projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac/nir: move ac_shader_variant_info and friends to radv folder
[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 #include "vk_debug_report.h"
53
54 #include "radv_radeon_winsys.h"
55 #include "ac_binary.h"
56 #include "ac_nir_to_llvm.h"
57 #include "ac_gpu_info.h"
58 #include "ac_surface.h"
59 #include "radv_descriptor_set.h"
60 #include "radv_extensions.h"
61
62 #include <llvm-c/TargetMachine.h>
63
64 /* Pre-declarations needed for WSI entrypoints */
65 struct wl_surface;
66 struct wl_display;
67 typedef struct xcb_connection_t xcb_connection_t;
68 typedef uint32_t xcb_visualid_t;
69 typedef uint32_t xcb_window_t;
70
71 #include <vulkan/vulkan.h>
72 #include <vulkan/vulkan_intel.h>
73 #include <vulkan/vk_icd.h>
74 #include <vulkan/vk_android_native_buffer.h>
75
76 #include "radv_entrypoints.h"
77
78 #include "wsi_common.h"
79
80 #define ATI_VENDOR_ID 0x1002
81
82 #define MAX_VBS 32
83 #define MAX_VERTEX_ATTRIBS 32
84 #define MAX_RTS 8
85 #define MAX_VIEWPORTS 16
86 #define MAX_SCISSORS 16
87 #define MAX_DISCARD_RECTANGLES 4
88 #define MAX_PUSH_CONSTANTS_SIZE 128
89 #define MAX_PUSH_DESCRIPTORS 32
90 #define MAX_DYNAMIC_UNIFORM_BUFFERS 16
91 #define MAX_DYNAMIC_STORAGE_BUFFERS 8
92 #define MAX_DYNAMIC_BUFFERS (MAX_DYNAMIC_UNIFORM_BUFFERS + MAX_DYNAMIC_STORAGE_BUFFERS)
93 #define MAX_SAMPLES_LOG2 4
94 #define NUM_META_FS_KEYS 13
95 #define RADV_MAX_DRM_DEVICES 8
96 #define MAX_VIEWS 8
97
98 #define NUM_DEPTH_CLEAR_PIPELINES 3
99
100 /*
101 * This is the point we switch from using CP to compute shader
102 * for certain buffer operations.
103 */
104 #define RADV_BUFFER_OPS_CS_THRESHOLD 4096
105
106 enum radv_mem_heap {
107 RADV_MEM_HEAP_VRAM,
108 RADV_MEM_HEAP_VRAM_CPU_ACCESS,
109 RADV_MEM_HEAP_GTT,
110 RADV_MEM_HEAP_COUNT
111 };
112
113 enum radv_mem_type {
114 RADV_MEM_TYPE_VRAM,
115 RADV_MEM_TYPE_GTT_WRITE_COMBINE,
116 RADV_MEM_TYPE_VRAM_CPU_ACCESS,
117 RADV_MEM_TYPE_GTT_CACHED,
118 RADV_MEM_TYPE_COUNT
119 };
120
121 #define radv_printflike(a, b) __attribute__((__format__(__printf__, a, b)))
122
123 static inline uint32_t
124 align_u32(uint32_t v, uint32_t a)
125 {
126 assert(a != 0 && a == (a & -a));
127 return (v + a - 1) & ~(a - 1);
128 }
129
130 static inline uint32_t
131 align_u32_npot(uint32_t v, uint32_t a)
132 {
133 return (v + a - 1) / a * a;
134 }
135
136 static inline uint64_t
137 align_u64(uint64_t v, uint64_t a)
138 {
139 assert(a != 0 && a == (a & -a));
140 return (v + a - 1) & ~(a - 1);
141 }
142
143 static inline int32_t
144 align_i32(int32_t v, int32_t a)
145 {
146 assert(a != 0 && a == (a & -a));
147 return (v + a - 1) & ~(a - 1);
148 }
149
150 /** Alignment must be a power of 2. */
151 static inline bool
152 radv_is_aligned(uintmax_t n, uintmax_t a)
153 {
154 assert(a == (a & -a));
155 return (n & (a - 1)) == 0;
156 }
157
158 static inline uint32_t
159 round_up_u32(uint32_t v, uint32_t a)
160 {
161 return (v + a - 1) / a;
162 }
163
164 static inline uint64_t
165 round_up_u64(uint64_t v, uint64_t a)
166 {
167 return (v + a - 1) / a;
168 }
169
170 static inline uint32_t
171 radv_minify(uint32_t n, uint32_t levels)
172 {
173 if (unlikely(n == 0))
174 return 0;
175 else
176 return MAX2(n >> levels, 1);
177 }
178 static inline float
179 radv_clamp_f(float f, float min, float max)
180 {
