projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv/meta: fix blit paths for depth/stencil (v2.1)
[mesa.git] / src / amd / vulkan / radv_private.h
1 /*
2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
4 *
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
17 * Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 * IN THE SOFTWARE.
26 */
27
28 #ifndef RADV_PRIVATE_H
29 #define RADV_PRIVATE_H
30
31 #include <stdlib.h>
32 #include <stdio.h>
33 #include <stdbool.h>
34 #include <pthread.h>
35 #include <assert.h>
36 #include <stdint.h>
37 #include <string.h>
38 #ifdef HAVE_VALGRIND
39 #include <valgrind.h>
40 #include <memcheck.h>
41 #define VG(x) x
42 #else
43 #define VG(x)
44 #endif
45
46 #include <amdgpu.h>
47 #include "compiler/shader_enums.h"
48 #include "util/macros.h"
49 #include "util/list.h"
50 #include "main/macros.h"
51 #include "vk_alloc.h"
52
53 #include "radv_radeon_winsys.h"
54 #include "ac_binary.h"
55 #include "ac_nir_to_llvm.h"
56 #include "ac_gpu_info.h"
57 #include "ac_surface.h"
58 #include "radv_descriptor_set.h"
59
60 #include <llvm-c/TargetMachine.h>
61
62 /* Pre-declarations needed for WSI entrypoints */
63 struct wl_surface;
64 struct wl_display;
65 typedef struct xcb_connection_t xcb_connection_t;
66 typedef uint32_t xcb_visualid_t;
67 typedef uint32_t xcb_window_t;
68
69 #include <vulkan/vulkan.h>
70 #include <vulkan/vulkan_intel.h>
71 #include <vulkan/vk_icd.h>
72
73 #include "radv_entrypoints.h"
74
75 #include "wsi_common.h"
76
77 #define ATI_VENDOR_ID 0x1002
78
79 #define MAX_VBS 32
80 #define MAX_VERTEX_ATTRIBS 32
81 #define MAX_RTS 8
82 #define MAX_VIEWPORTS 16
83 #define MAX_SCISSORS 16
84 #define MAX_PUSH_CONSTANTS_SIZE 128
85 #define MAX_PUSH_DESCRIPTORS 32
86 #define MAX_DYNAMIC_BUFFERS 16
87 #define MAX_SAMPLES_LOG2 4
88 #define NUM_META_FS_KEYS 13
89 #define RADV_MAX_DRM_DEVICES 8
90 #define MAX_VIEWS 8
91
92 #define NUM_DEPTH_CLEAR_PIPELINES 3
93
94 enum radv_mem_heap {
95 RADV_MEM_HEAP_VRAM,
96 RADV_MEM_HEAP_VRAM_CPU_ACCESS,
97 RADV_MEM_HEAP_GTT,
98 RADV_MEM_HEAP_COUNT
99 };
100
101 enum radv_mem_type {
102 RADV_MEM_TYPE_VRAM,
103 RADV_MEM_TYPE_GTT_WRITE_COMBINE,
104 RADV_MEM_TYPE_VRAM_CPU_ACCESS,
105 RADV_MEM_TYPE_GTT_CACHED,
106 RADV_MEM_TYPE_COUNT
107 };
108
109 #define radv_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
110
111 static inline uint32_t
112 align_u32(uint32_t v, uint32_t a)
113 {
114 assert(a != 0 && a == (a & -a));
115 return (v + a - 1) & ~(a - 1);
116 }
117
118 static inline uint32_t
119 align_u32_npot(uint32_t v, uint32_t a)
120 {
121 return (v + a - 1) / a * a;
122 }
123
124 static inline uint64_t
125 align_u64(uint64_t v, uint64_t a)
126 {
127 assert(a != 0 && a == (a & -a));
128 return (v + a - 1) & ~(a - 1);
129 }
130
131 static inline int32_t
132 align_i32(int32_t v, int32_t a)
133 {
134 assert(a != 0 && a == (a & -a));
135 return (v + a - 1) & ~(a - 1);
136 }
137
138 /** Alignment must be a power of 2. */
139 static inline bool
140 radv_is_aligned(uintmax_t n, uintmax_t a)
141 {
142 assert(a == (a & -a));
143 return (n & (a - 1)) == 0;
144 }
145
146 static inline uint32_t
147 round_up_u32(uint32_t v, uint32_t a)
148 {
149 return (v + a - 1) / a;
150 }
151
152 static inline uint64_t
153 round_up_u64(uint64_t v, uint64_t a)
154 {
155 return (v + a - 1) / a;
156 }
157
158 static inline uint32_t
159 radv_minify(uint32_t n, uint32_t levels)
160 {
161 if (unlikely(n == 0))
162 return 0;
163 else
164 return MAX2(n >> levels, 1);
165 }
166 static inline float
167 radv_clamp_f(float f, float min, float max)
168 {
169 assert(min < max);
170
171 if (f > max)
172 return max;
173 else if (f < min)
174 return min;
175 else
176 return f;
177 }
178
179 static inline bool
180 radv_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
181 {
182 if (*inout_mask & clear_mask) {
183 *inout_mask &= ~clear_mask;
184 return true;
185 } else {
186 return false;
187 }
188 }
189
190 #define for_each_bit(b, dword) \
191 for (uint32_t __dword = (dword); \
192 (b) = __builtin_ffs(__dword) - 1, __dword; \
193 __dword &= ~(1 << (b)))
194
195 #define typed_memcpy(dest, src, count) ({ \
196 STATIC_ASSERT(sizeof(*src) == sizeof(*dest)); \
197 memcpy((dest), (src), (count) * sizeof(*(src))); \
198 })
199
200 /* Whenever we generate an error, pass it through this function. Useful for
201 * debugging, where we can break on it. Only call at error site, not when
202 * propagating errors. Might be useful to plug in a stack trace here.
203 */
204
205 VkResult __vk_errorf(VkResult error, const char *file, int line, const char *format, ...);
206
207 #ifdef DEBUG
208 #define vk_error(error) __vk_errorf(error, __FILE__, __LINE__, NULL);
209 #define vk_errorf(error, format, ...) __vk_errorf(error, __FILE__, __LINE__, format, ## __VA_ARGS__);
210 #else
211 #define vk_error(error) error
212 #define vk_errorf(error, format, ...) error
213 #endif
214
215 void __radv_finishme(const char *file, int line, const char *format, ...)
216 radv_printflike(3, 4);
217 void radv_loge(const char *format, ...) radv_printflike(1, 2);
218 void radv_loge_v(const char *format, va_list va);
219
220 /**
221 * Print a FINISHME message, including its source location.
222 */
223 #define radv_finishme(format, ...) \
224 do { \
225 static bool reported = false; \
226 if (!reported) { \
227 __radv_finishme(__FILE__, __LINE__, format, ##__VA_ARGS__); \
228 reported = true; \
229 } \
230 } while (0)
231
232 /* A non-fatal assert. Useful for debugging. */
233 #ifdef DEBUG
234 #define radv_assert(x) ({ \
235 if (unlikely(!(x))) \
236 fprintf(stderr, "%s:%d ASSERT: %s\n", __FILE__, __LINE__, #x); \
237 })
238 #else
239 #define radv_assert(x)
240 #endif
241
242 #define stub_return(v) \
243 do { \
244 radv_finishme("stub %s", __func__); \
245 return (v); \
246 } while (0)
247
248 #define stub() \
249 do { \
250 radv_finishme("stub %s", __func__); \
251 return; \
252 } while (0)
253
254 void *radv_lookup_entrypoint(const char *name);
255
256 struct radv_physical_device {
257 VK_LOADER_DATA _loader_data;
258
259 struct radv_instance * instance;
260
261 struct radeon_winsys *ws;
262 struct radeon_info rad_info;
263 char path[20];
264 char name[VK_MAX_PHYSICAL_DEVICE_NAME_SIZE];
265 uint8_t driver_uuid[VK_UUID_SIZE];
266 uint8_t device_uuid[VK_UUID_SIZE];
267 uint8_t cache_uuid[VK_UUID_SIZE];
268
269 int local_fd;
270 struct wsi_device wsi_device;
271
272 bool has_rbplus; /* if RB+ register exist */
273 bool rbplus_allowed; /* if RB+ is allowed */
274 bool has_clear_state;
275
276 /* This is the drivers on-disk cache used as a fallback as opposed to
277 * the pipeline cache defined by apps.
