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
33 #include "drm-uapi/i915_drm.h"
40 #define __gen_validate_value(x) VALGRIND_CHECK_MEM_IS_DEFINED(&(x), sizeof(x))
43 #define VG(x) ((void)0)
46 #include "common/gen_clflush.h"
47 #include "common/gen_decoder.h"
48 #include "common/gen_gem.h"
49 #include "common/gen_l3_config.h"
50 #include "dev/gen_device_info.h"
51 #include "blorp/blorp.h"
52 #include "compiler/brw_compiler.h"
53 #include "util/bitset.h"
54 #include "util/macros.h"
55 #include "util/hash_table.h"
56 #include "util/list.h"
57 #include "util/sparse_array.h"
58 #include "util/u_atomic.h"
59 #include "util/u_vector.h"
60 #include "util/u_math.h"
62 #include "util/xmlconfig.h"
64 #include "vk_debug_report.h"
65 #include "vk_object.h"
67 /* Pre-declarations needed for WSI entrypoints */
70 typedef struct xcb_connection_t xcb_connection_t
;
71 typedef uint32_t xcb_visualid_t
;
72 typedef uint32_t xcb_window_t
;
76 struct anv_buffer_view
;
77 struct anv_image_view
;
80 struct gen_aux_map_context
;
81 struct gen_perf_config
;
82 struct gen_perf_counter_pass
;
83 struct gen_perf_query_result
;
85 #include <vulkan/vulkan.h>
86 #include <vulkan/vulkan_intel.h>
87 #include <vulkan/vk_icd.h>
89 #include "anv_android.h"
90 #include "anv_entrypoints.h"
91 #include "anv_extensions.h"
94 #include "dev/gen_debug.h"
95 #include "common/intel_log.h"
96 #include "wsi_common.h"
98 #define NSEC_PER_SEC 1000000000ull
100 /* anv Virtual Memory Layout
101 * =========================
103 * When the anv driver is determining the virtual graphics addresses of memory
104 * objects itself using the softpin mechanism, the following memory ranges
107 * Three special considerations to notice:
109 * (1) the dynamic state pool is located within the same 4 GiB as the low
110 * heap. This is to work around a VF cache issue described in a comment in
111 * anv_physical_device_init_heaps.
113 * (2) the binding table pool is located at lower addresses than the surface
114 * state pool, within a 4 GiB range. This allows surface state base addresses
115 * to cover both binding tables (16 bit offsets) and surface states (32 bit
118 * (3) the last 4 GiB of the address space is withheld from the high
119 * heap. Various hardware units will read past the end of an object for
120 * various reasons. This healthy margin prevents reads from wrapping around
123 #define LOW_HEAP_MIN_ADDRESS 0x000000001000ULL /* 4 KiB */
124 #define LOW_HEAP_MAX_ADDRESS 0x0000bfffffffULL
125 #define DYNAMIC_STATE_POOL_MIN_ADDRESS 0x0000c0000000ULL /* 3 GiB */
126 #define DYNAMIC_STATE_POOL_MAX_ADDRESS 0x0000ffffffffULL
127 #define BINDING_TABLE_POOL_MIN_ADDRESS 0x000100000000ULL /* 4 GiB */
128 #define BINDING_TABLE_POOL_MAX_ADDRESS 0x00013fffffffULL
129 #define SURFACE_STATE_POOL_MIN_ADDRESS 0x000140000000ULL /* 5 GiB */
130 #define SURFACE_STATE_POOL_MAX_ADDRESS 0x00017fffffffULL
131 #define INSTRUCTION_STATE_POOL_MIN_ADDRESS 0x000180000000ULL /* 6 GiB */
132 #define INSTRUCTION_STATE_POOL_MAX_ADDRESS 0x0001bfffffffULL
133 #define CLIENT_VISIBLE_HEAP_MIN_ADDRESS 0x0001c0000000ULL /* 7 GiB */
134 #define CLIENT_VISIBLE_HEAP_MAX_ADDRESS 0x0002bfffffffULL
135 #define HIGH_HEAP_MIN_ADDRESS 0x0002c0000000ULL /* 11 GiB */
137 #define LOW_HEAP_SIZE \
138 (LOW_HEAP_MAX_ADDRESS - LOW_HEAP_MIN_ADDRESS + 1)
139 #define DYNAMIC_STATE_POOL_SIZE \
140 (DYNAMIC_STATE_POOL_MAX_ADDRESS - DYNAMIC_STATE_POOL_MIN_ADDRESS + 1)
141 #define BINDING_TABLE_POOL_SIZE \
142 (BINDING_TABLE_POOL_MAX_ADDRESS - BINDING_TABLE_POOL_MIN_ADDRESS + 1)
143 #define SURFACE_STATE_POOL_SIZE \
144 (SURFACE_STATE_POOL_MAX_ADDRESS - SURFACE_STATE_POOL_MIN_ADDRESS + 1)
145 #define INSTRUCTION_STATE_POOL_SIZE \
146 (INSTRUCTION_STATE_POOL_MAX_ADDRESS - INSTRUCTION_STATE_POOL_MIN_ADDRESS + 1)
147 #define CLIENT_VISIBLE_HEAP_SIZE \
148 (CLIENT_VISIBLE_HEAP_MAX_ADDRESS - CLIENT_VISIBLE_HEAP_MIN_ADDRESS + 1)
150 /* Allowing different clear colors requires us to perform a depth resolve at
151 * the end of certain render passes. This is because while slow clears store
152 * the clear color in the HiZ buffer, fast clears (without a resolve) don't.
153 * See the PRMs for examples describing when additional resolves would be
154 * necessary. To enable fast clears without requiring extra resolves, we set
155 * the clear value to a globally-defined one. We could allow different values
156 * if the user doesn't expect coherent data during or after a render passes
157 * (VK_ATTACHMENT_STORE_OP_DONT_CARE), but such users (aside from the CTS)
158 * don't seem to exist yet. In almost all Vulkan applications tested thus far,
159 * 1.0f seems to be the only value used. The only application that doesn't set
160 * this value does so through the usage of an seemingly uninitialized clear
163 #define ANV_HZ_FC_VAL 1.0f
166 #define MAX_XFB_BUFFERS 4
167 #define MAX_XFB_STREAMS 4
170 #define MAX_VIEWPORTS 16
171 #define MAX_SCISSORS 16
172 #define MAX_PUSH_CONSTANTS_SIZE 128
173 #define MAX_DYNAMIC_BUFFERS 16
174 #define MAX_IMAGES 64
175 #define MAX_PUSH_DESCRIPTORS 32 /* Minimum requirement */
176 #define MAX_INLINE_UNIFORM_BLOCK_SIZE 4096
177 #define MAX_INLINE_UNIFORM_BLOCK_DESCRIPTORS 32
178 /* We need 16 for UBO block reads to work and 32 for push UBOs. However, we
179 * use 64 here to avoid cache issues. This could most likely bring it back to
180 * 32 if we had different virtual addresses for the different views on a given
183 #define ANV_UBO_ALIGNMENT 64
184 #define ANV_SSBO_BOUNDS_CHECK_ALIGNMENT 4
185 #define MAX_VIEWS_FOR_PRIMITIVE_REPLICATION 16
187 /* From the Skylake PRM Vol. 7 "Binding Table Surface State Model":
189 * "The surface state model is used when a Binding Table Index (specified
190 * in the message descriptor) of less than 240 is specified. In this model,
191 * the Binding Table Index is used to index into the binding table, and the
192 * binding table entry contains a pointer to the SURFACE_STATE."
194 * Binding table values above 240 are used for various things in the hardware
195 * such as stateless, stateless with incoherent cache, SLM, and bindless.
197 #define MAX_BINDING_TABLE_SIZE 240
199 /* The kernel relocation API has a limitation of a 32-bit delta value
200 * applied to the address before it is written which, in spite of it being
201 * unsigned, is treated as signed . Because of the way that this maps to
202 * the Vulkan API, we cannot handle an offset into a buffer that does not
203 * fit into a signed 32 bits. The only mechanism we have for dealing with
204 * this at the moment is to limit all VkDeviceMemory objects to a maximum
205 * of 2GB each. The Vulkan spec allows us to do this:
207 * "Some platforms may have a limit on the maximum size of a single
208 * allocation. For example, certain systems may fail to create
209 * allocations with a size greater than or equal to 4GB. Such a limit is
210 * implementation-dependent, and if such a failure occurs then the error
211 * VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
213 * We don't use vk_error here because it's not an error so much as an
214 * indication to the application that the allocation is too large.
216 #define MAX_MEMORY_ALLOCATION_SIZE (1ull << 31)
218 #define ANV_SVGS_VB_INDEX MAX_VBS
219 #define ANV_DRAWID_VB_INDEX (MAX_VBS + 1)
221 /* We reserve this MI ALU register for the purpose of handling predication.
222 * Other code which uses the MI ALU should leave it alone.
224 #define ANV_PREDICATE_RESULT_REG 0x2678 /* MI_ALU_REG15 */
226 /* We reserve this MI ALU register to pass around an offset computed from
227 * VkPerformanceQuerySubmitInfoKHR::counterPassIndex VK_KHR_performance_query.
228 * Other code which uses the MI ALU should leave it alone.
230 #define ANV_PERF_QUERY_OFFSET_REG 0x2670 /* MI_ALU_REG14 */
232 /* For gen12 we set the streamout buffers using 4 separate commands
233 * (3DSTATE_SO_BUFFER_INDEX_*) instead of 3DSTATE_SO_BUFFER. However the layout
234 * of the 3DSTATE_SO_BUFFER_INDEX_* commands is identical to that of
235 * 3DSTATE_SO_BUFFER apart from the SOBufferIndex field, so for now we use the
236 * 3DSTATE_SO_BUFFER command, but change the 3DCommandSubOpcode.
237 * SO_BUFFER_INDEX_0_CMD is actually the 3DCommandSubOpcode for
238 * 3DSTATE_SO_BUFFER_INDEX_0.
240 #define SO_BUFFER_INDEX_0_CMD 0x60
241 #define anv_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
243 static inline uint32_t
244 align_down_npot_u32(uint32_t v
, uint32_t a
)
249 static inline uint32_t
250 align_down_u32(uint32_t v
, uint32_t a
)
252 assert(a
!= 0 && a
== (a
& -a
));
256 static inline uint32_t
257 align_u32(uint32_t v
, uint32_t a
)
259 assert(a
!= 0 && a
== (a
& -a
));
260 return align_down_u32(v
+ a
- 1, a
);
263 static inline uint64_t
264 align_down_u64(uint64_t v
, uint64_t a
)
266 assert(a
!= 0 && a
== (a
& -a
));
270 static inline uint64_t
271 align_u64(uint64_t v
, uint64_t a
)
273 return align_down_u64(v
+ a
- 1, a
);
276 static inline int32_t
277 align_i32(int32_t v
, int32_t a
)
279 assert(a
!= 0 && a
== (a
& -a
));
280 return (v
+ a
- 1) & ~(a
- 1);
283 /** Alignment must be a power of 2. */
285 anv_is_aligned(uintmax_t n
, uintmax_t a
)
287 assert(a
== (a
& -a
));
288 return (n
& (a
- 1)) == 0;
291 static inline uint32_t
292 anv_minify(uint32_t n
, uint32_t levels
)
294 if (unlikely(n
== 0))
297 return MAX2(n
>> levels
, 1);
301 anv_clamp_f(float f
, float min
, float max
)
314 anv_clear_mask(uint32_t *inout_mask
, uint32_t clear_mask
)
316 if (*inout_mask
& clear_mask
) {
317 *inout_mask
&= ~clear_mask
;
324 static inline union isl_color_value
325 vk_to_isl_color(VkClearColorValue color
)
327 return (union isl_color_value
) {
337 static inline void *anv_unpack_ptr(uintptr_t ptr
, int bits
, int *flags
)
339 uintptr_t mask
= (1ull << bits
) - 1;
341 return (void *) (ptr
& ~mask
);
344 static inline uintptr_t anv_pack_ptr(void *ptr
, int bits
, int flags
)
346 uintptr_t value
= (uintptr_t) ptr
;
347 uintptr_t mask
= (1ull << bits
) - 1;
348 return value
| (mask
& flags
);
351 #define for_each_bit(b, dword) \
352 for (uint32_t __dword = (dword); \
353 (b) = __builtin_ffs(__dword) - 1, __dword; \
354 __dword &= ~(1 << (b)))
356 #define typed_memcpy(dest, src, count) ({ \
357 STATIC_ASSERT(sizeof(*src) == sizeof(*dest)); \
358 memcpy((dest), (src), (count) * sizeof(*(src))); \
361 /* Mapping from anv object to VkDebugReportObjectTypeEXT. New types need
362 * to be added here in order to utilize mapping in debug/error/perf macros.
364 #define REPORT_OBJECT_TYPE(o) \
365 __builtin_choose_expr ( \
366 __builtin_types_compatible_p (__typeof (o), struct anv_instance*), \
367 VK_DEBUG_REPORT_OBJECT_TYPE_INSTANCE_EXT, \
368 __builtin_choose_expr ( \
369 __builtin_types_compatible_p (__typeof (o), struct anv_physical_device*), \
370 VK_DEBUG_REPORT_OBJECT_TYPE_PHYSICAL_DEVICE_EXT, \
371 __builtin_choose_expr ( \
372 __builtin_types_compatible_p (__typeof (o), struct anv_device*), \
373 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT, \
374 __builtin_choose_expr ( \
375 __builtin_types_compatible_p (__typeof (o), const struct anv_device*), \
376 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_EXT, \
377 __builtin_choose_expr ( \
378 __builtin_types_compatible_p (__typeof (o), struct anv_queue*), \
379 VK_DEBUG_REPORT_OBJECT_TYPE_QUEUE_EXT, \
380 __builtin_choose_expr ( \
381 __builtin_types_compatible_p (__typeof (o), struct anv_semaphore*), \
382 VK_DEBUG_REPORT_OBJECT_TYPE_SEMAPHORE_EXT, \
383 __builtin_choose_expr ( \
384 __builtin_types_compatible_p (__typeof (o), struct anv_cmd_buffer*), \
385 VK_DEBUG_REPORT_OBJECT_TYPE_COMMAND_BUFFER_EXT, \
386 __builtin_choose_expr ( \
387 __builtin_types_compatible_p (__typeof (o), struct anv_fence*), \
388 VK_DEBUG_REPORT_OBJECT_TYPE_FENCE_EXT, \
389 __builtin_choose_expr ( \
390 __builtin_types_compatible_p (__typeof (o), struct anv_device_memory*), \
391 VK_DEBUG_REPORT_OBJECT_TYPE_DEVICE_MEMORY_EXT, \
392 __builtin_choose_expr ( \
393 __builtin_types_compatible_p (__typeof (o), struct anv_buffer*), \
394 VK_DEBUG_REPORT_OBJECT_TYPE_BUFFER_EXT, \
395 __builtin_choose_expr ( \
396 __builtin_types_compatible_p (__typeof (o), struct anv_image*), \
397 VK_DEBUG_REPORT_OBJECT_TYPE_IMAGE_EXT, \
398 __builtin_choose_expr ( \
399 __builtin_types_compatible_p (__typeof (o), const struct anv_image*), \
400 VK_DEBUG_REPORT_OBJECT_TYPE_IMAGE_EXT, \
401 __builtin_choose_expr ( \
402 __builtin_types_compatible_p (__typeof (o), struct anv_event*), \
403 VK_DEBUG_REPORT_OBJECT_TYPE_EVENT_EXT, \
404 __builtin_choose_expr ( \
405 __builtin_types_compatible_p (__typeof (o), struct anv_query_pool*), \
406 VK_DEBUG_REPORT_OBJECT_TYPE_QUERY_POOL_EXT, \
407 __builtin_choose_expr ( \
408 __builtin_types_compatible_p (__typeof (o), struct anv_buffer_view*), \
409 VK_DEBUG_REPORT_OBJECT_TYPE_BUFFER_VIEW_EXT, \
410 __builtin_choose_expr ( \
411 __builtin_types_compatible_p (__typeof (o), struct anv_image_view*), \
412 VK_DEBUG_REPORT_OBJECT_TYPE_IMAGE_VIEW_EXT, \
413 __builtin_choose_expr ( \
414 __builtin_types_compatible_p (__typeof (o), struct anv_shader_module*), \
415 VK_DEBUG_REPORT_OBJECT_TYPE_SHADER_MODULE_EXT, \
416 __builtin_choose_expr ( \
417 __builtin_types_compatible_p (__typeof (o), struct anv_pipeline_cache*), \
418 VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_CACHE_EXT, \
419 __builtin_choose_expr ( \
420 __builtin_types_compatible_p (__typeof (o), struct anv_pipeline_layout*), \
421 VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_LAYOUT_EXT, \
422 __builtin_choose_expr ( \
423 __builtin_types_compatible_p (__typeof (o), struct anv_render_pass*), \
424 VK_DEBUG_REPORT_OBJECT_TYPE_RENDER_PASS_EXT, \
425 __builtin_choose_expr ( \
426 __builtin_types_compatible_p (__typeof (o), struct anv_pipeline*), \
427 VK_DEBUG_REPORT_OBJECT_TYPE_PIPELINE_EXT, \
428 __builtin_choose_expr ( \
429 __builtin_types_compatible_p (__typeof (o), struct anv_descriptor_set_layout*), \
430 VK_DEBUG_REPORT_OBJECT_TYPE_DESCRIPTOR_SET_LAYOUT_EXT, \
431 __builtin_choose_expr ( \
432 __builtin_types_compatible_p (__typeof (o), struct anv_sampler*), \
433 VK_DEBUG_REPORT_OBJECT_TYPE_SAMPLER_EXT, \
434 __builtin_choose_expr ( \
435 __builtin_types_compatible_p (__typeof (o), struct anv_descriptor_pool*), \
436 VK_DEBUG_REPORT_OBJECT_TYPE_DESCRIPTOR_POOL_EXT, \
437 __builtin_choose_expr ( \
438 __builtin_types_compatible_p (__typeof (o), struct anv_descriptor_set*), \
439 VK_DEBUG_REPORT_OBJECT_TYPE_DESCRIPTOR_SET_EXT, \
440 __builtin_choose_expr ( \
441 __builtin_types_compatible_p (__typeof (o), struct anv_framebuffer*), \
442 VK_DEBUG_REPORT_OBJECT_TYPE_FRAMEBUFFER_EXT, \
443 __builtin_choose_expr ( \
444 __builtin_types_compatible_p (__typeof (o), struct anv_cmd_pool*), \
445 VK_DEBUG_REPORT_OBJECT_TYPE_COMMAND_POOL_EXT, \
446 __builtin_choose_expr ( \
447 __builtin_types_compatible_p (__typeof (o), struct anv_surface*), \
448 VK_DEBUG_REPORT_OBJECT_TYPE_SURFACE_KHR_EXT, \
449 __builtin_choose_expr ( \
450 __builtin_types_compatible_p (__typeof (o), struct wsi_swapchain*), \
451 VK_DEBUG_REPORT_OBJECT_TYPE_SWAPCHAIN_KHR_EXT, \
452 __builtin_choose_expr ( \
453 __builtin_types_compatible_p (__typeof (o), struct vk_debug_callback*), \
454 VK_DEBUG_REPORT_OBJECT_TYPE_DEBUG_REPORT_CALLBACK_EXT_EXT, \
455 __builtin_choose_expr ( \
456 __builtin_types_compatible_p (__typeof (o), void*), \
457 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT, \
458 /* The void expression results in a compile-time error \
459 when assigning the result to something. */ \
460 (void)0)))))))))))))))))))))))))))))))
462 /* Whenever we generate an error, pass it through this function. Useful for
463 * debugging, where we can break on it. Only call at error site, not when
464 * propagating errors. Might be useful to plug in a stack trace here.
467 VkResult
__vk_errorv(struct anv_instance
*instance
, const void *object
,
468 VkDebugReportObjectTypeEXT type
, VkResult error
,
469 const char *file
, int line
, const char *format
,
472 VkResult
__vk_errorf(struct anv_instance
*instance
, const void *object
,
473 VkDebugReportObjectTypeEXT type
, VkResult error
,
474 const char *file
, int line
, const char *format
, ...)