181 assert(min < max);
182
183 if (f > max)
184 return max;
185 else if (f < min)
186 return min;
187 else
188 return f;
189 }
190
191 static inline bool
192 radv_clear_mask(uint32_t *inout_mask, uint32_t clear_mask)
193 {
194 if (*inout_mask & clear_mask) {
195 *inout_mask &= ~clear_mask;
196 return true;
197 } else {
198 return false;
199 }
200 }
201
202 #define for_each_bit(b, dword) \
203 for (uint32_t __dword = (dword); \
204 (b) = __builtin_ffs(__dword) - 1, __dword; \
205 __dword &= ~(1 << (b)))
206
207 #define typed_memcpy(dest, src, count) ({ \
208 STATIC_ASSERT(sizeof(*src) == sizeof(*dest)); \
209 memcpy((dest), (src), (count) * sizeof(*(src))); \
210 })
211
212 /* Whenever we generate an error, pass it through this function. Useful for
213 * debugging, where we can break on it. Only call at error site, not when
214 * propagating errors. Might be useful to plug in a stack trace here.
215 */
216
217 VkResult __vk_errorf(VkResult error, const char *file, int line, const char *format, ...);
218
219 #ifdef DEBUG
220 #define vk_error(error) __vk_errorf(error, __FILE__, __LINE__, NULL);
221 #define vk_errorf(error, format, ...) __vk_errorf(error, __FILE__, __LINE__, format, ## __VA_ARGS__);
222 #else
223 #define vk_error(error) error
224 #define vk_errorf(error, format, ...) error
225 #endif
226
227 void __radv_finishme(const char *file, int line, const char *format, ...)
228 radv_printflike(3, 4);
229 void radv_loge(const char *format, ...) radv_printflike(1, 2);
230 void radv_loge_v(const char *format, va_list va);
231
232 /**
233 * Print a FINISHME message, including its source location.
234 */
235 #define radv_finishme(format, ...) \
236 do { \
237 static bool reported = false; \
238 if (!reported) { \
239 __radv_finishme(__FILE__, __LINE__, format, ##__VA_ARGS__); \
240 reported = true; \
241 } \
242 } while (0)
243
244 /* A non-fatal assert. Useful for debugging. */
245 #ifdef DEBUG
246 #define radv_assert(x) ({ \
247 if (unlikely(!(x))) \
248 fprintf(stderr, "%s:%d ASSERT: %s\n", __FILE__, __LINE__, #x); \
249 })
250 #else
251 #define radv_assert(x)
252 #endif
253
254 #define stub_return(v) \
255 do { \
256 radv_finishme("stub %s", __func__); \
257 return (v); \
258 } while (0)
259
260 #define stub() \
261 do { \
262 radv_finishme("stub %s", __func__); \
263 return; \
264 } while (0)
265
266 void *radv_lookup_entrypoint_unchecked(const char *name);
267 void *radv_lookup_entrypoint_checked(const char *name,
268 uint32_t core_version,
269 const struct radv_instance_extension_table *instance,
270 const struct radv_device_extension_table *device);
271
272 struct radv_physical_device {
273 VK_LOADER_DATA _loader_data;
274
275 struct radv_instance * instance;
276
277 struct radeon_winsys *ws;
278 struct radeon_info rad_info;
279 char path[20];
280 char name[VK_MAX_PHYSICAL_DEVICE_NAME_SIZE];
281 uint8_t driver_uuid[VK_UUID_SIZE];
282 uint8_t device_uuid[VK_UUID_SIZE];
283 uint8_t cache_uuid[VK_UUID_SIZE];
284
285 int local_fd;
286 struct wsi_device wsi_device;
287
288 bool has_rbplus; /* if RB+ register exist */
289 bool rbplus_allowed; /* if RB+ is allowed */
290 bool has_clear_state;
291 bool cpdma_prefetch_writes_memory;
292 bool has_scissor_bug;
293
294 /* This is the drivers on-disk cache used as a fallback as opposed to
295 * the pipeline cache defined by apps.
