2 * Copyright © 2015 Intel Corporation
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 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * 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 NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
27 #include "anv_private.h"
28 #include "vk_format_info.h"
30 #include "util/fast_idiv_by_const.h"
32 #include "common/gen_l3_config.h"
33 #include "genxml/gen_macros.h"
34 #include "genxml/genX_pack.h"
36 /* We reserve GPR 14 and 15 for conditional rendering */
37 #define GEN_MI_BUILDER_NUM_ALLOC_GPRS 14
38 #define __gen_get_batch_dwords anv_batch_emit_dwords
39 #define __gen_address_offset anv_address_add
40 #include "common/gen_mi_builder.h"
43 emit_lri(struct anv_batch
*batch
, uint32_t reg
, uint32_t imm
)
45 anv_batch_emit(batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
46 lri
.RegisterOffset
= reg
;
52 genX(cmd_buffer_emit_state_base_address
)(struct anv_cmd_buffer
*cmd_buffer
)
54 struct anv_device
*device
= cmd_buffer
->device
;
56 /* If we are emitting a new state base address we probably need to re-emit
59 cmd_buffer
->state
.descriptors_dirty
|= ~0;
61 /* Emit a render target cache flush.
63 * This isn't documented anywhere in the PRM. However, it seems to be
64 * necessary prior to changing the surface state base adress. Without
65 * this, we get GPU hangs when using multi-level command buffers which
66 * clear depth, reset state base address, and then go render stuff.
68 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
69 pc
.DCFlushEnable
= true;
70 pc
.RenderTargetCacheFlushEnable
= true;
71 pc
.CommandStreamerStallEnable
= true;
74 anv_batch_emit(&cmd_buffer
->batch
, GENX(STATE_BASE_ADDRESS
), sba
) {
75 sba
.GeneralStateBaseAddress
= (struct anv_address
) { NULL
, 0 };
76 sba
.GeneralStateMOCS
= GENX(MOCS
);
77 sba
.GeneralStateBaseAddressModifyEnable
= true;
79 sba
.StatelessDataPortAccessMOCS
= GENX(MOCS
);
81 sba
.SurfaceStateBaseAddress
=
82 anv_cmd_buffer_surface_base_address(cmd_buffer
);
83 sba
.SurfaceStateMOCS
= GENX(MOCS
);
84 sba
.SurfaceStateBaseAddressModifyEnable
= true;
86 sba
.DynamicStateBaseAddress
=
87 (struct anv_address
) { device
->dynamic_state_pool
.block_pool
.bo
, 0 };
88 sba
.DynamicStateMOCS
= GENX(MOCS
);
89 sba
.DynamicStateBaseAddressModifyEnable
= true;
91 sba
.IndirectObjectBaseAddress
= (struct anv_address
) { NULL
, 0 };
92 sba
.IndirectObjectMOCS
= GENX(MOCS
);
93 sba
.IndirectObjectBaseAddressModifyEnable
= true;
95 sba
.InstructionBaseAddress
=
96 (struct anv_address
) { device
->instruction_state_pool
.block_pool
.bo
, 0 };
97 sba
.InstructionMOCS
= GENX(MOCS
);
98 sba
.InstructionBaseAddressModifyEnable
= true;
101 /* Broadwell requires that we specify a buffer size for a bunch of
102 * these fields. However, since we will be growing the BO's live, we
103 * just set them all to the maximum.
105 sba
.GeneralStateBufferSize
= 0xfffff;
106 sba
.GeneralStateBufferSizeModifyEnable
= true;
107 sba
.DynamicStateBufferSize
= 0xfffff;
108 sba
.DynamicStateBufferSizeModifyEnable
= true;
109 sba
.IndirectObjectBufferSize
= 0xfffff;
110 sba
.IndirectObjectBufferSizeModifyEnable
= true;
111 sba
.InstructionBufferSize
= 0xfffff;
112 sba
.InstructionBuffersizeModifyEnable
= true;
114 /* On gen7, we have upper bounds instead. According to the docs,
115 * setting an upper bound of zero means that no bounds checking is
116 * performed so, in theory, we should be able to leave them zero.
117 * However, border color is broken and the GPU bounds-checks anyway.
118 * To avoid this and other potential problems, we may as well set it
121 sba
.GeneralStateAccessUpperBound
=
122 (struct anv_address
) { .bo
= NULL
, .offset
= 0xfffff000 };
123 sba
.GeneralStateAccessUpperBoundModifyEnable
= true;
124 sba
.DynamicStateAccessUpperBound
=
125 (struct anv_address
) { .bo
= NULL
, .offset
= 0xfffff000 };
126 sba
.DynamicStateAccessUpperBoundModifyEnable
= true;
127 sba
.InstructionAccessUpperBound
=
128 (struct anv_address
) { .bo
= NULL
, .offset
= 0xfffff000 };
129 sba
.InstructionAccessUpperBoundModifyEnable
= true;
132 if (cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
) {
133 sba
.BindlessSurfaceStateBaseAddress
= (struct anv_address
) {
134 .bo
= device
->surface_state_pool
.block_pool
.bo
,
137 sba
.BindlessSurfaceStateSize
= (1 << 20) - 1;
139 sba
.BindlessSurfaceStateBaseAddress
= ANV_NULL_ADDRESS
;
140 sba
.BindlessSurfaceStateSize
= 0;
142 sba
.BindlessSurfaceStateMOCS
= GENX(MOCS
);
143 sba
.BindlessSurfaceStateBaseAddressModifyEnable
= true;
146 sba
.BindlessSamplerStateBaseAddress
= (struct anv_address
) { NULL
, 0 };
147 sba
.BindlessSamplerStateMOCS
= GENX(MOCS
);
148 sba
.BindlessSamplerStateBaseAddressModifyEnable
= true;
149 sba
.BindlessSamplerStateBufferSize
= 0;
153 /* After re-setting the surface state base address, we have to do some
154 * cache flusing so that the sampler engine will pick up the new
155 * SURFACE_STATE objects and binding tables. From the Broadwell PRM,
156 * Shared Function > 3D Sampler > State > State Caching (page 96):
158 * Coherency with system memory in the state cache, like the texture
159 * cache is handled partially by software. It is expected that the
160 * command stream or shader will issue Cache Flush operation or
161 * Cache_Flush sampler message to ensure that the L1 cache remains
162 * coherent with system memory.
166 * Whenever the value of the Dynamic_State_Base_Addr,
167 * Surface_State_Base_Addr are altered, the L1 state cache must be
168 * invalidated to ensure the new surface or sampler state is fetched
169 * from system memory.
171 * The PIPE_CONTROL command has a "State Cache Invalidation Enable" bit
172 * which, according the PIPE_CONTROL instruction documentation in the
175 * Setting this bit is independent of any other bit in this packet.
176 * This bit controls the invalidation of the L1 and L2 state caches
177 * at the top of the pipe i.e. at the parsing time.
179 * Unfortunately, experimentation seems to indicate that state cache
180 * invalidation through a PIPE_CONTROL does nothing whatsoever in
181 * regards to surface state and binding tables. In stead, it seems that
182 * invalidating the texture cache is what is actually needed.
184 * XXX: As far as we have been able to determine through
185 * experimentation, shows that flush the texture cache appears to be
186 * sufficient. The theory here is that all of the sampling/rendering
187 * units cache the binding table in the texture cache. However, we have
188 * yet to be able to actually confirm this.
190 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
191 pc
.TextureCacheInvalidationEnable
= true;
192 pc
.ConstantCacheInvalidationEnable
= true;
193 pc
.StateCacheInvalidationEnable
= true;
198 add_surface_reloc(struct anv_cmd_buffer
*cmd_buffer
,
199 struct anv_state state
, struct anv_address addr
)
201 const struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
204 anv_reloc_list_add(&cmd_buffer
->surface_relocs
, &cmd_buffer
->pool
->alloc
,
205 state
.offset
+ isl_dev
->ss
.addr_offset
,
206 addr
.bo
, addr
.offset
);
207 if (result
!= VK_SUCCESS
)
208 anv_batch_set_error(&cmd_buffer
->batch
, result
);
212 add_surface_state_relocs(struct anv_cmd_buffer
*cmd_buffer
,
213 struct anv_surface_state state
)
215 const struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
217 assert(!anv_address_is_null(state
.address
));
218 add_surface_reloc(cmd_buffer
, state
.state
, state
.address
);
220 if (!anv_address_is_null(state
.aux_address
)) {
222 anv_reloc_list_add(&cmd_buffer
->surface_relocs
,
223 &cmd_buffer
->pool
->alloc
,
224 state
.state
.offset
+ isl_dev
->ss
.aux_addr_offset
,
225 state
.aux_address
.bo
, state
.aux_address
.offset
);
226 if (result
!= VK_SUCCESS
)
227 anv_batch_set_error(&cmd_buffer
->batch
, result
);
230 if (!anv_address_is_null(state
.clear_address
)) {
232 anv_reloc_list_add(&cmd_buffer
->surface_relocs
,
233 &cmd_buffer
->pool
->alloc
,
235 isl_dev
->ss
.clear_color_state_offset
,
236 state
.clear_address
.bo
, state
.clear_address
.offset
);
237 if (result
!= VK_SUCCESS
)
238 anv_batch_set_error(&cmd_buffer
->batch
, result
);
243 color_attachment_compute_aux_usage(struct anv_device
* device
,
244 struct anv_cmd_state
* cmd_state
,
245 uint32_t att
, VkRect2D render_area
,
246 union isl_color_value
*fast_clear_color
)
248 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[att
];
249 struct anv_image_view
*iview
= cmd_state
->attachments
[att
].image_view
;
251 assert(iview
->n_planes
== 1);
253 if (iview
->planes
[0].isl
.base_array_layer
>=
254 anv_image_aux_layers(iview
->image
, VK_IMAGE_ASPECT_COLOR_BIT
,
255 iview
->planes
[0].isl
.base_level
)) {
256 /* There is no aux buffer which corresponds to the level and layer(s)
259 att_state
->aux_usage
= ISL_AUX_USAGE_NONE
;
260 att_state
->input_aux_usage
= ISL_AUX_USAGE_NONE
;
261 att_state
->fast_clear
= false;
265 att_state
->aux_usage
=
266 anv_layout_to_aux_usage(&device
->info
, iview
->image
,
267 VK_IMAGE_ASPECT_COLOR_BIT
,
268 VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
);
270 /* If we don't have aux, then we should have returned early in the layer
271 * check above. If we got here, we must have something.
273 assert(att_state
->aux_usage
!= ISL_AUX_USAGE_NONE
);
275 if (att_state
->aux_usage
== ISL_AUX_USAGE_CCS_E
||
276 att_state
->aux_usage
== ISL_AUX_USAGE_MCS
) {
277 att_state
->input_aux_usage
= att_state
->aux_usage
;
279 /* From the Sky Lake PRM, RENDER_SURFACE_STATE::AuxiliarySurfaceMode:
281 * "If Number of Multisamples is MULTISAMPLECOUNT_1, AUX_CCS_D
282 * setting is only allowed if Surface Format supported for Fast
283 * Clear. In addition, if the surface is bound to the sampling
284 * engine, Surface Format must be supported for Render Target
285 * Compression for surfaces bound to the sampling engine."
287 * In other words, we can only sample from a fast-cleared image if it
288 * also supports color compression.
290 if (isl_format_supports_ccs_e(&device
->info
, iview
->planes
[0].isl
.format
)) {
291 att_state
->input_aux_usage
= ISL_AUX_USAGE_CCS_D
;
293 /* While fast-clear resolves and partial resolves are fairly cheap in the
294 * case where you render to most of the pixels, full resolves are not
295 * because they potentially involve reading and writing the entire
296 * framebuffer. If we can't texture with CCS_E, we should leave it off and
297 * limit ourselves to fast clears.
299 if (cmd_state
->pass
->attachments
[att
].first_subpass_layout
==
300 VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
) {
301 anv_perf_warn(device
->instance
, iview
->image
,
302 "Not temporarily enabling CCS_E.");
305 att_state
->input_aux_usage
= ISL_AUX_USAGE_NONE
;
309 assert(iview
->image
->planes
[0].aux_surface
.isl
.usage
&
310 (ISL_SURF_USAGE_CCS_BIT
| ISL_SURF_USAGE_MCS_BIT
));
312 union isl_color_value clear_color
= {};
313 anv_clear_color_from_att_state(&clear_color
, att_state
, iview
);
315 att_state
->clear_color_is_zero_one
=
316 isl_color_value_is_zero_one(clear_color
, iview
->planes
[0].isl
.format
);
317 att_state
->clear_color_is_zero
=
318 isl_color_value_is_zero(clear_color
, iview
->planes
[0].isl
.format
);
320 if (att_state
->pending_clear_aspects
== VK_IMAGE_ASPECT_COLOR_BIT
) {
321 /* Start by getting the fast clear type. We use the first subpass
322 * layout here because we don't want to fast-clear if the first subpass
323 * to use the attachment can't handle fast-clears.
325 enum anv_fast_clear_type fast_clear_type
=
326 anv_layout_to_fast_clear_type(&device
->info
, iview
->image
,
327 VK_IMAGE_ASPECT_COLOR_BIT
,
328 cmd_state
->pass
->attachments
[att
].first_subpass_layout
);
329 switch (fast_clear_type
) {
330 case ANV_FAST_CLEAR_NONE
:
331 att_state
->fast_clear
= false;
333 case ANV_FAST_CLEAR_DEFAULT_VALUE
:
334 att_state
->fast_clear
= att_state
->clear_color_is_zero
;
336 case ANV_FAST_CLEAR_ANY
:
337 att_state
->fast_clear
= true;
341 /* Potentially, we could do partial fast-clears but doing so has crazy
342 * alignment restrictions. It's easier to just restrict to full size
343 * fast clears for now.
345 if (render_area
.offset
.x
!= 0 ||
346 render_area
.offset
.y
!= 0 ||
347 render_area
.extent
.width
!= iview
->extent
.width
||
348 render_area
.extent
.height
!= iview
->extent
.height
)
349 att_state
->fast_clear
= false;
351 /* On Broadwell and earlier, we can only handle 0/1 clear colors */
352 if (GEN_GEN
<= 8 && !att_state
->clear_color_is_zero_one
)
353 att_state
->fast_clear
= false;
355 /* We only allow fast clears to the first slice of an image (level 0,
356 * layer 0) and only for the entire slice. This guarantees us that, at
357 * any given time, there is only one clear color on any given image at
358 * any given time. At the time of our testing (Jan 17, 2018), there
359 * were no known applications which would benefit from fast-clearing
360 * more than just the first slice.
362 if (att_state
->fast_clear
&&
363 (iview
->planes
[0].isl
.base_level
> 0 ||
364 iview
->planes
[0].isl
.base_array_layer
> 0)) {
365 anv_perf_warn(device
->instance
, iview
->image
,
366 "Rendering with multi-lod or multi-layer framebuffer "
367 "with LOAD_OP_LOAD and baseMipLevel > 0 or "
368 "baseArrayLayer > 0. Not fast clearing.");
369 att_state
->fast_clear
= false;
370 } else if (att_state
->fast_clear
&& cmd_state
->framebuffer
->layers
> 1) {
371 anv_perf_warn(device
->instance
, iview
->image
,
372 "Rendering to a multi-layer framebuffer with "
373 "LOAD_OP_CLEAR. Only fast-clearing the first slice");
376 if (att_state
->fast_clear
)
377 *fast_clear_color
= clear_color
;
379 att_state
->fast_clear
= false;
384 depth_stencil_attachment_compute_aux_usage(struct anv_device
*device
,
385 struct anv_cmd_state
*cmd_state
,
386 uint32_t att
, VkRect2D render_area
)
388 struct anv_render_pass_attachment
*pass_att
=
389 &cmd_state
->pass
->attachments
[att
];
390 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[att
];
391 struct anv_image_view
*iview
= cmd_state
->attachments
[att
].image_view
;
393 /* These will be initialized after the first subpass transition. */
394 att_state
->aux_usage
= ISL_AUX_USAGE_NONE
;
395 att_state
->input_aux_usage
= ISL_AUX_USAGE_NONE
;
398 /* We don't do any HiZ or depth fast-clears on gen7 yet */
399 att_state
->fast_clear
= false;
403 if (!(att_state
->pending_clear_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)) {
404 /* If we're just clearing stencil, we can always HiZ clear */
405 att_state
->fast_clear
= true;
409 /* Default to false for now */
410 att_state
->fast_clear
= false;
412 /* We must have depth in order to have HiZ */
413 if (!(iview
->image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
))
416 const enum isl_aux_usage first_subpass_aux_usage
=
417 anv_layout_to_aux_usage(&device
->info
, iview
->image
,
418 VK_IMAGE_ASPECT_DEPTH_BIT
,
419 pass_att
->first_subpass_layout
);
420 if (first_subpass_aux_usage
!= ISL_AUX_USAGE_HIZ
)
423 if (!blorp_can_hiz_clear_depth(GEN_GEN
,
424 iview
->planes
[0].isl
.format
,
425 iview
->image
->samples
,
426 render_area
.offset
.x
,
427 render_area
.offset
.y
,
428 render_area
.offset
.x
+
429 render_area
.extent
.width
,
430 render_area
.offset
.y
+
431 render_area
.extent
.height
))
434 if (att_state
->clear_value
.depthStencil
.depth
!= ANV_HZ_FC_VAL
)
437 if (GEN_GEN
== 8 && anv_can_sample_with_hiz(&device
->info
, iview
->image
)) {
438 /* Only gen9+ supports returning ANV_HZ_FC_VAL when sampling a
439 * fast-cleared portion of a HiZ buffer. Testing has revealed that Gen8
440 * only supports returning 0.0f. Gens prior to gen8 do not support this
446 /* If we got here, then we can fast clear */
447 att_state
->fast_clear
= true;
451 need_input_attachment_state(const struct anv_render_pass_attachment
*att
)
453 if (!(att
->usage
& VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT
))
456 /* We only allocate input attachment states for color surfaces. Compression
457 * is not yet enabled for depth textures and stencil doesn't allow
458 * compression so we can just use the texture surface state from the view.
460 return vk_format_is_color(att
->format
);
463 /* Transitions a HiZ-enabled depth buffer from one layout to another. Unless
464 * the initial layout is undefined, the HiZ buffer and depth buffer will
465 * represent the same data at the end of this operation.
468 transition_depth_buffer(struct anv_cmd_buffer
*cmd_buffer
,
469 const struct anv_image
*image
,
470 VkImageLayout initial_layout
,
471 VkImageLayout final_layout
)
473 const bool hiz_enabled
= ISL_AUX_USAGE_HIZ
==
474 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, image
,
475 VK_IMAGE_ASPECT_DEPTH_BIT
, initial_layout
);
476 const bool enable_hiz
= ISL_AUX_USAGE_HIZ
==
477 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, image
,
478 VK_IMAGE_ASPECT_DEPTH_BIT
, final_layout
);
480 enum isl_aux_op hiz_op
;
481 if (hiz_enabled
&& !enable_hiz
) {
482 hiz_op
= ISL_AUX_OP_FULL_RESOLVE
;
483 } else if (!hiz_enabled
&& enable_hiz
) {
484 hiz_op
= ISL_AUX_OP_AMBIGUATE
;
486 assert(hiz_enabled
== enable_hiz
);
487 /* If the same buffer will be used, no resolves are necessary. */
488 hiz_op
= ISL_AUX_OP_NONE
;
491 if (hiz_op
!= ISL_AUX_OP_NONE
)
492 anv_image_hiz_op(cmd_buffer
, image
, VK_IMAGE_ASPECT_DEPTH_BIT
,
497 vk_image_layout_stencil_write_optimal(VkImageLayout layout
)
499 return layout
== VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
||
500 layout
== VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
;
503 /* Transitions a HiZ-enabled depth buffer from one layout to another. Unless
504 * the initial layout is undefined, the HiZ buffer and depth buffer will
505 * represent the same data at the end of this operation.
508 transition_stencil_buffer(struct anv_cmd_buffer
*cmd_buffer
,
509 const struct anv_image
*image
,
510 uint32_t base_level
, uint32_t level_count
,
511 uint32_t base_layer
, uint32_t layer_count
,
512 VkImageLayout initial_layout
,
513 VkImageLayout final_layout
)
516 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
517 VK_IMAGE_ASPECT_STENCIL_BIT
);
519 /* On gen7, we have to store a texturable version of the stencil buffer in
520 * a shadow whenever VK_IMAGE_USAGE_SAMPLED_BIT is set and copy back and
521 * forth at strategic points. Stencil writes are only allowed in three
524 * - VK_IMAGE_LAYOUT_GENERAL
525 * - VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
526 * - VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
527 * - VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
529 * For general, we have no nice opportunity to transition so we do the copy
530 * to the shadow unconditionally at the end of the subpass. For transfer
531 * destinations, we can update it as part of the transfer op. For the
532 * other two, we delay the copy until a transition into some other layout.
534 if (image
->planes
[plane
].shadow_surface
.isl
.size_B
> 0 &&
535 vk_image_layout_stencil_write_optimal(initial_layout
) &&
536 !vk_image_layout_stencil_write_optimal(final_layout
)) {
537 anv_image_copy_to_shadow(cmd_buffer
, image
,
538 VK_IMAGE_ASPECT_STENCIL_BIT
,
539 base_level
, level_count
,
540 base_layer
, layer_count
);
542 #endif /* GEN_GEN == 7 */
545 #define MI_PREDICATE_SRC0 0x2400
546 #define MI_PREDICATE_SRC1 0x2408
547 #define MI_PREDICATE_RESULT 0x2418
550 set_image_compressed_bit(struct anv_cmd_buffer
*cmd_buffer
,
551 const struct anv_image
*image
,
552 VkImageAspectFlagBits aspect
,
554 uint32_t base_layer
, uint32_t layer_count
,
557 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
559 /* We only have compression tracking for CCS_E */
560 if (image
->planes
[plane
].aux_usage
!= ISL_AUX_USAGE_CCS_E
)
563 for (uint32_t a
= 0; a
< layer_count
; a
++) {
564 uint32_t layer
= base_layer
+ a
;
565 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
566 sdi
.Address
= anv_image_get_compression_state_addr(cmd_buffer
->device
,
569 sdi
.ImmediateData
= compressed
? UINT32_MAX
: 0;
575 set_image_fast_clear_state(struct anv_cmd_buffer
*cmd_buffer
,
576 const struct anv_image
*image
,
577 VkImageAspectFlagBits aspect
,
578 enum anv_fast_clear_type fast_clear
)
580 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
581 sdi
.Address
= anv_image_get_fast_clear_type_addr(cmd_buffer
->device
,
583 sdi
.ImmediateData
= fast_clear
;
586 /* Whenever we have fast-clear, we consider that slice to be compressed.
587 * This makes building predicates much easier.
589 if (fast_clear
!= ANV_FAST_CLEAR_NONE
)
590 set_image_compressed_bit(cmd_buffer
, image
, aspect
, 0, 0, 1, true);
593 /* This is only really practical on haswell and above because it requires
594 * MI math in order to get it correct.
596 #if GEN_GEN >= 8 || GEN_IS_HASWELL
598 anv_cmd_compute_resolve_predicate(struct anv_cmd_buffer
*cmd_buffer
,
599 const struct anv_image
*image
,
600 VkImageAspectFlagBits aspect
,
601 uint32_t level
, uint32_t array_layer
,
602 enum isl_aux_op resolve_op
,
603 enum anv_fast_clear_type fast_clear_supported
)
605 struct gen_mi_builder b
;
606 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
608 const struct gen_mi_value fast_clear_type
=
609 gen_mi_mem32(anv_image_get_fast_clear_type_addr(cmd_buffer
->device
,
612 if (resolve_op
== ISL_AUX_OP_FULL_RESOLVE
) {
613 /* In this case, we're doing a full resolve which means we want the
614 * resolve to happen if any compression (including fast-clears) is
617 * In order to simplify the logic a bit, we make the assumption that,
618 * if the first slice has been fast-cleared, it is also marked as
619 * compressed. See also set_image_fast_clear_state.
621 const struct gen_mi_value compression_state
=
622 gen_mi_mem32(anv_image_get_compression_state_addr(cmd_buffer
->device
,
624 level
, array_layer
));
625 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
),
627 gen_mi_store(&b
, compression_state
, gen_mi_imm(0));
629 if (level
== 0 && array_layer
== 0) {
630 /* If the predicate is true, we want to write 0 to the fast clear type
631 * and, if it's false, leave it alone. We can do this by writing
633 * clear_type = clear_type & ~predicate;
635 struct gen_mi_value new_fast_clear_type
=
636 gen_mi_iand(&b
, fast_clear_type
,
637 gen_mi_inot(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
)));
638 gen_mi_store(&b
, fast_clear_type
, new_fast_clear_type
);
640 } else if (level
== 0 && array_layer
== 0) {
641 /* In this case, we are doing a partial resolve to get rid of fast-clear
642 * colors. We don't care about the compression state but we do care
643 * about how much fast clear is allowed by the final layout.
645 assert(resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
);
646 assert(fast_clear_supported
< ANV_FAST_CLEAR_ANY
);
648 /* We need to compute (fast_clear_supported < image->fast_clear) */
649 struct gen_mi_value pred
=
650 gen_mi_ult(&b
, gen_mi_imm(fast_clear_supported
), fast_clear_type
);
651 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
),
652 gen_mi_value_ref(&b
, pred
));
654 /* If the predicate is true, we want to write 0 to the fast clear type
655 * and, if it's false, leave it alone. We can do this by writing
657 * clear_type = clear_type & ~predicate;
659 struct gen_mi_value new_fast_clear_type
=
660 gen_mi_iand(&b
, fast_clear_type
, gen_mi_inot(&b
, pred
));
661 gen_mi_store(&b
, fast_clear_type
, new_fast_clear_type
);
663 /* In this case, we're trying to do a partial resolve on a slice that
664 * doesn't have clear color. There's nothing to do.