475 anv_printflike(7, 8);
478 #define vk_error(error) __vk_errorf(NULL, NULL,\
479 VK_DEBUG_REPORT_OBJECT_TYPE_UNKNOWN_EXT,\
480 error, __FILE__, __LINE__, NULL)
481 #define vk_errorfi(instance, obj, error, format, ...)\
482 __vk_errorf(instance, obj, REPORT_OBJECT_TYPE(obj), error,\
483 __FILE__, __LINE__, format, ## __VA_ARGS__)
484 #define vk_errorf(device, obj, error, format, ...)\
485 vk_errorfi(anv_device_instance_or_null(device),\
486 obj, error, format, ## __VA_ARGS__)
488 #define vk_error(error) error
489 #define vk_errorfi(instance, obj, error, format, ...) error
490 #define vk_errorf(device, obj, error, format, ...) error
494 * Warn on ignored extension structs.
496 * The Vulkan spec requires us to ignore unsupported or unknown structs in
497 * a pNext chain. In debug mode, emitting warnings for ignored structs may
498 * help us discover structs that we should not have ignored.
501 * From the Vulkan 1.0.38 spec:
503 * Any component of the implementation (the loader, any enabled layers,
504 * and drivers) must skip over, without processing (other than reading the
505 * sType and pNext members) any chained structures with sType values not
506 * defined by extensions supported by that component.
508 #define anv_debug_ignored_stype(sType) \
509 intel_logd("%s: ignored VkStructureType %u\n", __func__, (sType))
511 void __anv_perf_warn(struct anv_device
*device
, const void *object
,
512 VkDebugReportObjectTypeEXT type
, const char *file
,
513 int line
, const char *format
, ...)
514 anv_printflike(6, 7);
515 void anv_loge(const char *format
, ...) anv_printflike(1, 2);
516 void anv_loge_v(const char *format
, va_list va
);
519 * Print a FINISHME message, including its source location.
521 #define anv_finishme(format, ...) \
523 static bool reported = false; \
525 intel_logw("%s:%d: FINISHME: " format, __FILE__, __LINE__, \
532 * Print a perf warning message. Set INTEL_DEBUG=perf to see these.
534 #define anv_perf_warn(instance, obj, format, ...) \
536 static bool reported = false; \
537 if (!reported && unlikely(INTEL_DEBUG & DEBUG_PERF)) { \
538 __anv_perf_warn(instance, obj, REPORT_OBJECT_TYPE(obj), __FILE__, __LINE__,\
539 format, ##__VA_ARGS__); \
544 /* A non-fatal assert. Useful for debugging. */
546 #define anv_assert(x) ({ \
547 if (unlikely(!(x))) \
548 intel_loge("%s:%d ASSERT: %s", __FILE__, __LINE__, #x); \
551 #define anv_assert(x)
554 /* A multi-pointer allocator
556 * When copying data structures from the user (such as a render pass), it's
557 * common to need to allocate data for a bunch of different things. Instead
558 * of doing several allocations and having to handle all of the error checking
559 * that entails, it can be easier to do a single allocation. This struct
560 * helps facilitate that. The intended usage looks like this:
563 * anv_multialloc_add(&ma, &main_ptr, 1);
564 * anv_multialloc_add(&ma, &substruct1, substruct1Count);
565 * anv_multialloc_add(&ma, &substruct2, substruct2Count);
567 * if (!anv_multialloc_alloc(&ma, pAllocator, VK_ALLOCATION_SCOPE_FOO))
568 * return vk_error(VK_ERROR_OUT_OF_HOST_MEORY);
570 struct anv_multialloc
{
578 #define ANV_MULTIALLOC_INIT \
579 ((struct anv_multialloc) { 0, })
581 #define ANV_MULTIALLOC(_name) \
582 struct anv_multialloc _name = ANV_MULTIALLOC_INIT
584 __attribute__((always_inline
))
586 _anv_multialloc_add(struct anv_multialloc
*ma
,
587 void **ptr
, size_t size
, size_t align
)
589 size_t offset
= align_u64(ma
->size
, align
);
590 ma
->size
= offset
+ size
;
591 ma
->align
= MAX2(ma
->align
, align
);
593 /* Store the offset in the pointer. */
594 *ptr
= (void *)(uintptr_t)offset
;
596 assert(ma
->ptr_count
< ARRAY_SIZE(ma
->ptrs
));
597 ma
->ptrs
[ma
->ptr_count
++] = ptr
;
600 #define anv_multialloc_add_size(_ma, _ptr, _size) \
601 _anv_multialloc_add((_ma), (void **)(_ptr), (_size), __alignof__(**(_ptr)))
603 #define anv_multialloc_add(_ma, _ptr, _count) \
604 anv_multialloc_add_size(_ma, _ptr, (_count) * sizeof(**(_ptr)));
606 __attribute__((always_inline
))
608 anv_multialloc_alloc(struct anv_multialloc
*ma
,
609 const VkAllocationCallbacks
*alloc
,
610 VkSystemAllocationScope scope
)
612 void *ptr
= vk_alloc(alloc
, ma
->size
, ma
->align
, scope
);
616 /* Fill out each of the pointers with their final value.
618 * for (uint32_t i = 0; i < ma->ptr_count; i++)
619 * *ma->ptrs[i] = ptr + (uintptr_t)*ma->ptrs[i];
621 * Unfortunately, even though ma->ptr_count is basically guaranteed to be a
622 * constant, GCC is incapable of figuring this out and unrolling the loop
623 * so we have to give it a little help.
625 STATIC_ASSERT(ARRAY_SIZE(ma
->ptrs
) == 8);
626 #define _ANV_MULTIALLOC_UPDATE_POINTER(_i) \
627 if ((_i) < ma->ptr_count) \
628 *ma->ptrs[_i] = ptr + (uintptr_t)*ma->ptrs[_i]
629 _ANV_MULTIALLOC_UPDATE_POINTER(0);
630 _ANV_MULTIALLOC_UPDATE_POINTER(1);
631 _ANV_MULTIALLOC_UPDATE_POINTER(2);
632 _ANV_MULTIALLOC_UPDATE_POINTER(3);
633 _ANV_MULTIALLOC_UPDATE_POINTER(4);
634 _ANV_MULTIALLOC_UPDATE_POINTER(5);
635 _ANV_MULTIALLOC_UPDATE_POINTER(6);
636 _ANV_MULTIALLOC_UPDATE_POINTER(7);
637 #undef _ANV_MULTIALLOC_UPDATE_POINTER
642 __attribute__((always_inline
))
644 anv_multialloc_alloc2(struct anv_multialloc
*ma
,
645 const VkAllocationCallbacks
*parent_alloc
,
646 const VkAllocationCallbacks
*alloc
,
647 VkSystemAllocationScope scope
)
649 return anv_multialloc_alloc(ma
, alloc
? alloc
: parent_alloc
, scope
);
657 /* Index into the current validation list. This is used by the
658 * validation list building alrogithm to track which buffers are already
659 * in the validation list so that we can ensure uniqueness.
663 /* Index for use with util_sparse_array_free_list */
666 /* Last known offset. This value is provided by the kernel when we
667 * execbuf and is used as the presumed offset for the next bunch of
672 /** Size of the buffer not including implicit aux */
675 /* Map for internally mapped BOs.
677 * If ANV_BO_WRAPPER is set in flags, map points to the wrapped BO.
681 /** Size of the implicit CCS range at the end of the buffer
683 * On Gen12, CCS data is always a direct 1/256 scale-down. A single 64K
684 * page of main surface data maps to a 256B chunk of CCS data and that
685 * mapping is provided on TGL-LP by the AUX table which maps virtual memory
686 * addresses in the main surface to virtual memory addresses for CCS data.
688 * Because we can't change these maps around easily and because Vulkan
689 * allows two VkImages to be bound to overlapping memory regions (as long
690 * as the app is careful), it's not feasible to make this mapping part of
691 * the image. (On Gen11 and earlier, the mapping was provided via
692 * RENDER_SURFACE_STATE so each image had its own main -> CCS mapping.)
693 * Instead, we attach the CCS data directly to the buffer object and setup
694 * the AUX table mapping at BO creation time.
696 * This field is for internal tracking use by the BO allocator only and
697 * should not be touched by other parts of the code. If something wants to
698 * know if a BO has implicit CCS data, it should instead look at the
699 * has_implicit_ccs boolean below.
701 * This data is not included in maps of this buffer.
705 /** Flags to pass to the kernel through drm_i915_exec_object2::flags */
708 /** True if this BO may be shared with other processes */
711 /** True if this BO is a wrapper
713 * When set to true, none of the fields in this BO are meaningful except
714 * for anv_bo::is_wrapper and anv_bo::map which points to the actual BO.
715 * See also anv_bo_unwrap(). Wrapper BOs are not allowed when use_softpin
716 * is set in the physical device.
720 /** See also ANV_BO_ALLOC_FIXED_ADDRESS */
721 bool has_fixed_address
:1;
723 /** True if this BO wraps a host pointer */
724 bool from_host_ptr
:1;
726 /** See also ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS */
727 bool has_client_visible_address
:1;
729 /** True if this BO has implicit CCS data attached to it */
730 bool has_implicit_ccs
:1;
733 static inline struct anv_bo
*
734 anv_bo_ref(struct anv_bo
*bo
)
736 p_atomic_inc(&bo
->refcount
);
740 static inline struct anv_bo
*
741 anv_bo_unwrap(struct anv_bo
*bo
)
743 while (bo
->is_wrapper
)
748 /* Represents a lock-free linked list of "free" things. This is used by
749 * both the block pool and the state pools. Unfortunately, in order to
750 * solve the ABA problem, we can't use a single uint32_t head.
752 union anv_free_list
{
756 /* A simple count that is incremented every time the head changes. */
759 /* Make sure it's aligned to 64 bits. This will make atomic operations
760 * faster on 32 bit platforms.
762 uint64_t u64
__attribute__ ((aligned (8)));
765 #define ANV_FREE_LIST_EMPTY ((union anv_free_list) { { UINT32_MAX, 0 } })
767 struct anv_block_state
{
773 /* Make sure it's aligned to 64 bits. This will make atomic operations
774 * faster on 32 bit platforms.
776 uint64_t u64
__attribute__ ((aligned (8)));
780 #define anv_block_pool_foreach_bo(bo, pool) \
781 for (struct anv_bo **_pp_bo = (pool)->bos, *bo; \
782 _pp_bo != &(pool)->bos[(pool)->nbos] && (bo = *_pp_bo, true); \
785 #define ANV_MAX_BLOCK_POOL_BOS 20
787 struct anv_block_pool
{
788 struct anv_device
*device
;
791 /* Wrapper BO for use in relocation lists. This BO is simply a wrapper
792 * around the actual BO so that we grow the pool after the wrapper BO has
793 * been put in a relocation list. This is only used in the non-softpin
796 struct anv_bo wrapper_bo
;
798 struct anv_bo
*bos
[ANV_MAX_BLOCK_POOL_BOS
];
804 /* The address where the start of the pool is pinned. The various bos that
805 * are created as the pool grows will have addresses in the range
806 * [start_address, start_address + BLOCK_POOL_MEMFD_SIZE).
808 uint64_t start_address
;
810 /* The offset from the start of the bo to the "center" of the block
811 * pool. Pointers to allocated blocks are given by
812 * bo.map + center_bo_offset + offsets.
814 uint32_t center_bo_offset
;
816 /* Current memory map of the block pool. This pointer may or may not
817 * point to the actual beginning of the block pool memory. If
818 * anv_block_pool_alloc_back has ever been called, then this pointer
819 * will point to the "center" position of the buffer and all offsets
820 * (negative or positive) given out by the block pool alloc functions
821 * will be valid relative to this pointer.
823 * In particular, map == bo.map + center_offset
825 * DO NOT access this pointer directly. Use anv_block_pool_map() instead,
826 * since it will handle the softpin case as well, where this points to NULL.
832 * Array of mmaps and gem handles owned by the block pool, reclaimed when
833 * the block pool is destroyed.
835 struct u_vector mmap_cleanups
;
837 struct anv_block_state state
;
839 struct anv_block_state back_state
;
842 /* Block pools are backed by a fixed-size 1GB memfd */
843 #define BLOCK_POOL_MEMFD_SIZE (1ul << 30)
845 /* The center of the block pool is also the middle of the memfd. This may
846 * change in the future if we decide differently for some reason.
848 #define BLOCK_POOL_MEMFD_CENTER (BLOCK_POOL_MEMFD_SIZE / 2)
850 static inline uint32_t
851 anv_block_pool_size(struct anv_block_pool
*pool
)
853 return pool
->state
.end
+ pool
->back_state
.end
;
863 #define ANV_STATE_NULL ((struct anv_state) { .alloc_size = 0 })
865 struct anv_fixed_size_state_pool
{
866 union anv_free_list free_list
;
867 struct anv_block_state block
;
870 #define ANV_MIN_STATE_SIZE_LOG2 6
871 #define ANV_MAX_STATE_SIZE_LOG2 21
873 #define ANV_STATE_BUCKETS (ANV_MAX_STATE_SIZE_LOG2 - ANV_MIN_STATE_SIZE_LOG2 + 1)
875 struct anv_free_entry
{
877 struct anv_state state
;
880 struct anv_state_table
{
881 struct anv_device
*device
;
883 struct anv_free_entry
*map
;
885 struct anv_block_state state
;
886 struct u_vector cleanups
;
889 struct anv_state_pool
{
890 struct anv_block_pool block_pool
;
892 /* Offset into the relevant state base address where the state pool starts
895 int32_t start_offset
;
897 struct anv_state_table table
;
899 /* The size of blocks which will be allocated from the block pool */
902 /** Free list for "back" allocations */
903 union anv_free_list back_alloc_free_list
;
905 struct anv_fixed_size_state_pool buckets
[ANV_STATE_BUCKETS
];
908 struct anv_state_reserved_pool
{
909 struct anv_state_pool
*pool
;
910 union anv_free_list reserved_blocks
;
914 struct anv_state_stream
{
915 struct anv_state_pool
*state_pool
;
917 /* The size of blocks to allocate from the state pool */
920 /* Current block we're allocating from */
921 struct anv_state block
;
923 /* Offset into the current block at which to allocate the next state */
926 /* List of all blocks allocated from this pool */
927 struct util_dynarray all_blocks
;
930 /* The block_pool functions exported for testing only. The block pool should
931 * only be used via a state pool (see below).
933 VkResult
anv_block_pool_init(struct anv_block_pool
*pool
,
934 struct anv_device
*device
,
935 uint64_t start_address
,
936 uint32_t initial_size
);
937 void anv_block_pool_finish(struct anv_block_pool
*pool
);
938 int32_t anv_block_pool_alloc(struct anv_block_pool
*pool
,
939 uint32_t block_size
, uint32_t *padding
);
940 int32_t anv_block_pool_alloc_back(struct anv_block_pool
*pool
,
941 uint32_t block_size
);
942 void* anv_block_pool_map(struct anv_block_pool
*pool
, int32_t offset
, uint32_t
945 VkResult
anv_state_pool_init(struct anv_state_pool
*pool
,
946 struct anv_device
*device
,
947 uint64_t base_address
,
948 int32_t start_offset
,
949 uint32_t block_size
);
950 void anv_state_pool_finish(struct anv_state_pool
*pool
);
951 struct anv_state
anv_state_pool_alloc(struct anv_state_pool
*pool
,
952 uint32_t state_size
, uint32_t alignment
);
953 struct anv_state
anv_state_pool_alloc_back(struct anv_state_pool
*pool
);
954 void anv_state_pool_free(struct anv_state_pool
*pool
, struct anv_state state
);
955 void anv_state_stream_init(struct anv_state_stream
*stream
,
956 struct anv_state_pool
*state_pool
,
957 uint32_t block_size
);
958 void anv_state_stream_finish(struct anv_state_stream
*stream
);
959 struct anv_state
anv_state_stream_alloc(struct anv_state_stream
*stream
,
960 uint32_t size
, uint32_t alignment
);
962 void anv_state_reserved_pool_init(struct anv_state_reserved_pool
*pool
,
963 struct anv_state_pool
*parent
,
964 uint32_t count
, uint32_t size
,
966 void anv_state_reserved_pool_finish(struct anv_state_reserved_pool
*pool
);
967 struct anv_state
anv_state_reserved_pool_alloc(struct anv_state_reserved_pool
*pool
);
968 void anv_state_reserved_pool_free(struct anv_state_reserved_pool
*pool
,
969 struct anv_state state
);
971 VkResult
anv_state_table_init(struct anv_state_table
*table
,
972 struct anv_device
*device
,
973 uint32_t initial_entries
);
974 void anv_state_table_finish(struct anv_state_table
*table
);
975 VkResult
anv_state_table_add(struct anv_state_table
*table
, uint32_t *idx
,
977 void anv_free_list_push(union anv_free_list
*list
,
978 struct anv_state_table
*table
,
979 uint32_t idx
, uint32_t count
);
980 struct anv_state
* anv_free_list_pop(union anv_free_list
*list
,
981 struct anv_state_table
*table
);
984 static inline struct anv_state
*
985 anv_state_table_get(struct anv_state_table
*table
, uint32_t idx
)
987 return &table
->map
[idx
].state
;
990 * Implements a pool of re-usable BOs. The interface is identical to that
991 * of block_pool except that each block is its own BO.
994 struct anv_device
*device
;
996 struct util_sparse_array_free_list free_list
[16];
999 void anv_bo_pool_init(struct anv_bo_pool
*pool
, struct anv_device
*device
);
1000 void anv_bo_pool_finish(struct anv_bo_pool
*pool
);
1001 VkResult
anv_bo_pool_alloc(struct anv_bo_pool
*pool
, uint32_t size
,
1002 struct anv_bo
**bo_out
);
1003 void anv_bo_pool_free(struct anv_bo_pool
*pool
, struct anv_bo
*bo
);
1005 struct anv_scratch_pool
{
1006 /* Indexed by Per-Thread Scratch Space number (the hardware value) and stage */
1007 struct anv_bo
*bos
[16][MESA_SHADER_STAGES
];
1010 void anv_scratch_pool_init(struct anv_device
*device
,
1011 struct anv_scratch_pool
*pool
);
1012 void anv_scratch_pool_finish(struct anv_device
*device
,
1013 struct anv_scratch_pool
*pool
);
1014 struct anv_bo
*anv_scratch_pool_alloc(struct anv_device
*device
,
1015 struct anv_scratch_pool
*pool
,
1016 gl_shader_stage stage
,
1017 unsigned per_thread_scratch
);
1019 /** Implements a BO cache that ensures a 1-1 mapping of GEM BOs to anv_bos */
1020 struct anv_bo_cache
{
1021 struct util_sparse_array bo_map
;
1022 pthread_mutex_t mutex
;
1025 VkResult
anv_bo_cache_init(struct anv_bo_cache
*cache
);
1026 void anv_bo_cache_finish(struct anv_bo_cache
*cache
);
1028 struct anv_memory_type
{
1029 /* Standard bits passed on to the client */
1030 VkMemoryPropertyFlags propertyFlags
;
1034 struct anv_memory_heap
{
1035 /* Standard bits passed on to the client */
1037 VkMemoryHeapFlags flags
;
1039 /** Driver-internal book-keeping.
1041 * Align it to 64 bits to make atomic operations faster on 32 bit platforms.
1043 VkDeviceSize used
__attribute__ ((aligned (8)));
1046 struct anv_physical_device
{
1047 struct vk_object_base base
;
1049 /* Link in anv_instance::physical_devices */
1050 struct list_head link
;
1052 struct anv_instance
* instance
;
1062 struct gen_device_info info
;
1063 /** Amount of "GPU memory" we want to advertise
1065 * Clearly, this value is bogus since Intel is a UMA architecture. On
1066 * gen7 platforms, we are limited by GTT size unless we want to implement
1067 * fine-grained tracking and GTT splitting. On Broadwell and above we are
1068 * practically unlimited. However, we will never report more than 3/4 of
1069 * the total system ram to try and avoid running out of RAM.