296 */
297 struct disk_cache * disk_cache;
298
299 VkPhysicalDeviceMemoryProperties memory_properties;
300 enum radv_mem_type mem_type_indices[RADV_MEM_TYPE_COUNT];
301
302 struct radv_device_extension_table supported_extensions;
303 };
304
305 struct radv_instance {
306 VK_LOADER_DATA _loader_data;
307
308 VkAllocationCallbacks alloc;
309
310 uint32_t apiVersion;
311 int physicalDeviceCount;
312 struct radv_physical_device physicalDevices[RADV_MAX_DRM_DEVICES];
313
314 uint64_t debug_flags;
315 uint64_t perftest_flags;
316
317 struct vk_debug_report_instance debug_report_callbacks;
318
319 struct radv_instance_extension_table enabled_extensions;
320 };
321
322 VkResult radv_init_wsi(struct radv_physical_device *physical_device);
323 void radv_finish_wsi(struct radv_physical_device *physical_device);
324
325 bool radv_instance_extension_supported(const char *name);
326 uint32_t radv_physical_device_api_version(struct radv_physical_device *dev);
327 bool radv_physical_device_extension_supported(struct radv_physical_device *dev,
328 const char *name);
329
330 struct cache_entry;
331
332 struct radv_pipeline_cache {
333 struct radv_device * device;
334 pthread_mutex_t mutex;
335
336 uint32_t total_size;
337 uint32_t table_size;
338 uint32_t kernel_count;
339 struct cache_entry ** hash_table;
340 bool modified;
341
342 VkAllocationCallbacks alloc;
343 };
344
345 struct radv_pipeline_key {
346 uint32_t instance_rate_inputs;
347 unsigned tess_input_vertices;
348 uint32_t col_format;
349 uint32_t is_int8;
350 uint32_t is_int10;
351 uint8_t log2_ps_iter_samples;
352 uint8_t log2_num_samples;
353 uint32_t multisample : 1;
354 uint32_t has_multiview_view_index : 1;
355 };
356
357 void
358 radv_pipeline_cache_init(struct radv_pipeline_cache *cache,
359 struct radv_device *device);
360 void
361 radv_pipeline_cache_finish(struct radv_pipeline_cache *cache);
362 void
363 radv_pipeline_cache_load(struct radv_pipeline_cache *cache,
364 const void *data, size_t size);
365
366 struct radv_shader_variant;
367
368 bool
369 radv_create_shader_variants_from_pipeline_cache(struct radv_device *device,
370 struct radv_pipeline_cache *cache,
371 const unsigned char *sha1,
372 struct radv_shader_variant **variants);
373
374 void
375 radv_pipeline_cache_insert_shaders(struct radv_device *device,
376 struct radv_pipeline_cache *cache,
377 const unsigned char *sha1,
378 struct radv_shader_variant **variants,
379 const void *const *codes,
380 const unsigned *code_sizes);
381
382 enum radv_blit_ds_layout {
383 RADV_BLIT_DS_LAYOUT_TILE_ENABLE,
384 RADV_BLIT_DS_LAYOUT_TILE_DISABLE,
385 RADV_BLIT_DS_LAYOUT_COUNT,
386 };
387
388 static inline enum radv_blit_ds_layout radv_meta_blit_ds_to_type(VkImageLayout layout)
389 {
390 return (layout == VK_IMAGE_LAYOUT_GENERAL) ? RADV_BLIT_DS_LAYOUT_TILE_DISABLE : RADV_BLIT_DS_LAYOUT_TILE_ENABLE;
391 }
392
393 static inline VkImageLayout radv_meta_blit_ds_to_layout(enum radv_blit_ds_layout ds_layout)
394 {
395 return ds_layout == RADV_BLIT_DS_LAYOUT_TILE_ENABLE ? VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL : VK_IMAGE_LAYOUT_GENERAL;
396 }
397
398 enum radv_meta_dst_layout {
399 RADV_META_DST_LAYOUT_GENERAL,
400 RADV_META_DST_LAYOUT_OPTIMAL,
401 RADV_META_DST_LAYOUT_COUNT,
402 };
403
404 static inline enum radv_meta_dst_layout radv_meta_dst_layout_from_layout(VkImageLayout layout)
405 {
406 return (layout == VK_IMAGE_LAYOUT_GENERAL) ? RADV_META_DST_LAYOUT_GENERAL : RADV_META_DST_LAYOUT_OPTIMAL;
407 }
408
409 static inline VkImageLayout radv_meta_dst_layout_to_layout(enum radv_meta_dst_layout layout)
410 {
411 return layout == RADV_META_DST_LAYOUT_OPTIMAL ? VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL : VK_IMAGE_LAYOUT_GENERAL;
412 }
413
414 struct radv_meta_state {
415 VkAllocationCallbacks alloc;
416
417 struct radv_pipeline_cache cache;
418
419 /**
420 * Use array element `i` for images with `2^i` samples.