666 assert(resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
);
670 /* Set src1 to 0 and use a != condition */
671 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
673 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
674 mip
.LoadOperation
= LOAD_LOADINV
;
675 mip
.CombineOperation
= COMBINE_SET
;
676 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
679 #endif /* GEN_GEN >= 8 || GEN_IS_HASWELL */
683 anv_cmd_simple_resolve_predicate(struct anv_cmd_buffer
*cmd_buffer
,
684 const struct anv_image
*image
,
685 VkImageAspectFlagBits aspect
,
686 uint32_t level
, uint32_t array_layer
,
687 enum isl_aux_op resolve_op
,
688 enum anv_fast_clear_type fast_clear_supported
)
690 struct gen_mi_builder b
;
691 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
693 struct gen_mi_value fast_clear_type_mem
=
694 gen_mi_mem32(anv_image_get_fast_clear_type_addr(cmd_buffer
->device
,
697 /* This only works for partial resolves and only when the clear color is
698 * all or nothing. On the upside, this emits less command streamer code
699 * and works on Ivybridge and Bay Trail.
701 assert(resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
);
702 assert(fast_clear_supported
!= ANV_FAST_CLEAR_ANY
);
704 /* We don't support fast clears on anything other than the first slice. */
705 if (level
> 0 || array_layer
> 0)
708 /* On gen8, we don't have a concept of default clear colors because we
709 * can't sample from CCS surfaces. It's enough to just load the fast clear
710 * state into the predicate register.
712 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
), fast_clear_type_mem
);
713 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
714 gen_mi_store(&b
, fast_clear_type_mem
, gen_mi_imm(0));
716 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
717 mip
.LoadOperation
= LOAD_LOADINV
;
718 mip
.CombineOperation
= COMBINE_SET
;
719 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
722 #endif /* GEN_GEN <= 8 */
725 anv_cmd_predicated_ccs_resolve(struct anv_cmd_buffer
*cmd_buffer
,
726 const struct anv_image
*image
,
727 enum isl_format format
,
728 VkImageAspectFlagBits aspect
,
729 uint32_t level
, uint32_t array_layer
,
730 enum isl_aux_op resolve_op
,
731 enum anv_fast_clear_type fast_clear_supported
)
733 const uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
736 anv_cmd_compute_resolve_predicate(cmd_buffer
, image
,
737 aspect
, level
, array_layer
,
738 resolve_op
, fast_clear_supported
);
739 #else /* GEN_GEN <= 8 */
740 anv_cmd_simple_resolve_predicate(cmd_buffer
, image
,
741 aspect
, level
, array_layer
,
742 resolve_op
, fast_clear_supported
);
745 /* CCS_D only supports full resolves and BLORP will assert on us if we try
746 * to do a partial resolve on a CCS_D surface.
748 if (resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
&&
749 image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_NONE
)
750 resolve_op
= ISL_AUX_OP_FULL_RESOLVE
;
752 anv_image_ccs_op(cmd_buffer
, image
, format
, aspect
, level
,
753 array_layer
, 1, resolve_op
, NULL
, true);
757 anv_cmd_predicated_mcs_resolve(struct anv_cmd_buffer
*cmd_buffer
,
758 const struct anv_image
*image
,
759 enum isl_format format
,
760 VkImageAspectFlagBits aspect
,
761 uint32_t array_layer
,
762 enum isl_aux_op resolve_op
,
763 enum anv_fast_clear_type fast_clear_supported
)
765 assert(aspect
== VK_IMAGE_ASPECT_COLOR_BIT
);
766 assert(resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
);
768 #if GEN_GEN >= 8 || GEN_IS_HASWELL
769 anv_cmd_compute_resolve_predicate(cmd_buffer
, image
,
770 aspect
, 0, array_layer
,
771 resolve_op
, fast_clear_supported
);
773 anv_image_mcs_op(cmd_buffer
, image
, format
, aspect
,
774 array_layer
, 1, resolve_op
, NULL
, true);
776 unreachable("MCS resolves are unsupported on Ivybridge and Bay Trail");
781 genX(cmd_buffer_mark_image_written
)(struct anv_cmd_buffer
*cmd_buffer
,
782 const struct anv_image
*image
,
783 VkImageAspectFlagBits aspect
,
784 enum isl_aux_usage aux_usage
,
787 uint32_t layer_count
)
789 /* The aspect must be exactly one of the image aspects. */
790 assert(util_bitcount(aspect
) == 1 && (aspect
& image
->aspects
));
792 /* The only compression types with more than just fast-clears are MCS,
793 * CCS_E, and HiZ. With HiZ we just trust the layout and don't actually
794 * track the current fast-clear and compression state. This leaves us
795 * with just MCS and CCS_E.
797 if (aux_usage
!= ISL_AUX_USAGE_CCS_E
&&
798 aux_usage
!= ISL_AUX_USAGE_MCS
)
801 set_image_compressed_bit(cmd_buffer
, image
, aspect
,
802 level
, base_layer
, layer_count
, true);
806 init_fast_clear_color(struct anv_cmd_buffer
*cmd_buffer
,
807 const struct anv_image
*image
,
808 VkImageAspectFlagBits aspect
)
810 assert(cmd_buffer
&& image
);
811 assert(image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
813 set_image_fast_clear_state(cmd_buffer
, image
, aspect
,
814 ANV_FAST_CLEAR_NONE
);
816 /* Initialize the struct fields that are accessed for fast-clears so that
817 * the HW restrictions on the field values are satisfied.
819 struct anv_address addr
=
820 anv_image_get_clear_color_addr(cmd_buffer
->device
, image
, aspect
);
823 const struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
824 const unsigned num_dwords
= GEN_GEN
>= 10 ?
825 isl_dev
->ss
.clear_color_state_size
/ 4 :
826 isl_dev
->ss
.clear_value_size
/ 4;
827 for (unsigned i
= 0; i
< num_dwords
; i
++) {
828 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
830 sdi
.Address
.offset
+= i
* 4;
831 sdi
.ImmediateData
= 0;
835 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_DATA_IMM
), sdi
) {
837 if (GEN_GEN
>= 8 || GEN_IS_HASWELL
) {
838 /* Pre-SKL, the dword containing the clear values also contains
839 * other fields, so we need to initialize those fields to match the
840 * values that would be in a color attachment.
842 sdi
.ImmediateData
= ISL_CHANNEL_SELECT_RED
<< 25 |
843 ISL_CHANNEL_SELECT_GREEN
<< 22 |
844 ISL_CHANNEL_SELECT_BLUE
<< 19 |
845 ISL_CHANNEL_SELECT_ALPHA
<< 16;
846 } else if (GEN_GEN
== 7) {
847 /* On IVB, the dword containing the clear values also contains
848 * other fields that must be zero or can be zero.
850 sdi
.ImmediateData
= 0;
856 /* Copy the fast-clear value dword(s) between a surface state object and an
857 * image's fast clear state buffer.
860 genX(copy_fast_clear_dwords
)(struct anv_cmd_buffer
*cmd_buffer
,
861 struct anv_state surface_state
,
862 const struct anv_image
*image
,
863 VkImageAspectFlagBits aspect
,
864 bool copy_from_surface_state
)
866 assert(cmd_buffer
&& image
);
867 assert(image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
869 struct anv_address ss_clear_addr
= {
870 .bo
= cmd_buffer
->device
->surface_state_pool
.block_pool
.bo
,
871 .offset
= surface_state
.offset
+
872 cmd_buffer
->device
->isl_dev
.ss
.clear_value_offset
,
874 const struct anv_address entry_addr
=
875 anv_image_get_clear_color_addr(cmd_buffer
->device
, image
, aspect
);
876 unsigned copy_size
= cmd_buffer
->device
->isl_dev
.ss
.clear_value_size
;
879 /* On gen7, the combination of commands used here(MI_LOAD_REGISTER_MEM
880 * and MI_STORE_REGISTER_MEM) can cause GPU hangs if any rendering is
881 * in-flight when they are issued even if the memory touched is not
882 * currently active for rendering. The weird bit is that it is not the
883 * MI_LOAD/STORE_REGISTER_MEM commands which hang but rather the in-flight
884 * rendering hangs such that the next stalling command after the
885 * MI_LOAD/STORE_REGISTER_MEM commands will catch the hang.
887 * It is unclear exactly why this hang occurs. Both MI commands come with
888 * warnings about the 3D pipeline but that doesn't seem to fully explain
889 * it. My (Jason's) best theory is that it has something to do with the
890 * fact that we're using a GPU state register as our temporary and that
891 * something with reading/writing it is causing problems.
893 * In order to work around this issue, we emit a PIPE_CONTROL with the
894 * command streamer stall bit set.
896 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
897 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
900 struct gen_mi_builder b
;
901 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
903 if (copy_from_surface_state
) {
904 gen_mi_memcpy(&b
, entry_addr
, ss_clear_addr
, copy_size
);
906 gen_mi_memcpy(&b
, ss_clear_addr
, entry_addr
, copy_size
);
908 /* Updating a surface state object may require that the state cache be
909 * invalidated. From the SKL PRM, Shared Functions -> State -> State
912 * Whenever the RENDER_SURFACE_STATE object in memory pointed to by
913 * the Binding Table Pointer (BTP) and Binding Table Index (BTI) is
914 * modified [...], the L1 state cache must be invalidated to ensure
915 * the new surface or sampler state is fetched from system memory.
917 * In testing, SKL doesn't actually seem to need this, but HSW does.
919 cmd_buffer
->state
.pending_pipe_bits
|=
920 ANV_PIPE_STATE_CACHE_INVALIDATE_BIT
;
925 * @brief Transitions a color buffer from one layout to another.
927 * See section 6.1.1. Image Layout Transitions of the Vulkan 1.0.50 spec for
930 * @param level_count VK_REMAINING_MIP_LEVELS isn't supported.
931 * @param layer_count VK_REMAINING_ARRAY_LAYERS isn't supported. For 3D images,
932 * this represents the maximum layers to transition at each
933 * specified miplevel.
936 transition_color_buffer(struct anv_cmd_buffer
*cmd_buffer
,
937 const struct anv_image
*image
,
938 VkImageAspectFlagBits aspect
,
939 const uint32_t base_level
, uint32_t level_count
,
940 uint32_t base_layer
, uint32_t layer_count
,
941 VkImageLayout initial_layout
,
942 VkImageLayout final_layout
)
944 const struct gen_device_info
*devinfo
= &cmd_buffer
->device
->info
;
945 /* Validate the inputs. */
947 assert(image
&& image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
948 /* These values aren't supported for simplicity's sake. */
949 assert(level_count
!= VK_REMAINING_MIP_LEVELS
&&
950 layer_count
!= VK_REMAINING_ARRAY_LAYERS
);
951 /* Ensure the subresource range is valid. */
952 UNUSED
uint64_t last_level_num
= base_level
+ level_count
;
953 const uint32_t max_depth
= anv_minify(image
->extent
.depth
, base_level
);
954 UNUSED
const uint32_t image_layers
= MAX2(image
->array_size
, max_depth
);
955 assert((uint64_t)base_layer
+ layer_count
<= image_layers
);
956 assert(last_level_num
<= image
->levels
);
957 /* The spec disallows these final layouts. */
958 assert(final_layout
!= VK_IMAGE_LAYOUT_UNDEFINED
&&
959 final_layout
!= VK_IMAGE_LAYOUT_PREINITIALIZED
);
961 /* No work is necessary if the layout stays the same or if this subresource
962 * range lacks auxiliary data.
964 if (initial_layout
== final_layout
)
967 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
969 if (image
->planes
[plane
].shadow_surface
.isl
.size_B
> 0 &&
970 final_layout
== VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
) {
971 /* This surface is a linear compressed image with a tiled shadow surface
972 * for texturing. The client is about to use it in READ_ONLY_OPTIMAL so
973 * we need to ensure the shadow copy is up-to-date.
975 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
976 assert(image
->planes
[plane
].surface
.isl
.tiling
== ISL_TILING_LINEAR
);
977 assert(image
->planes
[plane
].shadow_surface
.isl
.tiling
!= ISL_TILING_LINEAR
);
978 assert(isl_format_is_compressed(image
->planes
[plane
].surface
.isl
.format
));
980 anv_image_copy_to_shadow(cmd_buffer
, image
,
981 VK_IMAGE_ASPECT_COLOR_BIT
,
982 base_level
, level_count
,
983 base_layer
, layer_count
);
986 if (base_layer
>= anv_image_aux_layers(image
, aspect
, base_level
))
989 assert(image
->tiling
== VK_IMAGE_TILING_OPTIMAL
);
991 if (initial_layout
== VK_IMAGE_LAYOUT_UNDEFINED
||
992 initial_layout
== VK_IMAGE_LAYOUT_PREINITIALIZED
) {
993 /* A subresource in the undefined layout may have been aliased and
994 * populated with any arrangement of bits. Therefore, we must initialize
995 * the related aux buffer and clear buffer entry with desirable values.
996 * An initial layout of PREINITIALIZED is the same as UNDEFINED for
997 * images with VK_IMAGE_TILING_OPTIMAL.
999 * Initialize the relevant clear buffer entries.
1001 if (base_level
== 0 && base_layer
== 0)
1002 init_fast_clear_color(cmd_buffer
, image
, aspect
);
1004 /* Initialize the aux buffers to enable correct rendering. In order to
1005 * ensure that things such as storage images work correctly, aux buffers
1006 * need to be initialized to valid data.
1008 * Having an aux buffer with invalid data is a problem for two reasons:
1010 * 1) Having an invalid value in the buffer can confuse the hardware.
1011 * For instance, with CCS_E on SKL, a two-bit CCS value of 2 is
1012 * invalid and leads to the hardware doing strange things. It
1013 * doesn't hang as far as we can tell but rendering corruption can
1016 * 2) If this transition is into the GENERAL layout and we then use the
1017 * image as a storage image, then we must have the aux buffer in the
1018 * pass-through state so that, if we then go to texture from the
1019 * image, we get the results of our storage image writes and not the
1020 * fast clear color or other random data.
1022 * For CCS both of the problems above are real demonstrable issues. In
1023 * that case, the only thing we can do is to perform an ambiguate to
1024 * transition the aux surface into the pass-through state.
1026 * For MCS, (2) is never an issue because we don't support multisampled
1027 * storage images. In theory, issue (1) is a problem with MCS but we've
1028 * never seen it in the wild. For 4x and 16x, all bit patters could, in
1029 * theory, be interpreted as something but we don't know that all bit
1030 * patterns are actually valid. For 2x and 8x, you could easily end up
1031 * with the MCS referring to an invalid plane because not all bits of
1032 * the MCS value are actually used. Even though we've never seen issues
1033 * in the wild, it's best to play it safe and initialize the MCS. We
1034 * can use a fast-clear for MCS because we only ever touch from render
1035 * and texture (no image load store).
1037 if (image
->samples
== 1) {
1038 for (uint32_t l
= 0; l
< level_count
; l
++) {
1039 const uint32_t level
= base_level
+ l
;
1041 uint32_t aux_layers
= anv_image_aux_layers(image
, aspect
, level
);
1042 if (base_layer
>= aux_layers
)
1043 break; /* We will only get fewer layers as level increases */
1044 uint32_t level_layer_count
=
1045 MIN2(layer_count
, aux_layers
- base_layer
);
1047 anv_image_ccs_op(cmd_buffer
, image
,
1048 image
->planes
[plane
].surface
.isl
.format
,
1049 aspect
, level
, base_layer
, level_layer_count
,
1050 ISL_AUX_OP_AMBIGUATE
, NULL
, false);
1052 if (image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_CCS_E
) {
1053 set_image_compressed_bit(cmd_buffer
, image
, aspect
,
1054 level
, base_layer
, level_layer_count
,
1059 if (image
->samples
== 4 || image
->samples
== 16) {
1060 anv_perf_warn(cmd_buffer
->device
->instance
, image
,
1061 "Doing a potentially unnecessary fast-clear to "
1062 "define an MCS buffer.");
1065 assert(base_level
== 0 && level_count
== 1);
1066 anv_image_mcs_op(cmd_buffer
, image
,
1067 image
->planes
[plane
].surface
.isl
.format
,
1068 aspect
, base_layer
, layer_count
,
1069 ISL_AUX_OP_FAST_CLEAR
, NULL
, false);
1074 const enum isl_aux_usage initial_aux_usage
=
1075 anv_layout_to_aux_usage(devinfo
, image
, aspect
, initial_layout
);
1076 const enum isl_aux_usage final_aux_usage
=
1077 anv_layout_to_aux_usage(devinfo
, image
, aspect
, final_layout
);
1079 /* The current code assumes that there is no mixing of CCS_E and CCS_D.
1080 * We can handle transitions between CCS_D/E to and from NONE. What we
1081 * don't yet handle is switching between CCS_E and CCS_D within a given
1082 * image. Doing so in a performant way requires more detailed aux state
1083 * tracking such as what is done in i965. For now, just assume that we
1084 * only have one type of compression.
1086 assert(initial_aux_usage
== ISL_AUX_USAGE_NONE
||
1087 final_aux_usage
== ISL_AUX_USAGE_NONE
||
1088 initial_aux_usage
== final_aux_usage
);
1090 /* If initial aux usage is NONE, there is nothing to resolve */
1091 if (initial_aux_usage
== ISL_AUX_USAGE_NONE
)
1094 enum isl_aux_op resolve_op
= ISL_AUX_OP_NONE
;
1096 /* If the initial layout supports more fast clear than the final layout
1097 * then we need at least a partial resolve.
1099 const enum anv_fast_clear_type initial_fast_clear
=
1100 anv_layout_to_fast_clear_type(devinfo
, image
, aspect
, initial_layout
);
1101 const enum anv_fast_clear_type final_fast_clear
=
1102 anv_layout_to_fast_clear_type(devinfo
, image
, aspect
, final_layout
);
1103 if (final_fast_clear
< initial_fast_clear
)
1104 resolve_op
= ISL_AUX_OP_PARTIAL_RESOLVE
;
1106 if (initial_aux_usage
== ISL_AUX_USAGE_CCS_E
&&
1107 final_aux_usage
!= ISL_AUX_USAGE_CCS_E
)
1108 resolve_op
= ISL_AUX_OP_FULL_RESOLVE
;
1110 if (resolve_op
== ISL_AUX_OP_NONE
)
1113 /* Perform a resolve to synchronize data between the main and aux buffer.
1114 * Before we begin, we must satisfy the cache flushing requirement specified
1115 * in the Sky Lake PRM Vol. 7, "MCS Buffer for Render Target(s)":
1117 * Any transition from any value in {Clear, Render, Resolve} to a
1118 * different value in {Clear, Render, Resolve} requires end of pipe
1121 * We perform a flush of the write cache before and after the clear and
1122 * resolve operations to meet this requirement.
1124 * Unlike other drawing, fast clear operations are not properly
1125 * synchronized. The first PIPE_CONTROL here likely ensures that the
1126 * contents of the previous render or clear hit the render target before we
1127 * resolve and the second likely ensures that the resolve is complete before
1128 * we do any more rendering or clearing.
1130 cmd_buffer
->state
.pending_pipe_bits
|=
1131 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
| ANV_PIPE_CS_STALL_BIT
;
1133 for (uint32_t l
= 0; l
< level_count
; l
++) {
1134 uint32_t level
= base_level
+ l
;
1136 uint32_t aux_layers
= anv_image_aux_layers(image
, aspect
, level
);
1137 if (base_layer
>= aux_layers
)
1138 break; /* We will only get fewer layers as level increases */
1139 uint32_t level_layer_count
=
1140 MIN2(layer_count
, aux_layers
- base_layer
);
1142 for (uint32_t a
= 0; a
< level_layer_count
; a
++) {
1143 uint32_t array_layer
= base_layer
+ a
;
1144 if (image
->samples
== 1) {
1145 anv_cmd_predicated_ccs_resolve(cmd_buffer
, image
,
1146 image
->planes
[plane
].surface
.isl
.format
,
1147 aspect
, level
, array_layer
, resolve_op
,
1150 /* We only support fast-clear on the first layer so partial
1151 * resolves should not be used on other layers as they will use
1152 * the clear color stored in memory that is only valid for layer0.
1154 if (resolve_op
== ISL_AUX_OP_PARTIAL_RESOLVE
&&
1158 anv_cmd_predicated_mcs_resolve(cmd_buffer
, image
,
1159 image
->planes
[plane
].surface
.isl
.format
,
1160 aspect
, array_layer
, resolve_op
,
1166 cmd_buffer
->state
.pending_pipe_bits
|=
1167 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
| ANV_PIPE_CS_STALL_BIT
;
1171 * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
1174 genX(cmd_buffer_setup_attachments
)(struct anv_cmd_buffer
*cmd_buffer
,
1175 struct anv_render_pass
*pass
,
1176 const VkRenderPassBeginInfo
*begin
)
1178 const struct isl_device
*isl_dev
= &cmd_buffer
->device
->isl_dev
;
1179 struct anv_cmd_state
*state
= &cmd_buffer
->state
;
1180 struct anv_framebuffer
*framebuffer
= cmd_buffer
->state
.framebuffer
;
1182 vk_free(&cmd_buffer
->pool
->alloc
, state
->attachments
);
1184 if (pass
->attachment_count
> 0) {
1185 state
->attachments
= vk_alloc(&cmd_buffer
->pool
->alloc
,
1186 pass
->attachment_count
*
1187 sizeof(state
->attachments
[0]),
1188 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
1189 if (state
->attachments
== NULL
) {
1190 /* Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
1191 return anv_batch_set_error(&cmd_buffer
->batch
,
1192 VK_ERROR_OUT_OF_HOST_MEMORY
);
1195 state
->attachments
= NULL
;
1198 /* Reserve one for the NULL state. */
1199 unsigned num_states
= 1;
1200 for (uint32_t i
= 0; i
< pass
->attachment_count
; ++i
) {
1201 if (vk_format_is_color(pass
->attachments
[i
].format
))
1204 if (need_input_attachment_state(&pass
->attachments
[i
]))
1208 const uint32_t ss_stride
= align_u32(isl_dev
->ss
.size
, isl_dev
->ss
.align
);
1209 state
->render_pass_states
=
1210 anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
,
1211 num_states
* ss_stride
, isl_dev
->ss
.align
);
1213 struct anv_state next_state
= state
->render_pass_states
;
1214 next_state
.alloc_size
= isl_dev
->ss
.size
;
1216 state
->null_surface_state
= next_state
;
1217 next_state
.offset
+= ss_stride
;
1218 next_state
.map
+= ss_stride
;
1220 const VkRenderPassAttachmentBeginInfoKHR
*begin_attachment
=
1221 vk_find_struct_const(begin
, RENDER_PASS_ATTACHMENT_BEGIN_INFO_KHR
);
1223 if (begin
&& !begin_attachment
)
1224 assert(pass
->attachment_count
== framebuffer
->attachment_count
);
1226 for (uint32_t i
= 0; i
< pass
->attachment_count
; ++i
) {
1227 if (vk_format_is_color(pass
->attachments
[i
].format
)) {
1228 state
->attachments
[i
].color
.state
= next_state
;
1229 next_state
.offset
+= ss_stride
;
1230 next_state
.map
+= ss_stride
;
1233 if (need_input_attachment_state(&pass
->attachments
[i
])) {
1234 state
->attachments
[i
].input
.state
= next_state
;
1235 next_state
.offset
+= ss_stride
;
1236 next_state
.map
+= ss_stride
;
1239 if (begin_attachment
&& begin_attachment
->attachmentCount
!= 0) {
1240 assert(begin_attachment
->attachmentCount
== pass
->attachment_count
);
1241 ANV_FROM_HANDLE(anv_image_view
, iview
, begin_attachment
->pAttachments
[i
]);
1242 cmd_buffer
->state
.attachments
[i
].image_view
= iview
;
1243 } else if (framebuffer
&& i
< framebuffer
->attachment_count
) {
1244 cmd_buffer
->state
.attachments
[i
].image_view
= framebuffer
->attachments
[i
];
1247 assert(next_state
.offset
== state
->render_pass_states
.offset
+
1248 state
->render_pass_states
.alloc_size
);
1251 isl_null_fill_state(isl_dev
, state
->null_surface_state
.map
,
1252 isl_extent3d(framebuffer
->width
,
1253 framebuffer
->height
,
1254 framebuffer
->layers
));
1256 for (uint32_t i
= 0; i
< pass
->attachment_count
; ++i
) {
1257 struct anv_render_pass_attachment
*att
= &pass
->attachments
[i
];
1258 VkImageAspectFlags att_aspects
= vk_format_aspects(att
->format
);
1259 VkImageAspectFlags clear_aspects
= 0;
1260 VkImageAspectFlags load_aspects
= 0;
1262 if (att_aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
1263 /* color attachment */
1264 if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1265 clear_aspects
|= VK_IMAGE_ASPECT_COLOR_BIT
;
1266 } else if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_LOAD
) {
1267 load_aspects
|= VK_IMAGE_ASPECT_COLOR_BIT
;
1270 /* depthstencil attachment */
1271 if (att_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
1272 if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1273 clear_aspects
|= VK_IMAGE_ASPECT_DEPTH_BIT
;
1274 } else if (att
->load_op
== VK_ATTACHMENT_LOAD_OP_LOAD
) {
1275 load_aspects
|= VK_IMAGE_ASPECT_DEPTH_BIT
;
1278 if (att_aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
1279 if (att
->stencil_load_op
== VK_ATTACHMENT_LOAD_OP_CLEAR
) {
1280 clear_aspects
|= VK_IMAGE_ASPECT_STENCIL_BIT
;
1281 } else if (att
->stencil_load_op
== VK_ATTACHMENT_LOAD_OP_LOAD
) {
1282 load_aspects
|= VK_IMAGE_ASPECT_STENCIL_BIT
;
1287 state
->attachments
[i
].current_layout
= att
->initial_layout
;
1288 state
->attachments
[i
].pending_clear_aspects
= clear_aspects
;
1289 state
->attachments
[i
].pending_load_aspects
= load_aspects
;
1291 state
->attachments
[i
].clear_value
= begin
->pClearValues
[i
];
1293 struct anv_image_view
*iview
= cmd_buffer
->state
.attachments
[i
].image_view
;
1294 anv_assert(iview
->vk_format
== att
->format
);
1296 const uint32_t num_layers
= iview
->planes
[0].isl
.array_len
;
1297 state
->attachments
[i
].pending_clear_views
= (1 << num_layers
) - 1;
1299 union isl_color_value clear_color
= { .u32
= { 0, } };
1300 if (att_aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
1301 anv_assert(iview
->n_planes
== 1);
1302 assert(att_aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
1303 color_attachment_compute_aux_usage(cmd_buffer
->device
,
1304 state
, i
, begin
->renderArea
,
1307 anv_image_fill_surface_state(cmd_buffer
->device
,
1309 VK_IMAGE_ASPECT_COLOR_BIT
,
1310 &iview
->planes
[0].isl
,
1311 ISL_SURF_USAGE_RENDER_TARGET_BIT
,
1312 state
->attachments
[i
].aux_usage
,
1315 &state
->attachments
[i
].color
,
1318 add_surface_state_relocs(cmd_buffer
, state
->attachments
[i
].color
);
1320 depth_stencil_attachment_compute_aux_usage(cmd_buffer
->device
,
1325 if (need_input_attachment_state(&pass
->attachments
[i
])) {
1326 anv_image_fill_surface_state(cmd_buffer
->device
,
1328 VK_IMAGE_ASPECT_COLOR_BIT
,
1329 &iview
->planes
[0].isl
,
1330 ISL_SURF_USAGE_TEXTURE_BIT
,
1331 state
->attachments
[i
].input_aux_usage
,
1334 &state
->attachments
[i
].input
,
1337 add_surface_state_relocs(cmd_buffer
, state
->attachments
[i
].input
);
1346 genX(BeginCommandBuffer
)(
1347 VkCommandBuffer commandBuffer
,
1348 const VkCommandBufferBeginInfo
* pBeginInfo
)
1350 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1352 /* If this is the first vkBeginCommandBuffer, we must *initialize* the
1353 * command buffer's state. Otherwise, we must *reset* its state. In both
1354 * cases we reset it.