1071 bool supports_48bit_addresses
;
1072 struct brw_compiler
* compiler
;
1073 struct isl_device isl_dev
;
1074 struct gen_perf_config
* perf
;
1075 int cmd_parser_version
;
1077 bool has_exec_async
;
1078 bool has_exec_capture
;
1079 bool has_exec_fence
;
1081 bool has_syncobj_wait
;
1082 bool has_context_priority
;
1083 bool has_context_isolation
;
1084 bool has_mem_available
;
1085 bool has_mmap_offset
;
1089 bool always_use_bindless
;
1090 bool use_call_secondary
;
1092 /** True if we can access buffers using A64 messages */
1093 bool has_a64_buffer_access
;
1094 /** True if we can use bindless access for images */
1095 bool has_bindless_images
;
1096 /** True if we can use bindless access for samplers */
1097 bool has_bindless_samplers
;
1099 /** True if we can read the GPU timestamp register
1101 * When running in a virtual context, the timestamp register is unreadable
1104 bool has_reg_timestamp
;
1106 /** True if this device has implicit AUX
1108 * If true, CCS is handled as an implicit attachment to the BO rather than
1109 * as an explicitly bound surface.
1111 bool has_implicit_ccs
;
1113 bool always_flush_cache
;
1115 struct anv_device_extension_table supported_extensions
;
1118 uint32_t subslice_total
;
1121 uint32_t type_count
;
1122 struct anv_memory_type types
[VK_MAX_MEMORY_TYPES
];
1123 uint32_t heap_count
;
1124 struct anv_memory_heap heaps
[VK_MAX_MEMORY_HEAPS
];
1127 uint8_t driver_build_sha1
[20];
1128 uint8_t pipeline_cache_uuid
[VK_UUID_SIZE
];
1129 uint8_t driver_uuid
[VK_UUID_SIZE
];
1130 uint8_t device_uuid
[VK_UUID_SIZE
];
1132 struct disk_cache
* disk_cache
;
1134 struct wsi_device wsi_device
;
1139 struct anv_app_info
{
1140 const char* app_name
;
1141 uint32_t app_version
;
1142 const char* engine_name
;
1143 uint32_t engine_version
;
1144 uint32_t api_version
;
1147 struct anv_instance
{
1148 struct vk_object_base base
;
1150 VkAllocationCallbacks alloc
;
1152 struct anv_app_info app_info
;
1154 struct anv_instance_extension_table enabled_extensions
;
1155 struct anv_instance_dispatch_table dispatch
;
1156 struct anv_physical_device_dispatch_table physical_device_dispatch
;
1157 struct anv_device_dispatch_table device_dispatch
;
1159 bool physical_devices_enumerated
;
1160 struct list_head physical_devices
;
1162 bool pipeline_cache_enabled
;
1164 struct vk_debug_report_instance debug_report_callbacks
;
1166 struct driOptionCache dri_options
;
1167 struct driOptionCache available_dri_options
;
1170 VkResult
anv_init_wsi(struct anv_physical_device
*physical_device
);
1171 void anv_finish_wsi(struct anv_physical_device
*physical_device
);
1173 uint32_t anv_physical_device_api_version(struct anv_physical_device
*dev
);
1174 bool anv_physical_device_extension_supported(struct anv_physical_device
*dev
,
1177 struct anv_queue_submit
{
1178 struct anv_cmd_buffer
* cmd_buffer
;
1180 uint32_t fence_count
;
1181 uint32_t fence_array_length
;
1182 struct drm_i915_gem_exec_fence
* fences
;
1184 uint32_t temporary_semaphore_count
;
1185 uint32_t temporary_semaphore_array_length
;
1186 struct anv_semaphore_impl
* temporary_semaphores
;
1188 /* Semaphores to be signaled with a SYNC_FD. */
1189 struct anv_semaphore
** sync_fd_semaphores
;
1190 uint32_t sync_fd_semaphore_count
;
1191 uint32_t sync_fd_semaphore_array_length
;
1193 /* Allocated only with non shareable timelines. */
1194 struct anv_timeline
** wait_timelines
;
1195 uint32_t wait_timeline_count
;
1196 uint32_t wait_timeline_array_length
;
1197 uint64_t * wait_timeline_values
;
1199 struct anv_timeline
** signal_timelines
;
1200 uint32_t signal_timeline_count
;
1201 uint32_t signal_timeline_array_length
;
1202 uint64_t * signal_timeline_values
;
1205 bool need_out_fence
;
1208 uint32_t fence_bo_count
;
1209 uint32_t fence_bo_array_length
;
1210 /* An array of struct anv_bo pointers with lower bit used as a flag to
1211 * signal we will wait on that BO (see anv_(un)pack_ptr).
1213 uintptr_t * fence_bos
;
1215 int perf_query_pass
;
1217 const VkAllocationCallbacks
* alloc
;
1218 VkSystemAllocationScope alloc_scope
;
1220 struct anv_bo
* simple_bo
;
1221 uint32_t simple_bo_size
;
1223 struct list_head link
;
1227 struct vk_object_base base
;
1229 struct anv_device
* device
;
1232 * A list of struct anv_queue_submit to be submitted to i915.
1234 struct list_head queued_submits
;
1236 VkDeviceQueueCreateFlags flags
;
1239 struct anv_pipeline_cache
{
1240 struct vk_object_base base
;
1241 struct anv_device
* device
;
1242 pthread_mutex_t mutex
;
1244 struct hash_table
* nir_cache
;
1246 struct hash_table
* cache
;
1251 struct nir_xfb_info
;
1252 struct anv_pipeline_bind_map
;
1254 void anv_pipeline_cache_init(struct anv_pipeline_cache
*cache
,
1255 struct anv_device
*device
,
1257 bool external_sync
);
1258 void anv_pipeline_cache_finish(struct anv_pipeline_cache
*cache
);
1260 struct anv_shader_bin
*
1261 anv_pipeline_cache_search(struct anv_pipeline_cache
*cache
,
1262 const void *key
, uint32_t key_size
);
1263 struct anv_shader_bin
*
1264 anv_pipeline_cache_upload_kernel(struct anv_pipeline_cache
*cache
,
1265 gl_shader_stage stage
,
1266 const void *key_data
, uint32_t key_size
,
1267 const void *kernel_data
, uint32_t kernel_size
,
1268 const void *constant_data
,
1269 uint32_t constant_data_size
,
1270 const struct brw_stage_prog_data
*prog_data
,
1271 uint32_t prog_data_size
,
1272 const struct brw_compile_stats
*stats
,
1274 const struct nir_xfb_info
*xfb_info
,
1275 const struct anv_pipeline_bind_map
*bind_map
);
1277 struct anv_shader_bin
*
1278 anv_device_search_for_kernel(struct anv_device
*device
,
1279 struct anv_pipeline_cache
*cache
,
1280 const void *key_data
, uint32_t key_size
,
1281 bool *user_cache_bit
);
1283 struct anv_shader_bin
*
1284 anv_device_upload_kernel(struct anv_device
*device
,
1285 struct anv_pipeline_cache
*cache
,
1286 gl_shader_stage stage
,
1287 const void *key_data
, uint32_t key_size
,
1288 const void *kernel_data
, uint32_t kernel_size
,
1289 const void *constant_data
,
1290 uint32_t constant_data_size
,
1291 const struct brw_stage_prog_data
*prog_data
,
1292 uint32_t prog_data_size
,
1293 const struct brw_compile_stats
*stats
,
1295 const struct nir_xfb_info
*xfb_info
,
1296 const struct anv_pipeline_bind_map
*bind_map
);
1299 struct nir_shader_compiler_options
;
1302 anv_device_search_for_nir(struct anv_device
*device
,
1303 struct anv_pipeline_cache
*cache
,
1304 const struct nir_shader_compiler_options
*nir_options
,
1305 unsigned char sha1_key
[20],
1309 anv_device_upload_nir(struct anv_device
*device
,
1310 struct anv_pipeline_cache
*cache
,
1311 const struct nir_shader
*nir
,
1312 unsigned char sha1_key
[20]);
1314 struct anv_address
{
1320 struct vk_device vk
;
1322 struct anv_physical_device
* physical
;
1324 struct gen_device_info info
;
1325 struct isl_device isl_dev
;
1328 bool can_chain_batches
;
1329 bool robust_buffer_access
;
1330 struct anv_device_extension_table enabled_extensions
;
1331 struct anv_device_dispatch_table dispatch
;
1333 pthread_mutex_t vma_mutex
;
1334 struct util_vma_heap vma_lo
;
1335 struct util_vma_heap vma_cva
;
1336 struct util_vma_heap vma_hi
;
1338 /** List of all anv_device_memory objects */
1339 struct list_head memory_objects
;
1341 struct anv_bo_pool batch_bo_pool
;
1343 struct anv_bo_cache bo_cache
;
1345 struct anv_state_pool dynamic_state_pool
;
1346 struct anv_state_pool instruction_state_pool
;
1347 struct anv_state_pool binding_table_pool
;
1348 struct anv_state_pool surface_state_pool
;
1350 struct anv_state_reserved_pool custom_border_colors
;
1352 /** BO used for various workarounds
1354 * There are a number of workarounds on our hardware which require writing
1355 * data somewhere and it doesn't really matter where. For that, we use
1356 * this BO and just write to the first dword or so.
1358 * We also need to be able to handle NULL buffers bound as pushed UBOs.
1359 * For that, we use the high bytes (>= 1024) of the workaround BO.
1361 struct anv_bo
* workaround_bo
;
1362 struct anv_address workaround_address
;
1364 struct anv_bo
* trivial_batch_bo
;
1365 struct anv_bo
* hiz_clear_bo
;
1366 struct anv_state null_surface_state
;
1368 struct anv_pipeline_cache default_pipeline_cache
;
1369 struct blorp_context blorp
;
1371 struct anv_state border_colors
;
1373 struct anv_state slice_hash
;
1375 struct anv_queue queue
;
1377 struct anv_scratch_pool scratch_pool
;
1379 pthread_mutex_t mutex
;
1380 pthread_cond_t queue_submit
;
1383 struct gen_batch_decode_ctx decoder_ctx
;
1385 * When decoding a anv_cmd_buffer, we might need to search for BOs through
1386 * the cmd_buffer's list.
1388 struct anv_cmd_buffer
*cmd_buffer_being_decoded
;
1390 int perf_fd
; /* -1 if no opened */
1391 uint64_t perf_metric
; /* 0 if unset */
1393 struct gen_aux_map_context
*aux_map_ctx
;
1396 static inline struct anv_instance
*
1397 anv_device_instance_or_null(const struct anv_device
*device
)
1399 return device
? device
->physical
->instance
: NULL
;
1402 static inline struct anv_state_pool
*
1403 anv_binding_table_pool(struct anv_device
*device
)
1405 if (device
->physical
->use_softpin
)
1406 return &device
->binding_table_pool
;
1408 return &device
->surface_state_pool
;
1411 static inline struct anv_state
1412 anv_binding_table_pool_alloc(struct anv_device
*device
) {
1413 if (device
->physical
->use_softpin
)
1414 return anv_state_pool_alloc(&device
->binding_table_pool
,
1415 device
->binding_table_pool
.block_size
, 0);
1417 return anv_state_pool_alloc_back(&device
->surface_state_pool
);
1421 anv_binding_table_pool_free(struct anv_device
*device
, struct anv_state state
) {
1422 anv_state_pool_free(anv_binding_table_pool(device
), state
);
1425 static inline uint32_t
1426 anv_mocs_for_bo(const struct anv_device
*device
, const struct anv_bo
*bo
)
1428 if (bo
->is_external
)
1429 return device
->isl_dev
.mocs
.external
;
1431 return device
->isl_dev
.mocs
.internal
;
1434 void anv_device_init_blorp(struct anv_device
*device
);
1435 void anv_device_finish_blorp(struct anv_device
*device
);
1437 void _anv_device_set_all_queue_lost(struct anv_device
*device
);
1438 VkResult
_anv_device_set_lost(struct anv_device
*device
,
1439 const char *file
, int line
,
1440 const char *msg
, ...)
1441 anv_printflike(4, 5);
1442 VkResult
_anv_queue_set_lost(struct anv_queue
*queue
,
1443 const char *file
, int line
,
1444 const char *msg
, ...)
1445 anv_printflike(4, 5);
1446 #define anv_device_set_lost(dev, ...) \
1447 _anv_device_set_lost(dev, __FILE__, __LINE__, __VA_ARGS__)
1448 #define anv_queue_set_lost(queue, ...) \
1449 _anv_queue_set_lost(queue, __FILE__, __LINE__, __VA_ARGS__)
1452 anv_device_is_lost(struct anv_device
*device
)
1454 return unlikely(p_atomic_read(&device
->_lost
));
1457 VkResult
anv_device_query_status(struct anv_device
*device
);
1460 enum anv_bo_alloc_flags
{
1461 /** Specifies that the BO must have a 32-bit address
1463 * This is the opposite of EXEC_OBJECT_SUPPORTS_48B_ADDRESS.
1465 ANV_BO_ALLOC_32BIT_ADDRESS
= (1 << 0),
1467 /** Specifies that the BO may be shared externally */
1468 ANV_BO_ALLOC_EXTERNAL
= (1 << 1),
1470 /** Specifies that the BO should be mapped */
1471 ANV_BO_ALLOC_MAPPED
= (1 << 2),
1473 /** Specifies that the BO should be snooped so we get coherency */
1474 ANV_BO_ALLOC_SNOOPED
= (1 << 3),
1476 /** Specifies that the BO should be captured in error states */
1477 ANV_BO_ALLOC_CAPTURE
= (1 << 4),
1479 /** Specifies that the BO will have an address assigned by the caller
1481 * Such BOs do not exist in any VMA heap.
1483 ANV_BO_ALLOC_FIXED_ADDRESS
= (1 << 5),
1485 /** Enables implicit synchronization on the BO
1487 * This is the opposite of EXEC_OBJECT_ASYNC.
1489 ANV_BO_ALLOC_IMPLICIT_SYNC
= (1 << 6),
1491 /** Enables implicit synchronization on the BO
1493 * This is equivalent to EXEC_OBJECT_WRITE.
1495 ANV_BO_ALLOC_IMPLICIT_WRITE
= (1 << 7),
1497 /** Has an address which is visible to the client */
1498 ANV_BO_ALLOC_CLIENT_VISIBLE_ADDRESS
= (1 << 8),
1500 /** This buffer has implicit CCS data attached to it */
1501 ANV_BO_ALLOC_IMPLICIT_CCS
= (1 << 9),
1504 VkResult
anv_device_alloc_bo(struct anv_device
*device
, uint64_t size
,
1505 enum anv_bo_alloc_flags alloc_flags
,
1506 uint64_t explicit_address
,
1507 struct anv_bo
**bo
);
1508 VkResult
anv_device_import_bo_from_host_ptr(struct anv_device
*device
,
1509 void *host_ptr
, uint32_t size
,
1510 enum anv_bo_alloc_flags alloc_flags
,
1511 uint64_t client_address
,
1512 struct anv_bo
**bo_out
);
1513 VkResult
anv_device_import_bo(struct anv_device
*device
, int fd
,
1514 enum anv_bo_alloc_flags alloc_flags
,
1515 uint64_t client_address
,
1516 struct anv_bo
**bo
);
1517 VkResult
anv_device_export_bo(struct anv_device
*device
,
1518 struct anv_bo
*bo
, int *fd_out
);
1519 void anv_device_release_bo(struct anv_device
*device
,
1522 static inline struct anv_bo
*
1523 anv_device_lookup_bo(struct anv_device
*device
, uint32_t gem_handle
)
1525 return util_sparse_array_get(&device
->bo_cache
.bo_map
, gem_handle
);
1528 VkResult
anv_device_bo_busy(struct anv_device
*device
, struct anv_bo
*bo
);
1529 VkResult
anv_device_wait(struct anv_device
*device
, struct anv_bo
*bo
,
1532 VkResult
anv_queue_init(struct anv_device
*device
, struct anv_queue
*queue
);
1533 void anv_queue_finish(struct anv_queue
*queue
);
1535 VkResult
anv_queue_execbuf_locked(struct anv_queue
*queue
, struct anv_queue_submit
*submit
);
1536 VkResult
anv_queue_submit_simple_batch(struct anv_queue
*queue
,
1537 struct anv_batch
*batch
);
1539 uint64_t anv_gettime_ns(void);
1540 uint64_t anv_get_absolute_timeout(uint64_t timeout
);
1542 void* anv_gem_mmap(struct anv_device
*device
,
1543 uint32_t gem_handle
, uint64_t offset
, uint64_t size
, uint32_t flags
);
1544 void anv_gem_munmap(struct anv_device
*device
, void *p
, uint64_t size
);
1545 uint32_t anv_gem_create(struct anv_device
*device
, uint64_t size
);
1546 void anv_gem_close(struct anv_device
*device
, uint32_t gem_handle
);
1547 uint32_t anv_gem_userptr(struct anv_device
*device
, void *mem
, size_t size
);
1548 int anv_gem_busy(struct anv_device
*device
, uint32_t gem_handle
);
1549 int anv_gem_wait(struct anv_device
*device
, uint32_t gem_handle
, int64_t *timeout_ns
);
1550 int anv_gem_execbuffer(struct anv_device
*device
,
1551 struct drm_i915_gem_execbuffer2
*execbuf
);
1552 int anv_gem_set_tiling(struct anv_device
*device
, uint32_t gem_handle
,
1553 uint32_t stride
, uint32_t tiling
);
1554 int anv_gem_create_context(struct anv_device
*device
);
1555 bool anv_gem_has_context_priority(int fd
);
1556 int anv_gem_destroy_context(struct anv_device
*device
, int context
);
1557 int anv_gem_set_context_param(int fd
, int context
, uint32_t param
,
1559 int anv_gem_get_context_param(int fd
, int context
, uint32_t param
,
1561 int anv_gem_get_param(int fd
, uint32_t param
);
1562 int anv_gem_get_tiling(struct anv_device
*device
, uint32_t gem_handle
);
1563 bool anv_gem_get_bit6_swizzle(int fd
, uint32_t tiling
);
1564 int anv_gem_gpu_get_reset_stats(struct anv_device
*device
,
1565 uint32_t *active
, uint32_t *pending
);
1566 int anv_gem_handle_to_fd(struct anv_device
*device
, uint32_t gem_handle
);
1567 int anv_gem_reg_read(int fd
, uint32_t offset
, uint64_t *result
);
1568 uint32_t anv_gem_fd_to_handle(struct anv_device
*device
, int fd
);
1569 int anv_gem_set_caching(struct anv_device
*device
, uint32_t gem_handle
, uint32_t caching
);
1570 int anv_gem_set_domain(struct anv_device
*device
, uint32_t gem_handle
,
1571 uint32_t read_domains
, uint32_t write_domain
);
1572 int anv_gem_sync_file_merge(struct anv_device
*device
, int fd1
, int fd2
);
1573 uint32_t anv_gem_syncobj_create(struct anv_device
*device
, uint32_t flags
);
1574 void anv_gem_syncobj_destroy(struct anv_device
*device
, uint32_t handle
);
1575 int anv_gem_syncobj_handle_to_fd(struct anv_device
*device
, uint32_t handle
);
1576 uint32_t anv_gem_syncobj_fd_to_handle(struct anv_device
*device
, int fd
);
1577 int anv_gem_syncobj_export_sync_file(struct anv_device
*device
,
1579 int anv_gem_syncobj_import_sync_file(struct anv_device
*device
,
1580 uint32_t handle
, int fd
);
1581 void anv_gem_syncobj_reset(struct anv_device
*device
, uint32_t handle
);
1582 bool anv_gem_supports_syncobj_wait(int fd
);
1583 int anv_gem_syncobj_wait(struct anv_device
*device
,
1584 uint32_t *handles
, uint32_t num_handles
,
1585 int64_t abs_timeout_ns
, bool wait_all
);
1587 uint64_t anv_vma_alloc(struct anv_device
*device
,
1588 uint64_t size
, uint64_t align
,
1589 enum anv_bo_alloc_flags alloc_flags
,
1590 uint64_t client_address
);
1591 void anv_vma_free(struct anv_device
*device
,
1592 uint64_t address
, uint64_t size
);
1594 struct anv_reloc_list
{
1595 uint32_t num_relocs
;
1596 uint32_t array_length
;
1597 struct drm_i915_gem_relocation_entry
* relocs
;
1598 struct anv_bo
** reloc_bos
;
1603 VkResult
anv_reloc_list_init(struct anv_reloc_list
*list
,
1604 const VkAllocationCallbacks
*alloc
);
1605 void anv_reloc_list_finish(struct anv_reloc_list
*list
,
1606 const VkAllocationCallbacks
*alloc
);
1608 VkResult
anv_reloc_list_add(struct anv_reloc_list
*list
,
1609 const VkAllocationCallbacks
*alloc
,
1610 uint32_t offset
, struct anv_bo
*target_bo
,
1611 uint32_t delta
, uint64_t *address_u64_out
);
1613 struct anv_batch_bo
{
1614 /* Link in the anv_cmd_buffer.owned_batch_bos list */
1615 struct list_head link
;
1619 /* Bytes actually consumed in this batch BO */
1622 struct anv_reloc_list relocs
;
1626 const VkAllocationCallbacks
* alloc
;
1628 struct anv_address start_addr
;
1634 struct anv_reloc_list
* relocs
;
1636 /* This callback is called (with the associated user data) in the event
1637 * that the batch runs out of space.
