
2517 &sctx
->resident_tex_needs_depth_decompress
,
2518 struct si_texture_handle
*,
2522 if (color_needs_decompression(rtex
)) {
2523 util_dynarray_append(
2524 &sctx
->resident_tex_needs_color_decompress
,
2525 struct si_texture_handle
*,
2529 if (rtex
->dcc_offset
&&
2530 p_atomic_read(&rtex
->framebuffers_bound
))
2531 sctx
->need_check_render_feedback
= true;
2533 /* Re-upload the descriptor if it has been updated
2534 * while it wasn't resident.
2536 si_update_resident_texture_descriptor(sctx
, tex_handle
);
2537 if (tex_handle
->desc
->dirty
)
2538 sctx
->bindless_descriptors_dirty
= true;
2540 si_invalidate_bindless_buf_desc(sctx
, tex_handle
->desc
,
2541 sview
->base
.texture
,
2542 sview
->base
.u
.buf
.offset
);
2545 /* Add the texture handle to the per-context list. */
2546 util_dynarray_append(&sctx
->resident_tex_handles
,
2547 struct si_texture_handle
*, tex_handle
);
2549 /* Add the buffers to the current CS in case si_begin_new_cs()
2550 * is not going to be called.
2552 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2553 tex_handle
->desc
->buffer
,
2554 RADEON_USAGE_READWRITE
,
2555 RADEON_PRIO_DESCRIPTORS
);
2557 si_sampler_view_add_buffer(sctx
, sview
->base
.texture
,
2559 sview
->is_stencil_sampler
, false);
2561 /* Remove the texture handle from the per-context list. */
2562 util_dynarray_delete_unordered(&sctx
->resident_tex_handles
,
2563 struct si_texture_handle
*,
2566 if (sview
->base
.texture
->target
!= PIPE_BUFFER
) {
2567 util_dynarray_delete_unordered(
2568 &sctx
->resident_tex_needs_depth_decompress
,
2569 struct si_texture_handle
*, tex_handle
);
2571 util_dynarray_delete_unordered(
2572 &sctx
->resident_tex_needs_color_decompress
,
2573 struct si_texture_handle
*, tex_handle
);
2578 static uint64_t si_create_image_handle(struct pipe_context
*ctx
,
2579 const struct pipe_image_view
*view
)
2581 struct si_context
*sctx
= (struct si_context
*)ctx
;
2582 struct si_image_handle
*img_handle
;
2583 uint32_t desc_list
[16];
2586 if (!view
|| !view
->resource
)
2589 img_handle
= CALLOC_STRUCT(si_image_handle
);
2593 memset(desc_list
, 0, sizeof(desc_list
));
2594 si_init_descriptor_list(&desc_list
[0], 8, 1, null_image_descriptor
);
2596 si_set_shader_image_desc(sctx
, view
, false, &desc_list
[0]);
2598 img_handle
->desc
= si_create_bindless_descriptor(sctx
, desc_list
,
2600 if (!img_handle
->desc
) {
2605 handle
= img_handle
->desc
->buffer
->gpu_address
+
2606 img_handle
->desc
->offset
;
2608 if (!_mesa_hash_table_insert(sctx
->img_handles
, (void *)handle
,
2610 pb_slab_free(&sctx
->bindless_descriptor_slabs
,
2611 &img_handle
->desc
->entry
);
2616 util_copy_image_view(&img_handle
->view
, view
);
2618 r600_resource(view
->resource
)->image_handle_allocated
= true;
2623 static void si_delete_image_handle(struct pipe_context
*ctx
, uint64_t handle
)
2625 struct si_context
*sctx
= (struct si_context
*)ctx
;
2626 struct si_image_handle
*img_handle
;
2627 struct hash_entry
*entry
;
2629 entry