1356 * From the Vulkan 1.0 spec:
1358 * If a command buffer is in the executable state and the command buffer
1359 * was allocated from a command pool with the
1360 * VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT flag set, then
1361 * vkBeginCommandBuffer implicitly resets the command buffer, behaving
1362 * as if vkResetCommandBuffer had been called with
1363 * VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT not set. It then puts
1364 * the command buffer in the recording state.
1366 anv_cmd_buffer_reset(cmd_buffer
);
1368 cmd_buffer
->usage_flags
= pBeginInfo
->flags
;
1370 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
||
1371 !(cmd_buffer
->usage_flags
& VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT
));
1373 genX(cmd_buffer_emit_state_base_address
)(cmd_buffer
);
1375 /* We sometimes store vertex data in the dynamic state buffer for blorp
1376 * operations and our dynamic state stream may re-use data from previous
1377 * command buffers. In order to prevent stale cache data, we flush the VF
1378 * cache. We could do this on every blorp call but that's not really
1379 * needed as all of the data will get written by the CPU prior to the GPU
1380 * executing anything. The chances are fairly high that they will use
1381 * blorp at least once per primary command buffer so it shouldn't be
1384 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
)
1385 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_VF_CACHE_INVALIDATE_BIT
;
1387 /* We send an "Indirect State Pointers Disable" packet at
1388 * EndCommandBuffer, so all push contant packets are ignored during a
1389 * context restore. Documentation says after that command, we need to
1390 * emit push constants again before any rendering operation. So we
1391 * flag them dirty here to make sure they get emitted.
1393 cmd_buffer
->state
.push_constants_dirty
|= VK_SHADER_STAGE_ALL_GRAPHICS
;
1395 VkResult result
= VK_SUCCESS
;
1396 if (cmd_buffer
->usage_flags
&
1397 VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT
) {
1398 assert(pBeginInfo
->pInheritanceInfo
);
1399 cmd_buffer
->state
.pass
=
1400 anv_render_pass_from_handle(pBeginInfo
->pInheritanceInfo
->renderPass
);
1401 cmd_buffer
->state
.subpass
=
1402 &cmd_buffer
->state
.pass
->subpasses
[pBeginInfo
->pInheritanceInfo
->subpass
];
1404 /* This is optional in the inheritance info. */
1405 cmd_buffer
->state
.framebuffer
=
1406 anv_framebuffer_from_handle(pBeginInfo
->pInheritanceInfo
->framebuffer
);
1408 result
= genX(cmd_buffer_setup_attachments
)(cmd_buffer
,
1409 cmd_buffer
->state
.pass
, NULL
);
1411 /* Record that HiZ is enabled if we can. */
1412 if (cmd_buffer
->state
.framebuffer
) {
1413 const struct anv_image_view
* const iview
=
1414 anv_cmd_buffer_get_depth_stencil_view(cmd_buffer
);
1417 VkImageLayout layout
=
1418 cmd_buffer
->state
.subpass
->depth_stencil_attachment
->layout
;
1420 enum isl_aux_usage aux_usage
=
1421 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, iview
->image
,
1422 VK_IMAGE_ASPECT_DEPTH_BIT
, layout
);
1424 cmd_buffer
->state
.hiz_enabled
= aux_usage
== ISL_AUX_USAGE_HIZ
;
1428 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_RENDER_TARGETS
;
1431 #if GEN_GEN >= 8 || GEN_IS_HASWELL
1432 if (cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
) {
1433 const VkCommandBufferInheritanceConditionalRenderingInfoEXT
*conditional_rendering_info
=
1434 vk_find_struct_const(pBeginInfo
->pInheritanceInfo
->pNext
, COMMAND_BUFFER_INHERITANCE_CONDITIONAL_RENDERING_INFO_EXT
);
1436 /* If secondary buffer supports conditional rendering
1437 * we should emit commands as if conditional rendering is enabled.
1439 cmd_buffer
->state
.conditional_render_enabled
=
1440 conditional_rendering_info
&& conditional_rendering_info
->conditionalRenderingEnable
;
1447 /* From the PRM, Volume 2a:
1449 * "Indirect State Pointers Disable
1451 * At the completion of the post-sync operation associated with this pipe
1452 * control packet, the indirect state pointers in the hardware are
1453 * considered invalid; the indirect pointers are not saved in the context.
1454 * If any new indirect state commands are executed in the command stream
1455 * while the pipe control is pending, the new indirect state commands are
1458 * [DevIVB+]: Using Invalidate State Pointer (ISP) only inhibits context
1459 * restoring of Push Constant (3DSTATE_CONSTANT_*) commands. Push Constant
1460 * commands are only considered as Indirect State Pointers. Once ISP is
1461 * issued in a context, SW must initialize by programming push constant
1462 * commands for all the shaders (at least to zero length) before attempting
1463 * any rendering operation for the same context."
1465 * 3DSTATE_CONSTANT_* packets are restored during a context restore,
1466 * even though they point to a BO that has been already unreferenced at
1467 * the end of the previous batch buffer. This has been fine so far since
1468 * we are protected by these scratch page (every address not covered by
1469 * a BO should be pointing to the scratch page). But on CNL, it is
1470 * causing a GPU hang during context restore at the 3DSTATE_CONSTANT_*
1473 * The flag "Indirect State Pointers Disable" in PIPE_CONTROL tells the
1474 * hardware to ignore previous 3DSTATE_CONSTANT_* packets during a
1475 * context restore, so the mentioned hang doesn't happen. However,
1476 * software must program push constant commands for all stages prior to
1477 * rendering anything. So we flag them dirty in BeginCommandBuffer.
1479 * Finally, we also make sure to stall at pixel scoreboard to make sure the
1480 * constants have been loaded into the EUs prior to disable the push constants
1481 * so that it doesn't hang a previous 3DPRIMITIVE.
1484 emit_isp_disable(struct anv_cmd_buffer
*cmd_buffer
)
1486 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1487 pc
.StallAtPixelScoreboard
= true;
1488 pc
.CommandStreamerStallEnable
= true;
1490 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1491 pc
.IndirectStatePointersDisable
= true;
1492 pc
.CommandStreamerStallEnable
= true;
1497 genX(EndCommandBuffer
)(
1498 VkCommandBuffer commandBuffer
)
1500 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1502 if (anv_batch_has_error(&cmd_buffer
->batch
))
1503 return cmd_buffer
->batch
.status
;
1505 /* We want every command buffer to start with the PMA fix in a known state,
1506 * so we disable it at the end of the command buffer.
1508 genX(cmd_buffer_enable_pma_fix
)(cmd_buffer
, false);
1510 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
1512 emit_isp_disable(cmd_buffer
);
1514 anv_cmd_buffer_end_batch_buffer(cmd_buffer
);
1520 genX(CmdExecuteCommands
)(
1521 VkCommandBuffer commandBuffer
,
1522 uint32_t commandBufferCount
,
1523 const VkCommandBuffer
* pCmdBuffers
)
1525 ANV_FROM_HANDLE(anv_cmd_buffer
, primary
, commandBuffer
);
1527 assert(primary
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
1529 if (anv_batch_has_error(&primary
->batch
))
1532 /* The secondary command buffers will assume that the PMA fix is disabled
1533 * when they begin executing. Make sure this is true.
1535 genX(cmd_buffer_enable_pma_fix
)(primary
, false);
1537 /* The secondary command buffer doesn't know which textures etc. have been
1538 * flushed prior to their execution. Apply those flushes now.
1540 genX(cmd_buffer_apply_pipe_flushes
)(primary
);
1542 for (uint32_t i
= 0; i
< commandBufferCount
; i
++) {
1543 ANV_FROM_HANDLE(anv_cmd_buffer
, secondary
, pCmdBuffers
[i
]);
1545 assert(secondary
->level
== VK_COMMAND_BUFFER_LEVEL_SECONDARY
);
1546 assert(!anv_batch_has_error(&secondary
->batch
));
1548 #if GEN_GEN >= 8 || GEN_IS_HASWELL
1549 if (secondary
->state
.conditional_render_enabled
) {
1550 if (!primary
->state
.conditional_render_enabled
) {
1551 /* Secondary buffer is constructed as if it will be executed
1552 * with conditional rendering, we should satisfy this dependency
1553 * regardless of conditional rendering being enabled in primary.
1555 struct gen_mi_builder b
;
1556 gen_mi_builder_init(&b
, &primary
->batch
);
1557 gen_mi_store(&b
, gen_mi_reg64(ANV_PREDICATE_RESULT_REG
),
1558 gen_mi_imm(UINT64_MAX
));
1563 if (secondary
->usage_flags
&
1564 VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT
) {
1565 /* If we're continuing a render pass from the primary, we need to
1566 * copy the surface states for the current subpass into the storage
1567 * we allocated for them in BeginCommandBuffer.
1569 struct anv_bo
*ss_bo
=
1570 primary
->device
->surface_state_pool
.block_pool
.bo
;
1571 struct anv_state src_state
= primary
->state
.render_pass_states
;
1572 struct anv_state dst_state
= secondary
->state
.render_pass_states
;
1573 assert(src_state
.alloc_size
== dst_state
.alloc_size
);
1575 genX(cmd_buffer_so_memcpy
)(primary
,
1576 (struct anv_address
) {
1578 .offset
= dst_state
.offset
,
1580 (struct anv_address
) {
1582 .offset
= src_state
.offset
,
1584 src_state
.alloc_size
);
1587 anv_cmd_buffer_add_secondary(primary
, secondary
);
1590 /* The secondary may have selected a different pipeline (3D or compute) and
1591 * may have changed the current L3$ configuration. Reset our tracking
1592 * variables to invalid values to ensure that we re-emit these in the case
1593 * where we do any draws or compute dispatches from the primary after the
1594 * secondary has returned.
1596 primary
->state
.current_pipeline
= UINT32_MAX
;
1597 primary
->state
.current_l3_config
= NULL
;
1598 primary
->state
.current_hash_scale
= 0;
1600 /* Each of the secondary command buffers will use its own state base
1601 * address. We need to re-emit state base address for the primary after
1602 * all of the secondaries are done.
1604 * TODO: Maybe we want to make this a dirty bit to avoid extra state base
1607 genX(cmd_buffer_emit_state_base_address
)(primary
);
1610 #define IVB_L3SQCREG1_SQGHPCI_DEFAULT 0x00730000
1611 #define VLV_L3SQCREG1_SQGHPCI_DEFAULT 0x00d30000
1612 #define HSW_L3SQCREG1_SQGHPCI_DEFAULT 0x00610000
1615 * Program the hardware to use the specified L3 configuration.
1618 genX(cmd_buffer_config_l3
)(struct anv_cmd_buffer
*cmd_buffer
,
1619 const struct gen_l3_config
*cfg
)
1622 if (cfg
== cmd_buffer
->state
.current_l3_config
)
1625 if (unlikely(INTEL_DEBUG
& DEBUG_L3
)) {
1626 intel_logd("L3 config transition: ");
1627 gen_dump_l3_config(cfg
, stderr
);
1630 const bool has_slm
= cfg
->n
[GEN_L3P_SLM
];
1632 /* According to the hardware docs, the L3 partitioning can only be changed
1633 * while the pipeline is completely drained and the caches are flushed,
1634 * which involves a first PIPE_CONTROL flush which stalls the pipeline...
1636 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1637 pc
.DCFlushEnable
= true;
1638 pc
.PostSyncOperation
= NoWrite
;
1639 pc
.CommandStreamerStallEnable
= true;
1642 /* ...followed by a second pipelined PIPE_CONTROL that initiates
1643 * invalidation of the relevant caches. Note that because RO invalidation
1644 * happens at the top of the pipeline (i.e. right away as the PIPE_CONTROL
1645 * command is processed by the CS) we cannot combine it with the previous
1646 * stalling flush as the hardware documentation suggests, because that
1647 * would cause the CS to stall on previous rendering *after* RO
1648 * invalidation and wouldn't prevent the RO caches from being polluted by
1649 * concurrent rendering before the stall completes. This intentionally
1650 * doesn't implement the SKL+ hardware workaround suggesting to enable CS
1651 * stall on PIPE_CONTROLs with the texture cache invalidation bit set for
1652 * GPGPU workloads because the previous and subsequent PIPE_CONTROLs
1653 * already guarantee that there is no concurrent GPGPU kernel execution
1654 * (see SKL HSD 2132585).
1656 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1657 pc
.TextureCacheInvalidationEnable
= true;
1658 pc
.ConstantCacheInvalidationEnable
= true;
1659 pc
.InstructionCacheInvalidateEnable
= true;
1660 pc
.StateCacheInvalidationEnable
= true;
1661 pc
.PostSyncOperation
= NoWrite
;
1664 /* Now send a third stalling flush to make sure that invalidation is
1665 * complete when the L3 configuration registers are modified.
1667 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
1668 pc
.DCFlushEnable
= true;
1669 pc
.PostSyncOperation
= NoWrite
;
1670 pc
.CommandStreamerStallEnable
= true;
1675 assert(!cfg
->n
[GEN_L3P_IS
] && !cfg
->n
[GEN_L3P_C
] && !cfg
->n
[GEN_L3P_T
]);
1678 anv_pack_struct(&l3cr
, GENX(L3CNTLREG
),
1679 .SLMEnable
= has_slm
,
1681 /* WA_1406697149: Bit 9 "Error Detection Behavior Control" must be set
1682 * in L3CNTLREG register. The default setting of the bit is not the
1683 * desirable behavior.
1685 .ErrorDetectionBehaviorControl
= true,
1686 .UseFullWays
= true,
1688 .URBAllocation
= cfg
->n
[GEN_L3P_URB
],
1689 .ROAllocation
= cfg
->n
[GEN_L3P_RO
],
1690 .DCAllocation
= cfg
->n
[GEN_L3P_DC
],
1691 .AllAllocation
= cfg
->n
[GEN_L3P_ALL
]);
1693 /* Set up the L3 partitioning. */
1694 emit_lri(&cmd_buffer
->batch
, GENX(L3CNTLREG_num
), l3cr
);
1698 const bool has_dc
= cfg
->n
[GEN_L3P_DC
] || cfg
->n
[GEN_L3P_ALL
];
1699 const bool has_is
= cfg
->n
[GEN_L3P_IS
] || cfg
->n
[GEN_L3P_RO
] ||
1700 cfg
->n
[GEN_L3P_ALL
];
1701 const bool has_c
= cfg
->n
[GEN_L3P_C
] || cfg
->n
[GEN_L3P_RO
] ||
1702 cfg
->n
[GEN_L3P_ALL
];
1703 const bool has_t
= cfg
->n
[GEN_L3P_T
] || cfg
->n
[GEN_L3P_RO
] ||
1704 cfg
->n
[GEN_L3P_ALL
];
1706 assert(!cfg
->n
[GEN_L3P_ALL
]);
1708 /* When enabled SLM only uses a portion of the L3 on half of the banks,
1709 * the matching space on the remaining banks has to be allocated to a
1710 * client (URB for all validated configurations) set to the
1711 * lower-bandwidth 2-bank address hashing mode.
1713 const struct gen_device_info
*devinfo
= &cmd_buffer
->device
->info
;
1714 const bool urb_low_bw
= has_slm
&& !devinfo
->is_baytrail
;
1715 assert(!urb_low_bw
|| cfg
->n
[GEN_L3P_URB
] == cfg
->n
[GEN_L3P_SLM
]);
1717 /* Minimum number of ways that can be allocated to the URB. */
1718 const unsigned n0_urb
= devinfo
->is_baytrail
? 32 : 0;
1719 assert(cfg
->n
[GEN_L3P_URB
] >= n0_urb
);
1721 uint32_t l3sqcr1
, l3cr2
, l3cr3
;
1722 anv_pack_struct(&l3sqcr1
, GENX(L3SQCREG1
),
1723 .ConvertDC_UC
= !has_dc
,
1724 .ConvertIS_UC
= !has_is
,
1725 .ConvertC_UC
= !has_c
,
1726 .ConvertT_UC
= !has_t
);
1728 GEN_IS_HASWELL
? HSW_L3SQCREG1_SQGHPCI_DEFAULT
:
1729 devinfo
->is_baytrail
? VLV_L3SQCREG1_SQGHPCI_DEFAULT
:
1730 IVB_L3SQCREG1_SQGHPCI_DEFAULT
;
1732 anv_pack_struct(&l3cr2
, GENX(L3CNTLREG2
),
1733 .SLMEnable
= has_slm
,
1734 .URBLowBandwidth
= urb_low_bw
,
1735 .URBAllocation
= cfg
->n
[GEN_L3P_URB
] - n0_urb
,
1737 .ALLAllocation
= cfg
->n
[GEN_L3P_ALL
],
1739 .ROAllocation
= cfg
->n
[GEN_L3P_RO
],
1740 .DCAllocation
= cfg
->n
[GEN_L3P_DC
]);
1742 anv_pack_struct(&l3cr3
, GENX(L3CNTLREG3
),
1743 .ISAllocation
= cfg
->n
[GEN_L3P_IS
],
1744 .ISLowBandwidth
= 0,
1745 .CAllocation
= cfg
->n
[GEN_L3P_C
],
1747 .TAllocation
= cfg
->n
[GEN_L3P_T
],
1748 .TLowBandwidth
= 0);
1750 /* Set up the L3 partitioning. */
1751 emit_lri(&cmd_buffer
->batch
, GENX(L3SQCREG1_num
), l3sqcr1
);
1752 emit_lri(&cmd_buffer
->batch
, GENX(L3CNTLREG2_num
), l3cr2
);
1753 emit_lri(&cmd_buffer
->batch
, GENX(L3CNTLREG3_num
), l3cr3
);
1756 if (cmd_buffer
->device
->instance
->physicalDevice
.cmd_parser_version
>= 4) {
1757 /* Enable L3 atomics on HSW if we have a DC partition, otherwise keep
1758 * them disabled to avoid crashing the system hard.
1760 uint32_t scratch1
, chicken3
;
1761 anv_pack_struct(&scratch1
, GENX(SCRATCH1
),
1762 .L3AtomicDisable
= !has_dc
);
1763 anv_pack_struct(&chicken3
, GENX(CHICKEN3
),
1764 .L3AtomicDisableMask
= true,
1765 .L3AtomicDisable
= !has_dc
);
1766 emit_lri(&cmd_buffer
->batch
, GENX(SCRATCH1_num
), scratch1
);
1767 emit_lri(&cmd_buffer
->batch
, GENX(CHICKEN3_num
), chicken3
);
1773 cmd_buffer
->state
.current_l3_config
= cfg
;
1777 genX(cmd_buffer_apply_pipe_flushes
)(struct anv_cmd_buffer
*cmd_buffer
)
1779 enum anv_pipe_bits bits
= cmd_buffer
->state
.pending_pipe_bits
;
1781 /* Flushes are pipelined while invalidations are handled immediately.
1782 * Therefore, if we're flushing anything then we need to schedule a stall
1783 * before any invalidations can happen.
1785 if (bits
& ANV_PIPE_FLUSH_BITS
)
1786 bits
|= ANV_PIPE_NEEDS_CS_STALL_BIT
;
1788 /* If we're going to do an invalidate and we have a pending CS stall that
1789 * has yet to be resolved, we do the CS stall now.
1791 if ((bits
& ANV_PIPE_INVALIDATE_BITS
) &&
1792 (bits
& ANV_PIPE_NEEDS_CS_STALL_BIT
)) {
1793 bits
|= ANV_PIPE_CS_STALL_BIT
;
1794 bits
&= ~ANV_PIPE_NEEDS_CS_STALL_BIT
;
1797 if (bits
& (ANV_PIPE_FLUSH_BITS
| ANV_PIPE_CS_STALL_BIT
)) {
1798 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
1799 pipe
.DepthCacheFlushEnable
= bits
& ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
;
1800 pipe
.DCFlushEnable
= bits
& ANV_PIPE_DATA_CACHE_FLUSH_BIT
;
1801 pipe
.RenderTargetCacheFlushEnable
=
1802 bits
& ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
;
1804 pipe
.DepthStallEnable
= bits
& ANV_PIPE_DEPTH_STALL_BIT
;
1805 pipe
.CommandStreamerStallEnable
= bits
& ANV_PIPE_CS_STALL_BIT
;
1806 pipe
.StallAtPixelScoreboard
= bits
& ANV_PIPE_STALL_AT_SCOREBOARD_BIT
;
1809 * According to the Broadwell documentation, any PIPE_CONTROL with the
1810 * "Command Streamer Stall" bit set must also have another bit set,
1811 * with five different options:
1813 * - Render Target Cache Flush
1814 * - Depth Cache Flush
1815 * - Stall at Pixel Scoreboard
1816 * - Post-Sync Operation
1820 * I chose "Stall at Pixel Scoreboard" since that's what we use in
1821 * mesa and it seems to work fine. The choice is fairly arbitrary.
1823 if ((bits
& ANV_PIPE_CS_STALL_BIT
) &&
1824 !(bits
& (ANV_PIPE_FLUSH_BITS
| ANV_PIPE_DEPTH_STALL_BIT
|
1825 ANV_PIPE_STALL_AT_SCOREBOARD_BIT
)))
1826 pipe
.StallAtPixelScoreboard
= true;
1829 /* If a render target flush was emitted, then we can toggle off the bit
1830 * saying that render target writes are ongoing.
1832 if (bits
& ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
)
1833 bits
&= ~(ANV_PIPE_RENDER_TARGET_BUFFER_WRITES
);
1835 bits
&= ~(ANV_PIPE_FLUSH_BITS
| ANV_PIPE_CS_STALL_BIT
);
1838 if (bits
& ANV_PIPE_INVALIDATE_BITS
) {
1839 /* From the SKL PRM, Vol. 2a, "PIPE_CONTROL",
1841 * "If the VF Cache Invalidation Enable is set to a 1 in a
1842 * PIPE_CONTROL, a separate Null PIPE_CONTROL, all bitfields sets to
1843 * 0, with the VF Cache Invalidation Enable set to 0 needs to be sent
1844 * prior to the PIPE_CONTROL with VF Cache Invalidation Enable set to
1847 * This appears to hang Broadwell, so we restrict it to just gen9.
1849 if (GEN_GEN
== 9 && (bits
& ANV_PIPE_VF_CACHE_INVALIDATE_BIT
))
1850 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
);
1852 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
1853 pipe
.StateCacheInvalidationEnable
=
1854 bits
& ANV_PIPE_STATE_CACHE_INVALIDATE_BIT
;
1855 pipe
.ConstantCacheInvalidationEnable
=
1856 bits
& ANV_PIPE_CONSTANT_CACHE_INVALIDATE_BIT
;
1857 pipe
.VFCacheInvalidationEnable
=
1858 bits
& ANV_PIPE_VF_CACHE_INVALIDATE_BIT
;
1859 pipe
.TextureCacheInvalidationEnable
=
1860 bits
& ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
;
1861 pipe
.InstructionCacheInvalidateEnable
=
1862 bits
& ANV_PIPE_INSTRUCTION_CACHE_INVALIDATE_BIT
;
1864 /* From the SKL PRM, Vol. 2a, "PIPE_CONTROL",
1866 * "When VF Cache Invalidate is set “Post Sync Operation” must be
1867 * enabled to “Write Immediate Data” or “Write PS Depth Count” or
1868 * “Write Timestamp”.