1639 VkResult (*extend_cb
)(struct anv_batch
*, void *);
1643 * Current error status of the command buffer. Used to track inconsistent
1644 * or incomplete command buffer states that are the consequence of run-time
1645 * errors such as out of memory scenarios. We want to track this in the
1646 * batch because the command buffer object is not visible to some parts
1652 void *anv_batch_emit_dwords(struct anv_batch
*batch
, int num_dwords
);
1653 void anv_batch_emit_batch(struct anv_batch
*batch
, struct anv_batch
*other
);
1654 uint64_t anv_batch_emit_reloc(struct anv_batch
*batch
,
1655 void *location
, struct anv_bo
*bo
, uint32_t offset
);
1656 struct anv_address
anv_batch_address(struct anv_batch
*batch
, void *batch_location
);
1659 anv_batch_set_storage(struct anv_batch
*batch
, struct anv_address addr
,
1660 void *map
, size_t size
)
1662 batch
->start_addr
= addr
;
1663 batch
->next
= batch
->start
= map
;
1664 batch
->end
= map
+ size
;
1667 static inline VkResult
1668 anv_batch_set_error(struct anv_batch
*batch
, VkResult error
)
1670 assert(error
!= VK_SUCCESS
);
1671 if (batch
->status
== VK_SUCCESS
)
1672 batch
->status
= error
;
1673 return batch
->status
;
1677 anv_batch_has_error(struct anv_batch
*batch
)
1679 return batch
->status
!= VK_SUCCESS
;
1682 #define ANV_NULL_ADDRESS ((struct anv_address) { NULL, 0 })
1685 anv_address_is_null(struct anv_address addr
)
1687 return addr
.bo
== NULL
&& addr
.offset
== 0;
1690 static inline uint64_t
1691 anv_address_physical(struct anv_address addr
)
1693 if (addr
.bo
&& (addr
.bo
->flags
& EXEC_OBJECT_PINNED
))
1694 return gen_canonical_address(addr
.bo
->offset
+ addr
.offset
);
1696 return gen_canonical_address(addr
.offset
);
1699 static inline struct anv_address
1700 anv_address_add(struct anv_address addr
, uint64_t offset
)
1702 addr
.offset
+= offset
;
1707 write_reloc(const struct anv_device
*device
, void *p
, uint64_t v
, bool flush
)
1709 unsigned reloc_size
= 0;
1710 if (device
->info
.gen
>= 8) {
1711 reloc_size
= sizeof(uint64_t);
1712 *(uint64_t *)p
= gen_canonical_address(v
);
1714 reloc_size
= sizeof(uint32_t);
1718 if (flush
&& !device
->info
.has_llc
)
1719 gen_flush_range(p
, reloc_size
);
1722 static inline uint64_t
1723 _anv_combine_address(struct anv_batch
*batch
, void *location
,
1724 const struct anv_address address
, uint32_t delta
)
1726 if (address
.bo
== NULL
) {
1727 return address
.offset
+ delta
;
1729 assert(batch
->start
<= location
&& location
< batch
->end
);
1731 return anv_batch_emit_reloc(batch
, location
, address
.bo
, address
.offset
+ delta
);
1735 #define __gen_address_type struct anv_address
1736 #define __gen_user_data struct anv_batch
1737 #define __gen_combine_address _anv_combine_address
1739 /* Wrapper macros needed to work around preprocessor argument issues. In
1740 * particular, arguments don't get pre-evaluated if they are concatenated.
1741 * This means that, if you pass GENX(3DSTATE_PS) into the emit macro, the
1742 * GENX macro won't get evaluated if the emit macro contains "cmd ## foo".
1743 * We can work around this easily enough with these helpers.
1745 #define __anv_cmd_length(cmd) cmd ## _length
1746 #define __anv_cmd_length_bias(cmd) cmd ## _length_bias
1747 #define __anv_cmd_header(cmd) cmd ## _header
1748 #define __anv_cmd_pack(cmd) cmd ## _pack
1749 #define __anv_reg_num(reg) reg ## _num
1751 #define anv_pack_struct(dst, struc, ...) do { \
1752 struct struc __template = { \
1755 __anv_cmd_pack(struc)(NULL, dst, &__template); \
1756 VG(VALGRIND_CHECK_MEM_IS_DEFINED(dst, __anv_cmd_length(struc) * 4)); \
1759 #define anv_batch_emitn(batch, n, cmd, ...) ({ \
1760 void *__dst = anv_batch_emit_dwords(batch, n); \
1762 struct cmd __template = { \
1763 __anv_cmd_header(cmd), \
1764 .DWordLength = n - __anv_cmd_length_bias(cmd), \
1767 __anv_cmd_pack(cmd)(batch, __dst, &__template); \
1772 #define anv_batch_emit_merge(batch, dwords0, dwords1) \
1776 STATIC_ASSERT(ARRAY_SIZE(dwords0) == ARRAY_SIZE(dwords1)); \
1777 dw = anv_batch_emit_dwords((batch), ARRAY_SIZE(dwords0)); \
1780 for (uint32_t i = 0; i < ARRAY_SIZE(dwords0); i++) \
1781 dw[i] = (dwords0)[i] | (dwords1)[i]; \
1782 VG(VALGRIND_CHECK_MEM_IS_DEFINED(dw, ARRAY_SIZE(dwords0) * 4));\
1785 #define anv_batch_emit(batch, cmd, name) \
1786 for (struct cmd name = { __anv_cmd_header(cmd) }, \
1787 *_dst = anv_batch_emit_dwords(batch, __anv_cmd_length(cmd)); \
1788 __builtin_expect(_dst != NULL, 1); \
1789 ({ __anv_cmd_pack(cmd)(batch, _dst, &name); \
1790 VG(VALGRIND_CHECK_MEM_IS_DEFINED(_dst, __anv_cmd_length(cmd) * 4)); \
1794 /* #define __gen_get_batch_dwords anv_batch_emit_dwords */
1795 /* #define __gen_get_batch_address anv_batch_address */
1796 /* #define __gen_address_value anv_address_physical */
1797 /* #define __gen_address_offset anv_address_add */
1799 struct anv_device_memory
{
1800 struct vk_object_base base
;
1802 struct list_head link
;
1805 struct anv_memory_type
* type
;
1806 VkDeviceSize map_size
;
1809 /* If set, we are holding reference to AHardwareBuffer
1810 * which we must release when memory is freed.
1812 struct AHardwareBuffer
* ahw
;
1814 /* If set, this memory comes from a host pointer. */
1819 * Header for Vertex URB Entry (VUE)
1821 struct anv_vue_header
{
1823 uint32_t RTAIndex
; /* RenderTargetArrayIndex */
1824 uint32_t ViewportIndex
;
1828 /** Struct representing a sampled image descriptor
1830 * This descriptor layout is used for sampled images, bare sampler, and
1831 * combined image/sampler descriptors.
1833 struct anv_sampled_image_descriptor
{
1834 /** Bindless image handle
1836 * This is expected to already be shifted such that the 20-bit
1837 * SURFACE_STATE table index is in the top 20 bits.
1841 /** Bindless sampler handle
1843 * This is assumed to be a 32B-aligned SAMPLER_STATE pointer relative
1844 * to the dynamic state base address.
1849 struct anv_texture_swizzle_descriptor
{
1852 * See also nir_intrinsic_channel_select_intel
1856 /** Unused padding to ensure the struct is a multiple of 64 bits */
1860 /** Struct representing a storage image descriptor */
1861 struct anv_storage_image_descriptor
{
1862 /** Bindless image handles
1864 * These are expected to already be shifted such that the 20-bit
1865 * SURFACE_STATE table index is in the top 20 bits.
1867 uint32_t read_write
;
1868 uint32_t write_only
;
1871 /** Struct representing a address/range descriptor
1873 * The fields of this struct correspond directly to the data layout of
1874 * nir_address_format_64bit_bounded_global addresses. The last field is the
1875 * offset in the NIR address so it must be zero so that when you load the
1876 * descriptor you get a pointer to the start of the range.
1878 struct anv_address_range_descriptor
{
1884 enum anv_descriptor_data
{
1885 /** The descriptor contains a BTI reference to a surface state */
1886 ANV_DESCRIPTOR_SURFACE_STATE
= (1 << 0),
1887 /** The descriptor contains a BTI reference to a sampler state */
1888 ANV_DESCRIPTOR_SAMPLER_STATE
= (1 << 1),
1889 /** The descriptor contains an actual buffer view */
1890 ANV_DESCRIPTOR_BUFFER_VIEW
= (1 << 2),
1891 /** The descriptor contains auxiliary image layout data */
1892 ANV_DESCRIPTOR_IMAGE_PARAM
= (1 << 3),
1893 /** The descriptor contains auxiliary image layout data */
1894 ANV_DESCRIPTOR_INLINE_UNIFORM
= (1 << 4),
1895 /** anv_address_range_descriptor with a buffer address and range */
1896 ANV_DESCRIPTOR_ADDRESS_RANGE
= (1 << 5),
1897 /** Bindless surface handle */
1898 ANV_DESCRIPTOR_SAMPLED_IMAGE
= (1 << 6),
1899 /** Storage image handles */
1900 ANV_DESCRIPTOR_STORAGE_IMAGE
= (1 << 7),
1901 /** Storage image handles */
1902 ANV_DESCRIPTOR_TEXTURE_SWIZZLE
= (1 << 8),
1905 struct anv_descriptor_set_binding_layout
{
1907 /* The type of the descriptors in this binding */
1908 VkDescriptorType type
;
1911 /* Flags provided when this binding was created */
1912 VkDescriptorBindingFlagsEXT flags
;
1914 /* Bitfield representing the type of data this descriptor contains */
1915 enum anv_descriptor_data data
;
1917 /* Maximum number of YCbCr texture/sampler planes */
1918 uint8_t max_plane_count
;
1920 /* Number of array elements in this binding (or size in bytes for inline
1923 uint16_t array_size
;
1925 /* Index into the flattend descriptor set */
1926 uint16_t descriptor_index
;
1928 /* Index into the dynamic state array for a dynamic buffer */
1929 int16_t dynamic_offset_index
;
1931 /* Index into the descriptor set buffer views */
1932 int16_t buffer_view_index
;
1934 /* Offset into the descriptor buffer where this descriptor lives */
1935 uint32_t descriptor_offset
;
1937 /* Immutable samplers (or NULL if no immutable samplers) */
1938 struct anv_sampler
**immutable_samplers
;
1941 unsigned anv_descriptor_size(const struct anv_descriptor_set_binding_layout
*layout
);
1943 unsigned anv_descriptor_type_size(const struct anv_physical_device
*pdevice
,
1944 VkDescriptorType type
);
1946 bool anv_descriptor_supports_bindless(const struct anv_physical_device
*pdevice
,
1947 const struct anv_descriptor_set_binding_layout
*binding
,
1950 bool anv_descriptor_requires_bindless(const struct anv_physical_device
*pdevice
,
1951 const struct anv_descriptor_set_binding_layout
*binding
,
1954 struct anv_descriptor_set_layout
{
1955 struct vk_object_base base
;
1957 /* Descriptor set layouts can be destroyed at almost any time */
1960 /* Number of bindings in this descriptor set */
1961 uint16_t binding_count
;
1963 /* Total size of the descriptor set with room for all array entries */
1966 /* Shader stages affected by this descriptor set */
1967 uint16_t shader_stages
;
1969 /* Number of buffer views in this descriptor set */
1970 uint16_t buffer_view_count
;
1972 /* Number of dynamic offsets used by this descriptor set */
1973 uint16_t dynamic_offset_count
;
1975 /* For each shader stage, which offsets apply to that stage */
1976 uint16_t stage_dynamic_offsets
[MESA_SHADER_STAGES
];
1978 /* Size of the descriptor buffer for this descriptor set */
1979 uint32_t descriptor_buffer_size
;
1981 /* Bindings in this descriptor set */
1982 struct anv_descriptor_set_binding_layout binding
[0];
1985 void anv_descriptor_set_layout_destroy(struct anv_device
*device
,
1986 struct anv_descriptor_set_layout
*layout
);
1989 anv_descriptor_set_layout_ref(struct anv_descriptor_set_layout
*layout
)
1991 assert(layout
&& layout
->ref_cnt
>= 1);
1992 p_atomic_inc(&layout
->ref_cnt
);
1996 anv_descriptor_set_layout_unref(struct anv_device
*device
,
1997 struct anv_descriptor_set_layout
*layout
)
1999 assert(layout
&& layout
->ref_cnt
>= 1);
2000 if (p_atomic_dec_zero(&layout
->ref_cnt
))
2001 anv_descriptor_set_layout_destroy(device
, layout
);
2004 struct anv_descriptor
{
2005 VkDescriptorType type
;
2009 VkImageLayout layout
;
2010 struct anv_image_view
*image_view
;
2011 struct anv_sampler
*sampler
;
2015 struct anv_buffer
*buffer
;
2020 struct anv_buffer_view
*buffer_view
;
2024 struct anv_descriptor_set
{
2025 struct vk_object_base base
;
2027 struct anv_descriptor_pool
*pool
;
2028 struct anv_descriptor_set_layout
*layout
;
2030 /* Amount of space occupied in the the pool by this descriptor set. It can
2031 * be larger than the size of the descriptor set.
2035 /* State relative to anv_descriptor_pool::bo */
2036 struct anv_state desc_mem
;
2037 /* Surface state for the descriptor buffer */
2038 struct anv_state desc_surface_state
;
2040 uint32_t buffer_view_count
;
2041 struct anv_buffer_view
*buffer_views
;
2043 /* Link to descriptor pool's desc_sets list . */
2044 struct list_head pool_link
;
2046 struct anv_descriptor descriptors
[0];
2049 struct anv_buffer_view
{
2050 struct vk_object_base base
;
2052 enum isl_format format
; /**< VkBufferViewCreateInfo::format */
2053 uint64_t range
; /**< VkBufferViewCreateInfo::range */
2055 struct anv_address address
;
2057 struct anv_state surface_state
;
2058 struct anv_state storage_surface_state
;
2059 struct anv_state writeonly_storage_surface_state
;
2061 struct brw_image_param storage_image_param
;
2064 struct anv_push_descriptor_set
{
2065 struct anv_descriptor_set set
;
2067 /* Put this field right behind anv_descriptor_set so it fills up the
2068 * descriptors[0] field. */
2069 struct anv_descriptor descriptors
[MAX_PUSH_DESCRIPTORS
];
2071 /** True if the descriptor set buffer has been referenced by a draw or
2074 bool set_used_on_gpu
;
2076 struct anv_buffer_view buffer_views
[MAX_PUSH_DESCRIPTORS
];
2079 struct anv_descriptor_pool
{
2080 struct vk_object_base base
;
2087 struct util_vma_heap bo_heap
;
2089 struct anv_state_stream surface_state_stream
;
2090 void *surface_state_free_list
;
2092 struct list_head desc_sets
;
2097 enum anv_descriptor_template_entry_type
{
2098 ANV_DESCRIPTOR_TEMPLATE_ENTRY_TYPE_IMAGE
,
2099 ANV_DESCRIPTOR_TEMPLATE_ENTRY_TYPE_BUFFER
,
2100 ANV_DESCRIPTOR_TEMPLATE_ENTRY_TYPE_BUFFER_VIEW
2103 struct anv_descriptor_template_entry
{
2104 /* The type of descriptor in this entry */
2105 VkDescriptorType type
;
2107 /* Binding in the descriptor set */
2110 /* Offset at which to write into the descriptor set binding */
2111 uint32_t array_element
;
2113 /* Number of elements to write into the descriptor set binding */
2114 uint32_t array_count
;
2116 /* Offset into the user provided data */
2119 /* Stride between elements into the user provided data */
2123 struct anv_descriptor_update_template
{
2124 struct vk_object_base base
;
2126 VkPipelineBindPoint bind_point
;
2128 /* The descriptor set this template corresponds to. This value is only
2129 * valid if the template was created with the templateType
2130 * VK_DESCRIPTOR_UPDATE_TEMPLATE_TYPE_DESCRIPTOR_SET.
2134 /* Number of entries in this template */
2135 uint32_t entry_count
;
2137 /* Entries of the template */
2138 struct anv_descriptor_template_entry entries
[0];
2142 anv_descriptor_set_layout_size(const struct anv_descriptor_set_layout
*layout
);
2145 anv_descriptor_set_write_image_view(struct anv_device
*device
,
2146 struct anv_descriptor_set
*set
,
2147 const VkDescriptorImageInfo
* const info
,
2148 VkDescriptorType type
,
2153 anv_descriptor_set_write_buffer_view(struct anv_device
*device
,
2154 struct anv_descriptor_set
*set
,
2155 VkDescriptorType type
,
2156 struct anv_buffer_view
*buffer_view
,
2161 anv_descriptor_set_write_buffer(struct anv_device
*device
,
2162 struct anv_descriptor_set
*set
,
2163 struct anv_state_stream
*alloc_stream
,
2164 VkDescriptorType type
,
2165 struct anv_buffer
*buffer
,
2168 VkDeviceSize offset
,
2169 VkDeviceSize range
);
2171 anv_descriptor_set_write_inline_uniform_data(struct anv_device
*device
,
2172 struct anv_descriptor_set
*set
,
2179 anv_descriptor_set_write_template(struct anv_device
*device
,
2180 struct anv_descriptor_set
*set
,
2181 struct anv_state_stream
*alloc_stream
,
2182 const struct anv_descriptor_update_template
*template,
2186 anv_descriptor_set_create(struct anv_device
*device
,
2187 struct anv_descriptor_pool
*pool
,
2188 struct anv_descriptor_set_layout
*layout
,
2189 struct anv_descriptor_set
**out_set
);
2192 anv_descriptor_set_destroy(struct anv_device
*device
,
2193 struct anv_descriptor_pool
*pool
,
2194 struct anv_descriptor_set
*set
);
2196 #define ANV_DESCRIPTOR_SET_NULL (UINT8_MAX - 5)
2197 #define ANV_DESCRIPTOR_SET_PUSH_CONSTANTS (UINT8_MAX - 4)
2198 #define ANV_DESCRIPTOR_SET_DESCRIPTORS (UINT8_MAX - 3)
2199 #define ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS (UINT8_MAX - 2)
2200 #define ANV_DESCRIPTOR_SET_SHADER_CONSTANTS (UINT8_MAX - 1)
2201 #define ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS UINT8_MAX
2203 struct anv_pipeline_binding
{
2204 /** Index in the descriptor set
2206 * This is a flattened index; the descriptor set layout is already taken
2211 /** The descriptor set this surface corresponds to.
2213 * The special ANV_DESCRIPTOR_SET_* values above indicates that this
2214 * binding is not a normal descriptor set but something else.
