projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: pre-calculate user_data_0 registers and store in pipeline
[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 enum radv_mem_flags_bits {
110 /* enable implicit synchronization when accessing the underlying bo */
111 RADV_MEM_IMPLICIT_SYNC = 1 << 0,
112 };
113
114 #define radv_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
115
116 static inline uint32_t
117 align_u32(uint32_t v, uint32_t a)
118 {
119 assert(a != 0 && a == (a & -a));
120 return (v + a - 1) & ~(a - 1);
121 }
122
123 static inline uint32_t
124 align_u32_npot(uint32_t v, uint32_t a)
125 {
126 return (v + a - 1) / a * a;
127 }
128
129 static inline uint64_t
130 align_u64(uint64_t v, uint64_t a)
131 {
132 assert(a != 0 && a == (a & -a));
133 return (v + a - 1) & ~(a - 1);
134 }
135
136 static inline int32_t
137 align_i32(int32_t v, int32_t a)
138 {
139 assert(a != 0 && a == (a & -a));
140 return (v + a - 1) & ~(a - 1);
141 }
142
143 /** Alignment must be a power of 2. */
144 static inline bool
145 radv_is_aligned(uintmax_t n, uintmax_t a)
146 {
147 assert(a == (a & -a));
148 return (n & (a - 1)) == 0;
149 }
150
151 static inline uint32_t
152 round_up_u32(uint32_t v, uint32_t a)
153 {
154 return (v + a - 1) / a;
155 }
156
157 static inline uint64_t
158 round_up_u64(uint64_t v, uint64_t a)
159 {
160 return (v + a - 1) / a;
161 }
162
163 static inline uint32_t
164 radv_minify(uint32_t n, uint32_t levels)
165 {
166 if (unlikely(n == 0))
167 return 0;
168 else
169 return MAX2(n >> levels, 1);
170 }
171 static inline float
172 radv_clamp_f(float f, float min, float max)
173 {
174 assert(min < max);
175
176 if (f > max)
177 return max;
178 else if (f < min)
179 return min;
180 else
181 return f;
182 }
183
184 static inline bool
185 radv_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
186 {
187 if (*inout_mask & clear_mask) {
188 *inout_mask &= ~clear_mask;
189 return true;
190 } else {
191 return false;
192 }
193 }
194
195 #define for_each_bit(b, dword) \
196 for (uint32_t __dword = (dword); \
197 (b) = __builtin_ffs(__dword) - 1, __dword; \
198 __dword &= ~(1 << (b)))
199
200 #define typed_memcpy(dest, src, count) ({ \
201 STATIC_ASSERT(sizeof(*src) == sizeof(*dest)); \
202 memcpy((dest), (src), (count) * sizeof(*(src))); \
203 })
204
205 /* Whenever we generate an error, pass it through this function. Useful for
206 * debugging, where we can break on it. Only call at error site, not when
207 * propagating errors. Might be useful to plug in a stack trace here.
208 */
209
210 VkResult __vk_errorf(VkResult error, const char *file, int line, const char *format, ...);
211
212 #ifdef DEBUG
213 #define vk_error(error) __vk_errorf(error, __FILE__, __LINE__, NULL);
214 #define vk_errorf(error, format, ...) __vk_errorf(error, __FILE__, __LINE__, format, ## __VA_ARGS__);
215 #else
216 #define vk_error(error) error
217 #define vk_errorf(error, format, ...) error
218 #endif
219
220 void __radv_finishme(const char *file, int line, const char *format, ...)
221 radv_printflike(3, 4);
222 void radv_loge(const char *format, ...) radv_printflike(1, 2);
223 void radv_loge_v(const char *format, va_list va);
224
225 /**
226 * Print a FINISHME message, including its source location.
227 */
228 #define radv_finishme(format, ...) \
229 do { \
230 static bool reported = false; \
231 if (!reported) { \
232 __radv_finishme(__FILE__, __LINE__, format, ##__VA_ARGS__); \
233 reported = true; \
234 } \
235 } while (0)
236
237 /* A non-fatal assert. Useful for debugging. */
238 #ifdef DEBUG
239 #define radv_assert(x) ({ \
240 if (unlikely(!(x))) \
241 fprintf(stderr, "%s:%d ASSERT: %s\n", __FILE__, __LINE__, #x); \
242 })
243 #else
244 #define radv_assert(x)
245 #endif
246
247 #define stub_return(v) \
248 do { \
249 radv_finishme("stub %s", __func__); \
250 return (v); \
251 } while (0)
252
253 #define stub() \
254 do { \
255 radv_finishme("stub %s", __func__); \
256 return; \
257 } while (0)
258
259 void *radv_lookup_entrypoint(const char *name);
260
261 struct radv_physical_device {
262 VK_LOADER_DATA _loader_data;
263
264 struct radv_instance * instance;
265
266 struct radeon_winsys *ws;
267 struct radeon_info rad_info;
268 char path[20];
269 const char * name;
270 uint8_t driver_uuid[VK_UUID_SIZE];
271 uint8_t device_uuid[VK_UUID_SIZE];
272 uint8_t cache_uuid[VK_UUID_SIZE];
273
274 int local_fd;
275 struct wsi_device wsi_device;
276
277 bool has_rbplus; /* if RB+ register exist */
278 bool rbplus_allowed; /* if RB+ is allowed */
279 bool has_clear_state;
280
281 /* This is the drivers on-disk cache used as a fallback as opposed to
282 * the pipeline cache defined by apps.
283 */
284 struct disk_cache * disk_cache;
285
286 VkPhysicalDeviceMemoryProperties memory_properties;
287 enum radv_mem_type mem_type_indices[RADV_MEM_TYPE_COUNT];
288 };
289
290 struct radv_instance {
291 VK_LOADER_DATA _loader_data;
292
293 VkAllocationCallbacks alloc;
294
295 uint32_t apiVersion;
296 int physicalDeviceCount;
297 struct radv_physical_device physicalDevices[RADV_MAX_DRM_DEVICES];
298
299 uint64_t debug_flags;
300 uint64_t perftest_flags;
301 };
302
303 VkResult radv_init_wsi(struct radv_physical_device *physical_device);
304 void radv_finish_wsi(struct radv_physical_device *physical_device);
305
306 bool radv_instance_extension_supported(const char *name);
307 uint32_t radv_physical_device_api_version(struct radv_physical_device *dev);
308 bool radv_physical_device_extension_supported(struct radv_physical_device *dev,
309 const char *name);
310
311 struct cache_entry;
312
313 struct radv_pipeline_cache {
314 struct radv_device * device;
315 pthread_mutex_t mutex;
316
317 uint32_t total_size;
318 uint32_t table_size;
319 uint32_t kernel_count;
320 struct cache_entry ** hash_table;
321 bool modified;
322
323 VkAllocationCallbacks alloc;
324 };
325
326 struct radv_pipeline_key {
327 uint32_t instance_rate_inputs;
328 unsigned tess_input_vertices;
329 uint32_t col_format;
330 uint32_t is_int8;
331 uint32_t is_int10;
332 uint32_t multisample : 1;
333 uint32_t has_multiview_view_index : 1;
334 };
335
336 void
337 radv_pipeline_cache_init(struct radv_pipeline_cache *cache,
338 struct radv_device *device);
339 void
340 radv_pipeline_cache_finish(struct radv_pipeline_cache *cache);
341 void
342 radv_pipeline_cache_load(struct radv_pipeline_cache *cache,
343 const void *data, size_t size);
344
345 struct radv_shader_variant;
346
347 bool
348 radv_create_shader_variants_from_pipeline_cache(struct radv_device *device,
349 struct radv_pipeline_cache *cache,
350 const unsigned char *sha1,
351 struct radv_shader_variant **variants);
352
353 void
354 radv_pipeline_cache_insert_shaders(struct radv_device *device,
355 struct radv_pipeline_cache *cache,
356 const unsigned char *sha1,
357 struct radv_shader_variant **variants,
358 const void *const *codes,
359 const unsigned *code_sizes);
360
361 struct radv_meta_state {
362 VkAllocationCallbacks alloc;
363
364 struct radv_pipeline_cache cache;
365
366 /**
367 * Use array element `i` for images with `2^i` samples.