1870 if (GEN_GEN
== 9 && pipe
.VFCacheInvalidationEnable
) {
1871 pipe
.PostSyncOperation
= WriteImmediateData
;
1873 (struct anv_address
) { &cmd_buffer
->device
->workaround_bo
, 0 };
1877 bits
&= ~ANV_PIPE_INVALIDATE_BITS
;
1880 cmd_buffer
->state
.pending_pipe_bits
= bits
;
1883 void genX(CmdPipelineBarrier
)(
1884 VkCommandBuffer commandBuffer
,
1885 VkPipelineStageFlags srcStageMask
,
1886 VkPipelineStageFlags destStageMask
,
1888 uint32_t memoryBarrierCount
,
1889 const VkMemoryBarrier
* pMemoryBarriers
,
1890 uint32_t bufferMemoryBarrierCount
,
1891 const VkBufferMemoryBarrier
* pBufferMemoryBarriers
,
1892 uint32_t imageMemoryBarrierCount
,
1893 const VkImageMemoryBarrier
* pImageMemoryBarriers
)
1895 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
1897 /* XXX: Right now, we're really dumb and just flush whatever categories
1898 * the app asks for. One of these days we may make this a bit better
1899 * but right now that's all the hardware allows for in most areas.
1901 VkAccessFlags src_flags
= 0;
1902 VkAccessFlags dst_flags
= 0;
1904 for (uint32_t i
= 0; i
< memoryBarrierCount
; i
++) {
1905 src_flags
|= pMemoryBarriers
[i
].srcAccessMask
;
1906 dst_flags
|= pMemoryBarriers
[i
].dstAccessMask
;
1909 for (uint32_t i
= 0; i
< bufferMemoryBarrierCount
; i
++) {
1910 src_flags
|= pBufferMemoryBarriers
[i
].srcAccessMask
;
1911 dst_flags
|= pBufferMemoryBarriers
[i
].dstAccessMask
;
1914 for (uint32_t i
= 0; i
< imageMemoryBarrierCount
; i
++) {
1915 src_flags
|= pImageMemoryBarriers
[i
].srcAccessMask
;
1916 dst_flags
|= pImageMemoryBarriers
[i
].dstAccessMask
;
1917 ANV_FROM_HANDLE(anv_image
, image
, pImageMemoryBarriers
[i
].image
);
1918 const VkImageSubresourceRange
*range
=
1919 &pImageMemoryBarriers
[i
].subresourceRange
;
1921 uint32_t base_layer
, layer_count
;
1922 if (image
->type
== VK_IMAGE_TYPE_3D
) {
1924 layer_count
= anv_minify(image
->extent
.depth
, range
->baseMipLevel
);
1926 base_layer
= range
->baseArrayLayer
;
1927 layer_count
= anv_get_layerCount(image
, range
);
1930 if (range
->aspectMask
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
1931 transition_depth_buffer(cmd_buffer
, image
,
1932 pImageMemoryBarriers
[i
].oldLayout
,
1933 pImageMemoryBarriers
[i
].newLayout
);
1936 if (range
->aspectMask
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
1937 transition_stencil_buffer(cmd_buffer
, image
,
1938 range
->baseMipLevel
,
1939 anv_get_levelCount(image
, range
),
1940 base_layer
, layer_count
,
1941 pImageMemoryBarriers
[i
].oldLayout
,
1942 pImageMemoryBarriers
[i
].newLayout
);
1945 if (range
->aspectMask
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
1946 VkImageAspectFlags color_aspects
=
1947 anv_image_expand_aspects(image
, range
->aspectMask
);
1948 uint32_t aspect_bit
;
1949 anv_foreach_image_aspect_bit(aspect_bit
, image
, color_aspects
) {
1950 transition_color_buffer(cmd_buffer
, image
, 1UL << aspect_bit
,
1951 range
->baseMipLevel
,
1952 anv_get_levelCount(image
, range
),
1953 base_layer
, layer_count
,
1954 pImageMemoryBarriers
[i
].oldLayout
,
1955 pImageMemoryBarriers
[i
].newLayout
);
1960 cmd_buffer
->state
.pending_pipe_bits
|=
1961 anv_pipe_flush_bits_for_access_flags(src_flags
) |
1962 anv_pipe_invalidate_bits_for_access_flags(dst_flags
);
1966 cmd_buffer_alloc_push_constants(struct anv_cmd_buffer
*cmd_buffer
)
1968 VkShaderStageFlags stages
=
1969 cmd_buffer
->state
.gfx
.base
.pipeline
->active_stages
;
1971 /* In order to avoid thrash, we assume that vertex and fragment stages
1972 * always exist. In the rare case where one is missing *and* the other
1973 * uses push concstants, this may be suboptimal. However, avoiding stalls
1974 * seems more important.
1976 stages
|= VK_SHADER_STAGE_FRAGMENT_BIT
| VK_SHADER_STAGE_VERTEX_BIT
;
1978 if (stages
== cmd_buffer
->state
.push_constant_stages
)
1982 const unsigned push_constant_kb
= 32;
1983 #elif GEN_IS_HASWELL
1984 const unsigned push_constant_kb
= cmd_buffer
->device
->info
.gt
== 3 ? 32 : 16;
1986 const unsigned push_constant_kb
= 16;
1989 const unsigned num_stages
=
1990 util_bitcount(stages
& VK_SHADER_STAGE_ALL_GRAPHICS
);
1991 unsigned size_per_stage
= push_constant_kb
/ num_stages
;
1993 /* Broadwell+ and Haswell gt3 require that the push constant sizes be in
1994 * units of 2KB. Incidentally, these are the same platforms that have
1995 * 32KB worth of push constant space.
1997 if (push_constant_kb
== 32)
1998 size_per_stage
&= ~1u;
2000 uint32_t kb_used
= 0;
2001 for (int i
= MESA_SHADER_VERTEX
; i
< MESA_SHADER_FRAGMENT
; i
++) {
2002 unsigned push_size
= (stages
& (1 << i
)) ? size_per_stage
: 0;
2003 anv_batch_emit(&cmd_buffer
->batch
,
2004 GENX(3DSTATE_PUSH_CONSTANT_ALLOC_VS
), alloc
) {
2005 alloc
._3DCommandSubOpcode
= 18 + i
;
2006 alloc
.ConstantBufferOffset
= (push_size
> 0) ? kb_used
: 0;
2007 alloc
.ConstantBufferSize
= push_size
;
2009 kb_used
+= push_size
;
2012 anv_batch_emit(&cmd_buffer
->batch
,
2013 GENX(3DSTATE_PUSH_CONSTANT_ALLOC_PS
), alloc
) {
2014 alloc
.ConstantBufferOffset
= kb_used
;
2015 alloc
.ConstantBufferSize
= push_constant_kb
- kb_used
;
2018 cmd_buffer
->state
.push_constant_stages
= stages
;
2020 /* From the BDW PRM for 3DSTATE_PUSH_CONSTANT_ALLOC_VS:
2022 * "The 3DSTATE_CONSTANT_VS must be reprogrammed prior to
2023 * the next 3DPRIMITIVE command after programming the
2024 * 3DSTATE_PUSH_CONSTANT_ALLOC_VS"
2026 * Since 3DSTATE_PUSH_CONSTANT_ALLOC_VS is programmed as part of
2027 * pipeline setup, we need to dirty push constants.
2029 cmd_buffer
->state
.push_constants_dirty
|= VK_SHADER_STAGE_ALL_GRAPHICS
;
2032 static const struct anv_descriptor
*
2033 anv_descriptor_for_binding(const struct anv_cmd_pipeline_state
*pipe_state
,
2034 const struct anv_pipeline_binding
*binding
)
2036 assert(binding
->set
< MAX_SETS
);
2037 const struct anv_descriptor_set
*set
=
2038 pipe_state
->descriptors
[binding
->set
];
2039 const uint32_t offset
=
2040 set
->layout
->binding
[binding
->binding
].descriptor_index
;
2041 return &set
->descriptors
[offset
+ binding
->index
];
2045 dynamic_offset_for_binding(const struct anv_cmd_pipeline_state
*pipe_state
,
2046 const struct anv_pipeline_binding
*binding
)
2048 assert(binding
->set
< MAX_SETS
);
2049 const struct anv_descriptor_set
*set
=
2050 pipe_state
->descriptors
[binding
->set
];
2052 uint32_t dynamic_offset_idx
=
2053 pipe_state
->layout
->set
[binding
->set
].dynamic_offset_start
+
2054 set
->layout
->binding
[binding
->binding
].dynamic_offset_index
+
2057 return pipe_state
->dynamic_offsets
[dynamic_offset_idx
];
2060 static struct anv_address
2061 anv_descriptor_set_address(struct anv_cmd_buffer
*cmd_buffer
,
2062 struct anv_descriptor_set
*set
)
2065 /* This is a normal descriptor set */
2066 return (struct anv_address
) {
2067 .bo
= &set
->pool
->bo
,
2068 .offset
= set
->desc_mem
.offset
,
2071 /* This is a push descriptor set. We have to flag it as used on the GPU
2072 * so that the next time we push descriptors, we grab a new memory.
2074 struct anv_push_descriptor_set
*push_set
=
2075 (struct anv_push_descriptor_set
*)set
;
2076 push_set
->set_used_on_gpu
= true;
2078 return (struct anv_address
) {
2079 .bo
= cmd_buffer
->dynamic_state_stream
.state_pool
->block_pool
.bo
,
2080 .offset
= set
->desc_mem
.offset
,
2086 emit_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
2087 gl_shader_stage stage
,
2088 struct anv_state
*bt_state
)
2090 struct anv_subpass
*subpass
= cmd_buffer
->state
.subpass
;
2091 struct anv_cmd_pipeline_state
*pipe_state
;
2092 struct anv_pipeline
*pipeline
;
2093 uint32_t state_offset
;
2096 case MESA_SHADER_COMPUTE
:
2097 pipe_state
= &cmd_buffer
->state
.compute
.base
;
2100 pipe_state
= &cmd_buffer
->state
.gfx
.base
;
2103 pipeline
= pipe_state
->pipeline
;
2105 if (!anv_pipeline_has_stage(pipeline
, stage
)) {
2106 *bt_state
= (struct anv_state
) { 0, };
2110 struct anv_pipeline_bind_map
*map
= &pipeline
->shaders
[stage
]->bind_map
;
2111 if (map
->surface_count
== 0) {
2112 *bt_state
= (struct anv_state
) { 0, };
2116 *bt_state
= anv_cmd_buffer_alloc_binding_table(cmd_buffer
,
2119 uint32_t *bt_map
= bt_state
->map
;
2121 if (bt_state
->map
== NULL
)
2122 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2124 /* We only need to emit relocs if we're not using softpin. If we are using
2125 * softpin then we always keep all user-allocated memory objects resident.
2127 const bool need_client_mem_relocs
=
2128 !cmd_buffer
->device
->instance
->physicalDevice
.use_softpin
;
2130 for (uint32_t s
= 0; s
< map
->surface_count
; s
++) {
2131 struct anv_pipeline_binding
*binding
= &map
->surface_to_descriptor
[s
];
2133 struct anv_state surface_state
;
2135 if (binding
->set
== ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS
) {
2136 /* Color attachment binding */
2137 assert(stage
== MESA_SHADER_FRAGMENT
);
2138 assert(binding
->binding
== 0);
2139 if (binding
->index
< subpass
->color_count
) {
2140 const unsigned att
=
2141 subpass
->color_attachments
[binding
->index
].attachment
;
2143 /* From the Vulkan 1.0.46 spec:
2145 * "If any color or depth/stencil attachments are
2146 * VK_ATTACHMENT_UNUSED, then no writes occur for those
2149 if (att
== VK_ATTACHMENT_UNUSED
) {
2150 surface_state
= cmd_buffer
->state
.null_surface_state
;
2152 surface_state
= cmd_buffer
->state
.attachments
[att
].color
.state
;
2155 surface_state
= cmd_buffer
->state
.null_surface_state
;
2158 bt_map
[s
] = surface_state
.offset
+ state_offset
;
2160 } else if (binding
->set
== ANV_DESCRIPTOR_SET_SHADER_CONSTANTS
) {
2161 struct anv_state surface_state
=
2162 anv_cmd_buffer_alloc_surface_state(cmd_buffer
);
2164 struct anv_address constant_data
= {
2165 .bo
= pipeline
->device
->dynamic_state_pool
.block_pool
.bo
,
2166 .offset
= pipeline
->shaders
[stage
]->constant_data
.offset
,
2168 unsigned constant_data_size
=
2169 pipeline
->shaders
[stage
]->constant_data_size
;
2171 const enum isl_format format
=
2172 anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
);
2173 anv_fill_buffer_surface_state(cmd_buffer
->device
,
2174 surface_state
, format
,
2175 constant_data
, constant_data_size
, 1);
2177 bt_map
[s
] = surface_state
.offset
+ state_offset
;
2178 add_surface_reloc(cmd_buffer
, surface_state
, constant_data
);
2180 } else if (binding
->set
== ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS
) {
2181 /* This is always the first binding for compute shaders */
2182 assert(stage
== MESA_SHADER_COMPUTE
&& s
== 0);
2183 if (!get_cs_prog_data(pipeline
)->uses_num_work_groups
)
2186 struct anv_state surface_state
=
2187 anv_cmd_buffer_alloc_surface_state(cmd_buffer
);
2189 const enum isl_format format
=
2190 anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
);
2191 anv_fill_buffer_surface_state(cmd_buffer
->device
, surface_state
,
2193 cmd_buffer
->state
.compute
.num_workgroups
,
2195 bt_map
[s
] = surface_state
.offset
+ state_offset
;
2196 if (need_client_mem_relocs
) {
2197 add_surface_reloc(cmd_buffer
, surface_state
,
2198 cmd_buffer
->state
.compute
.num_workgroups
);
2201 } else if (binding
->set
== ANV_DESCRIPTOR_SET_DESCRIPTORS
) {
2202 /* This is a descriptor set buffer so the set index is actually
2203 * given by binding->binding. (Yes, that's confusing.)
2205 struct anv_descriptor_set
*set
=
2206 pipe_state
->descriptors
[binding
->binding
];
2207 assert(set
->desc_mem
.alloc_size
);
2208 assert(set
->desc_surface_state
.alloc_size
);
2209 bt_map
[s
] = set
->desc_surface_state
.offset
+ state_offset
;
2210 add_surface_reloc(cmd_buffer
, set
->desc_surface_state
,
2211 anv_descriptor_set_address(cmd_buffer
, set
));
2215 const struct anv_descriptor
*desc
=
2216 anv_descriptor_for_binding(pipe_state
, binding
);
2218 switch (desc
->type
) {
2219 case VK_DESCRIPTOR_TYPE_SAMPLER
:
2220 /* Nothing for us to do here */
2223 case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
:
2224 case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE
: {
2225 struct anv_surface_state sstate
=
2226 (desc
->layout
== VK_IMAGE_LAYOUT_GENERAL
) ?
2227 desc
->image_view
->planes
[binding
->plane
].general_sampler_surface_state
:
2228 desc
->image_view
->planes
[binding
->plane
].optimal_sampler_surface_state
;
2229 surface_state
= sstate
.state
;
2230 assert(surface_state
.alloc_size
);
2231 if (need_client_mem_relocs
)
2232 add_surface_state_relocs(cmd_buffer
, sstate
);
2235 case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT
:
2236 assert(stage
== MESA_SHADER_FRAGMENT
);
2237 if ((desc
->image_view
->aspect_mask
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) == 0) {
2238 /* For depth and stencil input attachments, we treat it like any
2239 * old texture that a user may have bound.
2241 struct anv_surface_state sstate
=
2242 (desc
->layout
== VK_IMAGE_LAYOUT_GENERAL
) ?
2243 desc
->image_view
->planes
[binding
->plane
].general_sampler_surface_state
:
2244 desc
->image_view
->planes
[binding
->plane
].optimal_sampler_surface_state
;
2245 surface_state
= sstate
.state
;
2246 assert(surface_state
.alloc_size
);
2247 if (need_client_mem_relocs
)
2248 add_surface_state_relocs(cmd_buffer
, sstate
);
2250 /* For color input attachments, we create the surface state at
2251 * vkBeginRenderPass time so that we can include aux and clear
2252 * color information.
2254 assert(binding
->input_attachment_index
< subpass
->input_count
);
2255 const unsigned subpass_att
= binding
->input_attachment_index
;
2256 const unsigned att
= subpass
->input_attachments
[subpass_att
].attachment
;
2257 surface_state
= cmd_buffer
->state
.attachments
[att
].input
.state
;
2261 case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE
: {
2262 struct anv_surface_state sstate
= (binding
->write_only
)
2263 ? desc
->image_view
->planes
[binding
->plane
].writeonly_storage_surface_state
2264 : desc
->image_view
->planes
[binding
->plane
].storage_surface_state
;
2265 surface_state
= sstate
.state
;
2266 assert(surface_state
.alloc_size
);
2267 if (need_client_mem_relocs
)
2268 add_surface_state_relocs(cmd_buffer
, sstate
);
2272 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
:
2273 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER
:
2274 case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER
:
2275 surface_state
= desc
->buffer_view
->surface_state
;
2276 assert(surface_state
.alloc_size
);
2277 if (need_client_mem_relocs
) {
2278 add_surface_reloc(cmd_buffer
, surface_state
,
2279 desc
->buffer_view
->address
);
2283 case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
:
2284 case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC
: {
2285 /* Compute the offset within the buffer */
2286 uint32_t dynamic_offset
=
2287 dynamic_offset_for_binding(pipe_state
, binding
);
2288 uint64_t offset
= desc
->offset
+ dynamic_offset
;
2289 /* Clamp to the buffer size */
2290 offset
= MIN2(offset
, desc
->buffer
->size
);
2291 /* Clamp the range to the buffer size */
2292 uint32_t range
= MIN2(desc
->range
, desc
->buffer
->size
- offset
);
2294 struct anv_address address
=
2295 anv_address_add(desc
->buffer
->address
, offset
);
2298 anv_state_stream_alloc(&cmd_buffer
->surface_state_stream
, 64, 64);
2299 enum isl_format format
=
2300 anv_isl_format_for_descriptor_type(desc
->type
);
2302 anv_fill_buffer_surface_state(cmd_buffer
->device
, surface_state
,
2303 format
, address
, range
, 1);
2304 if (need_client_mem_relocs
)
2305 add_surface_reloc(cmd_buffer
, surface_state
, address
);
2309 case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER
:
2310 surface_state
= (binding
->write_only
)
2311 ? desc
->buffer_view
->writeonly_storage_surface_state
2312 : desc
->buffer_view
->storage_surface_state
;
2313 assert(surface_state
.alloc_size
);
2314 if (need_client_mem_relocs
) {
2315 add_surface_reloc(cmd_buffer
, surface_state
,
2316 desc
->buffer_view
->address
);
2321 assert(!"Invalid descriptor type");
2325 bt_map
[s
] = surface_state
.offset
+ state_offset
;
2329 /* The PIPE_CONTROL command description says:
2331 * "Whenever a Binding Table Index (BTI) used by a Render Taget Message
2332 * points to a different RENDER_SURFACE_STATE, SW must issue a Render
2333 * Target Cache Flush by enabling this bit. When render target flush
2334 * is set due to new association of BTI, PS Scoreboard Stall bit must
2335 * be set in this packet."
2337 * FINISHME: Currently we shuffle around the surface states in the binding
2338 * table based on if they are getting used or not. So, we've to do below
2339 * pipe control flush for every binding table upload. Make changes so
2340 * that we do it only when we modify render target surface states.
2342 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
2343 pc
.RenderTargetCacheFlushEnable
= true;
2344 pc
.StallAtPixelScoreboard
= true;
2352 emit_samplers(struct anv_cmd_buffer
*cmd_buffer
,
2353 gl_shader_stage stage
,
2354 struct anv_state
*state
)
2356 struct anv_cmd_pipeline_state
*pipe_state
=
2357 stage
== MESA_SHADER_COMPUTE
? &cmd_buffer
->state
.compute
.base
:
2358 &cmd_buffer
->state
.gfx
.base
;
2359 struct anv_pipeline
*pipeline
= pipe_state
->pipeline
;
2361 if (!anv_pipeline_has_stage(pipeline
, stage
)) {
2362 *state
= (struct anv_state
) { 0, };
2366 struct anv_pipeline_bind_map
*map
= &pipeline
->shaders
[stage
]->bind_map
;
2367 if (map
->sampler_count
== 0) {
2368 *state
= (struct anv_state
) { 0, };
2372 uint32_t size
= map
->sampler_count
* 16;
2373 *state
= anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, size
, 32);
2375 if (state
->map
== NULL
)
2376 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
2378 for (uint32_t s
= 0; s
< map
->sampler_count
; s
++) {
2379 struct anv_pipeline_binding
*binding
= &map
->sampler_to_descriptor
[s
];
2380 const struct anv_descriptor
*desc
=
2381 anv_descriptor_for_binding(pipe_state
, binding
);
2383 if (desc
->type
!= VK_DESCRIPTOR_TYPE_SAMPLER
&&
2384 desc
->type
!= VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER
)
2387 struct anv_sampler
*sampler
= desc
->sampler
;
2389 /* This can happen if we have an unfilled slot since TYPE_SAMPLER
2390 * happens to be zero.
2392 if (sampler
== NULL
)
2395 memcpy(state
->map
+ (s
* 16),
2396 sampler
->state
[binding
->plane
], sizeof(sampler
->state
[0]));
2403 flush_descriptor_sets(struct anv_cmd_buffer
*cmd_buffer
)
2405 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
2407 VkShaderStageFlags dirty
= cmd_buffer
->state
.descriptors_dirty
&
2408 pipeline
->active_stages
;
2410 VkResult result
= VK_SUCCESS
;
2411 anv_foreach_stage(s
, dirty
) {
2412 result
= emit_samplers(cmd_buffer
, s
, &cmd_buffer
->state
.samplers
[s
]);
2413 if (result
!= VK_SUCCESS
)
2415 result
= emit_binding_table(cmd_buffer
, s
,
2416 &cmd_buffer
->state
.binding_tables
[s
]);
2417 if (result
!= VK_SUCCESS
)
2421 if (result
!= VK_SUCCESS
) {
2422 assert(result
== VK_ERROR_OUT_OF_DEVICE_MEMORY
);
2424 result
= anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
2425 if (result
!= VK_SUCCESS
)
2428 /* Re-emit state base addresses so we get the new surface state base
2429 * address before we start emitting binding tables etc.
2431 genX(cmd_buffer_emit_state_base_address
)(cmd_buffer
);
2433 /* Re-emit all active binding tables */
2434 dirty
|= pipeline
->active_stages
;
2435 anv_foreach_stage(s
, dirty
) {
2436 result
= emit_samplers(cmd_buffer
, s
, &cmd_buffer
->state
.samplers
[s
]);
2437 if (result
!= VK_SUCCESS
) {
2438 anv_batch_set_error(&cmd_buffer
->batch
, result
);
2441 result
= emit_binding_table(cmd_buffer
, s
,
2442 &cmd_buffer
->state
.binding_tables
[s
]);
2443 if (result
!= VK_SUCCESS
) {
2444 anv_batch_set_error(&cmd_buffer
->batch
, result
);
2450 cmd_buffer
->state
.descriptors_dirty
&= ~dirty
;
2456 cmd_buffer_emit_descriptor_pointers(struct anv_cmd_buffer
*cmd_buffer
,
2459 static const uint32_t sampler_state_opcodes
[] = {
2460 [MESA_SHADER_VERTEX
] = 43,
2461 [MESA_SHADER_TESS_CTRL
] = 44, /* HS */
2462 [MESA_SHADER_TESS_EVAL
] = 45, /* DS */
2463 [MESA_SHADER_GEOMETRY
] = 46,
2464 [MESA_SHADER_FRAGMENT
] = 47,
2465 [MESA_SHADER_COMPUTE
] = 0,
2468 static const uint32_t binding_table_opcodes
[] = {
2469 [MESA_SHADER_VERTEX
] = 38,
2470 [MESA_SHADER_TESS_CTRL
] = 39,
2471 [MESA_SHADER_TESS_EVAL
] = 40,
2472 [MESA_SHADER_GEOMETRY
] = 41,
2473 [MESA_SHADER_FRAGMENT
] = 42,
2474 [MESA_SHADER_COMPUTE
] = 0,
2477 anv_foreach_stage(s
, stages
) {
2478 assert(s
< ARRAY_SIZE(binding_table_opcodes
));
2479 assert(binding_table_opcodes
[s
] > 0);
2481 if (cmd_buffer
->state
.samplers
[s
].alloc_size
> 0) {
2482 anv_batch_emit(&cmd_buffer
->batch
,
2483 GENX(3DSTATE_SAMPLER_STATE_POINTERS_VS
), ssp
) {
2484 ssp
._3DCommandSubOpcode
= sampler_state_opcodes
[s
];
2485 ssp
.PointertoVSSamplerState
= cmd_buffer
->state
.samplers
[s
].offset
;
2489 /* Always emit binding table pointers if we're asked to, since on SKL
2490 * this is what flushes push constants. */
2491 anv_batch_emit(&cmd_buffer
->batch
,
2492 GENX(3DSTATE_BINDING_TABLE_POINTERS_VS
), btp
) {
2493 btp
._3DCommandSubOpcode
= binding_table_opcodes
[s
];
2494 btp
.PointertoVSBindingTable
= cmd_buffer
->state
.binding_tables
[s
].offset
;
2500 cmd_buffer_flush_push_constants(struct anv_cmd_buffer
*cmd_buffer
,
2501 VkShaderStageFlags dirty_stages
)
2503 const struct anv_cmd_graphics_state
*gfx_state
= &cmd_buffer
->state
.gfx
;
2504 const struct anv_pipeline
*pipeline
= gfx_state
->base
.pipeline
;
2506 static const uint32_t push_constant_opcodes
[] = {
2507 [MESA_SHADER_VERTEX
] = 21,
2508 [MESA_SHADER_TESS_CTRL
] = 25, /* HS */
2509 [MESA_SHADER_TESS_EVAL
] = 26, /* DS */
2510 [MESA_SHADER_GEOMETRY
] = 22,
2511 [MESA_SHADER_FRAGMENT
] = 23,
2512 [MESA_SHADER_COMPUTE
] = 0,
2515 VkShaderStageFlags flushed
= 0;
2517 anv_foreach_stage(stage
, dirty_stages
) {
2518 assert(stage
< ARRAY_SIZE(push_constant_opcodes
));
2519 assert(push_constant_opcodes
[stage
] > 0);
2521 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CONSTANT_VS
), c
) {
2522 c
._3DCommandSubOpcode
= push_constant_opcodes
[stage
];
2524 if (anv_pipeline_has_stage(pipeline
, stage
)) {
2525 #if GEN_GEN >= 8 || GEN_IS_HASWELL
2526 const struct brw_stage_prog_data
*prog_data
=
2527 pipeline
->shaders
[stage
]->prog_data
;
2528 const struct anv_pipeline_bind_map
*bind_map
=
2529 &pipeline
->shaders
[stage
]->bind_map
;
2531 /* The Skylake PRM contains the following restriction:
2533 * "The driver must ensure The following case does not occur
2534 * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
2535 * buffer 3 read length equal to zero committed followed by a
2536 * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to
2539 * To avoid this, we program the buffers in the highest slots.