421 */
422 struct {
423 VkRenderPass render_pass[NUM_META_FS_KEYS];
424 VkPipeline color_pipelines[NUM_META_FS_KEYS];
425
426 VkRenderPass depthstencil_rp;
427 VkPipeline depth_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
428 VkPipeline stencil_only_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
429 VkPipeline depthstencil_pipeline[NUM_DEPTH_CLEAR_PIPELINES];
430 } clear[1 + MAX_SAMPLES_LOG2];
431
432 VkPipelineLayout clear_color_p_layout;
433 VkPipelineLayout clear_depth_p_layout;
434 struct {
435 VkRenderPass render_pass[NUM_META_FS_KEYS][RADV_META_DST_LAYOUT_COUNT];
436
437 /** Pipeline that blits from a 1D image. */
438 VkPipeline pipeline_1d_src[NUM_META_FS_KEYS];
439
440 /** Pipeline that blits from a 2D image. */
441 VkPipeline pipeline_2d_src[NUM_META_FS_KEYS];
442
443 /** Pipeline that blits from a 3D image. */
444 VkPipeline pipeline_3d_src[NUM_META_FS_KEYS];
445
446 VkRenderPass depth_only_rp[RADV_BLIT_DS_LAYOUT_COUNT];
447 VkPipeline depth_only_1d_pipeline;
448 VkPipeline depth_only_2d_pipeline;
449 VkPipeline depth_only_3d_pipeline;
450
451 VkRenderPass stencil_only_rp[RADV_BLIT_DS_LAYOUT_COUNT];
452 VkPipeline stencil_only_1d_pipeline;
453 VkPipeline stencil_only_2d_pipeline;
454 VkPipeline stencil_only_3d_pipeline;
455 VkPipelineLayout pipeline_layout;
456 VkDescriptorSetLayout ds_layout;
457 } blit;
458
459 struct {
460 VkRenderPass render_passes[NUM_META_FS_KEYS][RADV_META_DST_LAYOUT_COUNT];
461
462 VkPipelineLayout p_layouts[3];
463 VkDescriptorSetLayout ds_layouts[3];
464 VkPipeline pipelines[3][NUM_META_FS_KEYS];
465
466 VkRenderPass depth_only_rp[RADV_BLIT_DS_LAYOUT_COUNT];
467 VkPipeline depth_only_pipeline[3];
468
469 VkRenderPass stencil_only_rp[RADV_BLIT_DS_LAYOUT_COUNT];
470 VkPipeline stencil_only_pipeline[3];
471 } blit2d;
472
473 struct {
474 VkPipelineLayout img_p_layout;
475 VkDescriptorSetLayout img_ds_layout;
476 VkPipeline pipeline;
477 VkPipeline pipeline_3d;
478 } itob;
479 struct {
480 VkPipelineLayout img_p_layout;
481 VkDescriptorSetLayout img_ds_layout;
482 VkPipeline pipeline;
483 VkPipeline pipeline_3d;
484 } btoi;
485 struct {
486 VkPipelineLayout img_p_layout;
487 VkDescriptorSetLayout img_ds_layout;
488 VkPipeline pipeline;
489 VkPipeline pipeline_3d;
490 } itoi;
491 struct {
492 VkPipelineLayout img_p_layout;
493 VkDescriptorSetLayout img_ds_layout;
494 VkPipeline pipeline;
495 VkPipeline pipeline_3d;
496 } cleari;
497
498 struct {
499 VkPipelineLayout p_layout;
500 VkPipeline pipeline[NUM_META_FS_KEYS];
501 VkRenderPass pass[NUM_META_FS_KEYS];
502 } resolve;
503
504 struct {
505 VkDescriptorSetLayout ds_layout;
506 VkPipelineLayout p_layout;
507 struct {
508 VkPipeline pipeline;
509 VkPipeline i_pipeline;
510 VkPipeline srgb_pipeline;
511 } rc[MAX_SAMPLES_LOG2];
512 } resolve_compute;
513
514 struct {
515 VkDescriptorSetLayout ds_layout;
516 VkPipelineLayout p_layout;
517
518 struct {
519 VkRenderPass render_pass[NUM_META_FS_KEYS][RADV_META_DST_LAYOUT_COUNT];