368 */
369 struct {
370 VkRenderPass render_pass[NUM_META_FS_KEYS];
371 VkPipeline color_pipelines[NUM_META_FS_KEYS];
372
373 VkRenderPass depthstencil_rp;
374 VkPipeline depth_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
375 VkPipeline stencil_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
376 VkPipeline depthstencil_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
377 } clear[1 + MAX_SAMPLES_LOG2];
378
379 VkPipelineLayout clear_color_p_layout;
380 VkPipelineLayout clear_depth_p_layout;
381 struct {
382 VkRenderPass render_pass[NUM_META_FS_KEYS];
383
384 /** Pipeline that blits from a 1D image. */
385 VkPipeline pipeline_1d_src[NUM_META_FS_KEYS];
386
387 /** Pipeline that blits from a 2D image. */
388 VkPipeline pipeline_2d_src[NUM_META_FS_KEYS];
389
390 /** Pipeline that blits from a 3D image. */
391 VkPipeline pipeline_3d_src[NUM_META_FS_KEYS];
392
393 VkRenderPass depth_only_rp;
394 VkPipeline depth_only_1d_pipeline;
395 VkPipeline depth_only_2d_pipeline;
396 VkPipeline depth_only_3d_pipeline;
397
398 VkRenderPass stencil_only_rp;
399 VkPipeline stencil_only_1d_pipeline;
400 VkPipeline stencil_only_2d_pipeline;
401 VkPipeline stencil_only_3d_pipeline;
402 VkPipelineLayout pipeline_layout;
403 VkDescriptorSetLayout ds_layout;
404 } blit;
405
406 struct {
407 VkRenderPass render_passes[NUM_META_FS_KEYS];
408
409 VkPipelineLayout p_layouts[2];
410 VkDescriptorSetLayout ds_layouts[2];
411 VkPipeline pipelines[2][NUM_META_FS_KEYS];
412
413 VkRenderPass depth_only_rp;
414 VkPipeline depth_only_pipeline[2];
415
416 VkRenderPass stencil_only_rp;
417 VkPipeline stencil_only_pipeline[2];
418 } blit2d;
419
420 struct {
421 VkPipelineLayout img_p_layout;
422 VkDescriptorSetLayout img_ds_layout;
423 VkPipeline pipeline;
424 } itob;
425 struct {
426 VkPipelineLayout img_p_layout;
427 VkDescriptorSetLayout img_ds_layout;
428 VkPipeline pipeline;
429 } btoi;
430 struct {
431 VkPipelineLayout img_p_layout;
432 VkDescriptorSetLayout img_ds_layout;
433 VkPipeline pipeline;
434 } itoi;
435 struct {
436 VkPipelineLayout img_p_layout;
437 VkDescriptorSetLayout img_ds_layout;
438 VkPipeline pipeline;
439 } cleari;
440
441 struct {
442 VkPipeline pipeline;
443 VkRenderPass pass;
444 } resolve;
445
446 struct {
447 VkDescriptorSetLayout ds_layout;
448 VkPipelineLayout p_layout;
449 struct {
450 VkPipeline pipeline;
451 VkPipeline i_pipeline;
452 VkPipeline srgb_pipeline;
453 } rc[MAX_SAMPLES_LOG2];
454 } resolve_compute;
455
456 struct {
457 VkDescriptorSetLayout ds_layout;
458 VkPipelineLayout p_layout;
459
460 struct {
461 VkRenderPass render_pass[NUM_META_FS_KEYS];
462 VkPipeline pipeline[NUM_META_FS_KEYS];
463 } rc[MAX_SAMPLES_LOG2];
464 } resolve_fragment;
465
466 struct {
467 VkPipeline decompress_pipeline;
468 VkPipeline resummarize_pipeline;
469 VkRenderPass pass;
470 } depth_decomp[1 + MAX_SAMPLES_LOG2];
471
472 struct {
473 VkPipeline cmask_eliminate_pipeline;
474 VkPipeline fmask_decompress_pipeline;
475 VkRenderPass pass;
476 } fast_clear_flush;
477
478 struct {
479 VkPipelineLayout fill_p_layout;
480 VkPipelineLayout copy_p_layout;
481 VkDescriptorSetLayout fill_ds_layout;
482 VkDescriptorSetLayout copy_ds_layout;
483 VkPipeline fill_pipeline;
484 VkPipeline copy_pipeline;
485 } buffer;
486
487 struct {
488 VkDescriptorSetLayout ds_layout;
489 VkPipelineLayout p_layout;
490 VkPipeline occlusion_query_pipeline;
491 VkPipeline pipeline_statistics_query_pipeline;
492 } query;
493 };
494
495 /* queue types */
496 #define RADV_QUEUE_GENERAL 0
497 #define RADV_QUEUE_COMPUTE 1
498 #define RADV_QUEUE_TRANSFER 2
499
500 #define RADV_MAX_QUEUE_FAMILIES 3
501
502 enum ring_type radv_queue_family_to_ring(int f);
503
504 struct radv_queue {
505 VK_LOADER_DATA _loader_data;
506 struct radv_device * device;
507 struct radeon_winsys_ctx *hw_ctx;
508 enum radeon_ctx_priority priority;
509 int queue_family_index;
510 int queue_idx;
511
512 uint32_t scratch_size;
513 uint32_t compute_scratch_size;
514 uint32_t esgs_ring_size;
515 uint32_t gsvs_ring_size;
516 bool has_tess_rings;
517 bool has_sample_positions;
518
519 struct radeon_winsys_bo *scratch_bo;
520 struct radeon_winsys_bo *descriptor_bo;
521 struct radeon_winsys_bo *compute_scratch_bo;
522 struct radeon_winsys_bo *esgs_ring_bo;
523 struct radeon_winsys_bo *gsvs_ring_bo;
524 struct radeon_winsys_bo *tess_factor_ring_bo;
525 struct radeon_winsys_bo *tess_offchip_ring_bo;
526 struct radeon_winsys_cs *initial_preamble_cs;
527 struct radeon_winsys_cs *initial_full_flush_preamble_cs;
528 struct radeon_winsys_cs *continue_preamble_cs;
529 };
530
531 struct radv_device {
532 VK_LOADER_DATA _loader_data;
533
534 VkAllocationCallbacks alloc;
535
536 struct radv_instance * instance;
537 struct radeon_winsys *ws;
538
539 struct radv_meta_state meta_state;
540
541 struct radv_queue *queues[RADV_MAX_QUEUE_FAMILIES];
542 int queue_count[RADV_MAX_QUEUE_FAMILIES];
543 struct radeon_winsys_cs *empty_cs[RADV_MAX_QUEUE_FAMILIES];
544
545 bool llvm_supports_spill;
546 bool has_distributed_tess;
547 bool dfsm_allowed;
548 uint32_t tess_offchip_block_dw_size;
549 uint32_t scratch_waves;
550
551 uint32_t gs_table_depth;
552
553 /* MSAA sample locations.