2540 * This way, slot 0 is only used if slot 3 is also used.
2544 for (int i
= 3; i
>= 0; i
--) {
2545 const struct brw_ubo_range
*range
= &prog_data
->ubo_ranges
[i
];
2546 if (range
->length
== 0)
2549 const unsigned surface
=
2550 prog_data
->binding_table
.ubo_start
+ range
->block
;
2552 assert(surface
<= bind_map
->surface_count
);
2553 const struct anv_pipeline_binding
*binding
=
2554 &bind_map
->surface_to_descriptor
[surface
];
2556 struct anv_address read_addr
;
2558 if (binding
->set
== ANV_DESCRIPTOR_SET_SHADER_CONSTANTS
) {
2559 struct anv_address constant_data
= {
2560 .bo
= pipeline
->device
->dynamic_state_pool
.block_pool
.bo
,
2561 .offset
= pipeline
->shaders
[stage
]->constant_data
.offset
,
2563 unsigned constant_data_size
=
2564 pipeline
->shaders
[stage
]->constant_data_size
;
2566 read_len
= MIN2(range
->length
,
2567 DIV_ROUND_UP(constant_data_size
, 32) - range
->start
);
2568 read_addr
= anv_address_add(constant_data
,
2570 } else if (binding
->set
== ANV_DESCRIPTOR_SET_DESCRIPTORS
) {
2571 /* This is a descriptor set buffer so the set index is
2572 * actually given by binding->binding. (Yes, that's
2575 struct anv_descriptor_set
*set
=
2576 gfx_state
->base
.descriptors
[binding
->binding
];
2577 struct anv_address desc_buffer_addr
=
2578 anv_descriptor_set_address(cmd_buffer
, set
);
2579 const unsigned desc_buffer_size
= set
->desc_mem
.alloc_size
;
2581 read_len
= MIN2(range
->length
,
2582 DIV_ROUND_UP(desc_buffer_size
, 32) - range
->start
);
2583 read_addr
= anv_address_add(desc_buffer_addr
,
2586 const struct anv_descriptor
*desc
=
2587 anv_descriptor_for_binding(&gfx_state
->base
, binding
);
2589 if (desc
->type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER
) {
2590 read_len
= MIN2(range
->length
,
2591 DIV_ROUND_UP(desc
->buffer_view
->range
, 32) - range
->start
);
2592 read_addr
= anv_address_add(desc
->buffer_view
->address
,
2595 assert(desc
->type
== VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC
);
2597 uint32_t dynamic_offset
=
2598 dynamic_offset_for_binding(&gfx_state
->base
, binding
);
2599 uint32_t buf_offset
=
2600 MIN2(desc
->offset
+ dynamic_offset
, desc
->buffer
->size
);
2601 uint32_t buf_range
=
2602 MIN2(desc
->range
, desc
->buffer
->size
- buf_offset
);
2604 read_len
= MIN2(range
->length
,
2605 DIV_ROUND_UP(buf_range
, 32) - range
->start
);
2606 read_addr
= anv_address_add(desc
->buffer
->address
,
2607 buf_offset
+ range
->start
* 32);
2612 c
.ConstantBody
.Buffer
[n
] = read_addr
;
2613 c
.ConstantBody
.ReadLength
[n
] = read_len
;
2618 struct anv_state state
=
2619 anv_cmd_buffer_push_constants(cmd_buffer
, stage
);
2621 if (state
.alloc_size
> 0) {
2622 c
.ConstantBody
.Buffer
[n
] = (struct anv_address
) {
2623 .bo
= cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
2624 .offset
= state
.offset
,
2626 c
.ConstantBody
.ReadLength
[n
] =
2627 DIV_ROUND_UP(state
.alloc_size
, 32);
2630 /* For Ivy Bridge, the push constants packets have a different
2631 * rule that would require us to iterate in the other direction
2632 * and possibly mess around with dynamic state base address.
2633 * Don't bother; just emit regular push constants at n = 0.
2635 struct anv_state state
=
2636 anv_cmd_buffer_push_constants(cmd_buffer
, stage
);
2638 if (state
.alloc_size
> 0) {
2639 c
.ConstantBody
.Buffer
[0].offset
= state
.offset
,
2640 c
.ConstantBody
.ReadLength
[0] =
2641 DIV_ROUND_UP(state
.alloc_size
, 32);
2647 flushed
|= mesa_to_vk_shader_stage(stage
);
2650 cmd_buffer
->state
.push_constants_dirty
&= ~flushed
;
2654 genX(cmd_buffer_flush_state
)(struct anv_cmd_buffer
*cmd_buffer
)
2656 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
2659 uint32_t vb_emit
= cmd_buffer
->state
.gfx
.vb_dirty
& pipeline
->vb_used
;
2660 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_PIPELINE
)
2661 vb_emit
|= pipeline
->vb_used
;
2663 assert((pipeline
->active_stages
& VK_SHADER_STAGE_COMPUTE_BIT
) == 0);
2665 genX(cmd_buffer_config_l3
)(cmd_buffer
, pipeline
->urb
.l3_config
);
2667 genX(cmd_buffer_emit_hashing_mode
)(cmd_buffer
, UINT_MAX
, UINT_MAX
, 1);
2669 genX(flush_pipeline_select_3d
)(cmd_buffer
);
2672 const uint32_t num_buffers
= __builtin_popcount(vb_emit
);
2673 const uint32_t num_dwords
= 1 + num_buffers
* 4;
2675 p
= anv_batch_emitn(&cmd_buffer
->batch
, num_dwords
,
2676 GENX(3DSTATE_VERTEX_BUFFERS
));
2678 for_each_bit(vb
, vb_emit
) {
2679 struct anv_buffer
*buffer
= cmd_buffer
->state
.vertex_bindings
[vb
].buffer
;
2680 uint32_t offset
= cmd_buffer
->state
.vertex_bindings
[vb
].offset
;
2682 struct GENX(VERTEX_BUFFER_STATE
) state
= {
2683 .VertexBufferIndex
= vb
,
2685 .MOCS
= anv_mocs_for_bo(cmd_buffer
->device
, buffer
->address
.bo
),
2687 .BufferAccessType
= pipeline
->vb
[vb
].instanced
? INSTANCEDATA
: VERTEXDATA
,
2688 .InstanceDataStepRate
= pipeline
->vb
[vb
].instance_divisor
,
2691 .AddressModifyEnable
= true,
2692 .BufferPitch
= pipeline
->vb
[vb
].stride
,
2693 .BufferStartingAddress
= anv_address_add(buffer
->address
, offset
),
2696 .BufferSize
= buffer
->size
- offset
2698 .EndAddress
= anv_address_add(buffer
->address
, buffer
->size
- 1),
2702 GENX(VERTEX_BUFFER_STATE_pack
)(&cmd_buffer
->batch
, &p
[1 + i
* 4], &state
);
2707 cmd_buffer
->state
.gfx
.vb_dirty
&= ~vb_emit
;
2710 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_XFB_ENABLE
) {
2711 /* We don't need any per-buffer dirty tracking because you're not
2712 * allowed to bind different XFB buffers while XFB is enabled.
2714 for (unsigned idx
= 0; idx
< MAX_XFB_BUFFERS
; idx
++) {
2715 struct anv_xfb_binding
*xfb
= &cmd_buffer
->state
.xfb_bindings
[idx
];
2716 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_SO_BUFFER
), sob
) {
2717 sob
.SOBufferIndex
= idx
;
2719 if (cmd_buffer
->state
.xfb_enabled
&& xfb
->buffer
&& xfb
->size
!= 0) {
2720 sob
.SOBufferEnable
= true;
2721 sob
.MOCS
= cmd_buffer
->device
->default_mocs
,
2722 sob
.StreamOffsetWriteEnable
= false;
2723 sob
.SurfaceBaseAddress
= anv_address_add(xfb
->buffer
->address
,
2725 /* Size is in DWords - 1 */
2726 sob
.SurfaceSize
= xfb
->size
/ 4 - 1;
2731 /* CNL and later require a CS stall after 3DSTATE_SO_BUFFER */
2733 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
2737 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_PIPELINE
) {
2738 anv_batch_emit_batch(&cmd_buffer
->batch
, &pipeline
->batch
);
2740 /* The exact descriptor layout is pulled from the pipeline, so we need
2741 * to re-emit binding tables on every pipeline change.
2743 cmd_buffer
->state
.descriptors_dirty
|= pipeline
->active_stages
;
2745 /* If the pipeline changed, we may need to re-allocate push constant
2748 cmd_buffer_alloc_push_constants(cmd_buffer
);
2752 if (cmd_buffer
->state
.descriptors_dirty
& VK_SHADER_STAGE_VERTEX_BIT
||
2753 cmd_buffer
->state
.push_constants_dirty
& VK_SHADER_STAGE_VERTEX_BIT
) {
2754 /* From the IVB PRM Vol. 2, Part 1, Section 3.2.1:
2756 * "A PIPE_CONTROL with Post-Sync Operation set to 1h and a depth
2757 * stall needs to be sent just prior to any 3DSTATE_VS,
2758 * 3DSTATE_URB_VS, 3DSTATE_CONSTANT_VS,
2759 * 3DSTATE_BINDING_TABLE_POINTER_VS,
2760 * 3DSTATE_SAMPLER_STATE_POINTER_VS command. Only one
2761 * PIPE_CONTROL needs to be sent before any combination of VS
2762 * associated 3DSTATE."
2764 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
2765 pc
.DepthStallEnable
= true;
2766 pc
.PostSyncOperation
= WriteImmediateData
;
2768 (struct anv_address
) { &cmd_buffer
->device
->workaround_bo
, 0 };
2773 /* Render targets live in the same binding table as fragment descriptors */
2774 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_RENDER_TARGETS
)
2775 cmd_buffer
->state
.descriptors_dirty
|= VK_SHADER_STAGE_FRAGMENT_BIT
;
2777 /* We emit the binding tables and sampler tables first, then emit push
2778 * constants and then finally emit binding table and sampler table
2779 * pointers. It has to happen in this order, since emitting the binding
2780 * tables may change the push constants (in case of storage images). After
2781 * emitting push constants, on SKL+ we have to emit the corresponding
2782 * 3DSTATE_BINDING_TABLE_POINTER_* for the push constants to take effect.
2785 if (cmd_buffer
->state
.descriptors_dirty
)
2786 dirty
= flush_descriptor_sets(cmd_buffer
);
2788 if (dirty
|| cmd_buffer
->state
.push_constants_dirty
) {
2789 /* Because we're pushing UBOs, we have to push whenever either
2790 * descriptors or push constants is dirty.
2792 dirty
|= cmd_buffer
->state
.push_constants_dirty
;
2793 dirty
&= ANV_STAGE_MASK
& VK_SHADER_STAGE_ALL_GRAPHICS
;
2794 cmd_buffer_flush_push_constants(cmd_buffer
, dirty
);
2798 cmd_buffer_emit_descriptor_pointers(cmd_buffer
, dirty
);
2800 if (cmd_buffer
->state
.gfx
.dirty
& ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
)
2801 gen8_cmd_buffer_emit_viewport(cmd_buffer
);
2803 if (cmd_buffer
->state
.gfx
.dirty
& (ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
|
2804 ANV_CMD_DIRTY_PIPELINE
)) {
2805 gen8_cmd_buffer_emit_depth_viewport(cmd_buffer
,
2806 pipeline
->depth_clamp_enable
);
2809 if (cmd_buffer
->state
.gfx
.dirty
& (ANV_CMD_DIRTY_DYNAMIC_SCISSOR
|
2810 ANV_CMD_DIRTY_RENDER_TARGETS
))
2811 gen7_cmd_buffer_emit_scissor(cmd_buffer
);
2813 genX(cmd_buffer_flush_dynamic_state
)(cmd_buffer
);
2815 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
2819 emit_vertex_bo(struct anv_cmd_buffer
*cmd_buffer
,
2820 struct anv_address addr
,
2821 uint32_t size
, uint32_t index
)
2823 uint32_t *p
= anv_batch_emitn(&cmd_buffer
->batch
, 5,
2824 GENX(3DSTATE_VERTEX_BUFFERS
));
2826 GENX(VERTEX_BUFFER_STATE_pack
)(&cmd_buffer
->batch
, p
+ 1,
2827 &(struct GENX(VERTEX_BUFFER_STATE
)) {
2828 .VertexBufferIndex
= index
,
2829 .AddressModifyEnable
= true,
2831 .MOCS
= anv_mocs_for_bo(cmd_buffer
->device
, addr
.bo
),
2833 .BufferStartingAddress
= addr
,
2836 .BufferStartingAddress
= addr
,
2837 .EndAddress
= anv_address_add(addr
, size
),
2843 emit_base_vertex_instance_bo(struct anv_cmd_buffer
*cmd_buffer
,
2844 struct anv_address addr
)
2846 emit_vertex_bo(cmd_buffer
, addr
, 8, ANV_SVGS_VB_INDEX
);
2850 emit_base_vertex_instance(struct anv_cmd_buffer
*cmd_buffer
,
2851 uint32_t base_vertex
, uint32_t base_instance
)
2853 struct anv_state id_state
=
2854 anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, 8, 4);
2856 ((uint32_t *)id_state
.map
)[0] = base_vertex
;
2857 ((uint32_t *)id_state
.map
)[1] = base_instance
;
2859 struct anv_address addr
= {
2860 .bo
= cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
2861 .offset
= id_state
.offset
,
2864 emit_base_vertex_instance_bo(cmd_buffer
, addr
);
2868 emit_draw_index(struct anv_cmd_buffer
*cmd_buffer
, uint32_t draw_index
)
2870 struct anv_state state
=
2871 anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, 4, 4);
2873 ((uint32_t *)state
.map
)[0] = draw_index
;
2875 struct anv_address addr
= {
2876 .bo
= cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
2877 .offset
= state
.offset
,
2880 emit_vertex_bo(cmd_buffer
, addr
, 4, ANV_DRAWID_VB_INDEX
);
2884 VkCommandBuffer commandBuffer
,
2885 uint32_t vertexCount
,
2886 uint32_t instanceCount
,
2887 uint32_t firstVertex
,
2888 uint32_t firstInstance
)
2890 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2891 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
2892 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
2894 if (anv_batch_has_error(&cmd_buffer
->batch
))
2897 genX(cmd_buffer_flush_state
)(cmd_buffer
);
2899 if (cmd_buffer
->state
.conditional_render_enabled
)
2900 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
2902 if (vs_prog_data
->uses_firstvertex
||
2903 vs_prog_data
->uses_baseinstance
)
2904 emit_base_vertex_instance(cmd_buffer
, firstVertex
, firstInstance
);
2905 if (vs_prog_data
->uses_drawid
)
2906 emit_draw_index(cmd_buffer
, 0);
2908 /* Our implementation of VK_KHR_multiview uses instancing to draw the
2909 * different views. We need to multiply instanceCount by the view count.
2911 instanceCount
*= anv_subpass_view_count(cmd_buffer
->state
.subpass
);
2913 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
2914 prim
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
2915 prim
.VertexAccessType
= SEQUENTIAL
;
2916 prim
.PrimitiveTopologyType
= pipeline
->topology
;
2917 prim
.VertexCountPerInstance
= vertexCount
;
2918 prim
.StartVertexLocation
= firstVertex
;
2919 prim
.InstanceCount
= instanceCount
;
2920 prim
.StartInstanceLocation
= firstInstance
;
2921 prim
.BaseVertexLocation
= 0;
2925 void genX(CmdDrawIndexed
)(
2926 VkCommandBuffer commandBuffer
,
2927 uint32_t indexCount
,
2928 uint32_t instanceCount
,
2929 uint32_t firstIndex
,
2930 int32_t vertexOffset
,
2931 uint32_t firstInstance
)
2933 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2934 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
2935 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
2937 if (anv_batch_has_error(&cmd_buffer
->batch
))
2940 genX(cmd_buffer_flush_state
)(cmd_buffer
);
2942 if (cmd_buffer
->state
.conditional_render_enabled
)
2943 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
2945 if (vs_prog_data
->uses_firstvertex
||
2946 vs_prog_data
->uses_baseinstance
)
2947 emit_base_vertex_instance(cmd_buffer
, vertexOffset
, firstInstance
);
2948 if (vs_prog_data
->uses_drawid
)
2949 emit_draw_index(cmd_buffer
, 0);
2951 /* Our implementation of VK_KHR_multiview uses instancing to draw the
2952 * different views. We need to multiply instanceCount by the view count.
2954 instanceCount
*= anv_subpass_view_count(cmd_buffer
->state
.subpass
);
2956 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
2957 prim
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
2958 prim
.VertexAccessType
= RANDOM
;
2959 prim
.PrimitiveTopologyType
= pipeline
->topology
;
2960 prim
.VertexCountPerInstance
= indexCount
;
2961 prim
.StartVertexLocation
= firstIndex
;
2962 prim
.InstanceCount
= instanceCount
;
2963 prim
.StartInstanceLocation
= firstInstance
;
2964 prim
.BaseVertexLocation
= vertexOffset
;
2968 /* Auto-Draw / Indirect Registers */
2969 #define GEN7_3DPRIM_END_OFFSET 0x2420
2970 #define GEN7_3DPRIM_START_VERTEX 0x2430
2971 #define GEN7_3DPRIM_VERTEX_COUNT 0x2434
2972 #define GEN7_3DPRIM_INSTANCE_COUNT 0x2438
2973 #define GEN7_3DPRIM_START_INSTANCE 0x243C
2974 #define GEN7_3DPRIM_BASE_VERTEX 0x2440
2976 void genX(CmdDrawIndirectByteCountEXT
)(
2977 VkCommandBuffer commandBuffer
,
2978 uint32_t instanceCount
,
2979 uint32_t firstInstance
,
2980 VkBuffer counterBuffer
,
2981 VkDeviceSize counterBufferOffset
,
2982 uint32_t counterOffset
,
2983 uint32_t vertexStride
)
2985 #if GEN_IS_HASWELL || GEN_GEN >= 8
2986 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
2987 ANV_FROM_HANDLE(anv_buffer
, counter_buffer
, counterBuffer
);
2988 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
2989 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
2991 /* firstVertex is always zero for this draw function */
2992 const uint32_t firstVertex
= 0;
2994 if (anv_batch_has_error(&cmd_buffer
->batch
))
2997 genX(cmd_buffer_flush_state
)(cmd_buffer
);
2999 if (vs_prog_data
->uses_firstvertex
||
3000 vs_prog_data
->uses_baseinstance
)
3001 emit_base_vertex_instance(cmd_buffer
, firstVertex
, firstInstance
);
3002 if (vs_prog_data
->uses_drawid
)
3003 emit_draw_index(cmd_buffer
, 0);
3005 /* Our implementation of VK_KHR_multiview uses instancing to draw the
3006 * different views. We need to multiply instanceCount by the view count.
3008 instanceCount
*= anv_subpass_view_count(cmd_buffer
->state
.subpass
);
3010 struct gen_mi_builder b
;
3011 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3012 struct gen_mi_value count
=
3013 gen_mi_mem32(anv_address_add(counter_buffer
->address
,
3014 counterBufferOffset
));
3016 count
= gen_mi_isub(&b
, count
, gen_mi_imm(counterOffset
));
3017 count
= gen_mi_udiv32_imm(&b
, count
, vertexStride
);
3018 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_VERTEX_COUNT
), count
);
3020 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_VERTEX
),
3021 gen_mi_imm(firstVertex
));
3022 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_INSTANCE_COUNT
),
3023 gen_mi_imm(instanceCount
));
3024 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE
),
3025 gen_mi_imm(firstInstance
));
3026 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX
), gen_mi_imm(0));
3028 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3029 prim
.IndirectParameterEnable
= true;
3030 prim
.VertexAccessType
= SEQUENTIAL
;
3031 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3033 #endif /* GEN_IS_HASWELL || GEN_GEN >= 8 */
3037 load_indirect_parameters(struct anv_cmd_buffer
*cmd_buffer
,
3038 struct anv_address addr
,
3041 struct gen_mi_builder b
;
3042 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3044 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_VERTEX_COUNT
),
3045 gen_mi_mem32(anv_address_add(addr
, 0)));
3047 struct gen_mi_value instance_count
= gen_mi_mem32(anv_address_add(addr
, 4));
3048 unsigned view_count
= anv_subpass_view_count(cmd_buffer
->state
.subpass
);
3049 if (view_count
> 1) {
3050 #if GEN_IS_HASWELL || GEN_GEN >= 8
3051 instance_count
= gen_mi_imul_imm(&b
, instance_count
, view_count
);
3053 anv_finishme("Multiview + indirect draw requires MI_MATH; "
3054 "MI_MATH is not supported on Ivy Bridge");
3057 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_INSTANCE_COUNT
), instance_count
);
3059 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_VERTEX
),
3060 gen_mi_mem32(anv_address_add(addr
, 8)));
3063 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX
),
3064 gen_mi_mem32(anv_address_add(addr
, 12)));
3065 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE
),
3066 gen_mi_mem32(anv_address_add(addr
, 16)));
3068 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE
),
3069 gen_mi_mem32(anv_address_add(addr
, 12)));
3070 gen_mi_store(&b
, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX
), gen_mi_imm(0));
3074 void genX(CmdDrawIndirect
)(
3075 VkCommandBuffer commandBuffer
,
3077 VkDeviceSize offset
,
3081 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3082 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3083 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3084 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3086 if (anv_batch_has_error(&cmd_buffer
->batch
))
3089 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3091 if (cmd_buffer
->state
.conditional_render_enabled
)
3092 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
3094 for (uint32_t i
= 0; i
< drawCount
; i
++) {
3095 struct anv_address draw
= anv_address_add(buffer
->address
, offset
);
3097 if (vs_prog_data
->uses_firstvertex
||
3098 vs_prog_data
->uses_baseinstance
)
3099 emit_base_vertex_instance_bo(cmd_buffer
, anv_address_add(draw
, 8));
3100 if (vs_prog_data
->uses_drawid
)
3101 emit_draw_index(cmd_buffer
, i
);
3103 load_indirect_parameters(cmd_buffer
, draw
, false);
3105 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3106 prim
.IndirectParameterEnable
= true;
3107 prim
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
3108 prim
.VertexAccessType
= SEQUENTIAL
;
3109 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3116 void genX(CmdDrawIndexedIndirect
)(
3117 VkCommandBuffer commandBuffer
,
3119 VkDeviceSize offset
,
3123 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3124 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3125 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.gfx
.base
.pipeline
;
3126 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3128 if (anv_batch_has_error(&cmd_buffer
->batch
))
3131 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3133 if (cmd_buffer
->state
.conditional_render_enabled
)
3134 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
3136 for (uint32_t i
= 0; i
< drawCount
; i
++) {
3137 struct anv_address draw
= anv_address_add(buffer
->address
, offset
);
3139 /* TODO: We need to stomp base vertex to 0 somehow */
3140 if (vs_prog_data
->uses_firstvertex
||
3141 vs_prog_data
->uses_baseinstance
)
3142 emit_base_vertex_instance_bo(cmd_buffer
, anv_address_add(draw
, 12));
3143 if (vs_prog_data
->uses_drawid
)
3144 emit_draw_index(cmd_buffer
, i
);
3146 load_indirect_parameters(cmd_buffer
, draw
, true);
3148 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3149 prim
.IndirectParameterEnable
= true;
3150 prim
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
3151 prim
.VertexAccessType
= RANDOM
;
3152 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3159 #define TMP_DRAW_COUNT_REG 0x2670 /* MI_ALU_REG14 */
3162 prepare_for_draw_count_predicate(struct anv_cmd_buffer
*cmd_buffer
,
3163 struct anv_address count_address
,
3164 const bool conditional_render_enabled
)
3166 struct gen_mi_builder b
;
3167 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3169 if (conditional_render_enabled
) {
3170 #if GEN_GEN >= 8 || GEN_IS_HASWELL
3171 gen_mi_store(&b
, gen_mi_reg64(TMP_DRAW_COUNT_REG
),
3172 gen_mi_mem32(count_address
));
3175 /* Upload the current draw count from the draw parameters buffer to
3176 * MI_PREDICATE_SRC0.