= _mesa_hash_table_search(sctx
->img_handles
, (void *)handle
);
2633 img_handle
= (struct si_image_handle
*)entry
->data
;
2635 util_copy_image_view(&img_handle
->view
, NULL
);
2636 _mesa_hash_table_remove(sctx
->img_handles
, entry
);
2637 pb_slab_free(&sctx
->bindless_descriptor_slabs
,
2638 &img_handle
->desc
->entry
);
2642 static void si_make_image_handle_resident(struct pipe_context
*ctx
,
2643 uint64_t handle
, unsigned access
,
2646 struct si_context
*sctx
= (struct si_context
*)ctx
;
2647 struct si_image_handle
*img_handle
;
2648 struct pipe_image_view
*view
;
2649 struct r600_resource
*res
;
2650 struct hash_entry
*entry
;
2652 entry
= _mesa_hash_table_search(sctx
->img_handles
, (void *)handle
);
2656 img_handle
= (struct si_image_handle
*)entry
->data
;
2657 view
= &img_handle
->view
;
2658 res
= (struct r600_resource
*)view
->resource
;
2661 if (res
->b
.b
.target
!= PIPE_BUFFER
) {
2662 struct r600_texture
*rtex
= (struct r600_texture
*)res
;
2663 unsigned level
= view
->u
.tex
.level
;
2665 if (color_needs_decompression(rtex
)) {
2666 util_dynarray_append(
2667 &sctx
->resident_img_needs_color_decompress
,
2668 struct si_image_handle
*,
2672 if (vi_dcc_enabled(rtex
, level
) &&
2673 p_atomic_read(&rtex
->framebuffers_bound
))
2674 sctx
->need_check_render_feedback
= true;
2676 /* Re-upload the descriptor if it has been updated
2677 * while it wasn't resident.
2679 si_update_resident_image_descriptor(sctx
, img_handle
);
2680 if (img_handle
->desc
->dirty
)
2681 sctx
->bindless_descriptors_dirty
= true;
2684 si_invalidate_bindless_buf_desc(sctx
, img_handle
->desc
,
2686 view
->u
.buf
.offset
);
2689 /* Add the image handle to the per-context list. */
2690 util_dynarray_append(&sctx
->resident_img_handles
,
2691 struct si_image_handle
*, img_handle
);
2693 /* Add the buffers to the current CS in case si_begin_new_cs()
2694 * is not going to be called.
2696 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
,
2697 img_handle
->desc
->buffer
,
2698 RADEON_USAGE_READWRITE
,
2699 RADEON_PRIO_DESCRIPTORS
);
2701 si_sampler_view_add_buffer(sctx
, view
->resource
,
2702 (access
& PIPE_IMAGE_ACCESS_WRITE
) ?
2703 RADEON_USAGE_READWRITE
:
2704 RADEON_USAGE_READ
, false, false);
2706 /* Remove the image handle from the per-context list. */
2707 util_dynarray_delete_unordered(&sctx
->resident_img_handles
,
2708 struct si_image_handle
*,
2711 if (res
->b
.b
.target
!= PIPE_BUFFER
) {
2712 util_dynarray_delete_unordered(
2713 &sctx
->resident_img_needs_color_decompress
,
2714 struct si_image_handle
*,
2721 void si_all_resident_buffers_begin_new_cs(struct si_context
*sctx
)
2723 unsigned num_resident_tex_handles
, num_resident_img_handles
;
2725 num_resident_tex_handles
= sctx
->resident_tex_handles
.size
/
2726 sizeof(struct si_texture_handle
*);
2727 num_resident_img_handles
= sctx
->resident_img_handles
.size
/
2728 sizeof(struct si_image_handle
*);
2730 /* Skip adding the bindless descriptors when no handles are resident.