554 * The first index is the sample index.
555 * The second index is the coordinate: X, Y. */
556 float sample_locations_1x[1][2];
557 float sample_locations_2x[2][2];
558 float sample_locations_4x[4][2];
559 float sample_locations_8x[8][2];
560 float sample_locations_16x[16][2];
561
562 /* CIK and later */
563 uint32_t gfx_init_size_dw;
564 struct radeon_winsys_bo *gfx_init;
565
566 struct radeon_winsys_bo *trace_bo;
567 uint32_t *trace_id_ptr;
568
569 /* Whether to keep shader debug info, for tracing or VK_AMD_shader_info */
570 bool keep_shader_info;
571
572 struct radv_physical_device *physical_device;
573
574 /* Backup in-memory cache to be used if the app doesn't provide one */
575 struct radv_pipeline_cache * mem_cache;
576
577 /*
578 * use different counters so MSAA MRTs get consecutive surface indices,
579 * even if MASK is allocated in between.
580 */
581 uint32_t image_mrt_offset_counter;
582 uint32_t fmask_mrt_offset_counter;
583 struct list_head shader_slabs;
584 mtx_t shader_slab_mutex;
585
586 /* For detecting VM faults reported by dmesg. */
587 uint64_t dmesg_timestamp;
588 };
589
590 struct radv_device_memory {
591 struct radeon_winsys_bo *bo;
592 /* for dedicated allocations */
593 struct radv_image *image;
594 struct radv_buffer *buffer;
595 uint32_t type_index;
596 VkDeviceSize map_size;
597 void * map;
598 };
599
600
601 struct radv_descriptor_range {
602 uint64_t va;
603 uint32_t size;
604 };
605
606 struct radv_descriptor_set {
607 const struct radv_descriptor_set_layout *layout;
608 uint32_t size;
609
610 struct radeon_winsys_bo *bo;
611 uint64_t va;
612 uint32_t *mapped_ptr;
613 struct radv_descriptor_range *dynamic_descriptors;
614
615 struct radeon_winsys_bo *descriptors[0];
616 };
617
618 struct radv_push_descriptor_set
619 {
620 struct radv_descriptor_set set;
621 uint32_t capacity;
622 };
623
624 struct radv_descriptor_pool_entry {
625 uint32_t offset;
626 uint32_t size;
627 struct radv_descriptor_set *set;
628 };
629
630 struct radv_descriptor_pool {
631 struct radeon_winsys_bo *bo;
632 uint8_t *mapped_ptr;
633 uint64_t current_offset;
634 uint64_t size;
635
636 uint8_t *host_memory_base;
637 uint8_t *host_memory_ptr;
638 uint8_t *host_memory_end;
639
640 uint32_t entry_count;
641 uint32_t max_entry_count;
642 struct radv_descriptor_pool_entry entries[0];
643 };
644
645 struct radv_descriptor_update_template_entry {
646 VkDescriptorType descriptor_type;
647
648 /* The number of descriptors to update */
649 uint32_t descriptor_count;
650
651 /* Into mapped_ptr or dynamic_descriptors, in units of the respective array */
652 uint32_t dst_offset;
653
654 /* In dwords. Not valid/used for dynamic descriptors */
655 uint32_t dst_stride;
656
657 uint32_t buffer_offset;
658
659 /* Only valid for combined image samplers and samplers */
660 uint16_t has_sampler;
661
662 /* In bytes */
663 size_t src_offset;
664 size_t src_stride;
665
666 /* For push descriptors */
667 const uint32_t *immutable_samplers;
668 };
669
670 struct radv_descriptor_update_template {
671 uint32_t entry_count;
672 struct radv_descriptor_update_template_entry entry[0];
673 };
674
675 struct radv_buffer {
676 struct radv_device * device;
677 VkDeviceSize size;
678
679 VkBufferUsageFlags usage;
680 VkBufferCreateFlags flags;
681
682 /* Set when bound */
683 struct radeon_winsys_bo * bo;
684 VkDeviceSize offset;
685
686 bool shareable;
687 };
688
689
690 enum radv_cmd_dirty_bits {
691 RADV_CMD_DIRTY_DYNAMIC_VIEWPORT = 1 << 0, /* VK_DYNAMIC_STATE_VIEWPORT */
692 RADV_CMD_DIRTY_DYNAMIC_SCISSOR = 1 << 1, /* VK_DYNAMIC_STATE_SCISSOR */
693 RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH = 1 << 2, /* VK_DYNAMIC_STATE_LINE_WIDTH */
694 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS = 1 << 3, /* VK_DYNAMIC_STATE_DEPTH_BIAS */
695 RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS = 1 << 4, /* VK_DYNAMIC_STATE_BLEND_CONSTANTS */
696 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS = 1 << 5, /* VK_DYNAMIC_STATE_DEPTH_BOUNDS */
697 RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6, /* VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK */
698 RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7, /* VK_DYNAMIC_STATE_STENCIL_WRITE_MASK */
699 RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE = 1 << 8, /* VK_DYNAMIC_STATE_STENCIL_REFERENCE */
700 RADV_CMD_DIRTY_DYNAMIC_ALL = (1 << 9) - 1,
701 RADV_CMD_DIRTY_PIPELINE = 1 << 9,
702 RADV_CMD_DIRTY_INDEX_BUFFER = 1 << 10,
703 RADV_CMD_DIRTY_FRAMEBUFFER = 1 << 11,
704 };
705 typedef uint32_t radv_cmd_dirty_mask_t;
706
707 enum radv_cmd_flush_bits {
708 RADV_CMD_FLAG_INV_ICACHE = 1 << 0,
709 /* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
710 RADV_CMD_FLAG_INV_SMEM_L1 = 1 << 1,
711 /* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
712 RADV_CMD_FLAG_INV_VMEM_L1 = 1 << 2,
713 /* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
714 RADV_CMD_FLAG_INV_GLOBAL_L2 = 1 << 3,
715 /* Same as above, but only writes back and doesn't invalidate */
716 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2 = 1 << 4,
717 /* Framebuffer caches */
718 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META = 1 << 5,
719 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META = 1 << 6,
720 RADV_CMD_FLAG_FLUSH_AND_INV_DB = 1 << 7,
721 RADV_CMD_FLAG_FLUSH_AND_INV_CB = 1 << 8,
722 /* Engine synchronization. */
723 RADV_CMD_FLAG_VS_PARTIAL_FLUSH = 1 << 9,
724 RADV_CMD_FLAG_PS_PARTIAL_FLUSH = 1 << 10,
725 RADV_CMD_FLAG_CS_PARTIAL_FLUSH = 1 << 11,
726 RADV_CMD_FLAG_VGT_FLUSH = 1 << 12,
727
728 RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER = (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
729 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
730 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
731 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META)
732 };
733
734 struct radv_vertex_binding {
735 struct radv_buffer * buffer;
736 VkDeviceSize offset;
737 };
738
739 struct radv_viewport_state {
740 uint32_t count;
741 VkViewport viewports[MAX_VIEWPORTS];
742 };
743
744 struct radv_scissor_state {
745 uint32_t count;
746 VkRect2D scissors[MAX_SCISSORS];
747 };
748
749 struct radv_dynamic_state {
750 /**
751 * Bitmask of (1 << VK_DYNAMIC_STATE_*).