2219 /** Plane in the binding index for images */
2222 /** Input attachment index (relative to the subpass) */
2223 uint8_t input_attachment_index
;
2225 /** Dynamic offset index (for dynamic UBOs and SSBOs) */
2226 uint8_t dynamic_offset_index
;
2229 /** For a storage image, whether it is write-only */
2232 /** Pad to 64 bits so that there are no holes and we can safely memcmp
2233 * assuming POD zero-initialization.
2238 struct anv_push_range
{
2239 /** Index in the descriptor set */
2242 /** Descriptor set index */
2245 /** Dynamic offset index (for dynamic UBOs) */
2246 uint8_t dynamic_offset_index
;
2248 /** Start offset in units of 32B */
2251 /** Range in units of 32B */
2255 struct anv_pipeline_layout
{
2256 struct vk_object_base base
;
2259 struct anv_descriptor_set_layout
*layout
;
2260 uint32_t dynamic_offset_start
;
2265 unsigned char sha1
[20];
2269 struct vk_object_base base
;
2271 struct anv_device
* device
;
2274 VkBufferUsageFlags usage
;
2276 /* Set when bound */
2277 struct anv_address address
;
2280 static inline uint64_t
2281 anv_buffer_get_range(struct anv_buffer
*buffer
, uint64_t offset
, uint64_t range
)
2283 assert(offset
<= buffer
->size
);
2284 if (range
== VK_WHOLE_SIZE
) {
2285 return buffer
->size
- offset
;
2287 assert(range
+ offset
>= range
);
2288 assert(range
+ offset
<= buffer
->size
);
2293 enum anv_cmd_dirty_bits
{
2294 ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
= 1 << 0, /* VK_DYNAMIC_STATE_VIEWPORT */
2295 ANV_CMD_DIRTY_DYNAMIC_SCISSOR
= 1 << 1, /* VK_DYNAMIC_STATE_SCISSOR */
2296 ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH
= 1 << 2, /* VK_DYNAMIC_STATE_LINE_WIDTH */
2297 ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS
= 1 << 3, /* VK_DYNAMIC_STATE_DEPTH_BIAS */
2298 ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS
= 1 << 4, /* VK_DYNAMIC_STATE_BLEND_CONSTANTS */
2299 ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS
= 1 << 5, /* VK_DYNAMIC_STATE_DEPTH_BOUNDS */
2300 ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK
= 1 << 6, /* VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK */
2301 ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK
= 1 << 7, /* VK_DYNAMIC_STATE_STENCIL_WRITE_MASK */
2302 ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE
= 1 << 8, /* VK_DYNAMIC_STATE_STENCIL_REFERENCE */
2303 ANV_CMD_DIRTY_PIPELINE
= 1 << 9,
2304 ANV_CMD_DIRTY_INDEX_BUFFER
= 1 << 10,
2305 ANV_CMD_DIRTY_RENDER_TARGETS
= 1 << 11,
2306 ANV_CMD_DIRTY_XFB_ENABLE
= 1 << 12,
2307 ANV_CMD_DIRTY_DYNAMIC_LINE_STIPPLE
= 1 << 13, /* VK_DYNAMIC_STATE_LINE_STIPPLE_EXT */
2309 typedef uint32_t anv_cmd_dirty_mask_t
;
2311 #define ANV_CMD_DIRTY_DYNAMIC_ALL \
2312 (ANV_CMD_DIRTY_DYNAMIC_VIEWPORT | \
2313 ANV_CMD_DIRTY_DYNAMIC_SCISSOR | \
2314 ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH | \
2315 ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS | \
2316 ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS | \
2317 ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS | \
2318 ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK | \
2319 ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK | \
2320 ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE | \
2321 ANV_CMD_DIRTY_DYNAMIC_LINE_STIPPLE)
2323 static inline enum anv_cmd_dirty_bits
2324 anv_cmd_dirty_bit_for_vk_dynamic_state(VkDynamicState vk_state
)
2327 case VK_DYNAMIC_STATE_VIEWPORT
:
2328 return ANV_CMD_DIRTY_DYNAMIC_VIEWPORT
;
2329 case VK_DYNAMIC_STATE_SCISSOR
:
2330 return ANV_CMD_DIRTY_DYNAMIC_SCISSOR
;
2331 case VK_DYNAMIC_STATE_LINE_WIDTH
:
2332 return ANV_CMD_DIRTY_DYNAMIC_LINE_WIDTH
;
2333 case VK_DYNAMIC_STATE_DEPTH_BIAS
:
2334 return ANV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS
;
2335 case VK_DYNAMIC_STATE_BLEND_CONSTANTS
:
2336 return ANV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS
;
2337 case VK_DYNAMIC_STATE_DEPTH_BOUNDS
:
2338 return ANV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS
;
2339 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
:
2340 return ANV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK
;
2341 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
:
2342 return ANV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK
;
2343 case VK_DYNAMIC_STATE_STENCIL_REFERENCE
:
2344 return ANV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE
;
2345 case VK_DYNAMIC_STATE_LINE_STIPPLE_EXT
:
2346 return ANV_CMD_DIRTY_DYNAMIC_LINE_STIPPLE
;
2348 assert(!"Unsupported dynamic state");
2354 enum anv_pipe_bits
{
2355 ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
= (1 << 0),
2356 ANV_PIPE_STALL_AT_SCOREBOARD_BIT
= (1 << 1),
2357 ANV_PIPE_STATE_CACHE_INVALIDATE_BIT
= (1 << 2),
2358 ANV_PIPE_CONSTANT_CACHE_INVALIDATE_BIT
= (1 << 3),
2359 ANV_PIPE_VF_CACHE_INVALIDATE_BIT
= (1 << 4),
2360 ANV_PIPE_DATA_CACHE_FLUSH_BIT
= (1 << 5),
2361 ANV_PIPE_TILE_CACHE_FLUSH_BIT
= (1 << 6),
2362 ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
= (1 << 10),
2363 ANV_PIPE_INSTRUCTION_CACHE_INVALIDATE_BIT
= (1 << 11),
2364 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
= (1 << 12),
2365 ANV_PIPE_DEPTH_STALL_BIT
= (1 << 13),
2366 ANV_PIPE_CS_STALL_BIT
= (1 << 20),
2367 ANV_PIPE_END_OF_PIPE_SYNC_BIT
= (1 << 21),
2369 /* This bit does not exist directly in PIPE_CONTROL. Instead it means that
2370 * a flush has happened but not a CS stall. The next time we do any sort
2371 * of invalidation we need to insert a CS stall at that time. Otherwise,
2372 * we would have to CS stall on every flush which could be bad.
2374 ANV_PIPE_NEEDS_END_OF_PIPE_SYNC_BIT
= (1 << 22),
2376 /* This bit does not exist directly in PIPE_CONTROL. It means that render
2377 * target operations related to transfer commands with VkBuffer as
2378 * destination are ongoing. Some operations like copies on the command
2379 * streamer might need to be aware of this to trigger the appropriate stall
2380 * before they can proceed with the copy.
2382 ANV_PIPE_RENDER_TARGET_BUFFER_WRITES
= (1 << 23),
2384 /* This bit does not exist directly in PIPE_CONTROL. It means that Gen12
2385 * AUX-TT data has changed and we need to invalidate AUX-TT data. This is
2386 * done by writing the AUX-TT register.
2388 ANV_PIPE_AUX_TABLE_INVALIDATE_BIT
= (1 << 24),
2390 /* This bit does not exist directly in PIPE_CONTROL. It means that a
2391 * PIPE_CONTROL with a post-sync operation will follow. This is used to
2392 * implement a workaround for Gen9.
2394 ANV_PIPE_POST_SYNC_BIT
= (1 << 25),
2397 #define ANV_PIPE_FLUSH_BITS ( \
2398 ANV_PIPE_DEPTH_CACHE_FLUSH_BIT | \
2399 ANV_PIPE_DATA_CACHE_FLUSH_BIT | \
2400 ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT | \
2401 ANV_PIPE_TILE_CACHE_FLUSH_BIT)
2403 #define ANV_PIPE_STALL_BITS ( \
2404 ANV_PIPE_STALL_AT_SCOREBOARD_BIT | \
2405 ANV_PIPE_DEPTH_STALL_BIT | \
2406 ANV_PIPE_CS_STALL_BIT)
2408 #define ANV_PIPE_INVALIDATE_BITS ( \
2409 ANV_PIPE_STATE_CACHE_INVALIDATE_BIT | \
2410 ANV_PIPE_CONSTANT_CACHE_INVALIDATE_BIT | \
2411 ANV_PIPE_VF_CACHE_INVALIDATE_BIT | \
2412 ANV_PIPE_DATA_CACHE_FLUSH_BIT | \
2413 ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT | \
2414 ANV_PIPE_INSTRUCTION_CACHE_INVALIDATE_BIT | \
2415 ANV_PIPE_AUX_TABLE_INVALIDATE_BIT)
2417 static inline enum anv_pipe_bits
2418 anv_pipe_flush_bits_for_access_flags(VkAccessFlags flags
)
2420 enum anv_pipe_bits pipe_bits
= 0;
2423 for_each_bit(b
, flags
) {
2424 switch ((VkAccessFlagBits
)(1 << b
)) {
2425 case VK_ACCESS_SHADER_WRITE_BIT
:
2426 /* We're transitioning a buffer that was previously used as write
2427 * destination through the data port. To make its content available
2428 * to future operations, flush the data cache.
2430 pipe_bits
|= ANV_PIPE_DATA_CACHE_FLUSH_BIT
;
2432 case VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT
:
2433 /* We're transitioning a buffer that was previously used as render
2434 * target. To make its content available to future operations, flush
2435 * the render target cache.
2437 pipe_bits
|= ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
;
2439 case VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT
:
2440 /* We're transitioning a buffer that was previously used as depth
2441 * buffer. To make its content available to future operations, flush
2444 pipe_bits
|= ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
;
2446 case VK_ACCESS_TRANSFER_WRITE_BIT
:
2447 /* We're transitioning a buffer that was previously used as a
2448 * transfer write destination. Generic write operations include color
2449 * & depth operations as well as buffer operations like :
2450 * - vkCmdClearColorImage()
2451 * - vkCmdClearDepthStencilImage()
2452 * - vkCmdBlitImage()
2453 * - vkCmdCopy*(), vkCmdUpdate*(), vkCmdFill*()
2455 * Most of these operations are implemented using Blorp which writes
2456 * through the render target, so flush that cache to make it visible
2457 * to future operations. And for depth related operations we also
2458 * need to flush the depth cache.
2460 pipe_bits
|= ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT
;
2461 pipe_bits
|= ANV_PIPE_DEPTH_CACHE_FLUSH_BIT
;
2463 case VK_ACCESS_MEMORY_WRITE_BIT
:
2464 /* We're transitioning a buffer for generic write operations. Flush
2467 pipe_bits
|= ANV_PIPE_FLUSH_BITS
;
2470 break; /* Nothing to do */
2477 static inline enum anv_pipe_bits
2478 anv_pipe_invalidate_bits_for_access_flags(VkAccessFlags flags
)
2480 enum anv_pipe_bits pipe_bits
= 0;
2483 for_each_bit(b
, flags
) {
2484 switch ((VkAccessFlagBits
)(1 << b
)) {
2485 case VK_ACCESS_INDIRECT_COMMAND_READ_BIT
:
2486 /* Indirect draw commands take a buffer as input that we're going to
2487 * read from the command streamer to load some of the HW registers
2488 * (see genX_cmd_buffer.c:load_indirect_parameters). This requires a
2489 * command streamer stall so that all the cache flushes have
2490 * completed before the command streamer loads from memory.
2492 pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
2493 /* Indirect draw commands also set gl_BaseVertex & gl_BaseIndex
2494 * through a vertex buffer, so invalidate that cache.
2496 pipe_bits
|= ANV_PIPE_VF_CACHE_INVALIDATE_BIT
;
2497 /* For CmdDipatchIndirect, we also load gl_NumWorkGroups through a
2498 * UBO from the buffer, so we need to invalidate constant cache.
2500 pipe_bits
|= ANV_PIPE_CONSTANT_CACHE_INVALIDATE_BIT
;
2502 case VK_ACCESS_INDEX_READ_BIT
:
2503 case VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT
:
2504 /* We transitioning a buffer to be used for as input for vkCmdDraw*
2505 * commands, so we invalidate the VF cache to make sure there is no
2506 * stale data when we start rendering.
2508 pipe_bits
|= ANV_PIPE_VF_CACHE_INVALIDATE_BIT
;
2510 case VK_ACCESS_UNIFORM_READ_BIT
:
2511 /* We transitioning a buffer to be used as uniform data. Because
2512 * uniform is accessed through the data port & sampler, we need to
2513 * invalidate the texture cache (sampler) & constant cache (data
2514 * port) to avoid stale data.
2516 pipe_bits
|= ANV_PIPE_CONSTANT_CACHE_INVALIDATE_BIT
;
2517 pipe_bits
|= ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
;
2519 case VK_ACCESS_SHADER_READ_BIT
:
2520 case VK_ACCESS_INPUT_ATTACHMENT_READ_BIT
:
2521 case VK_ACCESS_TRANSFER_READ_BIT
:
2522 /* Transitioning a buffer to be read through the sampler, so
2523 * invalidate the texture cache, we don't want any stale data.
2525 pipe_bits
|= ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT
;
2527 case VK_ACCESS_MEMORY_READ_BIT
:
2528 /* Transitioning a buffer for generic read, invalidate all the
2531 pipe_bits
|= ANV_PIPE_INVALIDATE_BITS
;
2533 case VK_ACCESS_MEMORY_WRITE_BIT
:
2534 /* Generic write, make sure all previously written things land in
2537 pipe_bits
|= ANV_PIPE_FLUSH_BITS
;
2539 case VK_ACCESS_CONDITIONAL_RENDERING_READ_BIT_EXT
:
2540 /* Transitioning a buffer for conditional rendering. We'll load the
2541 * content of this buffer into HW registers using the command
2542 * streamer, so we need to stall the command streamer to make sure
2543 * any in-flight flush operations have completed.
2545 pipe_bits
|= ANV_PIPE_CS_STALL_BIT
;
2548 break; /* Nothing to do */
2555 #define VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV ( \
2556 VK_IMAGE_ASPECT_COLOR_BIT | \
2557 VK_IMAGE_ASPECT_PLANE_0_BIT | \
2558 VK_IMAGE_ASPECT_PLANE_1_BIT | \
2559 VK_IMAGE_ASPECT_PLANE_2_BIT)
2560 #define VK_IMAGE_ASPECT_PLANES_BITS_ANV ( \
2561 VK_IMAGE_ASPECT_PLANE_0_BIT | \
2562 VK_IMAGE_ASPECT_PLANE_1_BIT | \
2563 VK_IMAGE_ASPECT_PLANE_2_BIT)
2565 struct anv_vertex_binding
{
2566 struct anv_buffer
* buffer
;
2567 VkDeviceSize offset
;
2570 struct anv_xfb_binding
{
2571 struct anv_buffer
* buffer
;
2572 VkDeviceSize offset
;
2576 struct anv_push_constants
{
2577 /** Push constant data provided by the client through vkPushConstants */
2578 uint8_t client_data
[MAX_PUSH_CONSTANTS_SIZE
];
2580 /** Dynamic offsets for dynamic UBOs and SSBOs */
2581 uint32_t dynamic_offsets
[MAX_DYNAMIC_BUFFERS
];
2583 uint64_t push_reg_mask
;
2585 /** Pad out to a multiple of 32 bytes */
2589 /** Base workgroup ID
2591 * Used for vkCmdDispatchBase.
2593 uint32_t base_work_group_id
[3];
2597 * This is never set by software but is implicitly filled out when
2598 * uploading the push constants for compute shaders.
2600 uint32_t subgroup_id
;
2604 struct anv_dynamic_state
{
2607 VkViewport viewports
[MAX_VIEWPORTS
];
2612 VkRect2D scissors
[MAX_SCISSORS
];
2623 float blend_constants
[4];
2633 } stencil_compare_mask
;
2638 } stencil_write_mask
;
2643 } stencil_reference
;
2651 extern const struct anv_dynamic_state default_dynamic_state
;
2653 uint32_t anv_dynamic_state_copy(struct anv_dynamic_state
*dest
,
2654 const struct anv_dynamic_state
*src
,
2655 uint32_t copy_mask
);
2657 struct anv_surface_state
{
2658 struct anv_state state
;
2659 /** Address of the surface referred to by this state
2661 * This address is relative to the start of the BO.
2663 struct anv_address address
;
2664 /* Address of the aux surface, if any
2666 * This field is ANV_NULL_ADDRESS if and only if no aux surface exists.
2668 * With the exception of gen8, the bottom 12 bits of this address' offset
2669 * include extra aux information.
2671 struct anv_address aux_address
;
2672 /* Address of the clear color, if any
2674 * This address is relative to the start of the BO.
2676 struct anv_address clear_address
;
2680 * Attachment state when recording a renderpass instance.
2682 * The clear value is valid only if there exists a pending clear.
2684 struct anv_attachment_state
{
2685 enum isl_aux_usage aux_usage
;
2686 struct anv_surface_state color
;
2687 struct anv_surface_state input
;
2689 VkImageLayout current_layout
;
2690 VkImageLayout current_stencil_layout
;
2691 VkImageAspectFlags pending_clear_aspects
;
2692 VkImageAspectFlags pending_load_aspects
;
2694 VkClearValue clear_value
;
2696 /* When multiview is active, attachments with a renderpass clear
2697 * operation have their respective layers cleared on the first
2698 * subpass that uses them, and only in that subpass. We keep track
2699 * of this using a bitfield to indicate which layers of an attachment
2700 * have not been cleared yet when multiview is active.
2702 uint32_t pending_clear_views
;
2703 struct anv_image_view
* image_view
;
2706 /** State tracking for vertex buffer flushes
2708 * On Gen8-9, the VF cache only considers the bottom 32 bits of memory
2709 * addresses. If you happen to have two vertex buffers which get placed
2710 * exactly 4 GiB apart and use them in back-to-back draw calls, you can get
2711 * collisions. In order to solve this problem, we track vertex address ranges
2712 * which are live in the cache and invalidate the cache if one ever exceeds 32
2715 struct anv_vb_cache_range
{
2716 /* Virtual address at which the live vertex buffer cache range starts for
2717 * this vertex buffer index.
2721 /* Virtual address of the byte after where vertex buffer cache range ends.
2722 * This is exclusive such that end - start is the size of the range.
2727 /** State tracking for particular pipeline bind point
2729 * This struct is the base struct for anv_cmd_graphics_state and
2730 * anv_cmd_compute_state. These are used to track state which is bound to a
2731 * particular type of pipeline. Generic state that applies per-stage such as
2732 * binding table offsets and push constants is tracked generically with a
2733 * per-stage array in anv_cmd_state.
2735 struct anv_cmd_pipeline_state
{
2736 struct anv_descriptor_set
*descriptors
[MAX_SETS
];
2737 struct anv_push_descriptor_set
*push_descriptors
[MAX_SETS
];
2740 /** State tracking for graphics pipeline
2742 * This has anv_cmd_pipeline_state as a base struct to track things which get
2743 * bound to a graphics pipeline. Along with general pipeline bind point state
2744 * which is in the anv_cmd_pipeline_state base struct, it also contains other
2745 * state which is graphics-specific.
2747 struct anv_cmd_graphics_state
{
2748 struct anv_cmd_pipeline_state base
;
2750 struct anv_graphics_pipeline
*pipeline
;
2752 anv_cmd_dirty_mask_t dirty
;
2755 struct anv_vb_cache_range ib_bound_range
;
2756 struct anv_vb_cache_range ib_dirty_range
;
2757 struct anv_vb_cache_range vb_bound_ranges
[33];
2758 struct anv_vb_cache_range vb_dirty_ranges
[33];
2760 struct anv_dynamic_state dynamic
;
2763 struct anv_buffer
*index_buffer
;
2764 uint32_t index_type
; /**< 3DSTATE_INDEX_BUFFER.IndexFormat */
2765 uint32_t index_offset
;
2769 /** State tracking for compute pipeline
2771 * This has anv_cmd_pipeline_state as a base struct to track things which get
2772 * bound to a compute pipeline. Along with general pipeline bind point state
2773 * which is in the anv_cmd_pipeline_state base struct, it also contains other
2774 * state which is compute-specific.