2732 if (!num_resident_tex_handles
&& !num_resident_img_handles
)
2735 /* Add all bindless descriptors. */
2736 util_dynarray_foreach(&sctx
->bindless_descriptors
,
2737 struct r600_resource
*, desc
) {
2739 radeon_add_to_buffer_list(&sctx
->b
, &sctx
->b
.gfx
, *desc
,
2740 RADEON_USAGE_READWRITE
,
2741 RADEON_PRIO_DESCRIPTORS
);
2744 /* Add all resident texture handles. */
2745 util_dynarray_foreach(&sctx
->resident_tex_handles
,
2746 struct si_texture_handle
*, tex_handle
) {
2747 struct si_sampler_view
*sview
=
2748 (struct si_sampler_view
*)(*tex_handle
)->view
;
2750 si_sampler_view_add_buffer(sctx
, sview
->base
.texture
,
2752 sview
->is_stencil_sampler
, false);
2755 /* Add all resident image handles. */
2756 util_dynarray_foreach(&sctx
->resident_img_handles
,
2757 struct si_image_handle
*, img_handle
) {
2758 struct pipe_image_view
*view
= &(*img_handle
)->view
;
2760 si_sampler_view_add_buffer(sctx
, view
->resource
,
2761 RADEON_USAGE_READWRITE
,
2765 sctx
->b
.num_resident_handles
+= num_resident_tex_handles
+
2766 num_resident_img_handles
;
2769 /* INIT/DEINIT/UPLOAD */
2771 /* GFX9 has only 4KB of CE, while previous chips had 32KB. In order
2772 * to make CE RAM as useful as possible, this defines limits
2773 * for the number slots that can be in CE RAM on GFX9. If a shader
2774 * is using more, descriptors will be uploaded to memory directly and
2777 * These numbers are based on shader-db.
2779 static unsigned gfx9_max_ce_samplers
[SI_NUM_SHADERS
] = {
2780 [PIPE_SHADER_VERTEX
] = 0,
2781 [PIPE_SHADER_TESS_CTRL
] = 0,
2782 [PIPE_SHADER_TESS_EVAL
] = 1,
2783 [PIPE_SHADER_GEOMETRY
] = 0,
2784 [PIPE_SHADER_FRAGMENT
] = 24,
2785 [PIPE_SHADER_COMPUTE
] = 16,
2787 static unsigned gfx9_max_ce_images
[SI_NUM_SHADERS
] = {
2788 /* these must be even due to slot alignment */
2789 [PIPE_SHADER_VERTEX
] = 0,
2790 [PIPE_SHADER_TESS_CTRL
] = 0,
2791 [PIPE_SHADER_TESS_EVAL
] = 0,
2792 [PIPE_SHADER_GEOMETRY
] = 0,
2793 [PIPE_SHADER_FRAGMENT
] = 2,
2794 [PIPE_SHADER_COMPUTE
] = 8,
2796 static unsigned gfx9_max_ce_const_buffers
[SI_NUM_SHADERS
] = {
2797 [PIPE_SHADER_VERTEX
] = 9,
2798 [PIPE_SHADER_TESS_CTRL
] = 3,
2799 [PIPE_SHADER_TESS_EVAL
] = 5,
2800 [PIPE_SHADER_GEOMETRY
] = 0,
2801 [PIPE_SHADER_FRAGMENT
] = 8,
2802 [PIPE_SHADER_COMPUTE
] = 6,
2804 static unsigned gfx9_max_ce_shader_buffers
[SI_NUM_SHADERS
] = {
2805 [PIPE_SHADER_VERTEX
] = 0,
2806 [PIPE_SHADER_TESS_CTRL
] = 0,
2807 [PIPE_SHADER_TESS_EVAL
] = 0,
2808 [PIPE_SHADER_GEOMETRY
] = 0,
2809 [PIPE_SHADER_FRAGMENT
] = 12,
2810 [PIPE_SHADER_COMPUTE
] = 13,
2813 void si_init_all_descriptors(struct si_context
*sctx
)
2816 unsigned ce_offset
= 0;
2818 STATIC_ASSERT(GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS
% 2 == 0);
2819 STATIC_ASSERT(GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS
% 2 == 0);
2821 for (i
= 0; i
< SI_NUM_SHADERS
; i
++) {
2822 bool gfx9_tcs
= false;
2823 bool gfx9_gs
= false;
2824 unsigned num_sampler_slots
= SI_NUM_IMAGES
/ 2 + SI_NUM_SAMPLERS
;
2825 unsigned num_buffer_slots
= SI_NUM_SHADER_BUFFERS
+ SI_NUM_CONST_BUFFERS
;
2827 unsigned first_sampler_ce_slot
= 0;
2828 unsigned num_sampler_ce_slots
= num_sampler_slots
;
2830 unsigned first_buffer_ce_slot
= 0;
2831 unsigned num_buffer_ce_slots
= num_buffer_slots
;
2833 /* Adjust CE slot ranges based on GFX9 CE RAM limits. */
2834 if (sctx
->b
.chip_class
>= GFX9
) {
2835 gfx9_tcs
= i
== PIPE_SHADER_TESS_CTRL
;
2836 gfx9_gs
= i
== PIPE_SHADER_GEOMETRY
;
2838 first_sampler_ce_slot
=
2839 si_get_image_slot(gfx9_max_ce_images
[i
] - 1) / 2;
2840 num_sampler_ce_slots
= gfx9_max_ce_images
[i
] / 2 +
2841 gfx9_max_ce_samplers
[i
];
2843 first_buffer_ce_slot
=
2844 si_get_shaderbuf_slot(gfx9_max_ce_shader_buffers
[i
] - 1);
2845 num_buffer_ce_slots
= gfx9_max_ce_shader_buffers
[i
] +
2846 gfx9_max_ce_const_buffers
[i
];
2849 si_init_buffer_resources(sctx
, &sctx
->const_and_shader_buffers
[i
],
2850 si_const_and_shader_buffer_descriptors(sctx
, i
),
2852 first_buffer_ce_slot
, num_buffer_ce_slots
,
2853 gfx9_tcs
? GFX9_SGPR_TCS_CONST_AND_SHADER_BUFFERS
:
2854 gfx9_gs
? GFX9_SGPR_GS_CONST_AND_SHADER_BUFFERS
:
2855 SI_SGPR_CONST_AND_SHADER_BUFFERS
,
2856 RADEON_USAGE_READWRITE
,
2858 RADEON_PRIO_SHADER_RW_BUFFER
,
2859 RADEON_PRIO_CONST_BUFFER
,
2862 struct si_descriptors
*desc
= si_sampler_and_image_descriptors(sctx
, i
);
2863 si_init_descriptors(sctx
, desc
,
2864 gfx9_tcs
? GFX9_SGPR_TCS_SAMPLERS_AND_IMAGES
:
2865 gfx9_gs
? GFX9_SGPR_GS_SAMPLERS_AND_IMAGES
:
2866 SI_SGPR_SAMPLERS_AND_IMAGES
,
2867 16, num_sampler_slots
,
2868 first_sampler_ce_slot
, num_sampler_ce_slots
,
2872 for (j
= 0; j
< SI_NUM_IMAGES
; j
++)
2873 memcpy(desc
->list
+ j
* 8, null_image_descriptor
, 8 * 4);
2874 for (; j
< SI_NUM_IMAGES
+ SI_NUM_SAMPLERS
* 2; j
++)
2875 memcpy(desc
->list
+ j
* 8, null_texture_descriptor
, 8 * 4);
2878 si_init_buffer_resources(sctx
, &sctx
->rw_buffers
,
2879 &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
],
2880 SI_NUM_RW_BUFFERS
, 0, SI_NUM_RW_BUFFERS
,
2882 /* The second set of usage/priority is used by
2883 * const buffers in RW buffer slots. */
2884 RADEON_USAGE_READWRITE
, RADEON_USAGE_READ
,
2885 RADEON_PRIO_SHADER_RINGS
, RADEON_PRIO_CONST_BUFFER
,
2887 sctx
->descriptors
[SI_DESCS_RW_BUFFERS
].num_active_slots
= SI_NUM_RW_BUFFERS
;
2889 si_init_descriptors(sctx
, &sctx
->vertex_buffers
, SI_SGPR_VERTEX_BUFFERS
,
2890 4, SI_NUM_VERTEX_BUFFERS
, 0, 0, NULL
);
2891 FREE(sctx
->vertex_buffers
.list
); /* not used */
2892 sctx
->vertex_buffers
.list
= NULL
;
2894 sctx
->descriptors_dirty
= u_bit_consecutive(0, SI_NUM_DESCS
);
2895 sctx
->total_ce_ram_allocated