752 * Defines the set of saved dynamic state.
753 */
754 uint32_t mask;
755
756 struct radv_viewport_state viewport;
757
758 struct radv_scissor_state scissor;
759
760 float line_width;
761
762 struct {
763 float bias;
764 float clamp;
765 float slope;
766 } depth_bias;
767
768 float blend_constants[4];
769
770 struct {
771 float min;
772 float max;
773 } depth_bounds;
774
775 struct {
776 uint32_t front;
777 uint32_t back;
778 } stencil_compare_mask;
779
780 struct {
781 uint32_t front;
782 uint32_t back;
783 } stencil_write_mask;
784
785 struct {
786 uint32_t front;
787 uint32_t back;
788 } stencil_reference;
789 };
790
791 extern const struct radv_dynamic_state default_dynamic_state;
792
793 const char *
794 radv_get_debug_option_name(int id);
795
796 const char *
797 radv_get_perftest_option_name(int id);
798
799 /**
800 * Attachment state when recording a renderpass instance.
801 *
802 * The clear value is valid only if there exists a pending clear.
803 */
804 struct radv_attachment_state {
805 VkImageAspectFlags pending_clear_aspects;
806 uint32_t cleared_views;
807 VkClearValue clear_value;
808 VkImageLayout current_layout;
809 };
810
811 struct radv_cmd_state {
812 bool vb_dirty;
813 bool push_descriptors_dirty;
814 bool predicating;
815 radv_cmd_dirty_mask_t dirty;
816
817 struct radv_pipeline * pipeline;
818 struct radv_pipeline * emitted_pipeline;
819 struct radv_pipeline * compute_pipeline;
820 struct radv_pipeline * emitted_compute_pipeline;
821 struct radv_framebuffer * framebuffer;
822 struct radv_render_pass * pass;
823 const struct radv_subpass * subpass;
824 struct radv_dynamic_state dynamic;
825 struct radv_attachment_state * attachments;
826 VkRect2D render_area;
827
828 /* Index buffer */
829 struct radv_buffer *index_buffer;
830 uint64_t index_offset;
831 uint32_t index_type;
832 uint32_t max_index_count;
833 uint64_t index_va;
834
835 int32_t last_primitive_reset_en;
836 uint32_t last_primitive_reset_index;
837 enum radv_cmd_flush_bits flush_bits;
838 unsigned active_occlusion_queries;
839 float offset_scale;
840 uint32_t descriptors_dirty;
841 uint32_t valid_descriptors;
842 uint32_t trace_id;
843 uint32_t last_ia_multi_vgt_param;
844 };
845
846 struct radv_cmd_pool {
847 VkAllocationCallbacks alloc;
848 struct list_head cmd_buffers;
849 struct list_head free_cmd_buffers;
850 uint32_t queue_family_index;
851 };
852
853 struct radv_cmd_buffer_upload {
854 uint8_t *map;
855 unsigned offset;
856 uint64_t size;
857 struct radeon_winsys_bo *upload_bo;
858 struct list_head list;
859 };
860
861 struct radv_cmd_buffer {
862 VK_LOADER_DATA _loader_data;
863
864 struct radv_device * device;
865
866 struct radv_cmd_pool * pool;
867 struct list_head pool_link;
868
869 VkCommandBufferUsageFlags usage_flags;
870 VkCommandBufferLevel level;
871 struct radeon_winsys_cs *cs;
872 struct radv_cmd_state state;
873 struct radv_vertex_binding vertex_bindings[MAX_VBS];
874 uint32_t queue_family_index;
875
876 uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
877 uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
878 VkShaderStageFlags push_constant_stages;
879 struct radv_push_descriptor_set push_descriptors;
880 struct radv_descriptor_set meta_push_descriptors;
881 struct radv_descriptor_set *descriptors[MAX_SETS];
882
883 struct radv_cmd_buffer_upload upload;
884
885 uint32_t scratch_size_needed;
886 uint32_t compute_scratch_size_needed;
887 uint32_t esgs_ring_size_needed;
888 uint32_t gsvs_ring_size_needed;
889 bool tess_rings_needed;
890 bool sample_positions_needed;
891
892 VkResult record_result;
893
894 int ring_offsets_idx; /* just used for verification */
895 uint32_t gfx9_fence_offset;
896 struct radeon_winsys_bo *gfx9_fence_bo;
897 uint32_t gfx9_fence_idx;
898 };
899
900 struct radv_image;
901
902 bool radv_cmd_buffer_uses_mec(struct radv_cmd_buffer *cmd_buffer);
903
904 void si_init_compute(struct radv_cmd_buffer *cmd_buffer);
905 void si_init_config(struct radv_cmd_buffer *cmd_buffer);
906
907 void cik_create_gfx_config(struct radv_device *device);
908
909 void si_write_viewport(struct radeon_winsys_cs *cs, int first_vp,
910 int count, const VkViewport *viewports);
911 void si_write_scissors(struct radeon_winsys_cs *cs, int first,
912 int count, const VkRect2D *scissors,
913 const VkViewport *viewports, bool can_use_guardband);
914 uint32_t si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer,
915 bool instanced_draw, bool indirect_draw,
916 uint32_t draw_vertex_count);
917 void si_cs_emit_write_event_eop(struct radeon_winsys_cs *cs,
918 bool predicated,
919 enum chip_class chip_class,
920 bool is_mec,
921 unsigned event, unsigned event_flags,
922 unsigned data_sel,
923 uint64_t va,
924 uint32_t old_fence,
925 uint32_t new_fence);
926
927 void si_emit_wait_fence(struct radeon_winsys_cs *cs,
928 bool predicated,
929 uint64_t va, uint32_t ref,
930 uint32_t mask);
931 void si_cs_emit_cache_flush(struct radeon_winsys_cs *cs,
932 bool predicated,
933 enum chip_class chip_class,
934 uint32_t *fence_ptr, uint64_t va,
935 bool is_mec,
936 enum radv_cmd_flush_bits flush_bits);
937 void si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer);
938 void si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer, uint64_t va);
939 void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
940 uint64_t src_va, uint64_t dest_va,
941 uint64_t size);
942 void si_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
943 unsigned size);
944 void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
945 uint64_t size, unsigned value);