520 VkPipeline pipeline[NUM_META_FS_KEYS];
521 } rc[MAX_SAMPLES_LOG2];
522 } resolve_fragment;
523
524 struct {
525 VkPipelineLayout p_layout;
526 VkPipeline decompress_pipeline;
527 VkPipeline resummarize_pipeline;
528 VkRenderPass pass;
529 } depth_decomp[1 + MAX_SAMPLES_LOG2];
530
531 struct {
532 VkPipelineLayout p_layout;
533 VkPipeline cmask_eliminate_pipeline;
534 VkPipeline fmask_decompress_pipeline;
535 VkPipeline dcc_decompress_pipeline;
536 VkRenderPass pass;
537
538 VkDescriptorSetLayout dcc_decompress_compute_ds_layout;
539 VkPipelineLayout dcc_decompress_compute_p_layout;
540 VkPipeline dcc_decompress_compute_pipeline;
541 } fast_clear_flush;
542
543 struct {
544 VkPipelineLayout fill_p_layout;
545 VkPipelineLayout copy_p_layout;
546 VkDescriptorSetLayout fill_ds_layout;
547 VkDescriptorSetLayout copy_ds_layout;
548 VkPipeline fill_pipeline;
549 VkPipeline copy_pipeline;
550 } buffer;
551
552 struct {
553 VkDescriptorSetLayout ds_layout;
554 VkPipelineLayout p_layout;
555 VkPipeline occlusion_query_pipeline;
556 VkPipeline pipeline_statistics_query_pipeline;
557 } query;
558 };
559
560 /* queue types */
561 #define RADV_QUEUE_GENERAL 0
562 #define RADV_QUEUE_COMPUTE 1
563 #define RADV_QUEUE_TRANSFER 2
564
565 #define RADV_MAX_QUEUE_FAMILIES 3
566
567 enum ring_type radv_queue_family_to_ring(int f);
568
569 struct radv_queue {
570 VK_LOADER_DATA _loader_data;
571 struct radv_device * device;
572 struct radeon_winsys_ctx *hw_ctx;
573 enum radeon_ctx_priority priority;
574 uint32_t queue_family_index;
575 int queue_idx;
576
577 uint32_t scratch_size;
578 uint32_t compute_scratch_size;
579 uint32_t esgs_ring_size;
580 uint32_t gsvs_ring_size;
581 bool has_tess_rings;
582 bool has_sample_positions;
583
584 struct radeon_winsys_bo *scratch_bo;
585 struct radeon_winsys_bo *descriptor_bo;
586 struct radeon_winsys_bo *compute_scratch_bo;
587 struct radeon_winsys_bo *esgs_ring_bo;
588 struct radeon_winsys_bo *gsvs_ring_bo;
589 struct radeon_winsys_bo *tess_rings_bo;
590 struct radeon_winsys_cs *initial_preamble_cs;
591 struct radeon_winsys_cs *initial_full_flush_preamble_cs;
592 struct radeon_winsys_cs *continue_preamble_cs;
593 };
594
595 struct radv_device {
596 VK_LOADER_DATA _loader_data;
597
598 VkAllocationCallbacks alloc;
599
600 struct radv_instance * instance;
601 struct radeon_winsys *ws;
602
603 struct radv_meta_state meta_state;
604
605 struct radv_queue *queues[RADV_MAX_QUEUE_FAMILIES];
606 int queue_count[RADV_MAX_QUEUE_FAMILIES];
607 struct radeon_winsys_cs *empty_cs[RADV_MAX_QUEUE_FAMILIES];
608
609 bool always_use_syncobj;
610 bool llvm_supports_spill;
611 bool has_distributed_tess;
612 bool pbb_allowed;
613 bool dfsm_allowed;
614 uint32_t tess_offchip_block_dw_size;
615 uint32_t scratch_waves;
616 uint32_t dispatch_initiator;
617
618 uint32_t gs_table_depth;
619
620 /* MSAA sample locations.
621 * The first index is the sample index.