2776 struct anv_cmd_compute_state
{
2777 struct anv_cmd_pipeline_state base
;
2779 struct anv_compute_pipeline
*pipeline
;
2781 bool pipeline_dirty
;
2783 struct anv_address num_workgroups
;
2786 /** State required while building cmd buffer */
2787 struct anv_cmd_state
{
2788 /* PIPELINE_SELECT.PipelineSelection */
2789 uint32_t current_pipeline
;
2790 const struct gen_l3_config
* current_l3_config
;
2791 uint32_t last_aux_map_state
;
2793 struct anv_cmd_graphics_state gfx
;
2794 struct anv_cmd_compute_state compute
;
2796 enum anv_pipe_bits pending_pipe_bits
;
2797 VkShaderStageFlags descriptors_dirty
;
2798 VkShaderStageFlags push_constants_dirty
;
2800 struct anv_framebuffer
* framebuffer
;
2801 struct anv_render_pass
* pass
;
2802 struct anv_subpass
* subpass
;
2803 VkRect2D render_area
;
2804 uint32_t restart_index
;
2805 struct anv_vertex_binding vertex_bindings
[MAX_VBS
];
2807 struct anv_xfb_binding xfb_bindings
[MAX_XFB_BUFFERS
];
2808 VkShaderStageFlags push_constant_stages
;
2809 struct anv_push_constants push_constants
[MESA_SHADER_STAGES
];
2810 struct anv_state binding_tables
[MESA_SHADER_STAGES
];
2811 struct anv_state samplers
[MESA_SHADER_STAGES
];
2813 unsigned char sampler_sha1s
[MESA_SHADER_STAGES
][20];
2814 unsigned char surface_sha1s
[MESA_SHADER_STAGES
][20];
2815 unsigned char push_sha1s
[MESA_SHADER_STAGES
][20];
2818 * Whether or not the gen8 PMA fix is enabled. We ensure that, at the top
2819 * of any command buffer it is disabled by disabling it in EndCommandBuffer
2820 * and before invoking the secondary in ExecuteCommands.
2822 bool pma_fix_enabled
;
2825 * Whether or not we know for certain that HiZ is enabled for the current
2826 * subpass. If, for whatever reason, we are unsure as to whether HiZ is
2827 * enabled or not, this will be false.
2831 bool conditional_render_enabled
;
2834 * Last rendering scale argument provided to
2835 * genX(cmd_buffer_emit_hashing_mode)().
2837 unsigned current_hash_scale
;
2840 * Array length is anv_cmd_state::pass::attachment_count. Array content is
2841 * valid only when recording a render pass instance.
2843 struct anv_attachment_state
* attachments
;
2846 * Surface states for color render targets. These are stored in a single
2847 * flat array. For depth-stencil attachments, the surface state is simply
2850 struct anv_state attachment_states
;
2853 * A null surface state of the right size to match the framebuffer. This
2854 * is one of the states in attachment_states.
2856 struct anv_state null_surface_state
;
2859 struct anv_cmd_pool
{
2860 struct vk_object_base base
;
2861 VkAllocationCallbacks alloc
;
2862 struct list_head cmd_buffers
;
2865 #define ANV_CMD_BUFFER_BATCH_SIZE 8192
2867 enum anv_cmd_buffer_exec_mode
{
2868 ANV_CMD_BUFFER_EXEC_MODE_PRIMARY
,
2869 ANV_CMD_BUFFER_EXEC_MODE_EMIT
,
2870 ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT
,
2871 ANV_CMD_BUFFER_EXEC_MODE_CHAIN
,
2872 ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN
,
2873 ANV_CMD_BUFFER_EXEC_MODE_CALL_AND_RETURN
,
2876 struct anv_cmd_buffer
{
2877 struct vk_object_base base
;
2879 struct anv_device
* device
;
2881 struct anv_cmd_pool
* pool
;
2882 struct list_head pool_link
;
2884 struct anv_batch batch
;
2886 /* Fields required for the actual chain of anv_batch_bo's.
2888 * These fields are initialized by anv_cmd_buffer_init_batch_bo_chain().
2890 struct list_head batch_bos
;
2891 enum anv_cmd_buffer_exec_mode exec_mode
;
2893 /* A vector of anv_batch_bo pointers for every batch or surface buffer
2894 * referenced by this command buffer
2896 * initialized by anv_cmd_buffer_init_batch_bo_chain()
2898 struct u_vector seen_bbos
;
2900 /* A vector of int32_t's for every block of binding tables.
2902 * initialized by anv_cmd_buffer_init_batch_bo_chain()
2904 struct u_vector bt_block_states
;
2905 struct anv_state bt_next
;
2907 struct anv_reloc_list surface_relocs
;
2908 /** Last seen surface state block pool center bo offset */
2909 uint32_t last_ss_pool_center
;
2911 /* Serial for tracking buffer completion */
2914 /* Stream objects for storing temporary data */
2915 struct anv_state_stream surface_state_stream
;
2916 struct anv_state_stream dynamic_state_stream
;
2918 VkCommandBufferUsageFlags usage_flags
;
2919 VkCommandBufferLevel level
;
2921 struct anv_query_pool
*perf_query_pool
;
2923 struct anv_cmd_state state
;
2925 struct anv_address return_addr
;
2927 /* Set by SetPerformanceMarkerINTEL, written into queries by CmdBeginQuery */
2928 uint64_t intel_perf_marker
;
2931 VkResult
anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
);
2932 void anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
);
2933 void anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer
*cmd_buffer
);
2934 void anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer
*cmd_buffer
);
2935 void anv_cmd_buffer_add_secondary(struct anv_cmd_buffer
*primary
,
2936 struct anv_cmd_buffer
*secondary
);
2937 void anv_cmd_buffer_prepare_execbuf(struct anv_cmd_buffer
*cmd_buffer
);
2938 VkResult
anv_cmd_buffer_execbuf(struct anv_queue
*queue
,
2939 struct anv_cmd_buffer
*cmd_buffer
,
2940 const VkSemaphore
*in_semaphores
,
2941 const uint64_t *in_wait_values
,
2942 uint32_t num_in_semaphores
,
2943 const VkSemaphore
*out_semaphores
,
2944 const uint64_t *out_signal_values
,
2945 uint32_t num_out_semaphores
,
2947 int perf_query_pass
);
2949 VkResult
anv_cmd_buffer_reset(struct anv_cmd_buffer
*cmd_buffer
);
2951 struct anv_state
anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer
*cmd_buffer
,
2952 const void *data
, uint32_t size
, uint32_t alignment
);
2953 struct anv_state
anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer
*cmd_buffer
,
2954 uint32_t *a
, uint32_t *b
,
2955 uint32_t dwords
, uint32_t alignment
);
2958 anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer
*cmd_buffer
);
2960 anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
2961 uint32_t entries
, uint32_t *state_offset
);
2963 anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer
*cmd_buffer
);
2965 anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer
*cmd_buffer
,
2966 uint32_t size
, uint32_t alignment
);
2969 anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer
*cmd_buffer
);
2971 void gen8_cmd_buffer_emit_viewport(struct anv_cmd_buffer
*cmd_buffer
);
2972 void gen8_cmd_buffer_emit_depth_viewport(struct anv_cmd_buffer
*cmd_buffer
,
2973 bool depth_clamp_enable
);
2974 void gen7_cmd_buffer_emit_scissor(struct anv_cmd_buffer
*cmd_buffer
);
2976 void anv_cmd_buffer_setup_attachments(struct anv_cmd_buffer
*cmd_buffer
,
2977 struct anv_render_pass
*pass
,
2978 struct anv_framebuffer
*framebuffer
,
2979 const VkClearValue
*clear_values
);
2981 void anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer
*cmd_buffer
);
2984 anv_cmd_buffer_push_constants(struct anv_cmd_buffer
*cmd_buffer
,
2985 gl_shader_stage stage
);
2987 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer
*cmd_buffer
);
2989 const struct anv_image_view
*
2990 anv_cmd_buffer_get_depth_stencil_view(const struct anv_cmd_buffer
*cmd_buffer
);
2993 anv_cmd_buffer_alloc_blorp_binding_table(struct anv_cmd_buffer
*cmd_buffer
,
2994 uint32_t num_entries
,
2995 uint32_t *state_offset
,
2996 struct anv_state
*bt_state
);
2998 void anv_cmd_buffer_dump(struct anv_cmd_buffer
*cmd_buffer
);
3000 void anv_cmd_emit_conditional_render_predicate(struct anv_cmd_buffer
*cmd_buffer
);
3002 enum anv_fence_type
{
3003 ANV_FENCE_TYPE_NONE
= 0,
3005 ANV_FENCE_TYPE_WSI_BO
,
3006 ANV_FENCE_TYPE_SYNCOBJ
,
3010 enum anv_bo_fence_state
{
3011 /** Indicates that this is a new (or newly reset fence) */
3012 ANV_BO_FENCE_STATE_RESET
,
3014 /** Indicates that this fence has been submitted to the GPU but is still
3015 * (as far as we know) in use by the GPU.
3017 ANV_BO_FENCE_STATE_SUBMITTED
,
3019 ANV_BO_FENCE_STATE_SIGNALED
,
3022 struct anv_fence_impl
{
3023 enum anv_fence_type type
;
3026 /** Fence implementation for BO fences
3028 * These fences use a BO and a set of CPU-tracked state flags. The BO
3029 * is added to the object list of the last execbuf call in a QueueSubmit
3030 * and is marked EXEC_WRITE. The state flags track when the BO has been
3031 * submitted to the kernel. We need to do this because Vulkan lets you
3032 * wait on a fence that has not yet been submitted and I915_GEM_BUSY
3033 * will say it's idle in this case.
3037 enum anv_bo_fence_state state
;
3040 /** DRM syncobj handle for syncobj-based fences */
3044 struct wsi_fence
*fence_wsi
;
3049 struct vk_object_base base
;
3051 /* Permanent fence state. Every fence has some form of permanent state
3052 * (type != ANV_SEMAPHORE_TYPE_NONE). This may be a BO to fence on (for
3053 * cross-process fences) or it could just be a dummy for use internally.
3055 struct anv_fence_impl permanent
;
3057 /* Temporary fence state. A fence *may* have temporary state. That state
3058 * is added to the fence by an import operation and is reset back to
3059 * ANV_SEMAPHORE_TYPE_NONE when the fence is reset. A fence with temporary
3060 * state cannot be signaled because the fence must already be signaled
3061 * before the temporary state can be exported from the fence in the other
3062 * process and imported here.
3064 struct anv_fence_impl temporary
;
3067 void anv_fence_reset_temporary(struct anv_device
*device
,
3068 struct anv_fence
*fence
);
3071 struct vk_object_base base
;
3073 struct anv_state state
;
3076 enum anv_semaphore_type
{
3077 ANV_SEMAPHORE_TYPE_NONE
= 0,
3078 ANV_SEMAPHORE_TYPE_DUMMY
,
3079 ANV_SEMAPHORE_TYPE_BO
,
3080 ANV_SEMAPHORE_TYPE_WSI_BO
,
3081 ANV_SEMAPHORE_TYPE_SYNC_FILE
,
3082 ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ
,
3083 ANV_SEMAPHORE_TYPE_TIMELINE
,
3086 struct anv_timeline_point
{
3087 struct list_head link
;
3091 /* Number of waiter on this point, when > 0 the point should not be garbage
3096 /* BO used for synchronization. */
3100 struct anv_timeline
{
3101 pthread_mutex_t mutex
;
3102 pthread_cond_t cond
;
3104 uint64_t highest_past
;
3105 uint64_t highest_pending
;
3107 struct list_head points
;
3108 struct list_head free_points
;
3111 struct anv_semaphore_impl
{
3112 enum anv_semaphore_type type
;
3115 /* A BO representing this semaphore when type == ANV_SEMAPHORE_TYPE_BO
3116 * or type == ANV_SEMAPHORE_TYPE_WSI_BO. This BO will be added to the
3117 * object list on any execbuf2 calls for which this semaphore is used as
3118 * a wait or signal fence. When used as a signal fence or when type ==
3119 * ANV_SEMAPHORE_TYPE_WSI_BO, the EXEC_OBJECT_WRITE flag will be set.
3123 /* The sync file descriptor when type == ANV_SEMAPHORE_TYPE_SYNC_FILE.
3124 * If the semaphore is in the unsignaled state due to either just being
3125 * created or because it has been used for a wait, fd will be -1.
3129 /* Sync object handle when type == ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ.
3130 * Unlike GEM BOs, DRM sync objects aren't deduplicated by the kernel on
3131 * import so we don't need to bother with a userspace cache.
3135 /* Non shareable timeline semaphore
3137 * Used when kernel don't have support for timeline semaphores.
3139 struct anv_timeline timeline
;
3143 struct anv_semaphore
{
3144 struct vk_object_base base
;
3148 /* Permanent semaphore state. Every semaphore has some form of permanent
3149 * state (type != ANV_SEMAPHORE_TYPE_NONE). This may be a BO to fence on
3150 * (for cross-process semaphores0 or it could just be a dummy for use
3153 struct anv_semaphore_impl permanent
;
3155 /* Temporary semaphore state. A semaphore *may* have temporary state.
3156 * That state is added to the semaphore by an import operation and is reset
3157 * back to ANV_SEMAPHORE_TYPE_NONE when the semaphore is waited on. A
3158 * semaphore with temporary state cannot be signaled because the semaphore
3159 * must already be signaled before the temporary state can be exported from
3160 * the semaphore in the other process and imported here.
3162 struct anv_semaphore_impl temporary
;
3165 void anv_semaphore_reset_temporary(struct anv_device
*device
,
3166 struct anv_semaphore
*semaphore
);
3168 struct anv_shader_module
{
3169 struct vk_object_base base
;
3171 unsigned char sha1
[20];
3176 static inline gl_shader_stage
3177 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage
)
3179 assert(__builtin_popcount(vk_stage
) == 1);
3180 return ffs(vk_stage
) - 1;
3183 static inline VkShaderStageFlagBits
3184 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage
)
3186 return (1 << mesa_stage
);
3189 #define ANV_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
3191 #define anv_foreach_stage(stage, stage_bits) \
3192 for (gl_shader_stage stage, \
3193 __tmp = (gl_shader_stage)((stage_bits) & ANV_STAGE_MASK); \
3194 stage = __builtin_ffs(__tmp) - 1, __tmp; \
3195 __tmp &= ~(1 << (stage)))
3197 struct anv_pipeline_bind_map
{
3198 unsigned char surface_sha1
[20];
3199 unsigned char sampler_sha1
[20];
3200 unsigned char push_sha1
[20];
3202 uint32_t surface_count
;
3203 uint32_t sampler_count
;
3205 struct anv_pipeline_binding
* surface_to_descriptor
;
3206 struct anv_pipeline_binding
* sampler_to_descriptor
;
3208 struct anv_push_range push_ranges
[4];
3211 struct anv_shader_bin_key
{
3216 struct anv_shader_bin
{
3219 gl_shader_stage stage
;
3221 const struct anv_shader_bin_key
*key
;
3223 struct anv_state kernel
;
3224 uint32_t kernel_size
;
3226 struct anv_state constant_data
;
3227 uint32_t constant_data_size
;
3229 const struct brw_stage_prog_data
*prog_data
;
3230 uint32_t prog_data_size
;
3232 struct brw_compile_stats stats
[3];
3235 struct nir_xfb_info
*xfb_info
;
3237 struct anv_pipeline_bind_map bind_map
;
3240 struct anv_shader_bin
*
3241 anv_shader_bin_create(struct anv_device
*device
,
3242 gl_shader_stage stage
,
3243 const void *key
, uint32_t key_size
,
3244 const void *kernel
, uint32_t kernel_size
,
3245 const void *constant_data
, uint32_t constant_data_size
,
3246 const struct brw_stage_prog_data
*prog_data
,
3247 uint32_t prog_data_size
,
3248 const struct brw_compile_stats
*stats
, uint32_t num_stats
,
3249 const struct nir_xfb_info
*xfb_info
,
3250 const struct anv_pipeline_bind_map
*bind_map
);
3253 anv_shader_bin_destroy(struct anv_device
*device
, struct anv_shader_bin
*shader
);
3256 anv_shader_bin_ref(struct anv_shader_bin
*shader
)
3258 assert(shader
&& shader
->ref_cnt
>= 1);
3259 p_atomic_inc(&shader
->ref_cnt
);
3263 anv_shader_bin_unref(struct anv_device
*device
, struct anv_shader_bin
*shader
)
3265 assert(shader
&& shader
->ref_cnt
>= 1);
3266 if (p_atomic_dec_zero(&shader
->ref_cnt
))
3267 anv_shader_bin_destroy(device
, shader
);
3270 struct anv_pipeline_executable
{
3271 gl_shader_stage stage
;
3273 struct brw_compile_stats stats
;
3279 enum anv_pipeline_type
{
3280 ANV_PIPELINE_GRAPHICS
,
3281 ANV_PIPELINE_COMPUTE
,
3284 struct anv_pipeline
{
3285 struct vk_object_base base
;
3287 struct anv_device
* device
;
3289 struct anv_batch batch
;
3290 struct anv_reloc_list batch_relocs
;
3294 enum anv_pipeline_type type
;
3295 VkPipelineCreateFlags flags
;
3297 struct util_dynarray executables
;
3299 const struct gen_l3_config
* l3_config
;
3302 struct anv_graphics_pipeline
{
3303 struct anv_pipeline base
;
3305 uint32_t batch_data
[512];
3307 anv_cmd_dirty_mask_t dynamic_state_mask
;
3308 struct anv_dynamic_state dynamic_state
;
3312 struct anv_subpass
* subpass
;
3314 struct anv_shader_bin
* shaders
[MESA_SHADER_STAGES
];
3316 VkShaderStageFlags active_stages
;
3318 bool primitive_restart
;
3320 bool depth_test_enable
;
3321 bool writes_stencil
;
3322 bool stencil_test_enable
;
3323 bool depth_clamp_enable
;
3324 bool depth_clip_enable
;
3325 bool sample_shading_enable
;
3327 bool depth_bounds_test_enable
;
3329 /* When primitive replication is used, subpass->view_mask will describe what
3330 * views to replicate.