= ce_offset
;
2897 assert(ce_offset
<= si_ce_ram_size(sctx
));
2899 /* Set pipe_context functions. */
2900 sctx
->b
.b
.bind_sampler_states
= si_bind_sampler_states
;
2901 sctx
->b
.b
.set_shader_images
= si_set_shader_images
;
2902 sctx
->b
.b
.set_constant_buffer
= si_pipe_set_constant_buffer
;
2903 sctx
->b
.b
.set_polygon_stipple
= si_set_polygon_stipple
;
2904 sctx
->b
.b
.set_shader_buffers
= si_set_shader_buffers
;
2905 sctx
->b
.b
.set_sampler_views
= si_set_sampler_views
;
2906 sctx
->b
.b
.set_stream_output_targets
= si_set_streamout_targets
;
2907 sctx
->b
.b
.create_texture_handle
= si_create_texture_handle
;
2908 sctx
->b
.b
.delete_texture_handle
= si_delete_texture_handle
;
2909 sctx
->b
.b
.make_texture_handle_resident
= si_make_texture_handle_resident
;
2910 sctx
->b
.b
.create_image_handle
= si_create_image_handle
;
2911 sctx
->b
.b
.delete_image_handle
= si_delete_image_handle
;
2912 sctx
->b
.b
.make_image_handle_resident
= si_make_image_handle_resident
;
2913 sctx
->b
.invalidate_buffer
= si_invalidate_buffer
;
2914 sctx
->b
.rebind_buffer
= si_rebind_buffer
;
2916 /* Shader user data. */
2917 si_init_atom(sctx
, &sctx
->shader_pointers
.atom
, &sctx
->atoms
.s
.shader_pointers
,
2918 si_emit_graphics_shader_pointers
);
2920 /* Set default and immutable mappings. */
2921 si_set_user_data_base(sctx
, PIPE_SHADER_VERTEX
, R_00B130_SPI_SHADER_USER_DATA_VS_0
);
2923 if (sctx
->b
.chip_class
>= GFX9
) {
2924 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_CTRL
,
2925 R_00B430_SPI_SHADER_USER_DATA_LS_0
);
2926 si_set_user_data_base(sctx
, PIPE_SHADER_GEOMETRY
,
2927 R_00B330_SPI_SHADER_USER_DATA_ES_0
);
2929 si_set_user_data_base(sctx
, PIPE_SHADER_TESS_CTRL
,
2930 R_00B430_SPI_SHADER_USER_DATA_HS_0
);
2931 si_set_user_data_base(sctx
, PIPE_SHADER_GEOMETRY
,
2932 R_00B230_SPI_SHADER_USER_DATA_GS_0
);
2934 si_set_user_data_base(sctx
, PIPE_SHADER_FRAGMENT
, R_00B030_SPI_SHADER_USER_DATA_PS_0
);
2937 static bool si_upload_shader_descriptors(struct si_context
*sctx
, unsigned mask
)
2939 unsigned dirty
= sctx
->descriptors_dirty
& mask
;
2941 /* Assume nothing will go wrong: */
2942 sctx
->shader_pointers_dirty
|= dirty
;
2945 unsigned i
= u_bit_scan(&dirty
);
2947 if (!si_upload_descriptors(sctx
, &sctx
->descriptors
[i
],
2948 &sctx
->shader_pointers
.atom
))
2952 sctx
->descriptors_dirty
&= ~mask
;
2954 si_upload_bindless_descriptors(sctx
);
2959 bool si_upload_graphics_shader_descriptors(struct si_context
*sctx
)
2961 const unsigned mask
= u_bit_consecutive(0, SI_DESCS_FIRST_COMPUTE
);
2962 return si_upload_shader_descriptors(sctx
, mask
);
2965 bool si_upload_compute_shader_descriptors(struct si_context
*sctx
)
2967 /* Does not update rw_buffers as that is not needed for compute shaders
2968 * and the input buffer is using the same SGPR's anyway.