946 void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer);
947 bool
948 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
949 unsigned size,
950 unsigned alignment,
951 unsigned *out_offset,
952 void **ptr);
953 void
954 radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
955 const struct radv_subpass *subpass,
956 bool transitions);
957 bool
958 radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
959 unsigned size, unsigned alignmnet,
960 const void *data, unsigned *out_offset);
961 void
962 radv_emit_framebuffer_state(struct radv_cmd_buffer *cmd_buffer);
963 void radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer);
964 void radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer);
965 void radv_cmd_buffer_resolve_subpass_cs(struct radv_cmd_buffer *cmd_buffer);
966 void radv_cmd_buffer_resolve_subpass_fs(struct radv_cmd_buffer *cmd_buffer);
967 void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs *cs, int nr_samples);
968 unsigned radv_cayman_get_maxdist(int log_samples);
969 void radv_device_init_msaa(struct radv_device *device);
970 void radv_set_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
971 struct radv_image *image,
972 VkClearDepthStencilValue ds_clear_value,
973 VkImageAspectFlags aspects);
974 void radv_set_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
975 struct radv_image *image,
976 int idx,
977 uint32_t color_values[2]);
978 void radv_set_dcc_need_cmask_elim_pred(struct radv_cmd_buffer *cmd_buffer,
979 struct radv_image *image,
980 bool value);
981 uint32_t radv_fill_buffer(struct radv_cmd_buffer *cmd_buffer,
982 struct radeon_winsys_bo *bo,
983 uint64_t offset, uint64_t size, uint32_t value);
984 void radv_cmd_buffer_trace_emit(struct radv_cmd_buffer *cmd_buffer);
985 bool radv_get_memory_fd(struct radv_device *device,
986 struct radv_device_memory *memory,
987 int *pFD);
988 VkResult radv_alloc_memory(VkDevice _device,
989 const VkMemoryAllocateInfo* pAllocateInfo,
990 const VkAllocationCallbacks* pAllocator,
991 enum radv_mem_flags_bits flags,
992 VkDeviceMemory* pMem);
993
994 /*
995 * Takes x,y,z as exact numbers of invocations, instead of blocks.
996 *
997 * Limitations: Can't call normal dispatch functions without binding or rebinding
998 * the compute pipeline.
999 */
1000 void radv_unaligned_dispatch(
1001 struct radv_cmd_buffer *cmd_buffer,
1002 uint32_t x,
1003 uint32_t y,
1004 uint32_t z);
1005
1006 struct radv_event {
1007 struct radeon_winsys_bo *bo;
1008 uint64_t *map;
1009 };
1010
1011 struct radv_shader_module;
1012
1013 #define RADV_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
1014 #define RADV_HASH_SHADER_SISCHED (1 << 1)
1015 #define RADV_HASH_SHADER_UNSAFE_MATH (1 << 2)
1016 void
1017 radv_hash_shaders(unsigned char *hash,
1018 const VkPipelineShaderStageCreateInfo **stages,
1019 const struct radv_pipeline_layout *layout,
1020 const struct radv_pipeline_key *key,
1021 uint32_t flags);
1022
1023 static inline gl_shader_stage
1024 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
1025 {
1026 assert(__builtin_popcount(vk_stage) == 1);
1027 return ffs(vk_stage) - 1;
1028 }
1029
1030 static inline VkShaderStageFlagBits
1031 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
1032 {
1033 return (1 << mesa_stage);
1034 }
1035
1036 #define RADV_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
1037
1038 #define radv_foreach_stage(stage, stage_bits) \
1039 for (gl_shader_stage stage, \
1040 __tmp = (gl_shader_stage)((stage_bits) & RADV_STAGE_MASK); \
1041 stage = __builtin_ffs(__tmp) - 1, __tmp; \
1042 __tmp &= ~(1 << (stage)))
1043
1044 struct radv_depth_stencil_state {
1045 uint32_t db_depth_control;
1046 uint32_t db_stencil_control;
1047 uint32_t db_render_control;
1048 uint32_t db_render_override2;
1049 };
1050
1051 struct radv_blend_state {
1052 uint32_t cb_color_control;
1053 uint32_t cb_target_mask;
1054 uint32_t sx_mrt_blend_opt[8];
1055 uint32_t cb_blend_control[8];
1056
1057 uint32_t spi_shader_col_format;
1058 uint32_t cb_shader_mask;
1059 uint32_t db_alpha_to_mask;
1060 };
1061
1062 unsigned radv_format_meta_fs_key(VkFormat format);
1063
1064 struct radv_raster_state {
1065 uint32_t pa_cl_clip_cntl;
1066 uint32_t spi_interp_control;
1067 uint32_t pa_su_vtx_cntl;
1068 uint32_t pa_su_sc_mode_cntl;
1069 };
1070
1071 struct radv_multisample_state {
1072 uint32_t db_eqaa;
1073 uint32_t pa_sc_line_cntl;
1074 uint32_t pa_sc_mode_cntl_0;
1075 uint32_t pa_sc_mode_cntl_1;
1076 uint32_t pa_sc_aa_config;
1077 uint32_t pa_sc_aa_mask[2];
1078 unsigned num_samples;
1079 };
1080
1081 struct radv_prim_vertex_count {
1082 uint8_t min;
1083 uint8_t incr;
1084 };
1085
1086 struct radv_tessellation_state {
1087 uint32_t ls_hs_config;
1088 uint32_t tcs_in_layout;
1089 uint32_t tcs_out_layout;
1090 uint32_t tcs_out_offsets;
1091 uint32_t offchip_layout;
1092 unsigned num_patches;
1093 unsigned lds_size;
1094 unsigned num_tcs_input_cp;
1095 uint32_t tf_param;
1096 };
1097
1098 struct radv_gs_state {
1099 uint32_t vgt_gs_onchip_cntl;
1100 uint32_t vgt_gs_max_prims_per_subgroup;
1101 uint32_t vgt_esgs_ring_itemsize;
1102 uint32_t lds_size;
1103 };
1104
1105 struct radv_vertex_elements_info {
1106 uint32_t rsrc_word3[MAX_VERTEX_ATTRIBS];
1107 uint32_t format_size[MAX_VERTEX_ATTRIBS];
1108 uint32_t binding[MAX_VERTEX_ATTRIBS];
1109 uint32_t offset[MAX_VERTEX_ATTRIBS];
1110 uint32_t count;
1111 };
1112
1113 #define SI_GS_PER_ES 128
1114
1115 struct radv_pipeline {
1116 struct radv_device * device;
1117 struct radv_dynamic_state dynamic_state;
1118
1119 struct radv_pipeline_layout * layout;
1120
1121 bool needs_data_cache;