3332 bool use_primitive_replication
;
3334 struct anv_state blend_state
;
3337 struct anv_pipeline_vertex_binding
{
3340 uint32_t instance_divisor
;
3345 uint32_t depth_stencil_state
[3];
3351 uint32_t wm_depth_stencil
[3];
3355 uint32_t wm_depth_stencil
[4];
3359 struct anv_compute_pipeline
{
3360 struct anv_pipeline base
;
3362 struct anv_shader_bin
* cs
;
3363 uint32_t cs_right_mask
;
3364 uint32_t batch_data
[9];
3365 uint32_t interface_descriptor_data
[8];
3368 #define ANV_DECL_PIPELINE_DOWNCAST(pipe_type, pipe_enum) \
3369 static inline struct anv_##pipe_type##_pipeline * \
3370 anv_pipeline_to_##pipe_type(struct anv_pipeline *pipeline) \
3372 assert(pipeline->type == pipe_enum); \
3373 return (struct anv_##pipe_type##_pipeline *) pipeline; \
3376 ANV_DECL_PIPELINE_DOWNCAST(graphics
, ANV_PIPELINE_GRAPHICS
)
3377 ANV_DECL_PIPELINE_DOWNCAST(compute
, ANV_PIPELINE_COMPUTE
)
3380 anv_pipeline_has_stage(const struct anv_graphics_pipeline
*pipeline
,
3381 gl_shader_stage stage
)
3383 return (pipeline
->active_stages
& mesa_to_vk_shader_stage(stage
)) != 0;
3386 #define ANV_DECL_GET_GRAPHICS_PROG_DATA_FUNC(prefix, stage) \
3387 static inline const struct brw_##prefix##_prog_data * \
3388 get_##prefix##_prog_data(const struct anv_graphics_pipeline *pipeline) \
3390 if (anv_pipeline_has_stage(pipeline, stage)) { \
3391 return (const struct brw_##prefix##_prog_data *) \
3392 pipeline->shaders[stage]->prog_data; \
3398 ANV_DECL_GET_GRAPHICS_PROG_DATA_FUNC(vs
, MESA_SHADER_VERTEX
)
3399 ANV_DECL_GET_GRAPHICS_PROG_DATA_FUNC(tcs
, MESA_SHADER_TESS_CTRL
)
3400 ANV_DECL_GET_GRAPHICS_PROG_DATA_FUNC(tes
, MESA_SHADER_TESS_EVAL
)
3401 ANV_DECL_GET_GRAPHICS_PROG_DATA_FUNC(gs
, MESA_SHADER_GEOMETRY
)
3402 ANV_DECL_GET_GRAPHICS_PROG_DATA_FUNC(wm
, MESA_SHADER_FRAGMENT
)
3404 static inline const struct brw_cs_prog_data
*
3405 get_cs_prog_data(const struct anv_compute_pipeline
*pipeline
)
3407 assert(pipeline
->cs
);
3408 return (const struct brw_cs_prog_data
*) pipeline
->cs
->prog_data
;
3411 static inline const struct brw_vue_prog_data
*
3412 anv_pipeline_get_last_vue_prog_data(const struct anv_graphics_pipeline
*pipeline
)
3414 if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_GEOMETRY
))
3415 return &get_gs_prog_data(pipeline
)->base
;
3416 else if (anv_pipeline_has_stage(pipeline
, MESA_SHADER_TESS_EVAL
))
3417 return &get_tes_prog_data(pipeline
)->base
;
3419 return &get_vs_prog_data(pipeline
)->base
;
3423 anv_pipeline_init(struct anv_pipeline
*pipeline
,
3424 struct anv_device
*device
,
3425 enum anv_pipeline_type type
,
3426 VkPipelineCreateFlags flags
,
3427 const VkAllocationCallbacks
*pAllocator
);
3430 anv_pipeline_finish(struct anv_pipeline
*pipeline
,
3431 struct anv_device
*device
,
3432 const VkAllocationCallbacks
*pAllocator
);
3435 anv_graphics_pipeline_init(struct anv_graphics_pipeline
*pipeline
, struct anv_device
*device
,
3436 struct anv_pipeline_cache
*cache
,
3437 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
3438 const VkAllocationCallbacks
*alloc
);
3441 anv_pipeline_compile_cs(struct anv_compute_pipeline
*pipeline
,
3442 struct anv_pipeline_cache
*cache
,
3443 const VkComputePipelineCreateInfo
*info
,
3444 const struct anv_shader_module
*module
,
3445 const char *entrypoint
,
3446 const VkSpecializationInfo
*spec_info
);
3448 struct anv_cs_parameters
{
3449 uint32_t group_size
;
3454 struct anv_cs_parameters
3455 anv_cs_parameters(const struct anv_compute_pipeline
*pipeline
);
3457 struct anv_format_plane
{
3458 enum isl_format isl_format
:16;
3459 struct isl_swizzle swizzle
;
3461 /* Whether this plane contains chroma channels */
3464 /* For downscaling of YUV planes */
3465 uint8_t denominator_scales
[2];
3467 /* How to map sampled ycbcr planes to a single 4 component element. */
3468 struct isl_swizzle ycbcr_swizzle
;
3470 /* What aspect is associated to this plane */
3471 VkImageAspectFlags aspect
;
3476 struct anv_format_plane planes
[3];
3483 * Return the aspect's _format_ plane, not its _memory_ plane (using the
3484 * vocabulary of VK_EXT_image_drm_format_modifier). As a consequence, \a
3485 * aspect_mask may contain VK_IMAGE_ASPECT_PLANE_*, but must not contain
3486 * VK_IMAGE_ASPECT_MEMORY_PLANE_* .
3488 static inline uint32_t
3489 anv_image_aspect_to_plane(VkImageAspectFlags image_aspects
,
3490 VkImageAspectFlags aspect_mask
)
3492 switch (aspect_mask
) {
3493 case VK_IMAGE_ASPECT_COLOR_BIT
:
3494 case VK_IMAGE_ASPECT_DEPTH_BIT
:
3495 case VK_IMAGE_ASPECT_PLANE_0_BIT
:
3497 case VK_IMAGE_ASPECT_STENCIL_BIT
:
3498 if ((image_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
) == 0)
3501 case VK_IMAGE_ASPECT_PLANE_1_BIT
:
3503 case VK_IMAGE_ASPECT_PLANE_2_BIT
:
3506 /* Purposefully assert with depth/stencil aspects. */
3507 unreachable("invalid image aspect");
3511 static inline VkImageAspectFlags
3512 anv_plane_to_aspect(VkImageAspectFlags image_aspects
,
3515 if (image_aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) {
3516 if (util_bitcount(image_aspects
) > 1)
3517 return VK_IMAGE_ASPECT_PLANE_0_BIT
<< plane
;
3518 return VK_IMAGE_ASPECT_COLOR_BIT
;
3520 if (image_aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
)
3521 return VK_IMAGE_ASPECT_DEPTH_BIT
<< plane
;
3522 assert(image_aspects
== VK_IMAGE_ASPECT_STENCIL_BIT
);
3523 return VK_IMAGE_ASPECT_STENCIL_BIT
;
3526 #define anv_foreach_image_aspect_bit(b, image, aspects) \
3527 for_each_bit(b, anv_image_expand_aspects(image, aspects))
3529 const struct anv_format
*
3530 anv_get_format(VkFormat format
);
3532 static inline uint32_t
3533 anv_get_format_planes(VkFormat vk_format
)
3535 const struct anv_format
*format
= anv_get_format(vk_format
);
3537 return format
!= NULL
? format
->n_planes
: 0;
3540 struct anv_format_plane
3541 anv_get_format_plane(const struct gen_device_info
*devinfo
, VkFormat vk_format
,
3542 VkImageAspectFlagBits aspect
, VkImageTiling tiling
);
3544 static inline enum isl_format
3545 anv_get_isl_format(const struct gen_device_info
*devinfo
, VkFormat vk_format
,
3546 VkImageAspectFlags aspect
, VkImageTiling tiling
)
3548 return anv_get_format_plane(devinfo
, vk_format
, aspect
, tiling
).isl_format
;
3551 bool anv_formats_ccs_e_compatible(const struct gen_device_info
*devinfo
,
3552 VkImageCreateFlags create_flags
,
3554 VkImageTiling vk_tiling
,
3555 const VkImageFormatListCreateInfoKHR
*fmt_list
);
3557 static inline struct isl_swizzle
3558 anv_swizzle_for_render(struct isl_swizzle swizzle
)
3560 /* Sometimes the swizzle will have alpha map to one. We do this to fake
3561 * RGB as RGBA for texturing
3563 assert(swizzle
.a
== ISL_CHANNEL_SELECT_ONE
||
3564 swizzle
.a
== ISL_CHANNEL_SELECT_ALPHA
);
3566 /* But it doesn't matter what we render to that channel */
3567 swizzle
.a
= ISL_CHANNEL_SELECT_ALPHA
;
3573 anv_pipeline_setup_l3_config(struct anv_pipeline
*pipeline
, bool needs_slm
);
3576 * Subsurface of an anv_image.
3578 struct anv_surface
{
3579 /** Valid only if isl_surf::size_B > 0. */
3580 struct isl_surf isl
;
3583 * Offset from VkImage's base address, as bound by vkBindImageMemory().
3589 struct vk_object_base base
;
3591 VkImageType type
; /**< VkImageCreateInfo::imageType */
3592 /* The original VkFormat provided by the client. This may not match any
3593 * of the actual surface formats.
3596 const struct anv_format
*format
;
3598 VkImageAspectFlags aspects
;
3601 uint32_t array_size
;
3602 uint32_t samples
; /**< VkImageCreateInfo::samples */
3604 VkImageUsageFlags usage
; /**< VkImageCreateInfo::usage. */
3605 VkImageUsageFlags stencil_usage
;
3606 VkImageCreateFlags create_flags
; /* Flags used when creating image. */
3607 VkImageTiling tiling
; /** VkImageCreateInfo::tiling */
3609 /** True if this is needs to be bound to an appropriately tiled BO.
3611 * When not using modifiers, consumers such as X11, Wayland, and KMS need
3612 * the tiling passed via I915_GEM_SET_TILING. When exporting these buffers
3613 * we require a dedicated allocation so that we can know to allocate a
3616 bool needs_set_tiling
;
3619 * Must be DRM_FORMAT_MOD_INVALID unless tiling is
3620 * VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT.
3622 uint64_t drm_format_mod
;
3627 /* Whether the image is made of several underlying buffer objects rather a
3628 * single one with different offsets.
3632 /* Image was created with external format. */
3633 bool external_format
;
3638 * For each foo, anv_image::planes[x].surface is valid if and only if
3639 * anv_image::aspects has a x aspect. Refer to anv_image_aspect_to_plane()
3640 * to figure the number associated with a given aspect.
3642 * The hardware requires that the depth buffer and stencil buffer be
3643 * separate surfaces. From Vulkan's perspective, though, depth and stencil
3644 * reside in the same VkImage. To satisfy both the hardware and Vulkan, we
3645 * allocate the depth and stencil buffers as separate surfaces in the same
3650 * -----------------------
3652 * ----------------------- |
3653 * | shadow surface0 | |
3654 * ----------------------- | Plane 0
3655 * | aux surface0 | |
3656 * ----------------------- |
3657 * | fast clear colors0 | \|/
3658 * -----------------------
3660 * ----------------------- |
3661 * | shadow surface1 | |
3662 * ----------------------- | Plane 1
3663 * | aux surface1 | |
3664 * ----------------------- |
3665 * | fast clear colors1 | \|/
3666 * -----------------------
3669 * -----------------------
3673 * Offset of the entire plane (whenever the image is disjoint this is
3681 struct anv_surface surface
;
3684 * A surface which shadows the main surface and may have different
3685 * tiling. This is used for sampling using a tiling that isn't supported
3686 * for other operations.
3688 struct anv_surface shadow_surface
;
3691 * The base aux usage for this image. For color images, this can be
3692 * either CCS_E or CCS_D depending on whether or not we can reliably
3693 * leave CCS on all the time.
3695 enum isl_aux_usage aux_usage
;
3697 struct anv_surface aux_surface
;
3700 * Offset of the fast clear state (used to compute the
3701 * fast_clear_state_offset of the following planes).
3703 uint32_t fast_clear_state_offset
;
3706 * BO associated with this plane, set when bound.
3708 struct anv_address address
;
3711 * When destroying the image, also free the bo.
3717 /* The ordering of this enum is important */
3718 enum anv_fast_clear_type
{
3719 /** Image does not have/support any fast-clear blocks */
3720 ANV_FAST_CLEAR_NONE
= 0,
3721 /** Image has/supports fast-clear but only to the default value */
3722 ANV_FAST_CLEAR_DEFAULT_VALUE
= 1,
3723 /** Image has/supports fast-clear with an arbitrary fast-clear value */
3724 ANV_FAST_CLEAR_ANY
= 2,
3727 /* Returns the number of auxiliary buffer levels attached to an image. */
3728 static inline uint8_t
3729 anv_image_aux_levels(const struct anv_image
* const image
,
3730 VkImageAspectFlagBits aspect
)
3732 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
3733 if (image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_NONE
)
3736 /* The Gen12 CCS aux surface is represented with only one level. */
3737 return image
->planes
[plane
].aux_surface
.isl
.tiling
== ISL_TILING_GEN12_CCS
?
3738 image
->planes
[plane
].surface
.isl
.levels
:
3739 image
->planes
[plane
].aux_surface
.isl
.levels
;
3742 /* Returns the number of auxiliary buffer layers attached to an image. */
3743 static inline uint32_t
3744 anv_image_aux_layers(const struct anv_image
* const image
,
3745 VkImageAspectFlagBits aspect
,
3746 const uint8_t miplevel
)
3750 /* The miplevel must exist in the main buffer. */
3751 assert(miplevel
< image
->levels
);
3753 if (miplevel
>= anv_image_aux_levels(image
, aspect
)) {
3754 /* There are no layers with auxiliary data because the miplevel has no
3759 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
3761 /* The Gen12 CCS aux surface is represented with only one layer. */
3762 const struct isl_extent4d
*aux_logical_level0_px
=
3763 image
->planes
[plane
].aux_surface
.isl
.tiling
== ISL_TILING_GEN12_CCS
?
3764 &image
->planes
[plane
].surface
.isl
.logical_level0_px
:
3765 &image
->planes
[plane
].aux_surface
.isl
.logical_level0_px
;
3767 return MAX2(aux_logical_level0_px
->array_len
,
3768 aux_logical_level0_px
->depth
>> miplevel
);
3772 static inline struct anv_address
3773 anv_image_get_clear_color_addr(UNUSED
const struct anv_device
*device
,
3774 const struct anv_image
*image
,
3775 VkImageAspectFlagBits aspect
)
3777 assert(image
->aspects
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
);
3779 uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
3780 return anv_address_add(image
->planes
[plane
].address
,
3781 image
->planes
[plane
].fast_clear_state_offset
);
3784 static inline struct anv_address
3785 anv_image_get_fast_clear_type_addr(const struct anv_device
*device
,
3786 const struct anv_image
*image
,
3787 VkImageAspectFlagBits aspect
)
3789 struct anv_address addr
=
3790 anv_image_get_clear_color_addr(device
, image
, aspect
);
3792 const unsigned clear_color_state_size
= device
->info
.gen
>= 10 ?
3793 device
->isl_dev
.ss
.clear_color_state_size
:
3794 device
->isl_dev
.ss
.clear_value_size
;
3795 return anv_address_add(addr
, clear_color_state_size
);
3798 static inline struct anv_address
3799 anv_image_get_compression_state_addr(const struct anv_device
*device
,
3800 const struct anv_image
*image
,
3801 VkImageAspectFlagBits aspect
,
3802 uint32_t level
, uint32_t array_layer
)
3804 assert(level
< anv_image_aux_levels(image
, aspect
));
3805 assert(array_layer
< anv_image_aux_layers(image
, aspect
, level
));
3806 UNUSED
uint32_t plane
= anv_image_aspect_to_plane(image
->aspects
, aspect
);
3807 assert(image
->planes
[plane
].aux_usage
== ISL_AUX_USAGE_CCS_E
);
3809 struct anv_address addr
=
3810 anv_image_get_fast_clear_type_addr(device
, image
, aspect
);
3811 addr
.offset
+= 4; /* Go past the fast clear type */
3813 if (image
->type
== VK_IMAGE_TYPE_3D
) {
3814 for (uint32_t l
= 0; l
< level
; l
++)
3815 addr
.offset
+= anv_minify(image
->extent
.depth
, l
) * 4;
3817 addr
.offset
+= level
* image
->array_size
* 4;
3819 addr
.offset
+= array_layer
* 4;
3821 assert(addr
.offset
<
3822 image
->planes
[plane
].address
.offset
+ image
->planes
[plane
].size
);
3826 /* Returns true if a HiZ-enabled depth buffer can be sampled from. */
3828 anv_can_sample_with_hiz(const struct gen_device_info
* const devinfo
,
3829 const struct anv_image
*image
)
3831 if (!(image
->aspects
& VK_IMAGE_ASPECT_DEPTH_BIT
))
3834 /* For Gen8-11, there are some restrictions around sampling from HiZ.
3835 * The Skylake PRM docs for RENDER_SURFACE_STATE::AuxiliarySurfaceMode
3838 * "If this field is set to AUX_HIZ, Number of Multisamples must
3839 * be MULTISAMPLECOUNT_1, and Surface Type cannot be SURFTYPE_3D."
3841 if (image
->type
== VK_IMAGE_TYPE_3D
)
3844 /* Allow this feature on BDW even though it is disabled in the BDW devinfo
3845 * struct. There's documentation which suggests that this feature actually
3846 * reduces performance on BDW, but it has only been observed to help so
3847 * far. Sampling fast-cleared blocks on BDW must also be handled with care
3848 * (see depth_stencil_attachment_compute_aux_usage() for more info).
3850 if (devinfo
->gen
!= 8 && !devinfo
->has_sample_with_hiz
)
3853 return image
->samples
== 1;
3857 anv_image_plane_uses_aux_map(const struct anv_device
*device
,
3858 const struct anv_image
*image
,
3861 return device
->info
.has_aux_map
&&
3862 isl_aux_usage_has_ccs(image
->planes
[plane
].aux_usage
);
3866 anv_cmd_buffer_mark_image_written(struct anv_cmd_buffer
*cmd_buffer
,
3867 const struct anv_image
*image
,
3868 VkImageAspectFlagBits aspect
,
3869 enum isl_aux_usage aux_usage
,
3871 uint32_t base_layer
,
3872 uint32_t layer_count
);
3875 anv_image_clear_color(struct anv_cmd_buffer
*cmd_buffer
,
3876 const struct anv_image
*image
,
3877 VkImageAspectFlagBits aspect
,
3878 enum isl_aux_usage aux_usage
,
3879 enum isl_format format
, struct isl_swizzle swizzle
,
3880 uint32_t level
, uint32_t base_layer
, uint32_t layer_count
,
3881 VkRect2D area
, union isl_color_value clear_color
);
3883 anv_image_clear_depth_stencil(struct anv_cmd_buffer
*cmd_buffer
,
3884 const struct anv_image
*image
,
3885 VkImageAspectFlags aspects
,
3886 enum isl_aux_usage depth_aux_usage
,
3888 uint32_t base_layer
, uint32_t layer_count
,
3890 float depth_value
, uint8_t stencil_value
);
3892 anv_image_msaa_resolve(struct anv_cmd_buffer
*cmd_buffer
,
3893 const struct anv_image
*src_image
,
3894 enum isl_aux_usage src_aux_usage
,
3895 uint32_t src_level
, uint32_t src_base_layer
,
3896 const struct anv_image
*dst_image
,
3897 enum isl_aux_usage dst_aux_usage
,
3898 uint32_t dst_level
, uint32_t dst_base_layer
,
3899 VkImageAspectFlagBits aspect
,
3900 uint32_t src_x
, uint32_t src_y
,
3901 uint32_t dst_x
, uint32_t dst_y
,
3902 uint32_t width
, uint32_t height
,
3903 uint32_t layer_count
,
3904 enum blorp_filter filter
);
3906 anv_image_hiz_op(struct anv_cmd_buffer
*cmd_buffer
,
3907 const struct anv_image
*image
,
3908 VkImageAspectFlagBits aspect
, uint32_t level
,
3909 uint32_t base_layer
, uint32_t layer_count
,
3910 enum isl_aux_op hiz_op
);
3912 anv_image_hiz_clear(struct anv_cmd_buffer
*cmd_buffer
,
3913 const struct anv_image
*image
,
3914 VkImageAspectFlags aspects
,
3916 uint32_t base_layer
, uint32_t layer_count
,
3917 VkRect2D area
, uint8_t stencil_value
);
3919 anv_image_mcs_op(struct anv_cmd_buffer
*cmd_buffer
,
3920 const struct anv_image
*image
,
3921 enum isl_format format
, struct isl_swizzle swizzle
,
3922 VkImageAspectFlagBits aspect
,
3923 uint32_t base_layer
, uint32_t layer_count
,
3924 enum isl_aux_op mcs_op
, union isl_color_value
*clear_value
,
3927 anv_image_ccs_op(struct anv_cmd_buffer
*cmd_buffer
,
3928 const struct anv_image
*image
,
3929 enum isl_format format
, struct isl_swizzle swizzle
,
3930 VkImageAspectFlagBits aspect
, uint32_t level
,
3931 uint32_t base_layer
, uint32_t layer_count
,
3932 enum isl_aux_op ccs_op
, union isl_color_value
*clear_value
,
3936 anv_image_copy_to_shadow(struct anv_cmd_buffer
*cmd_buffer
,
3937 const struct anv_image
*image
,
3938 VkImageAspectFlagBits aspect
,
3939 uint32_t base_level
, uint32_t level_count
,
3940 uint32_t base_layer
, uint32_t layer_count
);
3943 anv_layout_to_aux_state(const struct gen_device_info
* const devinfo
,
3944 const struct anv_image
*image
,
3945 const VkImageAspectFlagBits aspect
,
3946 const VkImageLayout layout
);
3949 anv_layout_to_aux_usage(const struct gen_device_info
* const devinfo
,
3950 const struct anv_image
*image
,
3951 const VkImageAspectFlagBits aspect
,
3952 const VkImageUsageFlagBits usage
,
3953 const VkImageLayout layout
);
3955 enum anv_fast_clear_type
3956 anv_layout_to_fast_clear_type(const struct gen_device_info
* const devinfo
,
3957 const struct anv_image
* const image
,
3958 const VkImageAspectFlagBits aspect
,
3959 const VkImageLayout layout
);
3961 /* This is defined as a macro so that it works for both
3962 * VkImageSubresourceRange and VkImageSubresourceLayers
3964 #define anv_get_layerCount(_image, _range) \
3965 ((_range)->layerCount == VK_REMAINING_ARRAY_LAYERS ? \
3966 (_image)->array_size - (_range)->baseArrayLayer : (_range)->layerCount)
3968 static inline uint32_t
3969 anv_get_levelCount(const struct anv_image
*image
,
3970 const VkImageSubresourceRange
*range
)
3972 return range
->levelCount
== VK_REMAINING_MIP_LEVELS
?