622 * The second index is the coordinate: X, Y. */
623 float sample_locations_1x[1][2];
624 float sample_locations_2x[2][2];
625 float sample_locations_4x[4][2];
626 float sample_locations_8x[8][2];
627 float sample_locations_16x[16][2];
628
629 /* CIK and later */
630 uint32_t gfx_init_size_dw;
631 struct radeon_winsys_bo *gfx_init;
632
633 struct radeon_winsys_bo *trace_bo;
634 uint32_t *trace_id_ptr;
635
636 /* Whether to keep shader debug info, for tracing or VK_AMD_shader_info */
637 bool keep_shader_info;
638
639 struct radv_physical_device *physical_device;
640
641 /* Backup in-memory cache to be used if the app doesn't provide one */
642 struct radv_pipeline_cache * mem_cache;
643
644 /*
645 * use different counters so MSAA MRTs get consecutive surface indices,
646 * even if MASK is allocated in between.
647 */
648 uint32_t image_mrt_offset_counter;
649 uint32_t fmask_mrt_offset_counter;
650 struct list_head shader_slabs;
651 mtx_t shader_slab_mutex;
652
653 /* For detecting VM faults reported by dmesg. */
654 uint64_t dmesg_timestamp;
655
656 struct radv_device_extension_table enabled_extensions;
657 };
658
659 struct radv_device_memory {
660 struct radeon_winsys_bo *bo;
661 /* for dedicated allocations */
662 struct radv_image *image;
663 struct radv_buffer *buffer;
664 uint32_t type_index;
665 VkDeviceSize map_size;
666 void * map;
667 void * user_ptr;
668 };
669
670
671 struct radv_descriptor_range {
672 uint64_t va;
673 uint32_t size;
674 };
675
676 struct radv_descriptor_set {
677 const struct radv_descriptor_set_layout *layout;
678 uint32_t size;
679
680 struct radeon_winsys_bo *bo;
681 uint64_t va;
682 uint32_t *mapped_ptr;
683 struct radv_descriptor_range *dynamic_descriptors;
684
685 struct radeon_winsys_bo *descriptors[0];
686 };
687
688 struct radv_push_descriptor_set
689 {
690 struct radv_descriptor_set set;
691 uint32_t capacity;
692 };
693
694 struct radv_descriptor_pool_entry {
695 uint32_t offset;
696 uint32_t size;
697 struct radv_descriptor_set *set;
698 };
699
700 struct radv_descriptor_pool {
701 struct radeon_winsys_bo *bo;
702 uint8_t *mapped_ptr;
703 uint64_t current_offset;
704 uint64_t size;
705
706 uint8_t *host_memory_base;
707 uint8_t *host_memory_ptr;
708 uint8_t *host_memory_end;
709
710 uint32_t entry_count;
711 uint32_t max_entry_count;
712 struct radv_descriptor_pool_entry entries[0];
713 };
714
715 struct radv_descriptor_update_template_entry {
716 VkDescriptorType descriptor_type;
717
718 /* The number of descriptors to update */
719 uint32_t descriptor_count;
720
721 /* Into mapped_ptr or dynamic_descriptors, in units of the respective array */
722 uint32_t dst_offset;
723
724 /* In dwords. Not valid/used for dynamic descriptors */
725 uint32_t dst_stride;
726
727 uint32_t buffer_offset;
728
729 /* Only valid for combined image samplers and samplers */
730 uint16_t has_sampler;
731
732 /* In bytes */
733 size_t src_offset;
734 size_t src_stride;
735
736 /* For push descriptors */
737 const uint32_t *immutable_samplers;
738 };
739
740 struct radv_descriptor_update_template {
741 uint32_t entry_count;
742 VkPipelineBindPoint bind_point;
743 struct radv_descriptor_update_template_entry entry[0];
744 };
745
746 struct radv_buffer {
747 VkDeviceSize size;
748
749 VkBufferUsageFlags usage;
750 VkBufferCreateFlags flags;
751
752 /* Set when bound */
753 struct radeon_winsys_bo * bo;
754 VkDeviceSize offset;
755
756 bool shareable;
757 };
758
759 enum radv_dynamic_state_bits {
760 RADV_DYNAMIC_VIEWPORT = 1 << 0,
761 RADV_DYNAMIC_SCISSOR = 1 << 1,
762 RADV_DYNAMIC_LINE_WIDTH = 1 << 2,
763 RADV_DYNAMIC_DEPTH_BIAS = 1 << 3,
764 RADV_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
765 RADV_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
766 RADV_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
767 RADV_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
768 RADV_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