2970 const unsigned mask
= u_bit_consecutive(SI_DESCS_FIRST_COMPUTE
,
2971 SI_NUM_DESCS
- SI_DESCS_FIRST_COMPUTE
);
2972 return si_upload_shader_descriptors(sctx
, mask
);
2975 void si_release_all_descriptors(struct si_context
*sctx
)
2979 for (i
= 0; i
< SI_NUM_SHADERS
; i
++) {
2980 si_release_buffer_resources(&sctx
->const_and_shader_buffers
[i
],
2981 si_const_and_shader_buffer_descriptors(sctx
, i
));
2982 si_release_sampler_views(&sctx
->samplers
[i
].views
);
2983 si_release_image_views(&sctx
->images
[i
]);
2985 si_release_buffer_resources(&sctx
->rw_buffers
,
2986 &sctx
->descriptors
[SI_DESCS_RW_BUFFERS
]);
2987 for (i
= 0; i
< SI_NUM_VERTEX_BUFFERS
; i
++)
2988 pipe_vertex_buffer_unreference(&sctx
->vertex_buffer
[i
]);
2990 for (i
= 0; i
< SI_NUM_DESCS
; ++i
)
2991 si_release_descriptors(&sctx
->descriptors
[i
]);
2993 sctx
->vertex_buffers
.list
= NULL
; /* points into a mapped buffer */
2994 si_release_descriptors(&sctx
->vertex_buffers
);
2997 void si_all_descriptors_begin_new_cs(struct si_context
*sctx
)
3001 for (i
= 0; i
< SI_NUM_SHADERS
; i
++) {
3002 si_buffer_resources_begin_new_cs(sctx
, &sctx
->const_and_shader_buffers
[i
]);
3003 si_sampler_views_begin_new_cs(sctx
, &sctx
->samplers
[i
].views
);
3004 si_image_views_begin_new_cs(sctx
, &sctx
->images
[i
]);
3006 si_buffer_resources_begin_new_cs(sctx
, &sctx
->rw_buffers
);
3007 si_vertex_buffers_begin_new_cs(sctx
);
3009 for (i
= 0; i
< SI_NUM_DESCS
; ++i
)
3010 si_descriptors_begin_new_cs(sctx
, &sctx
->descriptors
[i
]);
3012 si_shader_pointers_begin_new_cs(sctx
);
3015 void si_set_active_descriptors(struct si_context
*sctx
, unsigned desc_idx
,
3016 uint64_t new_active_mask
)
3018 struct si_descriptors
*desc
= &sctx
->descriptors
[desc_idx
];
3020 /* Ignore no-op updates and updates that disable all slots. */
3021 if (!new_active_mask
||
3022 new_active_mask
== u_bit_consecutive64(desc
->first_active_slot
,
3023 desc
->num_active_slots
))
3027 u_bit_scan_consecutive_range64(&new_active_mask
, &first
, &count
);
3028 assert(new_active_mask
== 0);
3030 /* Upload/dump descriptors if slots are being enabled. */
3031 if (first
< desc
->first_active_slot
||
3032 first
+ count
> desc
->first_active_slot
+ desc
->num_active_slots
)
3033 sctx
->descriptors_dirty
|= 1u << desc_idx
;
3035 /* Enable or disable CE for this descriptor array. */
3036 bool used_ce
= desc
->uses_ce
;
3037 desc
->uses_ce
= desc
->first_ce_slot
<= first
&&
3038 desc
->first_ce_slot
+ desc
->num_ce_slots
>= first
+ count
;
3040 if (desc
->uses_ce
!= used_ce
) {
3041 /* Upload or dump descriptors if we're disabling or enabling CE,
3043 sctx
->descriptors_dirty
|= 1u << desc_idx
;
3045 /* If we're enabling CE, re-upload all descriptors to CE RAM.
3046 * When CE was disabled, uploads to CE RAM stopped.
3048 if (desc
->uses_ce
) {
3050 u_bit_consecutive64(desc
->first_ce_slot
,
3051 desc
->num_ce_slots
);
3055 desc
->first_active_slot
= first
;
3056 desc
->num_active_slots
= count
;
3059 void si_set_active_descriptors_for_shader(struct si_context
*sctx
,
3060 struct si_shader_selector
*sel
)
3065 si_set_active_descriptors(sctx
,
3066 si_const_and_shader_buffer_descriptors_idx(sel
->type
),
3067 sel
->active_const_and_shader_buffers
);
3068 si_set_active_descriptors(sctx
,
3069 si_sampler_and_image_descriptors_idx(sel
->type
),
3070 sel
->active_samplers_and_images
);