3973 image
->levels
- range
->baseMipLevel
: range
->levelCount
;
3976 static inline VkImageAspectFlags
3977 anv_image_expand_aspects(const struct anv_image
*image
,
3978 VkImageAspectFlags aspects
)
3980 /* If the underlying image has color plane aspects and
3981 * VK_IMAGE_ASPECT_COLOR_BIT has been requested, then return the aspects of
3982 * the underlying image. */
3983 if ((image
->aspects
& VK_IMAGE_ASPECT_PLANES_BITS_ANV
) != 0 &&
3984 aspects
== VK_IMAGE_ASPECT_COLOR_BIT
)
3985 return image
->aspects
;
3991 anv_image_aspects_compatible(VkImageAspectFlags aspects1
,
3992 VkImageAspectFlags aspects2
)
3994 if (aspects1
== aspects2
)
3997 /* Only 1 color aspects are compatibles. */
3998 if ((aspects1
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) != 0 &&
3999 (aspects2
& VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV
) != 0 &&
4000 util_bitcount(aspects1
) == util_bitcount(aspects2
))
4006 struct anv_image_view
{
4007 struct vk_object_base base
;
4009 const struct anv_image
*image
; /**< VkImageViewCreateInfo::image */
4011 VkImageAspectFlags aspect_mask
;
4013 VkExtent3D extent
; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
4017 uint32_t image_plane
;
4019 struct isl_view isl
;
4022 * RENDER_SURFACE_STATE when using image as a sampler surface with an
4023 * image layout of SHADER_READ_ONLY_OPTIMAL or
4024 * DEPTH_STENCIL_READ_ONLY_OPTIMAL.
4026 struct anv_surface_state optimal_sampler_surface_state
;
4029 * RENDER_SURFACE_STATE when using image as a sampler surface with an
4030 * image layout of GENERAL.
4032 struct anv_surface_state general_sampler_surface_state
;
4035 * RENDER_SURFACE_STATE when using image as a storage image. Separate
4036 * states for write-only and readable, using the real format for
4037 * write-only and the lowered format for readable.
4039 struct anv_surface_state storage_surface_state
;
4040 struct anv_surface_state writeonly_storage_surface_state
;
4042 struct brw_image_param storage_image_param
;
4046 enum anv_image_view_state_flags
{
4047 ANV_IMAGE_VIEW_STATE_STORAGE_WRITE_ONLY
= (1 << 0),
4048 ANV_IMAGE_VIEW_STATE_TEXTURE_OPTIMAL
= (1 << 1),
4051 void anv_image_fill_surface_state(struct anv_device
*device
,
4052 const struct anv_image
*image
,
4053 VkImageAspectFlagBits aspect
,
4054 const struct isl_view
*view
,
4055 isl_surf_usage_flags_t view_usage
,
4056 enum isl_aux_usage aux_usage
,
4057 const union isl_color_value
*clear_color
,
4058 enum anv_image_view_state_flags flags
,
4059 struct anv_surface_state
*state_inout
,
4060 struct brw_image_param
*image_param_out
);
4062 struct anv_image_create_info
{
4063 const VkImageCreateInfo
*vk_info
;
4065 /** An opt-in bitmask which filters an ISL-mapping of the Vulkan tiling. */
4066 isl_tiling_flags_t isl_tiling_flags
;
4068 /** These flags will be added to any derived from VkImageCreateInfo. */
4069 isl_surf_usage_flags_t isl_extra_usage_flags
;
4072 bool external_format
;
4075 VkResult
anv_image_create(VkDevice _device
,
4076 const struct anv_image_create_info
*info
,
4077 const VkAllocationCallbacks
* alloc
,
4081 anv_isl_format_for_descriptor_type(VkDescriptorType type
);
4083 static inline VkExtent3D
4084 anv_sanitize_image_extent(const VkImageType imageType
,
4085 const VkExtent3D imageExtent
)
4087 switch (imageType
) {
4088 case VK_IMAGE_TYPE_1D
:
4089 return (VkExtent3D
) { imageExtent
.width
, 1, 1 };
4090 case VK_IMAGE_TYPE_2D
:
4091 return (VkExtent3D
) { imageExtent
.width
, imageExtent
.height
, 1 };
4092 case VK_IMAGE_TYPE_3D
:
4095 unreachable("invalid image type");
4099 static inline VkOffset3D
4100 anv_sanitize_image_offset(const VkImageType imageType
,
4101 const VkOffset3D imageOffset
)
4103 switch (imageType
) {
4104 case VK_IMAGE_TYPE_1D
:
4105 return (VkOffset3D
) { imageOffset
.x
, 0, 0 };
4106 case VK_IMAGE_TYPE_2D
:
4107 return (VkOffset3D
) { imageOffset
.x
, imageOffset
.y
, 0 };
4108 case VK_IMAGE_TYPE_3D
:
4111 unreachable("invalid image type");
4115 VkFormatFeatureFlags
4116 anv_get_image_format_features(const struct gen_device_info
*devinfo
,
4118 const struct anv_format
*anv_format
,
4119 VkImageTiling vk_tiling
);
4121 void anv_fill_buffer_surface_state(struct anv_device
*device
,
4122 struct anv_state state
,
4123 enum isl_format format
,
4124 struct anv_address address
,
4125 uint32_t range
, uint32_t stride
);
4128 anv_clear_color_from_att_state(union isl_color_value
*clear_color
,
4129 const struct anv_attachment_state
*att_state
,
4130 const struct anv_image_view
*iview
)
4132 const struct isl_format_layout
*view_fmtl
=
4133 isl_format_get_layout(iview
->planes
[0].isl
.format
);
4135 #define COPY_CLEAR_COLOR_CHANNEL(c, i) \
4136 if (view_fmtl->channels.c.bits) \
4137 clear_color->u32[i] = att_state->clear_value.color.uint32[i]
4139 COPY_CLEAR_COLOR_CHANNEL(r
, 0);
4140 COPY_CLEAR_COLOR_CHANNEL(g
, 1);
4141 COPY_CLEAR_COLOR_CHANNEL(b
, 2);
4142 COPY_CLEAR_COLOR_CHANNEL(a
, 3);
4144 #undef COPY_CLEAR_COLOR_CHANNEL
4148 /* Haswell border color is a bit of a disaster. Float and unorm formats use a
4149 * straightforward 32-bit float color in the first 64 bytes. Instead of using
4150 * a nice float/integer union like Gen8+, Haswell specifies the integer border
4151 * color as a separate entry /after/ the float color. The layout of this entry
4152 * also depends on the format's bpp (with extra hacks for RG32), and overlaps.
4154 * Since we don't know the format/bpp, we can't make any of the border colors
4155 * containing '1' work for all formats, as it would be in the wrong place for
4156 * some of them. We opt to make 32-bit integers work as this seems like the
4157 * most common option. Fortunately, transparent black works regardless, as
4158 * all zeroes is the same in every bit-size.
4160 struct hsw_border_color
{
4164 uint32_t _pad1
[108];
4167 struct gen8_border_color
{
4172 /* Pad out to 64 bytes */
4176 struct anv_ycbcr_conversion
{
4177 struct vk_object_base base
;
4179 const struct anv_format
* format
;
4180 VkSamplerYcbcrModelConversion ycbcr_model
;
4181 VkSamplerYcbcrRange ycbcr_range
;
4182 VkComponentSwizzle mapping
[4];
4183 VkChromaLocation chroma_offsets
[2];
4184 VkFilter chroma_filter
;
4185 bool chroma_reconstruction
;
4188 struct anv_sampler
{
4189 struct vk_object_base base
;
4191 uint32_t state
[3][4];
4193 struct anv_ycbcr_conversion
*conversion
;
4195 /* Blob of sampler state data which is guaranteed to be 32-byte aligned
4196 * and with a 32-byte stride for use as bindless samplers.
4198 struct anv_state bindless_state
;
4200 struct anv_state custom_border_color
;
4203 struct anv_framebuffer
{
4204 struct vk_object_base base
;
4210 uint32_t attachment_count
;
4211 struct anv_image_view
* attachments
[0];
4214 struct anv_subpass_attachment
{
4215 VkImageUsageFlagBits usage
;
4216 uint32_t attachment
;
4217 VkImageLayout layout
;
4219 /* Used only with attachment containing stencil data. */
4220 VkImageLayout stencil_layout
;
4223 struct anv_subpass
{
4224 uint32_t attachment_count
;
4227 * A pointer to all attachment references used in this subpass.
4228 * Only valid if ::attachment_count > 0.
4230 struct anv_subpass_attachment
* attachments
;
4231 uint32_t input_count
;
4232 struct anv_subpass_attachment
* input_attachments
;
4233 uint32_t color_count
;
4234 struct anv_subpass_attachment
* color_attachments
;
4235 struct anv_subpass_attachment
* resolve_attachments
;
4237 struct anv_subpass_attachment
* depth_stencil_attachment
;
4238 struct anv_subpass_attachment
* ds_resolve_attachment
;
4239 VkResolveModeFlagBitsKHR depth_resolve_mode
;
4240 VkResolveModeFlagBitsKHR stencil_resolve_mode
;
4244 /** Subpass has a depth/stencil self-dependency */
4245 bool has_ds_self_dep
;
4247 /** Subpass has at least one color resolve attachment */
4248 bool has_color_resolve
;
4251 static inline unsigned
4252 anv_subpass_view_count(const struct anv_subpass
*subpass
)
4254 return MAX2(1, util_bitcount(subpass
->view_mask
));
4257 struct anv_render_pass_attachment
{
4258 /* TODO: Consider using VkAttachmentDescription instead of storing each of
4259 * its members individually.
4263 VkImageUsageFlags usage
;
4264 VkAttachmentLoadOp load_op
;
4265 VkAttachmentStoreOp store_op
;
4266 VkAttachmentLoadOp stencil_load_op
;
4267 VkImageLayout initial_layout
;
4268 VkImageLayout final_layout
;
4269 VkImageLayout first_subpass_layout
;
4271 VkImageLayout stencil_initial_layout
;
4272 VkImageLayout stencil_final_layout
;
4274 /* The subpass id in which the attachment will be used last. */
4275 uint32_t last_subpass_idx
;
4278 struct anv_render_pass
{
4279 struct vk_object_base base
;
4281 uint32_t attachment_count
;
4282 uint32_t subpass_count
;
4283 /* An array of subpass_count+1 flushes, one per subpass boundary */
4284 enum anv_pipe_bits
* subpass_flushes
;
4285 struct anv_render_pass_attachment
* attachments
;
4286 struct anv_subpass subpasses
[0];
4289 #define ANV_PIPELINE_STATISTICS_MASK 0x000007ff
4291 #define OA_SNAPSHOT_SIZE (256)
4292 #define ANV_KHR_PERF_QUERY_SIZE (ALIGN(sizeof(uint64_t), 64) + 2 * OA_SNAPSHOT_SIZE)
4294 struct anv_query_pool
{
4295 struct vk_object_base base
;
4298 VkQueryPipelineStatisticFlags pipeline_statistics
;
4299 /** Stride between slots, in bytes */
4301 /** Number of slots in this query pool */
4305 /* Perf queries : */
4306 struct anv_bo reset_bo
;
4307 uint32_t n_counters
;
4308 struct gen_perf_counter_pass
*counter_pass
;
4310 struct gen_perf_query_info
**pass_query
;
4313 static inline uint32_t khr_perf_query_preamble_offset(struct anv_query_pool
*pool
,
4316 return pass
* ANV_KHR_PERF_QUERY_SIZE
+ 8;
4319 int anv_get_instance_entrypoint_index(const char *name
);
4320 int anv_get_device_entrypoint_index(const char *name
);
4321 int anv_get_physical_device_entrypoint_index(const char *name
);
4323 const char *anv_get_instance_entry_name(int index
);
4324 const char *anv_get_physical_device_entry_name(int index
);
4325 const char *anv_get_device_entry_name(int index
);
4328 anv_instance_entrypoint_is_enabled(int index
, uint32_t core_version
,
4329 const struct anv_instance_extension_table
*instance
);
4331 anv_physical_device_entrypoint_is_enabled(int index
, uint32_t core_version
,
4332 const struct anv_instance_extension_table
*instance
);
4334 anv_device_entrypoint_is_enabled(int index
, uint32_t core_version
,
4335 const struct anv_instance_extension_table
*instance
,
4336 const struct anv_device_extension_table
*device
);
4338 void *anv_resolve_device_entrypoint(const struct gen_device_info
*devinfo
,
4340 void *anv_lookup_entrypoint(const struct gen_device_info
*devinfo
,
4343 void anv_dump_image_to_ppm(struct anv_device
*device
,
4344 struct anv_image
*image
, unsigned miplevel
,
4345 unsigned array_layer
, VkImageAspectFlagBits aspect
,
4346 const char *filename
);
4348 enum anv_dump_action
{
4349 ANV_DUMP_FRAMEBUFFERS_BIT
= 0x1,
4352 void anv_dump_start(struct anv_device
*device
, enum anv_dump_action actions
);
4353 void anv_dump_finish(void);
4355 void anv_dump_add_attachments(struct anv_cmd_buffer
*cmd_buffer
);
4357 static inline uint32_t
4358 anv_get_subpass_id(const struct anv_cmd_state
* const cmd_state
)
4360 /* This function must be called from within a subpass. */
4361 assert(cmd_state
->pass
&& cmd_state
->subpass
);
4363 const uint32_t subpass_id
= cmd_state
->subpass
- cmd_state
->pass
->subpasses
;
4365 /* The id of this subpass shouldn't exceed the number of subpasses in this
4366 * render pass minus 1.
4368 assert(subpass_id
< cmd_state
->pass
->subpass_count
);
4372 struct gen_perf_config
*anv_get_perf(const struct gen_device_info
*devinfo
, int fd
);
4373 void anv_device_perf_init(struct anv_device
*device
);
4374 void anv_perf_write_pass_results(struct gen_perf_config
*perf
,
4375 struct anv_query_pool
*pool
, uint32_t pass
,
4376 const struct gen_perf_query_result
*accumulated_results
,
4377 union VkPerformanceCounterResultKHR
*results
);
4379 #define ANV_FROM_HANDLE(__anv_type, __name, __handle) \
4380 VK_FROM_HANDLE(__anv_type, __name, __handle)
4382 VK_DEFINE_HANDLE_CASTS(anv_cmd_buffer
, base
, VkCommandBuffer
,
4383 VK_OBJECT_TYPE_COMMAND_BUFFER
)
4384 VK_DEFINE_HANDLE_CASTS(anv_device
, vk
.base
, VkDevice
, VK_OBJECT_TYPE_DEVICE
)
4385 VK_DEFINE_HANDLE_CASTS(anv_instance
, base
, VkInstance
, VK_OBJECT_TYPE_INSTANCE
)
4386 VK_DEFINE_HANDLE_CASTS(anv_physical_device
, base
, VkPhysicalDevice
,
4387 VK_OBJECT_TYPE_PHYSICAL_DEVICE
)
4388 VK_DEFINE_HANDLE_CASTS(anv_queue
, base
, VkQueue
, VK_OBJECT_TYPE_QUEUE
)
4390 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_cmd_pool
, base
, VkCommandPool
,
4391 VK_OBJECT_TYPE_COMMAND_POOL
)
4392 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_buffer
, base
, VkBuffer
,
4393 VK_OBJECT_TYPE_BUFFER
)
4394 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_buffer_view
, base
, VkBufferView
,
4395 VK_OBJECT_TYPE_BUFFER_VIEW
)
4396 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_descriptor_pool
, base
, VkDescriptorPool
,
4397 VK_OBJECT_TYPE_DESCRIPTOR_POOL
)
4398 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_descriptor_set
, base
, VkDescriptorSet
,
4399 VK_OBJECT_TYPE_DESCRIPTOR_SET
)
4400 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_descriptor_set_layout
, base
,
4401 VkDescriptorSetLayout
,
4402 VK_OBJECT_TYPE_DESCRIPTOR_SET_LAYOUT
)
4403 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_descriptor_update_template
, base
,
4404 VkDescriptorUpdateTemplate
,
4405 VK_OBJECT_TYPE_DESCRIPTOR_UPDATE_TEMPLATE
)
4406 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_device_memory
, base
, VkDeviceMemory
,
4407 VK_OBJECT_TYPE_DEVICE_MEMORY
)
4408 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_fence
, base
, VkFence
, VK_OBJECT_TYPE_FENCE
)
4409 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_event
, base
, VkEvent
, VK_OBJECT_TYPE_EVENT
)
4410 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_framebuffer
, base
, VkFramebuffer
,
4411 VK_OBJECT_TYPE_FRAMEBUFFER
)
4412 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_image
, base
, VkImage
, VK_OBJECT_TYPE_IMAGE
)
4413 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_image_view
, base
, VkImageView
,
4414 VK_OBJECT_TYPE_IMAGE_VIEW
);
4415 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_pipeline_cache
, base
, VkPipelineCache
,
4416 VK_OBJECT_TYPE_PIPELINE_CACHE
)
4417 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_pipeline
, base
, VkPipeline
,
4418 VK_OBJECT_TYPE_PIPELINE
)
4419 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_pipeline_layout
, base
, VkPipelineLayout
,
4420 VK_OBJECT_TYPE_PIPELINE_LAYOUT
)
4421 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_query_pool
, base
, VkQueryPool
,
4422 VK_OBJECT_TYPE_QUERY_POOL
)
4423 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_render_pass
, base
, VkRenderPass
,
4424 VK_OBJECT_TYPE_RENDER_PASS
)
4425 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_sampler
, base
, VkSampler
,
4426 VK_OBJECT_TYPE_SAMPLER
)
4427 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_semaphore
, base
, VkSemaphore
,
4428 VK_OBJECT_TYPE_SEMAPHORE
)
4429 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_shader_module
, base
, VkShaderModule
,
4430 VK_OBJECT_TYPE_SHADER_MODULE
)
4431 VK_DEFINE_NONDISP_HANDLE_CASTS(anv_ycbcr_conversion
, base
,
4432 VkSamplerYcbcrConversion
,
4433 VK_OBJECT_TYPE_SAMPLER_YCBCR_CONVERSION
)
4435 /* Gen-specific function declarations */
4437 # include "anv_genX.h"
4439 # define genX(x) gen7_##x
4440 # include "anv_genX.h"
4442 # define genX(x) gen75_##x
4443 # include "anv_genX.h"
4445 # define genX(x) gen8_##x
4446 # include "anv_genX.h"
4448 # define genX(x) gen9_##x
4449 # include "anv_genX.h"
4451 # define genX(x) gen10_##x
4452 # include "anv_genX.h"
4454 # define genX(x) gen11_##x
4455 # include "anv_genX.h"
4457 # define genX(x) gen12_##x
4458 # include "anv_genX.h"
4462 #endif /* ANV_PRIVATE_H */