1122 bool need_indirect_descriptor_sets;
1123 struct radv_shader_variant * shaders[MESA_SHADER_STAGES];
1124 struct radv_shader_variant *gs_copy_shader;
1125 VkShaderStageFlags active_stages;
1126
1127 struct radv_vertex_elements_info vertex_elements;
1128
1129 uint32_t binding_stride[MAX_VBS];
1130
1131 uint32_t user_data_0[MESA_SHADER_STAGES];
1132 union {
1133 struct {
1134 struct radv_blend_state blend;
1135 struct radv_depth_stencil_state ds;
1136 struct radv_raster_state raster;
1137 struct radv_multisample_state ms;
1138 struct radv_tessellation_state tess;
1139 struct radv_gs_state gs;
1140 uint32_t db_shader_control;
1141 uint32_t shader_z_format;
1142 unsigned prim;
1143 unsigned gs_out;
1144 uint32_t vgt_gs_mode;
1145 bool vgt_primitiveid_en;
1146 bool prim_restart_enable;
1147 bool partial_es_wave;
1148 uint8_t primgroup_size;
1149 unsigned esgs_ring_size;
1150 unsigned gsvs_ring_size;
1151 uint32_t ps_input_cntl[32];
1152 uint32_t ps_input_cntl_num;
1153 uint32_t pa_cl_vs_out_cntl;
1154 uint32_t vgt_shader_stages_en;
1155 uint32_t vtx_base_sgpr;
1156 uint32_t base_ia_multi_vgt_param;
1157 bool wd_switch_on_eop;
1158 bool ia_switch_on_eoi;
1159 bool partial_vs_wave;
1160 uint8_t vtx_emit_num;
1161 uint32_t vtx_reuse_depth;
1162 struct radv_prim_vertex_count prim_vertex_count;
1163 bool can_use_guardband;
1164 } graphics;
1165 };
1166
1167 unsigned max_waves;
1168 unsigned scratch_bytes_per_wave;
1169 };
1170
1171 static inline bool radv_pipeline_has_gs(struct radv_pipeline *pipeline)
1172 {
1173 return pipeline->shaders[MESA_SHADER_GEOMETRY] ? true : false;
1174 }
1175
1176 static inline bool radv_pipeline_has_tess(struct radv_pipeline *pipeline)
1177 {
1178 return pipeline->shaders[MESA_SHADER_TESS_CTRL] ? true : false;
1179 }
1180
1181 struct ac_userdata_info *radv_lookup_user_sgpr(struct radv_pipeline *pipeline,
1182 gl_shader_stage stage,
1183 int idx);
1184
1185 struct radv_shader_variant *radv_get_vertex_shader(struct radv_pipeline *pipeline);
1186
1187 struct radv_graphics_pipeline_create_info {
1188 bool use_rectlist;
1189 bool db_depth_clear;
1190 bool db_stencil_clear;
1191 bool db_depth_disable_expclear;
1192 bool db_stencil_disable_expclear;
1193 bool db_flush_depth_inplace;
1194 bool db_flush_stencil_inplace;
1195 bool db_resummarize;
1196 uint32_t custom_blend_mode;
1197 };
1198
1199 VkResult
1200 radv_graphics_pipeline_create(VkDevice device,
1201 VkPipelineCache cache,
1202 const VkGraphicsPipelineCreateInfo *pCreateInfo,
1203 const struct radv_graphics_pipeline_create_info *extra,
1204 const VkAllocationCallbacks *alloc,
1205 VkPipeline *pPipeline);
1206
1207 struct vk_format_description;
1208 uint32_t radv_translate_buffer_dataformat(const struct vk_format_description *desc,
1209 int first_non_void);
1210 uint32_t radv_translate_buffer_numformat(const struct vk_format_description *desc,
1211 int first_non_void);
1212 uint32_t radv_translate_colorformat(VkFormat format);
1213 uint32_t radv_translate_color_numformat(VkFormat format,
1214 const struct vk_format_description *desc,
1215 int first_non_void);
1216 uint32_t radv_colorformat_endian_swap(uint32_t colorformat);
1217 unsigned radv_translate_colorswap(VkFormat format, bool do_endian_swap);
1218 uint32_t radv_translate_dbformat(VkFormat format);
1219 uint32_t radv_translate_tex_dataformat(VkFormat format,
1220 const struct vk_format_description *desc,
1221 int first_non_void);
1222 uint32_t radv_translate_tex_numformat(VkFormat format,
1223 const struct vk_format_description *desc,
1224 int first_non_void);
1225 bool radv_format_pack_clear_color(VkFormat format,
1226 uint32_t clear_vals[2],
1227 VkClearColorValue *value);
1228 bool radv_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
1229 bool radv_dcc_formats_compatible(VkFormat format1,
1230 VkFormat format2);
1231
1232 struct radv_fmask_info {
1233 uint64_t offset;
1234 uint64_t size;
1235 unsigned alignment;
1236 unsigned pitch_in_pixels;
1237 unsigned bank_height;
1238 unsigned slice_tile_max;
1239 unsigned tile_mode_index;
1240 unsigned tile_swizzle;
1241 };
1242
1243 struct radv_cmask_info {
1244 uint64_t offset;
1245 uint64_t size;
1246 unsigned alignment;
1247 unsigned slice_tile_max;
1248 unsigned base_address_reg;
1249 };
1250
1251 struct radv_image {
1252 VkImageType type;
1253 /* The original VkFormat provided by the client. This may not match any
1254 * of the actual surface formats.
1255 */
1256 VkFormat vk_format;
1257 VkImageAspectFlags aspects;
1258 VkImageUsageFlags usage; /**< Superset of VkImageCreateInfo::usage. */
1259 struct ac_surf_info info;
1260 VkImageTiling tiling; /** VkImageCreateInfo::tiling */
1261 VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
1262
1263 VkDeviceSize size;
1264 uint32_t alignment;
1265
1266 unsigned queue_family_mask;
1267 bool exclusive;
1268 bool shareable;
1269
1270 /* Set when bound */
1271 struct radeon_winsys_bo *bo;
1272 VkDeviceSize offset;
1273 uint64_t dcc_offset;
1274 uint64_t htile_offset;
1275 bool tc_compatible_htile;
1276 struct radeon_surf surface;
1277
1278 struct radv_fmask_info fmask;
1279 struct radv_cmask_info cmask;
1280 uint64_t clear_value_offset;
1281 uint64_t dcc_pred_offset;
1282 };
1283
1284 /* Whether the image has a htile that is known consistent with the contents of
1285 * the image. */
1286 bool radv_layout_has_htile(const struct radv_image *image,
1287 VkImageLayout layout,
1288 unsigned queue_mask);
1289
1290 /* Whether the image has a htile that is known consistent with the contents of
1291 * the image and is allowed to be in compressed form.