278 */
279 struct disk_cache * disk_cache;
280
281 VkPhysicalDeviceMemoryProperties memory_properties;
282 enum radv_mem_type mem_type_indices[RADV_MEM_TYPE_COUNT];
283 };
284
285 struct radv_instance {
286 VK_LOADER_DATA _loader_data;
287
288 VkAllocationCallbacks alloc;
289
290 uint32_t apiVersion;
291 int physicalDeviceCount;
292 struct radv_physical_device physicalDevices[RADV_MAX_DRM_DEVICES];
293
294 uint64_t debug_flags;
295 uint64_t perftest_flags;
296 };
297
298 VkResult radv_init_wsi(struct radv_physical_device *physical_device);
299 void radv_finish_wsi(struct radv_physical_device *physical_device);
300
301 bool radv_instance_extension_supported(const char *name);
302 uint32_t radv_physical_device_api_version(struct radv_physical_device *dev);
303 bool radv_physical_device_extension_supported(struct radv_physical_device *dev,
304 const char *name);
305
306 struct cache_entry;
307
308 struct radv_pipeline_cache {
309 struct radv_device * device;
310 pthread_mutex_t mutex;
311
312 uint32_t total_size;
313 uint32_t table_size;
314 uint32_t kernel_count;
315 struct cache_entry ** hash_table;
316 bool modified;
317
318 VkAllocationCallbacks alloc;
319 };
320
321 struct radv_pipeline_key {
322 uint32_t instance_rate_inputs;
323 unsigned tess_input_vertices;
324 uint32_t col_format;
325 uint32_t is_int8;
326 uint32_t is_int10;
327 uint32_t multisample : 1;
328 uint32_t has_multiview_view_index : 1;
329 };
330
331 void
332 radv_pipeline_cache_init(struct radv_pipeline_cache *cache,
333 struct radv_device *device);
334 void
335 radv_pipeline_cache_finish(struct radv_pipeline_cache *cache);
336 void
337 radv_pipeline_cache_load(struct radv_pipeline_cache *cache,
338 const void *data, size_t size);
339
340 struct radv_shader_variant;
341
342 bool
343 radv_create_shader_variants_from_pipeline_cache(struct radv_device *device,
344 struct radv_pipeline_cache *cache,
345 const unsigned char *sha1,
346 struct radv_shader_variant **variants);
347
348 void
349 radv_pipeline_cache_insert_shaders(struct radv_device *device,
350 struct radv_pipeline_cache *cache,
351 const unsigned char *sha1,
352 struct radv_shader_variant **variants,
353 const void *const *codes,
354 const unsigned *code_sizes);
355
356 enum radv_blit_ds_layout {
357 RADV_BLIT_DS_LAYOUT_TILE_ENABLE,
358 RADV_BLIT_DS_LAYOUT_TILE_DISABLE,
359 RADV_BLIT_DS_LAYOUT_COUNT,
360 };
361
362 static inline enum radv_blit_ds_layout radv_meta_blit_ds_to_type(VkImageLayout layout)
363 {
364 return (layout == VK_IMAGE_LAYOUT_GENERAL) ? RADV_BLIT_DS_LAYOUT_TILE_DISABLE : RADV_BLIT_DS_LAYOUT_TILE_ENABLE;
365 }
366
367 static inline VkImageLayout radv_meta_blit_ds_to_layout(enum radv_blit_ds_layout ds_layout)
368 {
369 return ds_layout == RADV_BLIT_DS_LAYOUT_TILE_ENABLE ? VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL : VK_IMAGE_LAYOUT_GENERAL;
370 }
371
372 struct radv_meta_state {
373 VkAllocationCallbacks alloc;
374
375 struct radv_pipeline_cache cache;
376
377 /**
378 * Use array element `i` for images with `2^i` samples.
379 */
380 struct {
381 VkRenderPass render_pass[NUM_META_FS_KEYS];
382 VkPipeline color_pipelines[NUM_META_FS_KEYS];
383
384 VkRenderPass depthstencil_rp;
385 VkPipeline depth_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
386 VkPipeline stencil_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
387 VkPipeline depthstencil_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
388 } clear[1 + MAX_SAMPLES_LOG2];
389
390 VkPipelineLayout clear_color_p_layout;
391 VkPipelineLayout clear_depth_p_layout;
392 struct {
393 VkRenderPass render_pass[NUM_META_FS_KEYS];
394
395 /** Pipeline that blits from a 1D image. */
396 VkPipeline pipeline_1d_src[NUM_META_FS_KEYS];
397
398 /** Pipeline that blits from a 2D image. */
399 VkPipeline pipeline_2d_src[NUM_META_FS_KEYS];
400
401 /** Pipeline that blits from a 3D image. */
402 VkPipeline pipeline_3d_src[NUM_META_FS_KEYS];
403
404 VkRenderPass depth_only_rp[RADV_BLIT_DS_LAYOUT_COUNT];
405 VkPipeline depth_only_1d_pipeline;
406 VkPipeline depth_only_2d_pipeline;
407 VkPipeline depth_only_3d_pipeline;
408
409 VkRenderPass stencil_only_rp[RADV_BLIT_DS_LAYOUT_COUNT];
410 VkPipeline stencil_only_1d_pipeline;
411 VkPipeline stencil_only_2d_pipeline;
412 VkPipeline stencil_only_3d_pipeline;
413 VkPipelineLayout pipeline_layout;
414 VkDescriptorSetLayout ds_layout;
415 } blit;
416
417 struct {
418 VkRenderPass render_passes[NUM_META_FS_KEYS];
419
420 VkPipelineLayout p_layouts[3];
421 VkDescriptorSetLayout ds_layouts[3];
422 VkPipeline pipelines[3][NUM_META_FS_KEYS];
423
424 VkRenderPass depth_only_rp[RADV_BLIT_DS_LAYOUT_COUNT];
425 VkPipeline depth_only_pipeline[3];
426