3178 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
),
3179 gen_mi_mem32(count_address
));
3181 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC1
+ 4), gen_mi_imm(0));
3186 emit_draw_count_predicate(struct anv_cmd_buffer
*cmd_buffer
,
3187 uint32_t draw_index
)
3189 struct gen_mi_builder b
;
3190 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3192 /* Upload the index of the current primitive to MI_PREDICATE_SRC1. */
3193 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC1
), gen_mi_imm(draw_index
));
3195 if (draw_index
== 0) {
3196 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
3197 mip
.LoadOperation
= LOAD_LOADINV
;
3198 mip
.CombineOperation
= COMBINE_SET
;
3199 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3202 /* While draw_index < draw_count the predicate's result will be
3203 * (draw_index == draw_count) ^ TRUE = TRUE
3204 * When draw_index == draw_count the result is
3205 * (TRUE) ^ TRUE = FALSE
3206 * After this all results will be:
3207 * (FALSE) ^ FALSE = FALSE
3209 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
3210 mip
.LoadOperation
= LOAD_LOAD
;
3211 mip
.CombineOperation
= COMBINE_XOR
;
3212 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3217 #if GEN_GEN >= 8 || GEN_IS_HASWELL
3219 emit_draw_count_predicate_with_conditional_render(
3220 struct anv_cmd_buffer
*cmd_buffer
,
3221 uint32_t draw_index
)
3223 struct gen_mi_builder b
;
3224 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3226 struct gen_mi_value pred
= gen_mi_ult(&b
, gen_mi_imm(draw_index
),
3227 gen_mi_reg64(TMP_DRAW_COUNT_REG
));
3228 pred
= gen_mi_iand(&b
, pred
, gen_mi_reg64(ANV_PREDICATE_RESULT_REG
));
3231 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_RESULT
), pred
);
3233 /* MI_PREDICATE_RESULT is not whitelisted in i915 command parser
3234 * so we emit MI_PREDICATE to set it.
3237 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
), pred
);
3238 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
3240 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
3241 mip
.LoadOperation
= LOAD_LOADINV
;
3242 mip
.CombineOperation
= COMBINE_SET
;
3243 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3249 void genX(CmdDrawIndirectCountKHR
)(
3250 VkCommandBuffer commandBuffer
,
3252 VkDeviceSize offset
,
3253 VkBuffer _countBuffer
,
3254 VkDeviceSize countBufferOffset
,
3255 uint32_t maxDrawCount
,
3258 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3259 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3260 ANV_FROM_HANDLE(anv_buffer
, count_buffer
, _countBuffer
);
3261 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
3262 struct anv_pipeline
*pipeline
= cmd_state
->gfx
.base
.pipeline
;
3263 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3265 if (anv_batch_has_error(&cmd_buffer
->batch
))
3268 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3270 struct anv_address count_address
=
3271 anv_address_add(count_buffer
->address
, countBufferOffset
);
3273 prepare_for_draw_count_predicate(cmd_buffer
, count_address
,
3274 cmd_state
->conditional_render_enabled
);
3276 for (uint32_t i
= 0; i
< maxDrawCount
; i
++) {
3277 struct anv_address draw
= anv_address_add(buffer
->address
, offset
);
3279 #if GEN_GEN >= 8 || GEN_IS_HASWELL
3280 if (cmd_state
->conditional_render_enabled
) {
3281 emit_draw_count_predicate_with_conditional_render(cmd_buffer
, i
);
3283 emit_draw_count_predicate(cmd_buffer
, i
);
3286 emit_draw_count_predicate(cmd_buffer
, i
);
3289 if (vs_prog_data
->uses_firstvertex
||
3290 vs_prog_data
->uses_baseinstance
)
3291 emit_base_vertex_instance_bo(cmd_buffer
, anv_address_add(draw
, 8));
3292 if (vs_prog_data
->uses_drawid
)
3293 emit_draw_index(cmd_buffer
, i
);
3295 load_indirect_parameters(cmd_buffer
, draw
, false);
3297 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3298 prim
.IndirectParameterEnable
= true;
3299 prim
.PredicateEnable
= true;
3300 prim
.VertexAccessType
= SEQUENTIAL
;
3301 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3308 void genX(CmdDrawIndexedIndirectCountKHR
)(
3309 VkCommandBuffer commandBuffer
,
3311 VkDeviceSize offset
,
3312 VkBuffer _countBuffer
,
3313 VkDeviceSize countBufferOffset
,
3314 uint32_t maxDrawCount
,
3317 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3318 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3319 ANV_FROM_HANDLE(anv_buffer
, count_buffer
, _countBuffer
);
3320 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
3321 struct anv_pipeline
*pipeline
= cmd_state
->gfx
.base
.pipeline
;
3322 const struct brw_vs_prog_data
*vs_prog_data
= get_vs_prog_data(pipeline
);
3324 if (anv_batch_has_error(&cmd_buffer
->batch
))
3327 genX(cmd_buffer_flush_state
)(cmd_buffer
);
3329 struct anv_address count_address
=
3330 anv_address_add(count_buffer
->address
, countBufferOffset
);
3332 prepare_for_draw_count_predicate(cmd_buffer
, count_address
,
3333 cmd_state
->conditional_render_enabled
);
3335 for (uint32_t i
= 0; i
< maxDrawCount
; i
++) {
3336 struct anv_address draw
= anv_address_add(buffer
->address
, offset
);
3338 #if GEN_GEN >= 8 || GEN_IS_HASWELL
3339 if (cmd_state
->conditional_render_enabled
) {
3340 emit_draw_count_predicate_with_conditional_render(cmd_buffer
, i
);
3342 emit_draw_count_predicate(cmd_buffer
, i
);
3345 emit_draw_count_predicate(cmd_buffer
, i
);
3348 /* TODO: We need to stomp base vertex to 0 somehow */
3349 if (vs_prog_data
->uses_firstvertex
||
3350 vs_prog_data
->uses_baseinstance
)
3351 emit_base_vertex_instance_bo(cmd_buffer
, anv_address_add(draw
, 12));
3352 if (vs_prog_data
->uses_drawid
)
3353 emit_draw_index(cmd_buffer
, i
);
3355 load_indirect_parameters(cmd_buffer
, draw
, true);
3357 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DPRIMITIVE
), prim
) {
3358 prim
.IndirectParameterEnable
= true;
3359 prim
.PredicateEnable
= true;
3360 prim
.VertexAccessType
= RANDOM
;
3361 prim
.PrimitiveTopologyType
= pipeline
->topology
;
3368 void genX(CmdBeginTransformFeedbackEXT
)(
3369 VkCommandBuffer commandBuffer
,
3370 uint32_t firstCounterBuffer
,
3371 uint32_t counterBufferCount
,
3372 const VkBuffer
* pCounterBuffers
,
3373 const VkDeviceSize
* pCounterBufferOffsets
)
3375 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3377 assert(firstCounterBuffer
< MAX_XFB_BUFFERS
);
3378 assert(counterBufferCount
<= MAX_XFB_BUFFERS
);
3379 assert(firstCounterBuffer
+ counterBufferCount
<= MAX_XFB_BUFFERS
);
3381 /* From the SKL PRM Vol. 2c, SO_WRITE_OFFSET:
3383 * "Ssoftware must ensure that no HW stream output operations can be in
3384 * process or otherwise pending at the point that the MI_LOAD/STORE
3385 * commands are processed. This will likely require a pipeline flush."
3387 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
3388 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3390 for (uint32_t idx
= 0; idx
< MAX_XFB_BUFFERS
; idx
++) {
3391 /* If we have a counter buffer, this is a resume so we need to load the
3392 * value into the streamout offset register. Otherwise, this is a begin
3393 * and we need to reset it to zero.
3395 if (pCounterBuffers
&&
3396 idx
>= firstCounterBuffer
&&
3397 idx
- firstCounterBuffer
< counterBufferCount
&&
3398 pCounterBuffers
[idx
- firstCounterBuffer
] != VK_NULL_HANDLE
) {
3399 uint32_t cb_idx
= idx
- firstCounterBuffer
;
3400 ANV_FROM_HANDLE(anv_buffer
, counter_buffer
, pCounterBuffers
[cb_idx
]);
3401 uint64_t offset
= pCounterBufferOffsets
?
3402 pCounterBufferOffsets
[cb_idx
] : 0;
3404 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_LOAD_REGISTER_MEM
), lrm
) {
3405 lrm
.RegisterAddress
= GENX(SO_WRITE_OFFSET0_num
) + idx
* 4;
3406 lrm
.MemoryAddress
= anv_address_add(counter_buffer
->address
,
3410 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_LOAD_REGISTER_IMM
), lri
) {
3411 lri
.RegisterOffset
= GENX(SO_WRITE_OFFSET0_num
) + idx
* 4;
3417 cmd_buffer
->state
.xfb_enabled
= true;
3418 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_XFB_ENABLE
;
3421 void genX(CmdEndTransformFeedbackEXT
)(
3422 VkCommandBuffer commandBuffer
,
3423 uint32_t firstCounterBuffer
,
3424 uint32_t counterBufferCount
,
3425 const VkBuffer
* pCounterBuffers
,
3426 const VkDeviceSize
* pCounterBufferOffsets
)
3428 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3430 assert(firstCounterBuffer
< MAX_XFB_BUFFERS
);
3431 assert(counterBufferCount
<= MAX_XFB_BUFFERS
);
3432 assert(firstCounterBuffer
+ counterBufferCount
<= MAX_XFB_BUFFERS
);
3434 /* From the SKL PRM Vol. 2c, SO_WRITE_OFFSET:
3436 * "Ssoftware must ensure that no HW stream output operations can be in
3437 * process or otherwise pending at the point that the MI_LOAD/STORE
3438 * commands are processed. This will likely require a pipeline flush."
3440 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
3441 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3443 for (uint32_t cb_idx
= 0; cb_idx
< counterBufferCount
; cb_idx
++) {
3444 unsigned idx
= firstCounterBuffer
+ cb_idx
;
3446 /* If we have a counter buffer, this is a resume so we need to load the
3447 * value into the streamout offset register. Otherwise, this is a begin
3448 * and we need to reset it to zero.
3450 if (pCounterBuffers
&&
3451 cb_idx
< counterBufferCount
&&
3452 pCounterBuffers
[cb_idx
] != VK_NULL_HANDLE
) {
3453 ANV_FROM_HANDLE(anv_buffer
, counter_buffer
, pCounterBuffers
[cb_idx
]);
3454 uint64_t offset
= pCounterBufferOffsets
?
3455 pCounterBufferOffsets
[cb_idx
] : 0;
3457 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_STORE_REGISTER_MEM
), srm
) {
3458 srm
.MemoryAddress
= anv_address_add(counter_buffer
->address
,
3460 srm
.RegisterAddress
= GENX(SO_WRITE_OFFSET0_num
) + idx
* 4;
3465 cmd_buffer
->state
.xfb_enabled
= false;
3466 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_XFB_ENABLE
;
3470 flush_compute_descriptor_set(struct anv_cmd_buffer
*cmd_buffer
)
3472 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute
.base
.pipeline
;
3473 struct anv_state surfaces
= { 0, }, samplers
= { 0, };
3476 result
= emit_binding_table(cmd_buffer
, MESA_SHADER_COMPUTE
, &surfaces
);
3477 if (result
!= VK_SUCCESS
) {
3478 assert(result
== VK_ERROR_OUT_OF_DEVICE_MEMORY
);
3480 result
= anv_cmd_buffer_new_binding_table_block(cmd_buffer
);
3481 if (result
!= VK_SUCCESS
)
3484 /* Re-emit state base addresses so we get the new surface state base
3485 * address before we start emitting binding tables etc.
3487 genX(cmd_buffer_emit_state_base_address
)(cmd_buffer
);
3489 result
= emit_binding_table(cmd_buffer
, MESA_SHADER_COMPUTE
, &surfaces
);
3490 if (result
!= VK_SUCCESS
) {
3491 anv_batch_set_error(&cmd_buffer
->batch
, result
);
3496 result
= emit_samplers(cmd_buffer
, MESA_SHADER_COMPUTE
, &samplers
);
3497 if (result
!= VK_SUCCESS
) {
3498 anv_batch_set_error(&cmd_buffer
->batch
, result
);
3502 uint32_t iface_desc_data_dw
[GENX(INTERFACE_DESCRIPTOR_DATA_length
)];
3503 struct GENX(INTERFACE_DESCRIPTOR_DATA
) desc
= {
3504 .BindingTablePointer
= surfaces
.offset
,
3505 .SamplerStatePointer
= samplers
.offset
,
3507 GENX(INTERFACE_DESCRIPTOR_DATA_pack
)(NULL
, iface_desc_data_dw
, &desc
);
3509 struct anv_state state
=
3510 anv_cmd_buffer_merge_dynamic(cmd_buffer
, iface_desc_data_dw
,
3511 pipeline
->interface_descriptor_data
,
3512 GENX(INTERFACE_DESCRIPTOR_DATA_length
),
3515 uint32_t size
= GENX(INTERFACE_DESCRIPTOR_DATA_length
) * sizeof(uint32_t);
3516 anv_batch_emit(&cmd_buffer
->batch
,
3517 GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD
), mid
) {
3518 mid
.InterfaceDescriptorTotalLength
= size
;
3519 mid
.InterfaceDescriptorDataStartAddress
= state
.offset
;
3526 genX(cmd_buffer_flush_compute_state
)(struct anv_cmd_buffer
*cmd_buffer
)
3528 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute
.base
.pipeline
;
3531 assert(pipeline
->active_stages
== VK_SHADER_STAGE_COMPUTE_BIT
);
3533 genX(cmd_buffer_config_l3
)(cmd_buffer
, pipeline
->urb
.l3_config
);
3535 genX(flush_pipeline_select_gpgpu
)(cmd_buffer
);
3537 if (cmd_buffer
->state
.compute
.pipeline_dirty
) {
3538 /* From the Sky Lake PRM Vol 2a, MEDIA_VFE_STATE:
3540 * "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
3541 * the only bits that are changed are scoreboard related: Scoreboard
3542 * Enable, Scoreboard Type, Scoreboard Mask, Scoreboard * Delta. For
3543 * these scoreboard related states, a MEDIA_STATE_FLUSH is
3546 cmd_buffer
->state
.pending_pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
3547 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3549 anv_batch_emit_batch(&cmd_buffer
->batch
, &pipeline
->batch
);
3552 if ((cmd_buffer
->state
.descriptors_dirty
& VK_SHADER_STAGE_COMPUTE_BIT
) ||
3553 cmd_buffer
->state
.compute
.pipeline_dirty
) {
3554 /* FIXME: figure out descriptors for gen7 */
3555 result
= flush_compute_descriptor_set(cmd_buffer
);
3556 if (result
!= VK_SUCCESS
)
3559 cmd_buffer
->state
.descriptors_dirty
&= ~VK_SHADER_STAGE_COMPUTE_BIT
;
3562 if (cmd_buffer
->state
.push_constants_dirty
& VK_SHADER_STAGE_COMPUTE_BIT
) {
3563 struct anv_state push_state
=
3564 anv_cmd_buffer_cs_push_constants(cmd_buffer
);
3566 if (push_state
.alloc_size
) {
3567 anv_batch_emit(&cmd_buffer
->batch
, GENX(MEDIA_CURBE_LOAD
), curbe
) {
3568 curbe
.CURBETotalDataLength
= push_state
.alloc_size
;
3569 curbe
.CURBEDataStartAddress
= push_state
.offset
;
3573 cmd_buffer
->state
.push_constants_dirty
&= ~VK_SHADER_STAGE_COMPUTE_BIT
;
3576 cmd_buffer
->state
.compute
.pipeline_dirty
= false;
3578 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
3584 verify_cmd_parser(const struct anv_device
*device
,
3585 int required_version
,
3586 const char *function
)
3588 if (device
->instance
->physicalDevice
.cmd_parser_version
< required_version
) {
3589 return vk_errorf(device
->instance
, device
->instance
,
3590 VK_ERROR_FEATURE_NOT_PRESENT
,
3591 "cmd parser version %d is required for %s",
3592 required_version
, function
);
3601 anv_cmd_buffer_push_base_group_id(struct anv_cmd_buffer
*cmd_buffer
,
3602 uint32_t baseGroupX
,
3603 uint32_t baseGroupY
,
3604 uint32_t baseGroupZ
)
3606 if (anv_batch_has_error(&cmd_buffer
->batch
))
3609 struct anv_push_constants
*push
=
3610 &cmd_buffer
->state
.push_constants
[MESA_SHADER_COMPUTE
];
3611 if (push
->base_work_group_id
[0] != baseGroupX
||
3612 push
->base_work_group_id
[1] != baseGroupY
||
3613 push
->base_work_group_id
[2] != baseGroupZ
) {
3614 push
->base_work_group_id
[0] = baseGroupX
;
3615 push
->base_work_group_id
[1] = baseGroupY
;
3616 push
->base_work_group_id
[2] = baseGroupZ
;
3618 cmd_buffer
->state
.push_constants_dirty
|= VK_SHADER_STAGE_COMPUTE_BIT
;
3622 void genX(CmdDispatch
)(
3623 VkCommandBuffer commandBuffer
,
3628 genX(CmdDispatchBase
)(commandBuffer
, 0, 0, 0, x
, y
, z
);
3631 void genX(CmdDispatchBase
)(
3632 VkCommandBuffer commandBuffer
,
3633 uint32_t baseGroupX
,
3634 uint32_t baseGroupY
,
3635 uint32_t baseGroupZ
,
3636 uint32_t groupCountX
,
3637 uint32_t groupCountY
,
3638 uint32_t groupCountZ
)
3640 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3641 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute
.base
.pipeline
;
3642 const struct brw_cs_prog_data
*prog_data
= get_cs_prog_data(pipeline
);
3644 anv_cmd_buffer_push_base_group_id(cmd_buffer
, baseGroupX
,
3645 baseGroupY
, baseGroupZ
);
3647 if (anv_batch_has_error(&cmd_buffer
->batch
))
3650 if (prog_data
->uses_num_work_groups
) {
3651 struct anv_state state
=
3652 anv_cmd_buffer_alloc_dynamic_state(cmd_buffer
, 12, 4);
3653 uint32_t *sizes
= state
.map
;
3654 sizes
[0] = groupCountX
;
3655 sizes
[1] = groupCountY
;
3656 sizes
[2] = groupCountZ
;
3657 cmd_buffer
->state
.compute
.num_workgroups
= (struct anv_address
) {
3658 .bo
= cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
3659 .offset
= state
.offset
,
3663 genX(cmd_buffer_flush_compute_state
)(cmd_buffer
);
3665 if (cmd_buffer
->state
.conditional_render_enabled
)
3666 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
3668 anv_batch_emit(&cmd_buffer
->batch
, GENX(GPGPU_WALKER
), ggw
) {
3669 ggw
.PredicateEnable
= cmd_buffer
->state
.conditional_render_enabled
;
3670 ggw
.SIMDSize
= prog_data
->simd_size
/ 16;
3671 ggw
.ThreadDepthCounterMaximum
= 0;
3672 ggw
.ThreadHeightCounterMaximum
= 0;
3673 ggw
.ThreadWidthCounterMaximum
= prog_data
->threads
- 1;
3674 ggw
.ThreadGroupIDXDimension
= groupCountX
;
3675 ggw
.ThreadGroupIDYDimension
= groupCountY
;
3676 ggw
.ThreadGroupIDZDimension
= groupCountZ
;
3677 ggw
.RightExecutionMask
= pipeline
->cs_right_mask
;
3678 ggw
.BottomExecutionMask
= 0xffffffff;
3681 anv_batch_emit(&cmd_buffer
->batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
3684 #define GPGPU_DISPATCHDIMX 0x2500
3685 #define GPGPU_DISPATCHDIMY 0x2504
3686 #define GPGPU_DISPATCHDIMZ 0x2508
3688 void genX(CmdDispatchIndirect
)(
3689 VkCommandBuffer commandBuffer
,
3691 VkDeviceSize offset
)
3693 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
3694 ANV_FROM_HANDLE(anv_buffer
, buffer
, _buffer
);
3695 struct anv_pipeline
*pipeline
= cmd_buffer
->state
.compute
.base
.pipeline
;
3696 const struct brw_cs_prog_data
*prog_data
= get_cs_prog_data(pipeline
);
3697 struct anv_address addr
= anv_address_add(buffer
->address
, offset
);
3698 struct anv_batch
*batch
= &cmd_buffer
->batch
;
3700 anv_cmd_buffer_push_base_group_id(cmd_buffer
, 0, 0, 0);
3703 /* Linux 4.4 added command parser version 5 which allows the GPGPU
3704 * indirect dispatch registers to be written.
3706 if (verify_cmd_parser(cmd_buffer
->device
, 5,
3707 "vkCmdDispatchIndirect") != VK_SUCCESS
)
3711 if (prog_data
->uses_num_work_groups
)
3712 cmd_buffer
->state
.compute
.num_workgroups
= addr
;
3714 genX(cmd_buffer_flush_compute_state
)(cmd_buffer
);
3716 struct gen_mi_builder b
;
3717 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
3719 struct gen_mi_value size_x
= gen_mi_mem32(anv_address_add(addr
, 0));
3720 struct gen_mi_value size_y
= gen_mi_mem32(anv_address_add(addr
, 4));
3721 struct gen_mi_value size_z
= gen_mi_mem32(anv_address_add(addr
, 8));
3723 gen_mi_store(&b
, gen_mi_reg32(GPGPU_DISPATCHDIMX
), size_x
);
3724 gen_mi_store(&b
, gen_mi_reg32(GPGPU_DISPATCHDIMY
), size_y
);
3725 gen_mi_store(&b
, gen_mi_reg32(GPGPU_DISPATCHDIMZ
), size_z
);
3728 /* predicate = (compute_dispatch_indirect_x_size == 0); */
3729 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
), size_x
);
3730 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
3731 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
3732 mip
.LoadOperation
= LOAD_LOAD
;
3733 mip
.CombineOperation
= COMBINE_SET
;
3734 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3737 /* predicate |= (compute_dispatch_indirect_y_size == 0); */
3738 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC0
), size_y
);
3739 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
3740 mip
.LoadOperation
= LOAD_LOAD
;
3741 mip
.CombineOperation
= COMBINE_OR
;
3742 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3745 /* predicate |= (compute_dispatch_indirect_z_size == 0); */
3746 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC0
), size_z
);
3747 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
3748 mip
.LoadOperation
= LOAD_LOAD
;
3749 mip
.CombineOperation
= COMBINE_OR
;
3750 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3753 /* predicate = !predicate; */
3754 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
3755 mip
.LoadOperation
= LOAD_LOADINV
;
3756 mip
.CombineOperation
= COMBINE_OR
;
3757 mip
.CompareOperation
= COMPARE_FALSE
;
3761 if (cmd_buffer
->state
.conditional_render_enabled
) {
3762 /* predicate &= !(conditional_rendering_predicate == 0); */
3763 gen_mi_store(&b
, gen_mi_reg32(MI_PREDICATE_SRC0
),
3764 gen_mi_reg32(ANV_PREDICATE_RESULT_REG
));
3765 anv_batch_emit(batch
, GENX(MI_PREDICATE
), mip
) {
3766 mip
.LoadOperation
= LOAD_LOADINV
;
3767 mip
.CombineOperation
= COMBINE_AND
;
3768 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
3773 #else /* GEN_GEN > 7 */
3774 if (cmd_buffer
->state
.conditional_render_enabled
)
3775 genX(cmd_emit_conditional_render_predicate
)(cmd_buffer
);
3778 anv_batch_emit(batch
, GENX(GPGPU_WALKER
), ggw
) {
3779 ggw
.IndirectParameterEnable
= true;
3780 ggw
.PredicateEnable
= GEN_GEN
<= 7 ||
3781 cmd_buffer
->state
.conditional_render_enabled
;
3782 ggw
.SIMDSize
= prog_data
->simd_size
/ 16;
3783 ggw
.ThreadDepthCounterMaximum
= 0;
3784 ggw
.ThreadHeightCounterMaximum
= 0;
3785 ggw
.ThreadWidthCounterMaximum
= prog_data
->threads
- 1;
3786 ggw
.RightExecutionMask
= pipeline
->cs_right_mask
;
3787 ggw
.BottomExecutionMask
= 0xffffffff;
3790 anv_batch_emit(batch
, GENX(MEDIA_STATE_FLUSH
), msf
);
3794 genX(flush_pipeline_select
)(struct anv_cmd_buffer
*cmd_buffer
,
3797 UNUSED
const struct gen_device_info
*devinfo
= &cmd_buffer
->device
->info
;
3799 if (cmd_buffer
->state
.current_pipeline
== pipeline
)
3802 #if GEN_GEN >= 8 && GEN_GEN < 10
3803 /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT:
3805 * Software must clear the COLOR_CALC_STATE Valid field in
3806 * 3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT
3807 * with Pipeline Select set to GPGPU.
3809 * The internal hardware docs recommend the same workaround for Gen9
3812 if (pipeline
== GPGPU
)
3813 anv_batch_emit(&cmd_buffer
->batch
, GENX(3DSTATE_CC_STATE_POINTERS
), t
);
3817 if (pipeline
== _3D
) {
3818 /* There is a mid-object preemption workaround which requires you to
3819 * re-emit MEDIA_VFE_STATE after switching from GPGPU to 3D. However,
3820 * even without preemption, we have issues with geometry flickering when
3821 * GPGPU and 3D are back-to-back and this seems to fix it. We don't
3824 const uint32_t subslices
=
3825 MAX2(cmd_buffer
->device
->instance
->physicalDevice
.subslice_total
, 1);
3826 anv_batch_emit(&cmd_buffer
->batch
, GENX(MEDIA_VFE_STATE
), vfe
) {
3827 vfe
.MaximumNumberofThreads
=
3828 devinfo
->max_cs_threads
* subslices
- 1;
3829 vfe
.NumberofURBEntries
= 2;
3830 vfe
.URBEntryAllocationSize
= 2;
3835 /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
3836 * PIPELINE_SELECT [DevBWR+]":
3840 * Software must ensure all the write caches are flushed through a
3841 * stalling PIPE_CONTROL command followed by another PIPE_CONTROL
3842 * command to invalidate read only caches prior to programming
3843 * MI_PIPELINE_SELECT command to change the Pipeline Select Mode.