1292 *
1293 * If this is false reads that don't use the htile should be able to return
1294 * correct results.
1295 */
1296 bool radv_layout_is_htile_compressed(const struct radv_image *image,
1297 VkImageLayout layout,
1298 unsigned queue_mask);
1299
1300 bool radv_layout_can_fast_clear(const struct radv_image *image,
1301 VkImageLayout layout,
1302 unsigned queue_mask);
1303
1304 static inline bool
1305 radv_vi_dcc_enabled(const struct radv_image *image, unsigned level)
1306 {
1307 return image->surface.dcc_size && level < image->surface.num_dcc_levels;
1308 }
1309
1310 static inline bool
1311 radv_htile_enabled(const struct radv_image *image, unsigned level)
1312 {
1313 return image->surface.htile_size && level == 0;
1314 }
1315
1316 unsigned radv_image_queue_family_mask(const struct radv_image *image, uint32_t family, uint32_t queue_family);
1317
1318 static inline uint32_t
1319 radv_get_layerCount(const struct radv_image *image,
1320 const VkImageSubresourceRange *range)
1321 {
1322 return range->layerCount == VK_REMAINING_ARRAY_LAYERS ?
1323 image->info.array_size - range->baseArrayLayer : range->layerCount;
1324 }
1325
1326 static inline uint32_t
1327 radv_get_levelCount(const struct radv_image *image,
1328 const VkImageSubresourceRange *range)
1329 {
1330 return range->levelCount == VK_REMAINING_MIP_LEVELS ?
1331 image->info.levels - range->baseMipLevel : range->levelCount;
1332 }
1333
1334 struct radeon_bo_metadata;
1335 void
1336 radv_init_metadata(struct radv_device *device,
1337 struct radv_image *image,
1338 struct radeon_bo_metadata *metadata);
1339
1340 struct radv_image_view {
1341 struct radv_image *image; /**< VkImageViewCreateInfo::image */
1342 struct radeon_winsys_bo *bo;
1343
1344 VkImageViewType type;
1345 VkImageAspectFlags aspect_mask;
1346 VkFormat vk_format;
1347 uint32_t base_layer;
1348 uint32_t layer_count;
1349 uint32_t base_mip;
1350 uint32_t level_count;
1351 VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
1352
1353 uint32_t descriptor[8];
1354 uint32_t fmask_descriptor[8];
1355
1356 /* Descriptor for use as a storage image as opposed to a sampled image.
1357 * This has a few differences for cube maps (e.g. type).
1358 */
1359 uint32_t storage_descriptor[8];
1360 uint32_t storage_fmask_descriptor[8];
1361 };
1362
1363 struct radv_image_create_info {
1364 const VkImageCreateInfo *vk_info;
1365 bool scanout;
1366 };
1367
1368 VkResult radv_image_create(VkDevice _device,
1369 const struct radv_image_create_info *info,
1370 const VkAllocationCallbacks* alloc,
1371 VkImage *pImage);
1372
1373 void radv_image_view_init(struct radv_image_view *view,
1374 struct radv_device *device,
1375 const VkImageViewCreateInfo* pCreateInfo);
1376
1377 struct radv_buffer_view {
1378 struct radeon_winsys_bo *bo;
1379 VkFormat vk_format;
1380 uint64_t range; /**< VkBufferViewCreateInfo::range */
1381 uint32_t state[4];
1382 };
1383 void radv_buffer_view_init(struct radv_buffer_view *view,
1384 struct radv_device *device,
1385 const VkBufferViewCreateInfo* pCreateInfo);
1386
1387 static inline struct VkExtent3D
1388 radv_sanitize_image_extent(const VkImageType imageType,
1389 const struct VkExtent3D imageExtent)
1390 {
1391 switch (imageType) {
1392 case VK_IMAGE_TYPE_1D:
1393 return (VkExtent3D) { imageExtent.width, 1, 1 };
1394 case VK_IMAGE_TYPE_2D:
1395 return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1396 case VK_IMAGE_TYPE_3D:
1397 return imageExtent;
1398 default:
1399 unreachable("invalid image type");
1400 }
1401 }
1402
1403 static inline struct VkOffset3D
1404 radv_sanitize_image_offset(const VkImageType imageType,
1405 const struct VkOffset3D imageOffset)
1406 {
1407 switch (imageType) {
1408 case VK_IMAGE_TYPE_1D:
1409 return (VkOffset3D) { imageOffset.x, 0, 0 };
1410 case VK_IMAGE_TYPE_2D:
1411 return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1412 case VK_IMAGE_TYPE_3D:
1413 return imageOffset;
1414 default:
1415 unreachable("invalid image type");
1416 }
1417 }
1418
1419 static inline bool
1420 radv_image_extent_compare(const struct radv_image *image,
1421 const VkExtent3D *extent)
1422 {
1423 if (extent->width != image->info.width ||
1424 extent->height != image->info.height ||
1425 extent->depth != image->info.depth)
1426 return false;
1427 return true;
1428 }
1429
1430 struct radv_sampler {
1431 uint32_t state[4];
1432 };
1433
1434 struct radv_color_buffer_info {
1435 uint64_t cb_color_base;
1436 uint64_t cb_color_cmask;
1437 uint64_t cb_color_fmask;
1438 uint64_t cb_dcc_base;
1439 uint32_t cb_color_pitch;
1440 uint32_t cb_color_slice;
1441 uint32_t cb_color_view;
1442 uint32_t cb_color_info;
1443 uint32_t cb_color_attrib;
1444 uint32_t cb_color_attrib2;
1445 uint32_t cb_dcc_control;
1446 uint32_t cb_color_cmask_slice;
1447 uint32_t cb_color_fmask_slice;
1448 uint32_t cb_clear_value0;
1449 uint32_t cb_clear_value1;
1450 uint32_t micro_tile_mode;
1451 uint32_t gfx9_epitch;
1452 };
1453
1454 struct radv_ds_buffer_info {
1455 uint64_t db_z_read_base;
1456 uint64_t db_stencil_read_base;
1457 uint64_t db_z_write_base;
1458 uint64_t db_stencil_write_base;
1459 uint64_t db_htile_data_base;
1460 uint32_t db_depth_info;
1461 uint32_t db_z_info;
1462 uint32_t db_stencil_info;
1463 uint32_t db_depth_view;
1464 uint32_t db_depth_size;
1465 uint32_t db_depth_slice;
1466 uint32_t db_htile_surface;
1467 uint32_t pa_su_poly_offset_db_fmt_cntl;
1468 uint32_t db_z_info2;
1469 uint32_t db_stencil_info2;
1470 float offset_scale;
1471 };
1472
1473 struct radv_attachment_info {
1474 union {
1475 struct radv_color_buffer_info cb;
1476 struct radv_ds_buffer_info ds;
1477 };
1478 struct radv_image_view *attachment;
1479 };
1480
1481 struct radv_framebuffer {
1482 uint32_t width;
1483 uint32_t height;
1484 uint32_t layers;
1485
1486 uint32_t attachment_count;
1487 struct radv_attachment_info attachments[0];
1488 };
1489
1490 struct radv_subpass_barrier {
1491 VkPipelineStageFlags src_stage_mask;
1492 VkAccessFlags src_access_mask;
1493 VkAccessFlags dst_access_mask;
1494 };
1495
1496 struct radv_subpass {
1497 uint32_t input_count;
1498 uint32_t color_count;