427 VkRenderPass stencil_only_rp[RADV_BLIT_DS_LAYOUT_COUNT];
428 VkPipeline stencil_only_pipeline[3];
429 } blit2d;
430
431 struct {
432 VkPipelineLayout img_p_layout;
433 VkDescriptorSetLayout img_ds_layout;
434 VkPipeline pipeline;
435 VkPipeline pipeline_3d;
436 } itob;
437 struct {
438 VkPipelineLayout img_p_layout;
439 VkDescriptorSetLayout img_ds_layout;
440 VkPipeline pipeline;
441 } btoi;
442 struct {
443 VkPipelineLayout img_p_layout;
444 VkDescriptorSetLayout img_ds_layout;
445 VkPipeline pipeline;
446 } itoi;
447 struct {
448 VkPipelineLayout img_p_layout;
449 VkDescriptorSetLayout img_ds_layout;
450 VkPipeline pipeline;
451 } cleari;
452
453 struct {
454 VkPipelineLayout p_layout;
455 VkPipeline pipeline;
456 VkRenderPass pass;
457 } resolve;
458
459 struct {
460 VkDescriptorSetLayout ds_layout;
461 VkPipelineLayout p_layout;
462 struct {
463 VkPipeline pipeline;
464 VkPipeline i_pipeline;
465 VkPipeline srgb_pipeline;
466 } rc[MAX_SAMPLES_LOG2];
467 } resolve_compute;
468
469 struct {
470 VkDescriptorSetLayout ds_layout;
471 VkPipelineLayout p_layout;
472
473 struct {
474 VkRenderPass render_pass[NUM_META_FS_KEYS];
475 VkPipeline pipeline[NUM_META_FS_KEYS];
476 } rc[MAX_SAMPLES_LOG2];
477 } resolve_fragment;
478
479 struct {
480 VkPipelineLayout p_layout;
481 VkPipeline decompress_pipeline;
482 VkPipeline resummarize_pipeline;
483 VkRenderPass pass;
484 } depth_decomp[1 + MAX_SAMPLES_LOG2];
485
486 struct {
487 VkPipelineLayout p_layout;
488 VkPipeline cmask_eliminate_pipeline;
489 VkPipeline fmask_decompress_pipeline;
490 VkRenderPass pass;
491 } fast_clear_flush;
492
493 struct {
494 VkPipelineLayout fill_p_layout;
495 VkPipelineLayout copy_p_layout;
496 VkDescriptorSetLayout fill_ds_layout;
497 VkDescriptorSetLayout copy_ds_layout;
498 VkPipeline fill_pipeline;
499 VkPipeline copy_pipeline;
500 } buffer;
501
502 struct {
503 VkDescriptorSetLayout ds_layout;
504 VkPipelineLayout p_layout;
505 VkPipeline occlusion_query_pipeline;
506 VkPipeline pipeline_statistics_query_pipeline;
507 } query;
508 };
509
510 /* queue types */
511 #define RADV_QUEUE_GENERAL 0
512 #define RADV_QUEUE_COMPUTE 1
513 #define RADV_QUEUE_TRANSFER 2
514
515 #define RADV_MAX_QUEUE_FAMILIES 3
516
517 enum ring_type radv_queue_family_to_ring(int f);
518
519 struct radv_queue {
520 VK_LOADER_DATA _loader_data;
521 struct radv_device * device;
522 struct radeon_winsys_ctx *hw_ctx;
523 enum radeon_ctx_priority priority;
524 uint32_t queue_family_index;
525 int queue_idx;
526
527 uint32_t scratch_size;
528 uint32_t compute_scratch_size;
529 uint32_t esgs_ring_size;
530 uint32_t gsvs_ring_size;
531 bool has_tess_rings;
532 bool has_sample_positions;
533
534 struct radeon_winsys_bo *scratch_bo;
535 struct radeon_winsys_bo *descriptor_bo;
536 struct radeon_winsys_bo *compute_scratch_bo;
537 struct radeon_winsys_bo *esgs_ring_bo;
538 struct radeon_winsys_bo *gsvs_ring_bo;
539 struct radeon_winsys_bo *tess_factor_ring_bo;
540 struct radeon_winsys_bo *tess_offchip_ring_bo;
541 struct radeon_winsys_cs *initial_preamble_cs;
542 struct radeon_winsys_cs *initial_full_flush_preamble_cs;
543 struct radeon_winsys_cs *continue_preamble_cs;
544 };
545
546 struct radv_device {
547 VK_LOADER_DATA _loader_data;
548
549 VkAllocationCallbacks alloc;
550
551 struct radv_instance * instance;
552 struct radeon_winsys *ws;
553
554 struct radv_meta_state meta_state;
555
556 struct radv_queue *queues[RADV_MAX_QUEUE_FAMILIES];
557 int queue_count[RADV_MAX_QUEUE_FAMILIES];
558 struct radeon_winsys_cs *empty_cs[RADV_MAX_QUEUE_FAMILIES];
559
560 bool llvm_supports_spill;
561 bool has_distributed_tess;
562 bool dfsm_allowed;
563 uint32_t tess_offchip_block_dw_size;
564 uint32_t scratch_waves;
565 uint32_t dispatch_initiator;
566
567 uint32_t gs_table_depth;
568
569 /* MSAA sample locations.
570 * The first index is the sample index.
571 * The second index is the coordinate: X, Y. */
572 float sample_locations_1x[1][2];
573 float sample_locations_2x[2][2];
574 float sample_locations_4x[4][2];
575 float sample_locations_8x[8][2];
576 float sample_locations_16x[16][2];
577
578 /* CIK and later */
579 uint32_t gfx_init_size_dw;
580 struct radeon_winsys_bo *gfx_init;
581
582 struct radeon_winsys_bo *trace_bo;
583 uint32_t *trace_id_ptr;
584
585 /* Whether to keep shader debug info, for tracing or VK_AMD_shader_info */
586 bool keep_shader_info;
587
588 struct radv_physical_device *physical_device;
589
590 /* Backup in-memory cache to be used if the app doesn't provide one */
591 struct radv_pipeline_cache * mem_cache;
592
593 /*
594 * use different counters so MSAA MRTs get consecutive surface indices,
595 * even if MASK is allocated in between.