3845 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
3846 pc
.RenderTargetCacheFlushEnable
= true;
3847 pc
.DepthCacheFlushEnable
= true;
3848 pc
.DCFlushEnable
= true;
3849 pc
.PostSyncOperation
= NoWrite
;
3850 pc
.CommandStreamerStallEnable
= true;
3853 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
3854 pc
.TextureCacheInvalidationEnable
= true;
3855 pc
.ConstantCacheInvalidationEnable
= true;
3856 pc
.StateCacheInvalidationEnable
= true;
3857 pc
.InstructionCacheInvalidateEnable
= true;
3858 pc
.PostSyncOperation
= NoWrite
;
3861 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPELINE_SELECT
), ps
) {
3865 ps
.PipelineSelection
= pipeline
;
3869 if (devinfo
->is_geminilake
) {
3872 * "This chicken bit works around a hardware issue with barrier logic
3873 * encountered when switching between GPGPU and 3D pipelines. To
3874 * workaround the issue, this mode bit should be set after a pipeline
3878 anv_pack_struct(&scec
, GENX(SLICE_COMMON_ECO_CHICKEN1
),
3880 pipeline
== GPGPU
? GLK_BARRIER_MODE_GPGPU
3881 : GLK_BARRIER_MODE_3D_HULL
,
3882 .GLKBarrierModeMask
= 1);
3883 emit_lri(&cmd_buffer
->batch
, GENX(SLICE_COMMON_ECO_CHICKEN1_num
), scec
);
3887 cmd_buffer
->state
.current_pipeline
= pipeline
;
3891 genX(flush_pipeline_select_3d
)(struct anv_cmd_buffer
*cmd_buffer
)
3893 genX(flush_pipeline_select
)(cmd_buffer
, _3D
);
3897 genX(flush_pipeline_select_gpgpu
)(struct anv_cmd_buffer
*cmd_buffer
)
3899 genX(flush_pipeline_select
)(cmd_buffer
, GPGPU
);
3903 genX(cmd_buffer_emit_gen7_depth_flush
)(struct anv_cmd_buffer
*cmd_buffer
)
3908 /* From the Haswell PRM, documentation for 3DSTATE_DEPTH_BUFFER:
3910 * "Restriction: Prior to changing Depth/Stencil Buffer state (i.e., any
3911 * combination of 3DSTATE_DEPTH_BUFFER, 3DSTATE_CLEAR_PARAMS,
3912 * 3DSTATE_STENCIL_BUFFER, 3DSTATE_HIER_DEPTH_BUFFER) SW must first
3913 * issue a pipelined depth stall (PIPE_CONTROL with Depth Stall bit
3914 * set), followed by a pipelined depth cache flush (PIPE_CONTROL with
3915 * Depth Flush Bit set, followed by another pipelined depth stall
3916 * (PIPE_CONTROL with Depth Stall Bit set), unless SW can otherwise
3917 * guarantee that the pipeline from WM onwards is already flushed (e.g.,
3918 * via a preceding MI_FLUSH)."
3920 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
3921 pipe
.DepthStallEnable
= true;
3923 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
3924 pipe
.DepthCacheFlushEnable
= true;
3926 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pipe
) {
3927 pipe
.DepthStallEnable
= true;
3932 * Update the pixel hashing modes that determine the balancing of PS threads
3933 * across subslices and slices.
3935 * \param width Width bound of the rendering area (already scaled down if \p
3936 * scale is greater than 1).
3937 * \param height Height bound of the rendering area (already scaled down if \p
3938 * scale is greater than 1).
3939 * \param scale The number of framebuffer samples that could potentially be
3940 * affected by an individual channel of the PS thread. This is
3941 * typically one for single-sampled rendering, but for operations
3942 * like CCS resolves and fast clears a single PS invocation may
3943 * update a huge number of pixels, in which case a finer
3944 * balancing is desirable in order to maximally utilize the
3945 * bandwidth available. UINT_MAX can be used as shorthand for
3946 * "finest hashing mode available".
3949 genX(cmd_buffer_emit_hashing_mode
)(struct anv_cmd_buffer
*cmd_buffer
,
3950 unsigned width
, unsigned height
,
3954 const struct gen_device_info
*devinfo
= &cmd_buffer
->device
->info
;
3955 const unsigned slice_hashing
[] = {
3956 /* Because all Gen9 platforms with more than one slice require
3957 * three-way subslice hashing, a single "normal" 16x16 slice hashing
3958 * block is guaranteed to suffer from substantial imbalance, with one
3959 * subslice receiving twice as much work as the other two in the
3962 * The performance impact of that would be particularly severe when
3963 * three-way hashing is also in use for slice balancing (which is the
3964 * case for all Gen9 GT4 platforms), because one of the slices
3965 * receives one every three 16x16 blocks in either direction, which
3966 * is roughly the periodicity of the underlying subslice imbalance
3967 * pattern ("roughly" because in reality the hardware's
3968 * implementation of three-way hashing doesn't do exact modulo 3
3969 * arithmetic, which somewhat decreases the magnitude of this effect
3970 * in practice). This leads to a systematic subslice imbalance
3971 * within that slice regardless of the size of the primitive. The
3972 * 32x32 hashing mode guarantees that the subslice imbalance within a
3973 * single slice hashing block is minimal, largely eliminating this
3977 /* Finest slice hashing mode available. */
3980 const unsigned subslice_hashing
[] = {
3981 /* 16x16 would provide a slight cache locality benefit especially
3982 * visible in the sampler L1 cache efficiency of low-bandwidth
3983 * non-LLC platforms, but it comes at the cost of greater subslice
3984 * imbalance for primitives of dimensions approximately intermediate
3985 * between 16x4 and 16x16.
3988 /* Finest subslice hashing mode available. */
3991 /* Dimensions of the smallest hashing block of a given hashing mode. If
3992 * the rendering area is smaller than this there can't possibly be any
3993 * benefit from switching to this mode, so we optimize out the
3996 const unsigned min_size
[][2] = {
4000 const unsigned idx
= scale
> 1;
4002 if (cmd_buffer
->state
.current_hash_scale
!= scale
&&
4003 (width
> min_size
[idx
][0] || height
> min_size
[idx
][1])) {
4006 anv_pack_struct(>_mode
, GENX(GT_MODE
),
4007 .SliceHashing
= (devinfo
->num_slices
> 1 ? slice_hashing
[idx
] : 0),
4008 .SliceHashingMask
= (devinfo
->num_slices
> 1 ? -1 : 0),
4009 .SubsliceHashing
= subslice_hashing
[idx
],
4010 .SubsliceHashingMask
= -1);
4012 cmd_buffer
->state
.pending_pipe_bits
|=
4013 ANV_PIPE_CS_STALL_BIT
| ANV_PIPE_STALL_AT_SCOREBOARD_BIT
;
4014 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
4016 emit_lri(&cmd_buffer
->batch
, GENX(GT_MODE_num
), gt_mode
);
4018 cmd_buffer
->state
.current_hash_scale
= scale
;
4024 cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer
*cmd_buffer
)
4026 struct anv_device
*device
= cmd_buffer
->device
;
4027 const struct anv_image_view
*iview
=
4028 anv_cmd_buffer_get_depth_stencil_view(cmd_buffer
);
4029 const struct anv_image
*image
= iview
? iview
->image
: NULL
;
4031 /* FIXME: Width and Height are wrong */
4033 genX(cmd_buffer_emit_gen7_depth_flush
)(cmd_buffer
);
4035 uint32_t *dw
= anv_batch_emit_dwords(&cmd_buffer
->batch
,
4036 device
->isl_dev
.ds
.size
/ 4);
4040 struct isl_depth_stencil_hiz_emit_info info
= { };
4043 info
.view
= &iview
->planes
[0].isl
;
4045 if (image
&& (image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)) {
4046 uint32_t depth_plane
=
4047 anv_image_aspect_to_plane(image
->aspects
, VK_IMAGE_ASPECT_DEPTH_BIT
);
4048 const struct anv_surface
*surface
= &image
->planes
[depth_plane
].surface
;
4050 info
.depth_surf
= &surface
->isl
;
4052 info
.depth_address
=
4053 anv_batch_emit_reloc(&cmd_buffer
->batch
,
4054 dw
+ device
->isl_dev
.ds
.depth_offset
/ 4,
4055 image
->planes
[depth_plane
].address
.bo
,
4056 image
->planes
[depth_plane
].address
.offset
+
4059 anv_mocs_for_bo(device
, image
->planes
[depth_plane
].address
.bo
);
4062 cmd_buffer
->state
.subpass
->depth_stencil_attachment
->attachment
;
4063 info
.hiz_usage
= cmd_buffer
->state
.attachments
[ds
].aux_usage
;
4064 if (info
.hiz_usage
== ISL_AUX_USAGE_HIZ
) {
4065 info
.hiz_surf
= &image
->planes
[depth_plane
].aux_surface
.isl
;
4068 anv_batch_emit_reloc(&cmd_buffer
->batch
,
4069 dw
+ device
->isl_dev
.ds
.hiz_offset
/ 4,
4070 image
->planes
[depth_plane
].address
.bo
,
4071 image
->planes
[depth_plane
].address
.offset
+
4072 image
->planes
[depth_plane
].aux_surface
.offset
);
4074 info
.depth_clear_value
= ANV_HZ_FC_VAL
;
4078 if (image
&& (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4079 uint32_t stencil_plane
=
4080 anv_image_aspect_to_plane(image
->aspects
, VK_IMAGE_ASPECT_STENCIL_BIT
);
4081 const struct anv_surface
*surface
= &image
->planes
[stencil_plane
].surface
;
4083 info
.stencil_surf
= &surface
->isl
;
4085 info
.stencil_address
=
4086 anv_batch_emit_reloc(&cmd_buffer
->batch
,
4087 dw
+ device
->isl_dev
.ds
.stencil_offset
/ 4,
4088 image
->planes
[stencil_plane
].address
.bo
,
4089 image
->planes
[stencil_plane
].address
.offset
+
4092 anv_mocs_for_bo(device
, image
->planes
[stencil_plane
].address
.bo
);
4095 isl_emit_depth_stencil_hiz_s(&device
->isl_dev
, dw
, &info
);
4097 cmd_buffer
->state
.hiz_enabled
= info
.hiz_usage
== ISL_AUX_USAGE_HIZ
;
4101 * This ANDs the view mask of the current subpass with the pending clear
4102 * views in the attachment to get the mask of views active in the subpass
4103 * that still need to be cleared.
4105 static inline uint32_t
4106 get_multiview_subpass_clear_mask(const struct anv_cmd_state
*cmd_state
,
4107 const struct anv_attachment_state
*att_state
)
4109 return cmd_state
->subpass
->view_mask
& att_state
->pending_clear_views
;
4113 do_first_layer_clear(const struct anv_cmd_state
*cmd_state
,
4114 const struct anv_attachment_state
*att_state
)
4116 if (!cmd_state
->subpass
->view_mask
)
4119 uint32_t pending_clear_mask
=
4120 get_multiview_subpass_clear_mask(cmd_state
, att_state
);
4122 return pending_clear_mask
& 1;
4126 current_subpass_is_last_for_attachment(const struct anv_cmd_state
*cmd_state
,
4129 const uint32_t last_subpass_idx
=
4130 cmd_state
->pass
->attachments
[att_idx
].last_subpass_idx
;
4131 const struct anv_subpass
*last_subpass
=
4132 &cmd_state
->pass
->subpasses
[last_subpass_idx
];
4133 return last_subpass
== cmd_state
->subpass
;
4137 cmd_buffer_begin_subpass(struct anv_cmd_buffer
*cmd_buffer
,
4138 uint32_t subpass_id
)
4140 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
4141 struct anv_subpass
*subpass
= &cmd_state
->pass
->subpasses
[subpass_id
];
4142 cmd_state
->subpass
= subpass
;
4144 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_RENDER_TARGETS
;
4146 /* Our implementation of VK_KHR_multiview uses instancing to draw the
4147 * different views. If the client asks for instancing, we need to use the
4148 * Instance Data Step Rate to ensure that we repeat the client's
4149 * per-instance data once for each view. Since this bit is in
4150 * VERTEX_BUFFER_STATE on gen7, we need to dirty vertex buffers at the top
4154 cmd_buffer
->state
.gfx
.vb_dirty
|= ~0;
4156 /* It is possible to start a render pass with an old pipeline. Because the
4157 * render pass and subpass index are both baked into the pipeline, this is
4158 * highly unlikely. In order to do so, it requires that you have a render
4159 * pass with a single subpass and that you use that render pass twice
4160 * back-to-back and use the same pipeline at the start of the second render
4161 * pass as at the end of the first. In order to avoid unpredictable issues
4162 * with this edge case, we just dirty the pipeline at the start of every
4165 cmd_buffer
->state
.gfx
.dirty
|= ANV_CMD_DIRTY_PIPELINE
;
4167 /* Accumulate any subpass flushes that need to happen before the subpass */
4168 cmd_buffer
->state
.pending_pipe_bits
|=
4169 cmd_buffer
->state
.pass
->subpass_flushes
[subpass_id
];
4171 VkRect2D render_area
= cmd_buffer
->state
.render_area
;
4172 struct anv_framebuffer
*fb
= cmd_buffer
->state
.framebuffer
;
4174 bool is_multiview
= subpass
->view_mask
!= 0;
4176 for (uint32_t i
= 0; i
< subpass
->attachment_count
; ++i
) {
4177 const uint32_t a
= subpass
->attachments
[i
].attachment
;
4178 if (a
== VK_ATTACHMENT_UNUSED
)
4181 assert(a
< cmd_state
->pass
->attachment_count
);
4182 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[a
];
4184 struct anv_image_view
*iview
= cmd_state
->attachments
[a
].image_view
;
4185 const struct anv_image
*image
= iview
->image
;
4187 /* A resolve is necessary before use as an input attachment if the clear
4188 * color or auxiliary buffer usage isn't supported by the sampler.
4190 const bool input_needs_resolve
=
4191 (att_state
->fast_clear
&& !att_state
->clear_color_is_zero_one
) ||
4192 att_state
->input_aux_usage
!= att_state
->aux_usage
;
4194 VkImageLayout target_layout
;
4195 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
&&
4196 !input_needs_resolve
) {
4197 /* Layout transitions before the final only help to enable sampling
4198 * as an input attachment. If the input attachment supports sampling
4199 * using the auxiliary surface, we can skip such transitions by
4200 * making the target layout one that is CCS-aware.
4202 target_layout
= VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL
;
4204 target_layout
= subpass
->attachments
[i
].layout
;
4207 uint32_t base_layer
, layer_count
;
4208 if (image
->type
== VK_IMAGE_TYPE_3D
) {
4210 layer_count
= anv_minify(iview
->image
->extent
.depth
,
4211 iview
->planes
[0].isl
.base_level
);
4213 base_layer
= iview
->planes
[0].isl
.base_array_layer
;
4214 layer_count
= fb
->layers
;
4217 if (image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
4218 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
4219 transition_color_buffer(cmd_buffer
, image
, VK_IMAGE_ASPECT_COLOR_BIT
,
4220 iview
->planes
[0].isl
.base_level
, 1,
4221 base_layer
, layer_count
,
4222 att_state
->current_layout
, target_layout
);
4225 if (image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
4226 transition_depth_buffer(cmd_buffer
, image
,
4227 att_state
->current_layout
, target_layout
);
4228 att_state
->aux_usage
=
4229 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, image
,
4230 VK_IMAGE_ASPECT_DEPTH_BIT
, target_layout
);
4233 if (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
4234 transition_stencil_buffer(cmd_buffer
, image
,
4235 iview
->planes
[0].isl
.base_level
, 1,
4236 base_layer
, layer_count
,
4237 att_state
->current_layout
, target_layout
);
4239 att_state
->current_layout
= target_layout
;
4241 if (att_state
->pending_clear_aspects
& VK_IMAGE_ASPECT_COLOR_BIT
) {
4242 assert(att_state
->pending_clear_aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
4244 /* Multi-planar images are not supported as attachments */
4245 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
4246 assert(image
->n_planes
== 1);
4248 uint32_t base_clear_layer
= iview
->planes
[0].isl
.base_array_layer
;
4249 uint32_t clear_layer_count
= fb
->layers
;
4251 if (att_state
->fast_clear
&&
4252 do_first_layer_clear(cmd_state
, att_state
)) {
4253 /* We only support fast-clears on the first layer */
4254 assert(iview
->planes
[0].isl
.base_level
== 0);
4255 assert(iview
->planes
[0].isl
.base_array_layer
== 0);
4257 union isl_color_value clear_color
= {};
4258 anv_clear_color_from_att_state(&clear_color
, att_state
, iview
);
4259 if (iview
->image
->samples
== 1) {
4260 anv_image_ccs_op(cmd_buffer
, image
,
4261 iview
->planes
[0].isl
.format
,
4262 VK_IMAGE_ASPECT_COLOR_BIT
,
4263 0, 0, 1, ISL_AUX_OP_FAST_CLEAR
,
4267 anv_image_mcs_op(cmd_buffer
, image
,
4268 iview
->planes
[0].isl
.format
,
4269 VK_IMAGE_ASPECT_COLOR_BIT
,
4270 0, 1, ISL_AUX_OP_FAST_CLEAR
,
4275 clear_layer_count
--;
4277 att_state
->pending_clear_views
&= ~1;
4279 if (att_state
->clear_color_is_zero
) {
4280 /* This image has the auxiliary buffer enabled. We can mark the
4281 * subresource as not needing a resolve because the clear color
4282 * will match what's in every RENDER_SURFACE_STATE object when
4283 * it's being used for sampling.
4285 set_image_fast_clear_state(cmd_buffer
, iview
->image
,
4286 VK_IMAGE_ASPECT_COLOR_BIT
,
4287 ANV_FAST_CLEAR_DEFAULT_VALUE
);
4289 set_image_fast_clear_state(cmd_buffer
, iview
->image
,
4290 VK_IMAGE_ASPECT_COLOR_BIT
,
4291 ANV_FAST_CLEAR_ANY
);
4295 /* From the VkFramebufferCreateInfo spec:
4297 * "If the render pass uses multiview, then layers must be one and each
4298 * attachment requires a number of layers that is greater than the
4299 * maximum bit index set in the view mask in the subpasses in which it
4302 * So if multiview is active we ignore the number of layers in the
4303 * framebuffer and instead we honor the view mask from the subpass.
4306 assert(image
->n_planes
== 1);
4307 uint32_t pending_clear_mask
=
4308 get_multiview_subpass_clear_mask(cmd_state
, att_state
);
4311 for_each_bit(layer_idx
, pending_clear_mask
) {
4313 iview
->planes
[0].isl
.base_array_layer
+ layer_idx
;
4315 anv_image_clear_color(cmd_buffer
, image
,
4316 VK_IMAGE_ASPECT_COLOR_BIT
,
4317 att_state
->aux_usage
,
4318 iview
->planes
[0].isl
.format
,
4319 iview
->planes
[0].isl
.swizzle
,
4320 iview
->planes
[0].isl
.base_level
,
4323 vk_to_isl_color(att_state
->clear_value
.color
));
4326 att_state
->pending_clear_views
&= ~pending_clear_mask
;
4327 } else if (clear_layer_count
> 0) {
4328 assert(image
->n_planes
== 1);
4329 anv_image_clear_color(cmd_buffer
, image
, VK_IMAGE_ASPECT_COLOR_BIT
,
4330 att_state
->aux_usage
,
4331 iview
->planes
[0].isl
.format
,
4332 iview
->planes
[0].isl
.swizzle
,
4333 iview
->planes
[0].isl
.base_level
,
4334 base_clear_layer
, clear_layer_count
,
4336 vk_to_isl_color(att_state
->clear_value
.color
));
4338 } else if (att_state
->pending_clear_aspects
& (VK_IMAGE_ASPECT_DEPTH_BIT
|
4339 VK_IMAGE_ASPECT_STENCIL_BIT
)) {
4340 if (att_state
->fast_clear
&& !is_multiview
) {
4341 /* We currently only support HiZ for single-layer images */
4342 if (att_state
->pending_clear_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
4343 assert(iview
->image
->planes
[0].aux_usage
== ISL_AUX_USAGE_HIZ
);
4344 assert(iview
->planes
[0].isl
.base_level
== 0);
4345 assert(iview
->planes
[0].isl
.base_array_layer
== 0);
4346 assert(fb
->layers
== 1);
4349 anv_image_hiz_clear(cmd_buffer
, image
,
4350 att_state
->pending_clear_aspects
,
4351 iview
->planes
[0].isl
.base_level
,
4352 iview
->planes
[0].isl
.base_array_layer
,
4353 fb
->layers
, render_area
,
4354 att_state
->clear_value
.depthStencil
.stencil
);
4355 } else if (is_multiview
) {
4356 uint32_t pending_clear_mask
=
4357 get_multiview_subpass_clear_mask(cmd_state
, att_state
);
4360 for_each_bit(layer_idx
, pending_clear_mask
) {
4362 iview
->planes
[0].isl
.base_array_layer
+ layer_idx
;
4364 anv_image_clear_depth_stencil(cmd_buffer
, image
,
4365 att_state
->pending_clear_aspects
,
4366 att_state
->aux_usage
,
4367 iview
->planes
[0].isl
.base_level
,
4370 att_state
->clear_value
.depthStencil
.depth
,
4371 att_state
->clear_value
.depthStencil
.stencil
);
4374 att_state
->pending_clear_views
&= ~pending_clear_mask
;
4376 anv_image_clear_depth_stencil(cmd_buffer
, image
,
4377 att_state
->pending_clear_aspects
,
4378 att_state
->aux_usage
,
4379 iview
->planes
[0].isl
.base_level
,
4380 iview
->planes
[0].isl
.base_array_layer
,
4381 fb
->layers
, render_area
,
4382 att_state
->clear_value
.depthStencil
.depth
,
4383 att_state
->clear_value
.depthStencil
.stencil
);
4386 assert(att_state
->pending_clear_aspects
== 0);
4390 (att_state
->pending_load_aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) &&
4391 image
->planes
[0].aux_surface
.isl
.size_B
> 0 &&
4392 iview
->planes
[0].isl
.base_level
== 0 &&
4393 iview
->planes
[0].isl
.base_array_layer
== 0) {
4394 if (att_state
->aux_usage
!= ISL_AUX_USAGE_NONE
) {
4395 genX(copy_fast_clear_dwords
)(cmd_buffer
, att_state
->color
.state
,
4396 image
, VK_IMAGE_ASPECT_COLOR_BIT
,
4397 false /* copy to ss */);
4400 if (need_input_attachment_state(&cmd_state
->pass
->attachments
[a
]) &&
4401 att_state
->input_aux_usage
!= ISL_AUX_USAGE_NONE
) {
4402 genX(copy_fast_clear_dwords
)(cmd_buffer
, att_state
->input
.state
,
4403 image
, VK_IMAGE_ASPECT_COLOR_BIT
,
4404 false /* copy to ss */);
4408 if (subpass
->attachments
[i
].usage
==
4409 VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT
) {
4410 /* We assume that if we're starting a subpass, we're going to do some
4411 * rendering so we may end up with compressed data.
4413 genX(cmd_buffer_mark_image_written
)(cmd_buffer
, iview
->image
,
4414 VK_IMAGE_ASPECT_COLOR_BIT
,
4415 att_state
->aux_usage
,
4416 iview
->planes
[0].isl
.base_level
,
4417 iview
->planes
[0].isl
.base_array_layer
,
4419 } else if (subpass
->attachments
[i
].usage
==
4420 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
) {
4421 /* We may be writing depth or stencil so we need to mark the surface.
4422 * Unfortunately, there's no way to know at this point whether the
4423 * depth or stencil tests used will actually write to the surface.
4425 * Even though stencil may be plane 1, it always shares a base_level
4428 const struct isl_view
*ds_view
= &iview
->planes
[0].isl
;
4429 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
4430 genX(cmd_buffer_mark_image_written
)(cmd_buffer
, image
,
4431 VK_IMAGE_ASPECT_DEPTH_BIT
,
4432 att_state
->aux_usage
,
4433 ds_view
->base_level
,
4434 ds_view
->base_array_layer
,
4437 if (iview
->aspect_mask
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
4438 /* Even though stencil may be plane 1, it always shares a
4439 * base_level with depth.
4441 genX(cmd_buffer_mark_image_written
)(cmd_buffer
, image
,
4442 VK_IMAGE_ASPECT_STENCIL_BIT
,
4444 ds_view
->base_level
,
4445 ds_view
->base_array_layer
,
4450 /* If multiview is enabled, then we are only done clearing when we no
4451 * longer have pending layers to clear, or when we have processed the
4452 * last subpass that uses this attachment.
4454 if (!is_multiview
||
4455 att_state
->pending_clear_views
== 0 ||
4456 current_subpass_is_last_for_attachment(cmd_state
, a
)) {
4457 att_state
->pending_clear_aspects
= 0;
4460 att_state
->pending_load_aspects
= 0;
4463 cmd_buffer_emit_depth_stencil(cmd_buffer
);
4466 static enum blorp_filter
4467 vk_to_blorp_resolve_mode(VkResolveModeFlagBitsKHR vk_mode
)
4470 case VK_RESOLVE_MODE_SAMPLE_ZERO_BIT_KHR
:
4471 return BLORP_FILTER_SAMPLE_0
;
4472 case VK_RESOLVE_MODE_AVERAGE_BIT_KHR
:
4473 return BLORP_FILTER_AVERAGE
;
4474 case VK_RESOLVE_MODE_MIN_BIT_KHR
:
4475 return BLORP_FILTER_MIN_SAMPLE
;
4476 case VK_RESOLVE_MODE_MAX_BIT_KHR
:
4477 return BLORP_FILTER_MAX_SAMPLE
;
4479 return BLORP_FILTER_NONE
;
4484 cmd_buffer_end_subpass(struct anv_cmd_buffer
*cmd_buffer
)
4486 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
4487 struct anv_subpass
*subpass
= cmd_state
->subpass
;
4488 uint32_t subpass_id
= anv_get_subpass_id(&cmd_buffer
->state
);
4489 struct anv_framebuffer
*fb
= cmd_buffer
->state
.framebuffer
;
4491 if (subpass
->has_color_resolve
) {
4492 /* We are about to do some MSAA resolves. We need to flush so that the
4493 * result of writes to the MSAA color attachments show up in the sampler
4494 * when we blit to the single-sampled resolve target.