1499 VkAttachmentReference * input_attachments;
1500 VkAttachmentReference * color_attachments;
1501 VkAttachmentReference * resolve_attachments;
1502 VkAttachmentReference depth_stencil_attachment;
1503
1504 /** Subpass has at least one resolve attachment */
1505 bool has_resolve;
1506
1507 struct radv_subpass_barrier start_barrier;
1508
1509 uint32_t view_mask;
1510 };
1511
1512 struct radv_render_pass_attachment {
1513 VkFormat format;
1514 uint32_t samples;
1515 VkAttachmentLoadOp load_op;
1516 VkAttachmentLoadOp stencil_load_op;
1517 VkImageLayout initial_layout;
1518 VkImageLayout final_layout;
1519 uint32_t view_mask;
1520 };
1521
1522 struct radv_render_pass {
1523 uint32_t attachment_count;
1524 uint32_t subpass_count;
1525 VkAttachmentReference * subpass_attachments;
1526 struct radv_render_pass_attachment * attachments;
1527 struct radv_subpass_barrier end_barrier;
1528 struct radv_subpass subpasses[0];
1529 };
1530
1531 VkResult radv_device_init_meta(struct radv_device *device);
1532 void radv_device_finish_meta(struct radv_device *device);
1533
1534 struct radv_query_pool {
1535 struct radeon_winsys_bo *bo;
1536 uint32_t stride;
1537 uint32_t availability_offset;
1538 char *ptr;
1539 VkQueryType type;
1540 uint32_t pipeline_stats_mask;
1541 };
1542
1543 struct radv_semaphore {
1544 /* use a winsys sem for non-exportable */
1545 struct radeon_winsys_sem *sem;
1546 uint32_t syncobj;
1547 uint32_t temp_syncobj;
1548 };
1549
1550 VkResult radv_alloc_sem_info(struct radv_winsys_sem_info *sem_info,
1551 int num_wait_sems,
1552 const VkSemaphore *wait_sems,
1553 int num_signal_sems,
1554 const VkSemaphore *signal_sems);
1555 void radv_free_sem_info(struct radv_winsys_sem_info *sem_info);
1556
1557 void radv_set_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
1558 struct radv_descriptor_set *set,
1559 unsigned idx);
1560
1561 void
1562 radv_update_descriptor_sets(struct radv_device *device,
1563 struct radv_cmd_buffer *cmd_buffer,
1564 VkDescriptorSet overrideSet,
1565 uint32_t descriptorWriteCount,
1566 const VkWriteDescriptorSet *pDescriptorWrites,
1567 uint32_t descriptorCopyCount,
1568 const VkCopyDescriptorSet *pDescriptorCopies);
1569
1570 void
1571 radv_update_descriptor_set_with_template(struct radv_device *device,
1572 struct radv_cmd_buffer *cmd_buffer,
1573 struct radv_descriptor_set *set,
1574 VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1575 const void *pData);
1576
1577 void radv_meta_push_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
1578 VkPipelineBindPoint pipelineBindPoint,
1579 VkPipelineLayout _layout,
1580 uint32_t set,
1581 uint32_t descriptorWriteCount,
1582 const VkWriteDescriptorSet *pDescriptorWrites);
1583
1584 void radv_initialise_cmask(struct radv_cmd_buffer *cmd_buffer,
1585 struct radv_image *image, uint32_t value);
1586 void radv_initialize_dcc(struct radv_cmd_buffer *cmd_buffer,
1587 struct radv_image *image, uint32_t value);
1588
1589 struct radv_fence {
1590 struct radeon_winsys_fence *fence;
1591 bool submitted;
1592 bool signalled;
1593 };
1594
1595 struct radeon_winsys_sem;
1596
1597 #define RADV_DEFINE_HANDLE_CASTS(__radv_type, __VkType) \
1598 \
1599 static inline struct __radv_type * \
1600 __radv_type ## _from_handle(__VkType _handle) \
1601 { \
1602 return (struct __radv_type *) _handle; \
1603 } \
1604 \
1605 static inline __VkType \
1606 __radv_type ## _to_handle(struct __radv_type *_obj) \
1607 { \
1608 return (__VkType) _obj; \
1609 }
1610
1611 #define RADV_DEFINE_NONDISP_HANDLE_CASTS(__radv_type, __VkType) \
1612 \
1613 static inline struct __radv_type * \
1614 __radv_type ## _from_handle(__VkType _handle) \
1615 { \
1616 return (struct __radv_type *)(uintptr_t) _handle; \
1617 } \
1618 \
1619 static inline __VkType \
1620 __radv_type ## _to_handle(struct __radv_type *_obj) \
1621 { \
1622 return (__VkType)(uintptr_t) _obj; \
1623 }
1624
1625 #define RADV_FROM_HANDLE(__radv_type, __name, __handle) \
1626 struct __radv_type *__name = __radv_type ## _from_handle(__handle)
1627
1628 RADV_DEFINE_HANDLE_CASTS(radv_cmd_buffer, VkCommandBuffer)
1629 RADV_DEFINE_HANDLE_CASTS(radv_device, VkDevice)
1630 RADV_DEFINE_HANDLE_CASTS(radv_instance, VkInstance)
1631 RADV_DEFINE_HANDLE_CASTS(radv_physical_device, VkPhysicalDevice)
1632 RADV_DEFINE_HANDLE_CASTS(radv_queue, VkQueue)
1633
1634 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_cmd_pool, VkCommandPool)
1635 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer, VkBuffer)
1636 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer_view, VkBufferView)
1637 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_pool, VkDescriptorPool)
1638 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set, VkDescriptorSet)
1639 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set_layout, VkDescriptorSetLayout)
1640 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_update_template, VkDescriptorUpdateTemplateKHR)
1641 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_device_memory, VkDeviceMemory)
1642 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_fence, VkFence)
1643 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_event, VkEvent)
1644 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_framebuffer, VkFramebuffer)
1645 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image, VkImage)
1646 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image_view, VkImageView);
1647 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_cache, VkPipelineCache)
1648 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline, VkPipeline)
1649 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_layout, VkPipelineLayout)
1650 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_query_pool, VkQueryPool)
1651 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_render_pass, VkRenderPass)
1652 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_sampler, VkSampler)
1653 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_shader_module, VkShaderModule)
1654 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_semaphore, VkSemaphore)
1655
1656 #endif /* RADV_PRIVATE_H */