596 */
597 uint32_t image_mrt_offset_counter;
598 uint32_t fmask_mrt_offset_counter;
599 struct list_head shader_slabs;
600 mtx_t shader_slab_mutex;
601
602 /* For detecting VM faults reported by dmesg. */
603 uint64_t dmesg_timestamp;
604 };
605
606 struct radv_device_memory {
607 struct radeon_winsys_bo *bo;
608 /* for dedicated allocations */
609 struct radv_image *image;
610 struct radv_buffer *buffer;
611 uint32_t type_index;
612 VkDeviceSize map_size;
613 void * map;
614 };
615
616
617 struct radv_descriptor_range {
618 uint64_t va;
619 uint32_t size;
620 };
621
622 struct radv_descriptor_set {
623 const struct radv_descriptor_set_layout *layout;
624 uint32_t size;
625
626 struct radeon_winsys_bo *bo;
627 uint64_t va;
628 uint32_t *mapped_ptr;
629 struct radv_descriptor_range *dynamic_descriptors;
630
631 struct radeon_winsys_bo *descriptors[0];
632 };
633
634 struct radv_push_descriptor_set
635 {
636 struct radv_descriptor_set set;
637 uint32_t capacity;
638 };
639
640 struct radv_descriptor_pool_entry {
641 uint32_t offset;
642 uint32_t size;
643 struct radv_descriptor_set *set;
644 };
645
646 struct radv_descriptor_pool {
647 struct radeon_winsys_bo *bo;
648 uint8_t *mapped_ptr;
649 uint64_t current_offset;
650 uint64_t size;
651
652 uint8_t *host_memory_base;
653 uint8_t *host_memory_ptr;
654 uint8_t *host_memory_end;
655
656 uint32_t entry_count;
657 uint32_t max_entry_count;
658 struct radv_descriptor_pool_entry entries[0];
659 };
660
661 struct radv_descriptor_update_template_entry {
662 VkDescriptorType descriptor_type;
663
664 /* The number of descriptors to update */
665 uint32_t descriptor_count;
666
667 /* Into mapped_ptr or dynamic_descriptors, in units of the respective array */
668 uint32_t dst_offset;
669
670 /* In dwords. Not valid/used for dynamic descriptors */
671 uint32_t dst_stride;
672
673 uint32_t buffer_offset;
674
675 /* Only valid for combined image samplers and samplers */
676 uint16_t has_sampler;
677
678 /* In bytes */
679 size_t src_offset;
680 size_t src_stride;
681
682 /* For push descriptors */
683 const uint32_t *immutable_samplers;
684 };
685
686 struct radv_descriptor_update_template {
687 uint32_t entry_count;
688 struct radv_descriptor_update_template_entry entry[0];
689 };
690
691 struct radv_buffer {
692 struct radv_device * device;
693 VkDeviceSize size;
694
695 VkBufferUsageFlags usage;
696 VkBufferCreateFlags flags;
697
698 /* Set when bound */
699 struct radeon_winsys_bo * bo;
700 VkDeviceSize offset;
701
702 bool shareable;
703 };
704
705
706 enum radv_cmd_dirty_bits {
707 RADV_CMD_DIRTY_DYNAMIC_VIEWPORT = 1 << 0, /* VK_DYNAMIC_STATE_VIEWPORT */
708 RADV_CMD_DIRTY_DYNAMIC_SCISSOR = 1 << 1, /* VK_DYNAMIC_STATE_SCISSOR */
709 RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH = 1 << 2, /* VK_DYNAMIC_STATE_LINE_WIDTH */
710 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS = 1 << 3, /* VK_DYNAMIC_STATE_DEPTH_BIAS */
711 RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS = 1 << 4, /* VK_DYNAMIC_STATE_BLEND_CONSTANTS */
712 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS = 1 << 5, /* VK_DYNAMIC_STATE_DEPTH_BOUNDS */
713 RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6, /* VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK */
714 RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7, /* VK_DYNAMIC_STATE_STENCIL_WRITE_MASK */
715 RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE = 1 << 8, /* VK_DYNAMIC_STATE_STENCIL_REFERENCE */
716 RADV_CMD_DIRTY_DYNAMIC_ALL = (1 << 9) - 1,
717 RADV_CMD_DIRTY_PIPELINE = 1 << 9,
718 RADV_CMD_DIRTY_INDEX_BUFFER = 1 << 10,
719 RADV_CMD_DIRTY_FRAMEBUFFER = 1 << 11,
720 RADV_CMD_DIRTY_VERTEX_BUFFER = 1 << 12,
721 };
722
723 enum radv_cmd_flush_bits {
724 RADV_CMD_FLAG_INV_ICACHE = 1 << 0,
725 /* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
726 RADV_CMD_FLAG_INV_SMEM_L1 = 1 << 1,
727 /* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
728 RADV_CMD_FLAG_INV_VMEM_L1 = 1 << 2,
729 /* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
730 RADV_CMD_FLAG_INV_GLOBAL_L2 = 1 << 3,
731 /* Same as above, but only writes back and doesn't invalidate */
732 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2 = 1 << 4,
733 /* Framebuffer caches */
734 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META = 1 << 5,
735 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META = 1 << 6,
736 RADV_CMD_FLAG_FLUSH_AND_INV_DB = 1 << 7,
737 RADV_CMD_FLAG_FLUSH_AND_INV_CB = 1 << 8,
738 /* Engine synchronization. */
739 RADV_CMD_FLAG_VS_PARTIAL_FLUSH = 1 << 9,
740 RADV_CMD_FLAG_PS_PARTIAL_FLUSH = 1 << 10,
741 RADV_CMD_FLAG_CS_PARTIAL_FLUSH = 1 << 11,
742 RADV_CMD_FLAG_VGT_FLUSH = 1 << 12,
743
744 RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER = (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
745 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
746 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
747 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META)
748 };
749
750 struct radv_vertex_binding {
751 struct radv_buffer * buffer;
752 VkDeviceSize offset;
753 };
754
755 struct radv_viewport_state {
756 uint32_t count;
757 VkViewport viewports[MAX_VIEWPORTS];
758 };
759
760 struct radv_scissor_state {
761 uint32_t count;
762 VkRect2D scissors[MAX_SCISSORS];
763 };
764
765 struct radv_dynamic_state {
766 /**
767 * Bitmask of (1 << VK_DYNAMIC_STATE_*).
768 * Defines the set of saved dynamic state.
769 */
770 uint32_t mask;
771
772 struct radv_viewport_state viewport;
773
774 struct radv_scissor_state scissor;
775
776 float line_width;
777
778 struct {
779 float bias;
780 float clamp;
781 float slope;
782 } depth_bias;
783
784 float blend_constants[4];
785
786 struct {
787 float min;
788 float max;
789 } depth_bounds;
790
791 struct {
792 uint32_t front;
793 uint32_t back;
794 } stencil_compare_mask;
795
796 struct {
797 uint32_t front;
798 uint32_t back;
799 } stencil_write_mask;
800
801 struct {
802 uint32_t front;
803 uint32_t back;
804 } stencil_reference;
805 };
806
807 extern const struct radv_dynamic_state default_dynamic_state;
808
809 const char *
810 radv_get_debug_option_name(int id);
811
812 const char *
813 radv_get_perftest_option_name(int id);
814
815 /**
816 * Attachment state when recording a renderpass instance.
817 *
818 * The clear value is valid only if there exists a pending clear.
819 */
820 struct radv_attachment_state {
821 VkImageAspectFlags pending_clear_aspects;
822 uint32_t cleared_views;
823 VkClearValue clear_value;
824 VkImageLayout current_layout;
825 };
826
827 struct radv_cmd_state {
828 /* Vertex descriptors */
829 bool vb_prefetch_dirty;
830 uint64_t vb_va;
831 unsigned vb_size;
832
833 bool push_descriptors_dirty;
834 bool predicating;
835 uint32_t dirty;
836
837 struct radv_pipeline * pipeline;
838 struct radv_pipeline * emitted_pipeline;
839 struct radv_pipeline * compute_pipeline;
840 struct radv_pipeline * emitted_compute_pipeline;
841 struct radv_framebuffer * framebuffer;
842 struct radv_render_pass * pass;
843 const struct radv_subpass * subpass;