4496 cmd_buffer
->state
.pending_pipe_bits
|=
4497 ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
|
4498 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
;
4500 for (uint32_t i
= 0; i
< subpass
->color_count
; ++i
) {
4501 uint32_t src_att
= subpass
->color_attachments
[i
].attachment
;
4502 uint32_t dst_att
= subpass
->resolve_attachments
[i
].attachment
;
4504 if (dst_att
== VK_ATTACHMENT_UNUSED
)
4507 assert(src_att
< cmd_buffer
->state
.pass
->attachment_count
);
4508 assert(dst_att
< cmd_buffer
->state
.pass
->attachment_count
);
4510 if (cmd_buffer
->state
.attachments
[dst_att
].pending_clear_aspects
) {
4511 /* From the Vulkan 1.0 spec:
4513 * If the first use of an attachment in a render pass is as a
4514 * resolve attachment, then the loadOp is effectively ignored
4515 * as the resolve is guaranteed to overwrite all pixels in the
4518 cmd_buffer
->state
.attachments
[dst_att
].pending_clear_aspects
= 0;
4521 struct anv_image_view
*src_iview
= cmd_state
->attachments
[src_att
].image_view
;
4522 struct anv_image_view
*dst_iview
= cmd_state
->attachments
[dst_att
].image_view
;
4524 const VkRect2D render_area
= cmd_buffer
->state
.render_area
;
4526 enum isl_aux_usage src_aux_usage
=
4527 cmd_buffer
->state
.attachments
[src_att
].aux_usage
;
4528 enum isl_aux_usage dst_aux_usage
=
4529 cmd_buffer
->state
.attachments
[dst_att
].aux_usage
;
4531 assert(src_iview
->aspect_mask
== VK_IMAGE_ASPECT_COLOR_BIT
&&
4532 dst_iview
->aspect_mask
== VK_IMAGE_ASPECT_COLOR_BIT
);
4534 anv_image_msaa_resolve(cmd_buffer
,
4535 src_iview
->image
, src_aux_usage
,
4536 src_iview
->planes
[0].isl
.base_level
,
4537 src_iview
->planes
[0].isl
.base_array_layer
,
4538 dst_iview
->image
, dst_aux_usage
,
4539 dst_iview
->planes
[0].isl
.base_level
,
4540 dst_iview
->planes
[0].isl
.base_array_layer
,
4541 VK_IMAGE_ASPECT_COLOR_BIT
,
4542 render_area
.offset
.x
, render_area
.offset
.y
,
4543 render_area
.offset
.x
, render_area
.offset
.y
,
4544 render_area
.extent
.width
,
4545 render_area
.extent
.height
,
4546 fb
->layers
, BLORP_FILTER_NONE
);
4550 if (subpass
->ds_resolve_attachment
) {
4551 /* We are about to do some MSAA resolves. We need to flush so that the
4552 * result of writes to the MSAA depth attachments show up in the sampler
4553 * when we blit to the single-sampled resolve target.
4555 cmd_buffer
->state
.pending_pipe_bits
|=
4556 ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
|
4557 ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
;
4559 uint32_t src_att
= subpass
->depth_stencil_attachment
->attachment
;
4560 uint32_t dst_att
= subpass
->ds_resolve_attachment
->attachment
;
4562 assert(src_att
< cmd_buffer
->state
.pass
->attachment_count
);
4563 assert(dst_att
< cmd_buffer
->state
.pass
->attachment_count
);
4565 if (cmd_buffer
->state
.attachments
[dst_att
].pending_clear_aspects
) {
4566 /* From the Vulkan 1.0 spec:
4568 * If the first use of an attachment in a render pass is as a
4569 * resolve attachment, then the loadOp is effectively ignored
4570 * as the resolve is guaranteed to overwrite all pixels in the
4573 cmd_buffer
->state
.attachments
[dst_att
].pending_clear_aspects
= 0;
4576 struct anv_image_view
*src_iview
= cmd_state
->attachments
[src_att
].image_view
;
4577 struct anv_image_view
*dst_iview
= cmd_state
->attachments
[dst_att
].image_view
;
4579 const VkRect2D render_area
= cmd_buffer
->state
.render_area
;
4581 if ((src_iview
->image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) &&
4582 subpass
->depth_resolve_mode
!= VK_RESOLVE_MODE_NONE_KHR
) {
4584 struct anv_attachment_state
*src_state
=
4585 &cmd_state
->attachments
[src_att
];
4586 struct anv_attachment_state
*dst_state
=
4587 &cmd_state
->attachments
[dst_att
];
4589 /* MSAA resolves sample from the source attachment. Transition the
4590 * depth attachment first to get rid of any HiZ that we may not be
4593 transition_depth_buffer(cmd_buffer
, src_iview
->image
,
4594 src_state
->current_layout
,
4595 VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
);
4596 src_state
->aux_usage
=
4597 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, src_iview
->image
,
4598 VK_IMAGE_ASPECT_DEPTH_BIT
,
4599 VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
);
4600 src_state
->current_layout
= VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
;
4602 /* MSAA resolves write to the resolve attachment as if it were any
4603 * other transfer op. Transition the resolve attachment accordingly.
4605 VkImageLayout dst_initial_layout
= dst_state
->current_layout
;
4607 /* If our render area is the entire size of the image, we're going to
4608 * blow it all away so we can claim the initial layout is UNDEFINED
4609 * and we'll get a HiZ ambiguate instead of a resolve.
4611 if (dst_iview
->image
->type
!= VK_IMAGE_TYPE_3D
&&
4612 render_area
.offset
.x
== 0 && render_area
.offset
.y
== 0 &&
4613 render_area
.extent
.width
== dst_iview
->extent
.width
&&
4614 render_area
.extent
.height
== dst_iview
->extent
.height
)
4615 dst_initial_layout
= VK_IMAGE_LAYOUT_UNDEFINED
;
4617 transition_depth_buffer(cmd_buffer
, dst_iview
->image
,
4619 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
);
4620 dst_state
->aux_usage
=
4621 anv_layout_to_aux_usage(&cmd_buffer
->device
->info
, dst_iview
->image
,
4622 VK_IMAGE_ASPECT_DEPTH_BIT
,
4623 VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
);
4624 dst_state
->current_layout
= VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
;
4626 enum blorp_filter filter
=
4627 vk_to_blorp_resolve_mode(subpass
->depth_resolve_mode
);
4629 anv_image_msaa_resolve(cmd_buffer
,
4630 src_iview
->image
, src_state
->aux_usage
,
4631 src_iview
->planes
[0].isl
.base_level
,
4632 src_iview
->planes
[0].isl
.base_array_layer
,
4633 dst_iview
->image
, dst_state
->aux_usage
,
4634 dst_iview
->planes
[0].isl
.base_level
,
4635 dst_iview
->planes
[0].isl
.base_array_layer
,
4636 VK_IMAGE_ASPECT_DEPTH_BIT
,
4637 render_area
.offset
.x
, render_area
.offset
.y
,
4638 render_area
.offset
.x
, render_area
.offset
.y
,
4639 render_area
.extent
.width
,
4640 render_area
.extent
.height
,
4641 fb
->layers
, filter
);
4644 if ((src_iview
->image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) &&
4645 subpass
->stencil_resolve_mode
!= VK_RESOLVE_MODE_NONE_KHR
) {
4647 enum isl_aux_usage src_aux_usage
= ISL_AUX_USAGE_NONE
;
4648 enum isl_aux_usage dst_aux_usage
= ISL_AUX_USAGE_NONE
;
4650 enum blorp_filter filter
=
4651 vk_to_blorp_resolve_mode(subpass
->stencil_resolve_mode
);
4653 anv_image_msaa_resolve(cmd_buffer
,
4654 src_iview
->image
, src_aux_usage
,
4655 src_iview
->planes
[0].isl
.base_level
,
4656 src_iview
->planes
[0].isl
.base_array_layer
,
4657 dst_iview
->image
, dst_aux_usage
,
4658 dst_iview
->planes
[0].isl
.base_level
,
4659 dst_iview
->planes
[0].isl
.base_array_layer
,
4660 VK_IMAGE_ASPECT_STENCIL_BIT
,
4661 render_area
.offset
.x
, render_area
.offset
.y
,
4662 render_area
.offset
.x
, render_area
.offset
.y
,
4663 render_area
.extent
.width
,
4664 render_area
.extent
.height
,
4665 fb
->layers
, filter
);
4670 /* On gen7, we have to store a texturable version of the stencil buffer in
4671 * a shadow whenever VK_IMAGE_USAGE_SAMPLED_BIT is set and copy back and
4672 * forth at strategic points. Stencil writes are only allowed in three
4675 * - VK_IMAGE_LAYOUT_GENERAL
4676 * - VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
4677 * - VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
4678 * - VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
4680 * For general, we have no nice opportunity to transition so we do the copy
4681 * to the shadow unconditionally at the end of the subpass. For transfer
4682 * destinations, we can update it as part of the transfer op. For the
4683 * other two, we delay the copy until a transition into some other layout.
4685 if (subpass
->depth_stencil_attachment
) {
4686 uint32_t a
= subpass
->depth_stencil_attachment
->attachment
;
4687 assert(a
!= VK_ATTACHMENT_UNUSED
);
4689 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[a
];
4690 struct anv_image_view
*iview
= cmd_state
->attachments
[a
].image_view
;;
4691 const struct anv_image
*image
= iview
->image
;
4693 if (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
4694 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
,
4695 VK_IMAGE_ASPECT_STENCIL_BIT
);
4697 if (image
->planes
[plane
].shadow_surface
.isl
.size_B
> 0 &&
4698 att_state
->current_layout
== VK_IMAGE_LAYOUT_GENERAL
) {
4699 assert(image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
);
4700 anv_image_copy_to_shadow(cmd_buffer
, image
,
4701 VK_IMAGE_ASPECT_STENCIL_BIT
,
4702 iview
->planes
[plane
].isl
.base_level
, 1,
4703 iview
->planes
[plane
].isl
.base_array_layer
,
4708 #endif /* GEN_GEN == 7 */
4710 for (uint32_t i
= 0; i
< subpass
->attachment_count
; ++i
) {
4711 const uint32_t a
= subpass
->attachments
[i
].attachment
;
4712 if (a
== VK_ATTACHMENT_UNUSED
)
4715 if (cmd_state
->pass
->attachments
[a
].last_subpass_idx
!= subpass_id
)
4718 assert(a
< cmd_state
->pass
->attachment_count
);
4719 struct anv_attachment_state
*att_state
= &cmd_state
->attachments
[a
];
4720 struct anv_image_view
*iview
= cmd_state
->attachments
[a
].image_view
;
4721 const struct anv_image
*image
= iview
->image
;
4723 if ((image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) &&
4724 image
->vk_format
!= iview
->vk_format
) {
4725 enum anv_fast_clear_type fast_clear_type
=
4726 anv_layout_to_fast_clear_type(&cmd_buffer
->device
->info
,
4727 image
, VK_IMAGE_ASPECT_COLOR_BIT
,
4728 att_state
->current_layout
);
4730 /* If any clear color was used, flush it down the aux surfaces. If we
4731 * don't do it now using the view's format we might use the clear
4732 * color incorrectly in the following resolves (for example with an
4733 * SRGB view & a UNORM image).
4735 if (fast_clear_type
!= ANV_FAST_CLEAR_NONE
) {
4736 anv_perf_warn(cmd_buffer
->device
->instance
, iview
,
4737 "Doing a partial resolve to get rid of clear color at the "
4738 "end of a renderpass due to an image/view format mismatch");
4740 uint32_t base_layer
, layer_count
;
4741 if (image
->type
== VK_IMAGE_TYPE_3D
) {
4743 layer_count
= anv_minify(iview
->image
->extent
.depth
,
4744 iview
->planes
[0].isl
.base_level
);
4746 base_layer
= iview
->planes
[0].isl
.base_array_layer
;
4747 layer_count
= fb
->layers
;
4750 for (uint32_t a
= 0; a
< layer_count
; a
++) {
4751 uint32_t array_layer
= base_layer
+ a
;
4752 if (image
->samples
== 1) {
4753 anv_cmd_predicated_ccs_resolve(cmd_buffer
, image
,
4754 iview
->planes
[0].isl
.format
,
4755 VK_IMAGE_ASPECT_COLOR_BIT
,
4756 iview
->planes
[0].isl
.base_level
,
4758 ISL_AUX_OP_PARTIAL_RESOLVE
,
4759 ANV_FAST_CLEAR_NONE
);
4761 anv_cmd_predicated_mcs_resolve(cmd_buffer
, image
,
4762 iview
->planes
[0].isl
.format
,
4763 VK_IMAGE_ASPECT_COLOR_BIT
,
4765 ISL_AUX_OP_PARTIAL_RESOLVE
,
4766 ANV_FAST_CLEAR_NONE
);
4772 /* Transition the image into the final layout for this render pass */
4773 VkImageLayout target_layout
=
4774 cmd_state
->pass
->attachments
[a
].final_layout
;
4776 uint32_t base_layer
, layer_count
;
4777 if (image
->type
== VK_IMAGE_TYPE_3D
) {
4779 layer_count
= anv_minify(iview
->image
->extent
.depth
,
4780 iview
->planes
[0].isl
.base_level
);
4782 base_layer
= iview
->planes
[0].isl
.base_array_layer
;
4783 layer_count
= fb
->layers
;
4786 if (image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
4787 assert(image
->aspects
== VK_IMAGE_ASPECT_COLOR_BIT
);
4788 transition_color_buffer(cmd_buffer
, image
, VK_IMAGE_ASPECT_COLOR_BIT
,
4789 iview
->planes
[0].isl
.base_level
, 1,
4790 base_layer
, layer_count
,
4791 att_state
->current_layout
, target_layout
);
4794 if (image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) {
4795 transition_depth_buffer(cmd_buffer
, image
,
4796 att_state
->current_layout
, target_layout
);
4799 if (image
->aspects
& VK_IMAGE_ASPECT_STENCIL_BIT
) {
4800 transition_stencil_buffer(cmd_buffer
, image
,
4801 iview
->planes
[0].isl
.base_level
, 1,
4802 base_layer
, layer_count
,
4803 att_state
->current_layout
, target_layout
);
4807 /* Accumulate any subpass flushes that need to happen after the subpass.
4808 * Yes, they do get accumulated twice in the NextSubpass case but since
4809 * genX_CmdNextSubpass just calls end/begin back-to-back, we just end up
4810 * ORing the bits in twice so it's harmless.
4812 cmd_buffer
->state
.pending_pipe_bits
|=
4813 cmd_buffer
->state
.pass
->subpass_flushes
[subpass_id
+ 1];
4816 void genX(CmdBeginRenderPass
)(
4817 VkCommandBuffer commandBuffer
,
4818 const VkRenderPassBeginInfo
* pRenderPassBegin
,
4819 VkSubpassContents contents
)
4821 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4822 ANV_FROM_HANDLE(anv_render_pass
, pass
, pRenderPassBegin
->renderPass
);
4823 ANV_FROM_HANDLE(anv_framebuffer
, framebuffer
, pRenderPassBegin
->framebuffer
);
4825 cmd_buffer
->state
.framebuffer
= framebuffer
;
4826 cmd_buffer
->state
.pass
= pass
;
4827 cmd_buffer
->state
.render_area
= pRenderPassBegin
->renderArea
;
4829 genX(cmd_buffer_setup_attachments
)(cmd_buffer
, pass
, pRenderPassBegin
);
4831 /* If we failed to setup the attachments we should not try to go further */
4832 if (result
!= VK_SUCCESS
) {
4833 assert(anv_batch_has_error(&cmd_buffer
->batch
));
4837 genX(flush_pipeline_select_3d
)(cmd_buffer
);
4839 cmd_buffer_begin_subpass(cmd_buffer
, 0);
4842 void genX(CmdBeginRenderPass2KHR
)(
4843 VkCommandBuffer commandBuffer
,
4844 const VkRenderPassBeginInfo
* pRenderPassBeginInfo
,
4845 const VkSubpassBeginInfoKHR
* pSubpassBeginInfo
)
4847 genX(CmdBeginRenderPass
)(commandBuffer
, pRenderPassBeginInfo
,
4848 pSubpassBeginInfo
->contents
);
4851 void genX(CmdNextSubpass
)(
4852 VkCommandBuffer commandBuffer
,
4853 VkSubpassContents contents
)
4855 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4857 if (anv_batch_has_error(&cmd_buffer
->batch
))
4860 assert(cmd_buffer
->level
== VK_COMMAND_BUFFER_LEVEL_PRIMARY
);
4862 uint32_t prev_subpass
= anv_get_subpass_id(&cmd_buffer
->state
);
4863 cmd_buffer_end_subpass(cmd_buffer
);
4864 cmd_buffer_begin_subpass(cmd_buffer
, prev_subpass
+ 1);
4867 void genX(CmdNextSubpass2KHR
)(
4868 VkCommandBuffer commandBuffer
,
4869 const VkSubpassBeginInfoKHR
* pSubpassBeginInfo
,
4870 const VkSubpassEndInfoKHR
* pSubpassEndInfo
)
4872 genX(CmdNextSubpass
)(commandBuffer
, pSubpassBeginInfo
->contents
);
4875 void genX(CmdEndRenderPass
)(
4876 VkCommandBuffer commandBuffer
)
4878 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4880 if (anv_batch_has_error(&cmd_buffer
->batch
))
4883 cmd_buffer_end_subpass(cmd_buffer
);
4885 cmd_buffer
->state
.hiz_enabled
= false;
4888 anv_dump_add_attachments(cmd_buffer
);
4891 /* Remove references to render pass specific state. This enables us to
4892 * detect whether or not we're in a renderpass.
4894 cmd_buffer
->state
.framebuffer
= NULL
;
4895 cmd_buffer
->state
.pass
= NULL
;
4896 cmd_buffer
->state
.subpass
= NULL
;
4899 void genX(CmdEndRenderPass2KHR
)(
4900 VkCommandBuffer commandBuffer
,
4901 const VkSubpassEndInfoKHR
* pSubpassEndInfo
)
4903 genX(CmdEndRenderPass
)(commandBuffer
);
4907 genX(cmd_emit_conditional_render_predicate
)(struct anv_cmd_buffer
*cmd_buffer
)
4909 #if GEN_GEN >= 8 || GEN_IS_HASWELL
4910 struct gen_mi_builder b
;
4911 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
4913 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC0
),
4914 gen_mi_reg32(ANV_PREDICATE_RESULT_REG
));
4915 gen_mi_store(&b
, gen_mi_reg64(MI_PREDICATE_SRC1
), gen_mi_imm(0));
4917 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_PREDICATE
), mip
) {
4918 mip
.LoadOperation
= LOAD_LOADINV
;
4919 mip
.CombineOperation
= COMBINE_SET
;
4920 mip
.CompareOperation
= COMPARE_SRCS_EQUAL
;
4925 #if GEN_GEN >= 8 || GEN_IS_HASWELL
4926 void genX(CmdBeginConditionalRenderingEXT
)(
4927 VkCommandBuffer commandBuffer
,
4928 const VkConditionalRenderingBeginInfoEXT
* pConditionalRenderingBegin
)
4930 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4931 ANV_FROM_HANDLE(anv_buffer
, buffer
, pConditionalRenderingBegin
->buffer
);
4932 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
4933 struct anv_address value_address
=
4934 anv_address_add(buffer
->address
, pConditionalRenderingBegin
->offset
);
4936 const bool isInverted
= pConditionalRenderingBegin
->flags
&
4937 VK_CONDITIONAL_RENDERING_INVERTED_BIT_EXT
;
4939 cmd_state
->conditional_render_enabled
= true;
4941 genX(cmd_buffer_apply_pipe_flushes
)(cmd_buffer
);
4943 struct gen_mi_builder b
;
4944 gen_mi_builder_init(&b
, &cmd_buffer
->batch
);
4946 /* Section 19.4 of the Vulkan 1.1.85 spec says:
4948 * If the value of the predicate in buffer memory changes
4949 * while conditional rendering is active, the rendering commands
4950 * may be discarded in an implementation-dependent way.
4951 * Some implementations may latch the value of the predicate
4952 * upon beginning conditional rendering while others
4953 * may read it before every rendering command.
4955 * So it's perfectly fine to read a value from the buffer once.
4957 struct gen_mi_value value
= gen_mi_mem32(value_address
);
4959 /* Precompute predicate result, it is necessary to support secondary
4960 * command buffers since it is unknown if conditional rendering is
4961 * inverted when populating them.
4963 gen_mi_store(&b
, gen_mi_reg64(ANV_PREDICATE_RESULT_REG
),
4964 isInverted
? gen_mi_uge(&b
, gen_mi_imm(0), value
) :
4965 gen_mi_ult(&b
, gen_mi_imm(0), value
));
4968 void genX(CmdEndConditionalRenderingEXT
)(
4969 VkCommandBuffer commandBuffer
)
4971 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
4972 struct anv_cmd_state
*cmd_state
= &cmd_buffer
->state
;
4974 cmd_state
->conditional_render_enabled
= false;
4978 /* Set of stage bits for which are pipelined, i.e. they get queued by the
4979 * command streamer for later execution.
4981 #define ANV_PIPELINE_STAGE_PIPELINED_BITS \
4982 (VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | \
4983 VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | \
4984 VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT | \
4985 VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT | \
4986 VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT | \
4987 VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | \
4988 VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT | \
4989 VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT | \
4990 VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | \
4991 VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | \
4992 VK_PIPELINE_STAGE_TRANSFER_BIT | \
4993 VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT | \
4994 VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT | \
4995 VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)
4997 void genX(CmdSetEvent
)(
4998 VkCommandBuffer commandBuffer
,
5000 VkPipelineStageFlags stageMask
)
5002 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5003 ANV_FROM_HANDLE(anv_event
, event
, _event
);
5005 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
5006 if (stageMask
& ANV_PIPELINE_STAGE_PIPELINED_BITS
) {
5007 pc
.StallAtPixelScoreboard
= true;
5008 pc
.CommandStreamerStallEnable
= true;
5011 pc
.DestinationAddressType
= DAT_PPGTT
,
5012 pc
.PostSyncOperation
= WriteImmediateData
,
5013 pc
.Address
= (struct anv_address
) {
5014 cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
5017 pc
.ImmediateData
= VK_EVENT_SET
;
5021 void genX(CmdResetEvent
)(
5022 VkCommandBuffer commandBuffer
,
5024 VkPipelineStageFlags stageMask
)
5026 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5027 ANV_FROM_HANDLE(anv_event
, event
, _event
);
5029 anv_batch_emit(&cmd_buffer
->batch
, GENX(PIPE_CONTROL
), pc
) {
5030 if (stageMask
& ANV_PIPELINE_STAGE_PIPELINED_BITS
) {
5031 pc
.StallAtPixelScoreboard
= true;
5032 pc
.CommandStreamerStallEnable
= true;
5035 pc
.DestinationAddressType
= DAT_PPGTT
;
5036 pc
.PostSyncOperation
= WriteImmediateData
;
5037 pc
.Address
= (struct anv_address
) {
5038 cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
5041 pc
.ImmediateData
= VK_EVENT_RESET
;
5045 void genX(CmdWaitEvents
)(
5046 VkCommandBuffer commandBuffer
,
5047 uint32_t eventCount
,
5048 const VkEvent
* pEvents
,
5049 VkPipelineStageFlags srcStageMask
,
5050 VkPipelineStageFlags destStageMask
,
5051 uint32_t memoryBarrierCount
,
5052 const VkMemoryBarrier
* pMemoryBarriers
,
5053 uint32_t bufferMemoryBarrierCount
,
5054 const VkBufferMemoryBarrier
* pBufferMemoryBarriers
,
5055 uint32_t imageMemoryBarrierCount
,
5056 const VkImageMemoryBarrier
* pImageMemoryBarriers
)
5059 ANV_FROM_HANDLE(anv_cmd_buffer
, cmd_buffer
, commandBuffer
);
5061 for (uint32_t i
= 0; i
< eventCount
; i
++) {
5062 ANV_FROM_HANDLE(anv_event
, event
, pEvents
[i
]);
5064 anv_batch_emit(&cmd_buffer
->batch
, GENX(MI_SEMAPHORE_WAIT
), sem
) {
5065 sem
.WaitMode
= PollingMode
,
5066 sem
.CompareOperation
= COMPARE_SAD_EQUAL_SDD
,
5067 sem
.SemaphoreDataDword
= VK_EVENT_SET
,
5068 sem
.SemaphoreAddress
= (struct anv_address
) {
5069 cmd_buffer
->device
->dynamic_state_pool
.block_pool
.bo
,
5075 anv_finishme("Implement events on gen7");
5078 genX(CmdPipelineBarrier
)(commandBuffer
, srcStageMask
, destStageMask
,
5079 false, /* byRegion */
5080 memoryBarrierCount
, pMemoryBarriers
,
5081 bufferMemoryBarrierCount
, pBufferMemoryBarriers
,
5082 imageMemoryBarrierCount
, pImageMemoryBarriers
);