844 struct radv_dynamic_state dynamic;
845 struct radv_attachment_state * attachments;
846 VkRect2D render_area;
847
848 /* Index buffer */
849 struct radv_buffer *index_buffer;
850 uint64_t index_offset;
851 uint32_t index_type;
852 uint32_t max_index_count;
853 uint64_t index_va;
854 int32_t last_index_type;
855
856 int32_t last_primitive_reset_en;
857 uint32_t last_primitive_reset_index;
858 enum radv_cmd_flush_bits flush_bits;
859 unsigned active_occlusion_queries;
860 float offset_scale;
861 uint32_t descriptors_dirty;
862 uint32_t valid_descriptors;
863 uint32_t trace_id;
864 uint32_t last_ia_multi_vgt_param;
865 };
866
867 struct radv_cmd_pool {
868 VkAllocationCallbacks alloc;
869 struct list_head cmd_buffers;
870 struct list_head free_cmd_buffers;
871 uint32_t queue_family_index;
872 };
873
874 struct radv_cmd_buffer_upload {
875 uint8_t *map;
876 unsigned offset;
877 uint64_t size;
878 struct radeon_winsys_bo *upload_bo;
879 struct list_head list;
880 };
881
882 enum radv_cmd_buffer_status {
883 RADV_CMD_BUFFER_STATUS_INVALID,
884 RADV_CMD_BUFFER_STATUS_INITIAL,
885 RADV_CMD_BUFFER_STATUS_RECORDING,
886 RADV_CMD_BUFFER_STATUS_EXECUTABLE,
887 RADV_CMD_BUFFER_STATUS_PENDING,
888 };
889
890 struct radv_cmd_buffer {
891 VK_LOADER_DATA _loader_data;
892
893 struct radv_device * device;
894
895 struct radv_cmd_pool * pool;
896 struct list_head pool_link;
897
898 VkCommandBufferUsageFlags usage_flags;
899 VkCommandBufferLevel level;
900 enum radv_cmd_buffer_status status;
901 struct radeon_winsys_cs *cs;
902 struct radv_cmd_state state;
903 struct radv_vertex_binding vertex_bindings[MAX_VBS];
904 uint32_t queue_family_index;
905
906 uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
907 uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
908 VkShaderStageFlags push_constant_stages;
909 struct radv_push_descriptor_set push_descriptors;
910 struct radv_descriptor_set meta_push_descriptors;
911 struct radv_descriptor_set *descriptors[MAX_SETS];
912
913 struct radv_cmd_buffer_upload upload;
914
915 uint32_t scratch_size_needed;
916 uint32_t compute_scratch_size_needed;
917 uint32_t esgs_ring_size_needed;
918 uint32_t gsvs_ring_size_needed;
919 bool tess_rings_needed;
920 bool sample_positions_needed;
921
922 VkResult record_result;
923
924 int ring_offsets_idx; /* just used for verification */
925 uint32_t gfx9_fence_offset;
926 struct radeon_winsys_bo *gfx9_fence_bo;
927 uint32_t gfx9_fence_idx;
928 };
929
930 struct radv_image;
931
932 bool radv_cmd_buffer_uses_mec(struct radv_cmd_buffer *cmd_buffer);
933
934 void si_init_compute(struct radv_cmd_buffer *cmd_buffer);
935 void si_init_config(struct radv_cmd_buffer *cmd_buffer);
936
937 void cik_create_gfx_config(struct radv_device *device);
938
939 void si_write_viewport(struct radeon_winsys_cs *cs, int first_vp,
940 int count, const VkViewport *viewports);
941 void si_write_scissors(struct radeon_winsys_cs *cs, int first,
942 int count, const VkRect2D *scissors,
943 const VkViewport *viewports, bool can_use_guardband);
944 uint32_t si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer,
945 bool instanced_draw, bool indirect_draw,
946 uint32_t draw_vertex_count);
947 void si_cs_emit_write_event_eop(struct radeon_winsys_cs *cs,
948 bool predicated,
949 enum chip_class chip_class,
950 bool is_mec,
951 unsigned event, unsigned event_flags,
952 unsigned data_sel,
953 uint64_t va,
954 uint32_t old_fence,
955 uint32_t new_fence);
956
957 void si_emit_wait_fence(struct radeon_winsys_cs *cs,
958 bool predicated,
959 uint64_t va, uint32_t ref,
960 uint32_t mask);
961 void si_cs_emit_cache_flush(struct radeon_winsys_cs *cs,
962 bool predicated,
963 enum chip_class chip_class,
964 uint32_t *fence_ptr, uint64_t va,
965 bool is_mec,
966 enum radv_cmd_flush_bits flush_bits);
967 void si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer);
968 void si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer, uint64_t va);
969 void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
970 uint64_t src_va, uint64_t dest_va,
971 uint64_t size);
972 void si_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
973 unsigned size);
974 void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
975 uint64_t size, unsigned value);
976 void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer);
977 bool
978 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
979 unsigned size,
980 unsigned alignment,
981 unsigned *out_offset,
982 void **ptr);
983 void
984 radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
985 const struct radv_subpass *subpass,
986 bool transitions);
987 bool
988 radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
989 unsigned size, unsigned alignmnet,
990 const void *data, unsigned *out_offset);
991
992 void radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer);
993 void radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer);
994 void radv_cmd_buffer_resolve_subpass_cs(struct radv_cmd_buffer *cmd_buffer);
995 void radv_cmd_buffer_resolve_subpass_fs(struct radv_cmd_buffer *cmd_buffer);
996 void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs *cs, int nr_samples);
997 unsigned radv_cayman_get_maxdist(int log_samples);
998 void radv_device_init_msaa(struct radv_device *device);
999 void radv_set_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1000 struct radv_image *image,
1001 VkClearDepthStencilValue ds_clear_value,
1002 VkImageAspectFlags aspects);
1003 void radv_set_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1004 struct radv_image *image,
1005 int idx,
1006 uint32_t color_values[2]);
1007 void radv_set_dcc_need_cmask_elim_pred(struct radv_cmd_buffer *cmd_buffer,
1008 struct radv_image *image,
1009 bool value);
1010 uint32_t radv_fill_buffer(struct radv_cmd_buffer *cmd_buffer,
1011 struct radeon_winsys_bo *bo,
1012 uint64_t offset, uint64_t size, uint32_t value);
1013 void radv_cmd_buffer_trace_emit(struct radv_cmd_buffer *cmd_buffer);
1014 bool radv_get_memory_fd(struct radv_device *device,
1015 struct radv_device_memory *memory,
1016 int *pFD);
1017
1018 /*
1019 * Takes x,y,z as exact numbers of invocations, instead of blocks.
1020 *
1021 * Limitations: Can't call normal dispatch functions without binding or rebinding
1022 * the compute pipeline.
1023 */
1024 void radv_unaligned_dispatch(
1025 struct radv_cmd_buffer *cmd_buffer,
1026 uint32_t x,
1027 uint32_t y,
1028 uint32_t z);
1029
1030 struct radv_event {
1031 struct radeon_winsys_bo *bo;
1032 uint64_t *map;
1033 };
1034
1035 struct radv_shader_module;
1036
1037 #define RADV_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
1038 #define RADV_HASH_SHADER_SISCHED (1 << 1)
1039 #define RADV_HASH_SHADER_UNSAFE_MATH (1 << 2)
1040 void
1041 radv_hash_shaders(unsigned char *hash,
1042 const VkPipelineShaderStageCreateInfo **stages,
1043 const struct radv_pipeline_layout *layout,
1044 const struct radv_pipeline_key *key,
1045 uint32_t flags);
1046
1047 static inline gl_shader_stage
1048 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
1049 {
1050 assert(__builtin_popcount(vk_stage) == 1);
1051 return ffs(vk_stage) - 1;
1052 }
1053
1054 static inline VkShaderStageFlagBits
1055 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
1056 {
1057 return (1 << mesa_stage);
1058 }
1059
1060 #define RADV_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
1061
1062 #define radv_foreach_stage(stage, stage_bits) \
1063 for (gl_shader_stage stage, \
1064 __tmp = (gl_shader_stage)((stage_bits) & RADV_STAGE_MASK); \
1065 stage = __builtin_ffs(__tmp) - 1, __tmp; \
1066 __tmp &= ~(1 << (stage)))
1067
1068 struct radv_depth_stencil_state {
1069 uint32_t db_depth_control;
1070 uint32_t db_stencil_control;
1071 uint32_t db_render_control;
1072 uint32_t db_render_override2;
1073 };
1074
1075 struct radv_blend_state {
1076 uint32_t cb_color_control;
1077 uint32_t cb_target_mask;
1078 uint32_t sx_mrt_blend_opt[8];
1079 uint32_t cb_blend_control[8];
1080
1081 uint32_t spi_shader_col_format;
1082 uint32_t cb_shader_mask;
1083 uint32_t db_alpha_to_mask;
1084 };
1085
1086 unsigned radv_format_meta_fs_key(VkFormat format);
1087
1088 struct radv_raster_state {
1089 uint32_t pa_cl_clip_cntl;
1090 uint32_t spi_interp_control;
1091 uint32_t pa_su_vtx_cntl;
1092 uint32_t pa_su_sc_mode_cntl;
1093 };
1094
1095 struct radv_multisample_state {
1096 uint32_t db_eqaa;
1097 uint32_t pa_sc_line_cntl;
1098 uint32_t pa_sc_mode_cntl_0;
1099 uint32_t pa_sc_mode_cntl_1;
1100 uint32_t pa_sc_aa_config;
1101 uint32_t pa_sc_aa_mask[2];
1102 unsigned num_samples;
1103 };
1104
1105 struct radv_prim_vertex_count {
1106 uint8_t min;
1107 uint8_t incr;
1108 };
1109
1110 struct radv_tessellation_state {
1111 uint32_t ls_hs_config;
1112 uint32_t tcs_in_layout;
1113 uint32_t tcs_out_layout;
1114 uint32_t tcs_out_offsets;
1115 uint32_t offchip_layout;
1116 unsigned num_patches;
1117 unsigned lds_size;
1118 unsigned num_tcs_input_cp;
1119 uint32_t tf_param;
1120 };
1121
1122 struct radv_gs_state {
1123 uint32_t vgt_gs_onchip_cntl;
1124 uint32_t vgt_gs_max_prims_per_subgroup;
1125 uint32_t vgt_esgs_ring_itemsize;
1126 uint32_t lds_size;
1127 };
1128
1129 struct radv_vertex_elements_info {
1130 uint32_t rsrc_word3[MAX_VERTEX_ATTRIBS];
1131 uint32_t format_size[MAX_VERTEX_ATTRIBS];
1132 uint32_t binding[MAX_VERTEX_ATTRIBS];
1133 uint32_t offset[MAX_VERTEX_ATTRIBS];
1134 uint32_t count;
1135 };
1136
1137 struct radv_vs_state {
1138 uint32_t pa_cl_vs_out_cntl;
1139 uint32_t spi_shader_pos_format;
1140 uint32_t spi_vs_out_config;
1141 uint32_t vgt_reuse_off;
1142 };
1143
1144 #define SI_GS_PER_ES 128
1145
1146 struct radv_pipeline {
1147 struct radv_device * device;
1148 struct radv_dynamic_state dynamic_state;
1149
1150 struct radv_pipeline_layout * layout;
1151
1152 bool needs_data_cache;
1153 bool need_indirect_descriptor_sets;
1154 struct radv_shader_variant * shaders[MESA_SHADER_STAGES];
1155 struct radv_shader_variant *gs_copy_shader;
1156 VkShaderStageFlags active_stages;
1157
1158 struct radv_vertex_elements_info vertex_elements;
1159
1160 uint32_t binding_stride[MAX_VBS];
1161
1162 uint32_t user_data_0[MESA_SHADER_STAGES];
1163 union {
1164 struct {
1165 struct radv_blend_state blend;
1166 struct radv_depth_stencil_state ds;
1167 struct radv_raster_state raster;
1168 struct radv_multisample_state ms;
1169 struct radv_tessellation_state tess;
1170 struct radv_gs_state gs;
1171 struct radv_vs_state vs;
1172 uint32_t db_shader_control;
1173 uint32_t shader_z_format;
1174 unsigned prim;
1175 unsigned gs_out;
1176 uint32_t vgt_gs_mode;
1177 bool vgt_primitiveid_en;
1178 bool prim_restart_enable;
1179 bool partial_es_wave;
1180 uint8_t primgroup_size;
1181 unsigned esgs_ring_size;
1182 unsigned gsvs_ring_size;
1183 uint32_t ps_input_cntl[32];
1184 uint32_t ps_input_cntl_num;
1185 uint32_t vgt_shader_stages_en;
1186 uint32_t vtx_base_sgpr;
1187 uint32_t base_ia_multi_vgt_param;
1188 bool wd_switch_on_eop;
1189 bool ia_switch_on_eoi;
1190 bool partial_vs_wave;
1191 uint8_t vtx_emit_num;
1192 uint32_t vtx_reuse_depth;
1193 struct radv_prim_vertex_count prim_vertex_count;
1194 bool can_use_guardband;
1195 } graphics;
1196 };
1197
1198 unsigned max_waves;
1199 unsigned scratch_bytes_per_wave;
1200 };
1201
1202 static inline bool radv_pipeline_has_gs(struct radv_pipeline *pipeline)
1203 {
1204 return pipeline->shaders[MESA_SHADER_GEOMETRY] ? true : false;
1205 }
1206
1207 static inline bool radv_pipeline_has_tess(struct radv_pipeline *pipeline)
1208 {
1209 return pipeline->shaders[MESA_SHADER_TESS_CTRL] ? true : false;
1210 }
1211
1212 struct ac_userdata_info *radv_lookup_user_sgpr(struct radv_pipeline *pipeline,
1213 gl_shader_stage stage,
1214 int idx);
1215
1216 struct radv_shader_variant *radv_get_vertex_shader(struct radv_pipeline *pipeline);
1217
1218 struct radv_graphics_pipeline_create_info {
1219 bool use_rectlist;
1220 bool db_depth_clear;
1221 bool db_stencil_clear;
1222 bool db_depth_disable_expclear;
1223 bool db_stencil_disable_expclear;
1224 bool db_flush_depth_inplace;
1225 bool db_flush_stencil_inplace;
1226 bool db_resummarize;
1227 uint32_t custom_blend_mode;
1228 };
1229
1230 VkResult
1231 radv_graphics_pipeline_create(VkDevice device,
1232 VkPipelineCache cache,
1233 const VkGraphicsPipelineCreateInfo *pCreateInfo,
1234 const struct radv_graphics_pipeline_create_info *extra,
1235 const VkAllocationCallbacks *alloc,
1236 VkPipeline *pPipeline);
1237
1238 struct vk_format_description;
1239 uint32_t radv_translate_buffer_dataformat(const struct vk_format_description *desc,
1240 int first_non_void);
1241 uint32_t radv_translate_buffer_numformat(const struct vk_format_description *desc,
1242 int first_non_void);
1243 uint32_t radv_translate_colorformat(VkFormat format);
1244 uint32_t radv_translate_color_numformat(VkFormat format,
1245 const struct vk_format_description *desc,
1246 int first_non_void);
1247 uint32_t radv_colorformat_endian_swap(uint32_t colorformat);
1248 unsigned radv_translate_colorswap(VkFormat format, bool do_endian_swap);
1249 uint32_t radv_translate_dbformat(VkFormat format);
1250 uint32_t radv_translate_tex_dataformat(VkFormat format,
1251 const struct vk_format_description *desc,
1252 int first_non_void);
1253 uint32_t radv_translate_tex_numformat(VkFormat format,
1254 const struct vk_format_description *desc,
1255 int first_non_void);
1256 bool radv_format_pack_clear_color(VkFormat format,
1257 uint32_t clear_vals[2],
1258 VkClearColorValue *value);
1259 bool radv_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
1260 bool radv_dcc_formats_compatible(VkFormat format1,
1261 VkFormat format2);
1262
1263 struct radv_fmask_info {
1264 uint64_t offset;
1265 uint64_t size;
1266 unsigned alignment;
1267 unsigned pitch_in_pixels;
1268 unsigned bank_height;
1269 unsigned slice_tile_max;
1270 unsigned tile_mode_index;
1271 unsigned tile_swizzle;
1272 };
1273
1274 struct radv_cmask_info {
1275 uint64_t offset;
1276 uint64_t size;
1277 unsigned alignment;
1278 unsigned slice_tile_max;
1279 };
1280
1281 struct radv_image {
1282 VkImageType type;
1283 /* The original VkFormat provided by the client. This may not match any
1284 * of the actual surface formats.
1285 */
1286 VkFormat vk_format;
1287 VkImageAspectFlags aspects;
1288 VkImageUsageFlags usage; /**< Superset of VkImageCreateInfo::usage. */
1289 struct ac_surf_info info;
1290 VkImageTiling tiling; /** VkImageCreateInfo::tiling */
1291 VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
1292
1293 VkDeviceSize size;
1294 uint32_t alignment;
1295
1296 unsigned queue_family_mask;
1297 bool exclusive;
1298 bool shareable;
1299
1300 /* Set when bound */
1301 struct radeon_winsys_bo *bo;
1302 VkDeviceSize offset;
1303 uint64_t dcc_offset;
1304 uint64_t htile_offset;
1305 bool tc_compatible_htile;
1306 struct radeon_surf surface;
1307
1308 struct radv_fmask_info fmask;
1309 struct radv_cmask_info cmask;
1310 uint64_t clear_value_offset;
1311 uint64_t dcc_pred_offset;
1312 };
1313
1314 /* Whether the image has a htile that is known consistent with the contents of
1315 * the image. */
1316 bool radv_layout_has_htile(const struct radv_image *image,
1317 VkImageLayout layout,
1318 unsigned queue_mask);
1319
1320 /* Whether the image has a htile that is known consistent with the contents of
1321 * the image and is allowed to be in compressed form.
1322 *
1323 * If this is false reads that don't use the htile should be able to return
1324 * correct results.
1325 */
1326 bool radv_layout_is_htile_compressed(const struct radv_image *image,
1327 VkImageLayout layout,
1328 unsigned queue_mask);
1329
1330 bool radv_layout_can_fast_clear(const struct radv_image *image,
1331 VkImageLayout layout,
1332 unsigned queue_mask);
1333
1334 static inline bool
1335 radv_vi_dcc_enabled(const struct radv_image *image, unsigned level)
1336 {
1337 return image->surface.dcc_size && level < image->surface.num_dcc_levels;
1338 }
1339
1340 static inline bool
1341 radv_htile_enabled(const struct radv_image *image, unsigned level)
1342 {
1343 return image->surface.htile_size && level == 0;
1344 }
1345
1346 unsigned radv_image_queue_family_mask(const struct radv_image *image, uint32_t family, uint32_t queue_family);
1347
1348 static inline uint32_t
1349 radv_get_layerCount(const struct radv_image *image,
1350 const VkImageSubresourceRange *range)
1351 {
1352 return range->layerCount == VK_REMAINING_ARRAY_LAYERS ?
1353 image->info.array_size - range->baseArrayLayer : range->layerCount;
1354 }
1355
1356 static inline uint32_t
1357 radv_get_levelCount(const struct radv_image *image,
1358 const VkImageSubresourceRange *range)
1359 {
1360 return range->levelCount == VK_REMAINING_MIP_LEVELS ?
1361 image->info.levels - range->baseMipLevel : range->levelCount;
1362 }
1363
1364 struct radeon_bo_metadata;
1365 void
1366 radv_init_metadata(struct radv_device *device,
1367 struct radv_image *image,
1368 struct radeon_bo_metadata *metadata);
1369
1370 struct radv_image_view {
1371 struct radv_image *image; /**< VkImageViewCreateInfo::image */
1372 struct radeon_winsys_bo *bo;
1373
1374 VkImageViewType type;
1375 VkImageAspectFlags aspect_mask;
1376 VkFormat vk_format;
1377 uint32_t base_layer;
1378 uint32_t layer_count;
1379 uint32_t base_mip;
1380 uint32_t level_count;
1381 VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
1382
1383 uint32_t descriptor[16];
1384
1385 /* Descriptor for use as a storage image as opposed to a sampled image.
1386 * This has a few differences for cube maps (e.g. type).
1387 */
1388 uint32_t storage_descriptor[16];
1389 };
1390
1391 struct radv_image_create_info {
1392 const VkImageCreateInfo *vk_info;
1393 bool scanout;
1394 };
1395
1396 VkResult radv_image_create(VkDevice _device,
1397 const struct radv_image_create_info *info,
1398 const VkAllocationCallbacks* alloc,
1399 VkImage *pImage);
1400
1401 void radv_image_view_init(struct radv_image_view *view,
1402 struct radv_device *device,
1403 const VkImageViewCreateInfo* pCreateInfo);
1404
1405 struct radv_buffer_view {
1406 struct radeon_winsys_bo *bo;
1407 VkFormat vk_format;
1408 uint64_t range; /**< VkBufferViewCreateInfo::range */
1409 uint32_t state[4];
1410 };
1411 void radv_buffer_view_init(struct radv_buffer_view *view,
1412 struct radv_device *device,
1413 const VkBufferViewCreateInfo* pCreateInfo);
1414
1415 static inline struct VkExtent3D
1416 radv_sanitize_image_extent(const VkImageType imageType,
1417 const struct VkExtent3D imageExtent)
1418 {
1419 switch (imageType) {
1420 case VK_IMAGE_TYPE_1D:
1421 return (VkExtent3D) { imageExtent.width, 1, 1 };
1422 case VK_IMAGE_TYPE_2D:
1423 return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1424 case VK_IMAGE_TYPE_3D:
1425 return imageExtent;
1426 default:
1427 unreachable("invalid image type");
1428 }
1429 }
1430
1431 static inline struct VkOffset3D
1432 radv_sanitize_image_offset(const VkImageType imageType,
1433 const struct VkOffset3D imageOffset)
1434 {
1435 switch (imageType) {
1436 case VK_IMAGE_TYPE_1D:
1437 return (VkOffset3D) { imageOffset.x, 0, 0 };
1438 case VK_IMAGE_TYPE_2D:
1439 return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1440 case VK_IMAGE_TYPE_3D:
1441 return imageOffset;
1442 default:
1443 unreachable("invalid image type");
1444 }
1445 }
1446
1447 static inline bool
1448 radv_image_extent_compare(const struct radv_image *image,
1449 const VkExtent3D *extent)
1450 {
1451 if (extent->width != image->info.width ||
1452 extent->height != image->info.height ||
1453 extent->depth != image->info.depth)
1454 return false;
1455 return true;
1456 }
1457
1458 struct radv_sampler {
1459 uint32_t state[4];
1460 };
1461
1462 struct radv_color_buffer_info {
1463 uint64_t cb_color_base;
1464 uint64_t cb_color_cmask;
1465 uint64_t cb_color_fmask;
1466 uint64_t cb_dcc_base;
1467 uint32_t cb_color_pitch;
1468 uint32_t cb_color_slice;
1469 uint32_t cb_color_view;
1470 uint32_t cb_color_info;
1471 uint32_t cb_color_attrib;
1472 uint32_t cb_color_attrib2;
1473 uint32_t cb_dcc_control;
1474 uint32_t cb_color_cmask_slice;
1475 uint32_t cb_color_fmask_slice;
1476 uint32_t cb_clear_value0;
1477 uint32_t cb_clear_value1;
1478 };
1479
1480 struct radv_ds_buffer_info {
1481 uint64_t db_z_read_base;
1482 uint64_t db_stencil_read_base;
1483 uint64_t db_z_write_base;
1484 uint64_t db_stencil_write_base;
1485 uint64_t db_htile_data_base;
1486 uint32_t db_depth_info;
1487 uint32_t db_z_info;
1488 uint32_t db_stencil_info;
1489 uint32_t db_depth_view;
1490 uint32_t db_depth_size;
1491 uint32_t db_depth_slice;
1492 uint32_t db_htile_surface;
1493 uint32_t pa_su_poly_offset_db_fmt_cntl;
1494 uint32_t db_z_info2;
1495 uint32_t db_stencil_info2;
1496 float offset_scale;
1497 };
1498
1499 struct radv_attachment_info {
1500 union {
1501 struct radv_color_buffer_info cb;
1502 struct radv_ds_buffer_info ds;
1503 };
1504 struct radv_image_view *attachment;
1505 };
1506
1507 struct radv_framebuffer {
1508 uint32_t width;
1509 uint32_t height;
1510 uint32_t layers;
1511
1512 uint32_t attachment_count;
1513 struct radv_attachment_info attachments[0];
1514 };
1515
1516 struct radv_subpass_barrier {
1517 VkPipelineStageFlags src_stage_mask;
1518 VkAccessFlags src_access_mask;
1519 VkAccessFlags dst_access_mask;
1520 };
1521
1522 struct radv_subpass {
1523 uint32_t input_count;
1524 uint32_t color_count;
1525 VkAttachmentReference * input_attachments;
1526 VkAttachmentReference * color_attachments;
1527 VkAttachmentReference * resolve_attachments;
1528 VkAttachmentReference depth_stencil_attachment;
1529
1530 /** Subpass has at least one resolve attachment */
1531 bool has_resolve;
1532
1533 struct radv_subpass_barrier start_barrier;
1534
1535 uint32_t view_mask;
1536 };
1537
1538 struct radv_render_pass_attachment {
1539 VkFormat format;
1540 uint32_t samples;
1541 VkAttachmentLoadOp load_op;
1542 VkAttachmentLoadOp stencil_load_op;
1543 VkImageLayout initial_layout;
1544 VkImageLayout final_layout;
1545 uint32_t view_mask;
1546 };
1547
1548 struct radv_render_pass {
1549 uint32_t attachment_count;
1550 uint32_t subpass_count;
1551 VkAttachmentReference * subpass_attachments;
1552 struct radv_render_pass_attachment * attachments;
1553 struct radv_subpass_barrier end_barrier;
1554 struct radv_subpass subpasses[0];
1555 };
1556
1557 VkResult radv_device_init_meta(struct radv_device *device);
1558 void radv_device_finish_meta(struct radv_device *device);
1559
1560 struct radv_query_pool {
1561 struct radeon_winsys_bo *bo;
1562 uint32_t stride;
1563 uint32_t availability_offset;
1564 char *ptr;
1565 VkQueryType type;
1566 uint32_t pipeline_stats_mask;
1567 };
1568
1569 struct radv_semaphore {
1570 /* use a winsys sem for non-exportable */
1571 struct radeon_winsys_sem *sem;
1572 uint32_t syncobj;
1573 uint32_t temp_syncobj;
1574 };
1575
1576 VkResult radv_alloc_sem_info(struct radv_winsys_sem_info *sem_info,
1577 int num_wait_sems,
1578 const VkSemaphore *wait_sems,
1579 int num_signal_sems,
1580 const VkSemaphore *signal_sems,
1581 VkFence fence);
1582 void radv_free_sem_info(struct radv_winsys_sem_info *sem_info);
1583
1584 void radv_set_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
1585 struct radv_descriptor_set *set,
1586 unsigned idx);
1587
1588 void
1589 radv_update_descriptor_sets(struct radv_device *device,
1590 struct radv_cmd_buffer *cmd_buffer,
1591 VkDescriptorSet overrideSet,
1592 uint32_t descriptorWriteCount,
1593 const VkWriteDescriptorSet *pDescriptorWrites,
1594 uint32_t descriptorCopyCount,
1595 const VkCopyDescriptorSet *pDescriptorCopies);
1596
1597 void
1598 radv_update_descriptor_set_with_template(struct radv_device *device,
1599 struct radv_cmd_buffer *cmd_buffer,
1600 struct radv_descriptor_set *set,
1601 VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1602 const void *pData);
1603
1604 void radv_meta_push_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
1605 VkPipelineBindPoint pipelineBindPoint,
1606 VkPipelineLayout _layout,
1607 uint32_t set,
1608 uint32_t descriptorWriteCount,
1609 const VkWriteDescriptorSet *pDescriptorWrites);
1610
1611 void radv_initialise_cmask(struct radv_cmd_buffer *cmd_buffer,
1612 struct radv_image *image, uint32_t value);
1613 void radv_initialize_dcc(struct radv_cmd_buffer *cmd_buffer,
1614 struct radv_image *image, uint32_t value);
1615
1616 struct radv_fence {
1617 struct radeon_winsys_fence *fence;
1618 bool submitted;
1619 bool signalled;
1620
1621 uint32_t syncobj;
1622 uint32_t temp_syncobj;
1623 };
1624
1625 struct radeon_winsys_sem;
1626
1627 #define RADV_DEFINE_HANDLE_CASTS(__radv_type, __VkType) \
1628 \
1629 static inline struct __radv_type * \
1630 __radv_type ## _from_handle(__VkType _handle) \
1631 { \
1632 return (struct __radv_type *) _handle; \
1633 } \
1634 \
1635 static inline __VkType \
1636 __radv_type ## _to_handle(struct __radv_type *_obj) \
1637 { \
1638 return (__VkType) _obj; \
1639 }
1640
1641 #define RADV_DEFINE_NONDISP_HANDLE_CASTS(__radv_type, __VkType) \
1642 \
1643 static inline struct __radv_type * \
1644 __radv_type ## _from_handle(__VkType _handle) \
1645 { \
1646 return (struct __radv_type *)(uintptr_t) _handle; \
1647 } \
1648 \
1649 static inline __VkType \
1650 __radv_type ## _to_handle(struct __radv_type *_obj) \
1651 { \
1652 return (__VkType)(uintptr_t) _obj; \
1653 }
1654
1655 #define RADV_FROM_HANDLE(__radv_type, __name, __handle) \
1656 struct __radv_type *__name = __radv_type ## _from_handle(__handle)
1657
1658 RADV_DEFINE_HANDLE_CASTS(radv_cmd_buffer, VkCommandBuffer)
1659 RADV_DEFINE_HANDLE_CASTS(radv_device, VkDevice)
1660 RADV_DEFINE_HANDLE_CASTS(radv_instance, VkInstance)
1661 RADV_DEFINE_HANDLE_CASTS(radv_physical_device, VkPhysicalDevice)
1662 RADV_DEFINE_HANDLE_CASTS(radv_queue, VkQueue)
1663
1664 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_cmd_pool, VkCommandPool)
1665 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer, VkBuffer)
1666 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer_view, VkBufferView)
1667 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_pool, VkDescriptorPool)
1668 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set, VkDescriptorSet)
1669 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set_layout, VkDescriptorSetLayout)
1670 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_update_template, VkDescriptorUpdateTemplateKHR)
1671 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_device_memory, VkDeviceMemory)
1672 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_fence, VkFence)
1673 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_event, VkEvent)
1674 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_framebuffer, VkFramebuffer)
1675 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image, VkImage)
1676 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image_view, VkImageView);
1677 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_cache, VkPipelineCache)
1678 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline, VkPipeline)
1679 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_layout, VkPipelineLayout)
1680 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_query_pool, VkQueryPool)
1681 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_render_pass, VkRenderPass)
1682 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_sampler, VkSampler)
1683 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_shader_module, VkShaderModule)
1684 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_semaphore, VkSemaphore)
1685
1686 #endif /* RADV_PRIVATE_H */