769 RADV_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
770 RADV_DYNAMIC_ALL = (1 << 10) - 1,
771 };
772
773 enum radv_cmd_dirty_bits {
774 /* Keep the dynamic state dirty bits in sync with
775 * enum radv_dynamic_state_bits */
776 RADV_CMD_DIRTY_DYNAMIC_VIEWPORT = 1 << 0,
777 RADV_CMD_DIRTY_DYNAMIC_SCISSOR = 1 << 1,
778 RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH = 1 << 2,
779 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS = 1 << 3,
780 RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
781 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
782 RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
783 RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
784 RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
785 RADV_CMD_DIRTY_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
786 RADV_CMD_DIRTY_DYNAMIC_ALL = (1 << 10) - 1,
787 RADV_CMD_DIRTY_PIPELINE = 1 << 10,
788 RADV_CMD_DIRTY_INDEX_BUFFER = 1 << 11,
789 RADV_CMD_DIRTY_FRAMEBUFFER = 1 << 12,
790 RADV_CMD_DIRTY_VERTEX_BUFFER = 1 << 13,
791 };
792
793 enum radv_cmd_flush_bits {
794 RADV_CMD_FLAG_INV_ICACHE = 1 << 0,
795 /* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
796 RADV_CMD_FLAG_INV_SMEM_L1 = 1 << 1,
797 /* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
798 RADV_CMD_FLAG_INV_VMEM_L1 = 1 << 2,
799 /* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
800 RADV_CMD_FLAG_INV_GLOBAL_L2 = 1 << 3,
801 /* Same as above, but only writes back and doesn't invalidate */
802 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2 = 1 << 4,
803 /* Framebuffer caches */
804 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META = 1 << 5,
805 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META = 1 << 6,
806 RADV_CMD_FLAG_FLUSH_AND_INV_DB = 1 << 7,
807 RADV_CMD_FLAG_FLUSH_AND_INV_CB = 1 << 8,
808 /* Engine synchronization. */
809 RADV_CMD_FLAG_VS_PARTIAL_FLUSH = 1 << 9,
810 RADV_CMD_FLAG_PS_PARTIAL_FLUSH = 1 << 10,
811 RADV_CMD_FLAG_CS_PARTIAL_FLUSH = 1 << 11,
812 RADV_CMD_FLAG_VGT_FLUSH = 1 << 12,
813
814 RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER = (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
815 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
816 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
817 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META)
818 };
819
820 struct radv_vertex_binding {
821 struct radv_buffer * buffer;
822 VkDeviceSize offset;
823 };
824
825 struct radv_viewport_state {
826 uint32_t count;
827 VkViewport viewports[MAX_VIEWPORTS];
828 };
829
830 struct radv_scissor_state {
831 uint32_t count;
832 VkRect2D scissors[MAX_SCISSORS];
833 };
834
835 struct radv_discard_rectangle_state {
836 uint32_t count;
837 VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
838 };
839
840 struct radv_dynamic_state {
841 /**
842 * Bitmask of (1 << VK_DYNAMIC_STATE_*).
843 * Defines the set of saved dynamic state.
844 */
845 uint32_t mask;
846
847 struct radv_viewport_state viewport;
848
849 struct radv_scissor_state scissor;
850
851 float line_width;
852
853 struct {
854 float bias;
855 float clamp;
856 float slope;
857 } depth_bias;
858
859 float blend_constants[4];
860
861 struct {
862 float min;
863 float max;
864 } depth_bounds;
865
866 struct {
867 uint32_t front;
868 uint32_t back;
869 } stencil_compare_mask;
870
871 struct {
872 uint32_t front;
873 uint32_t back;
874 } stencil_write_mask;
875
876 struct {
877 uint32_t front;
878 uint32_t back;
879 } stencil_reference;
880
881 struct radv_discard_rectangle_state discard_rectangle;
882 };
883
884 extern const struct radv_dynamic_state default_dynamic_state;
885
886 const char *
887 radv_get_debug_option_name(int id);
888
889 const char *
890 radv_get_perftest_option_name(int id);
891
892 /**
893 * Attachment state when recording a renderpass instance.
894 *
895 * The clear value is valid only if there exists a pending clear.
896 */
897 struct radv_attachment_state {
898 VkImageAspectFlags pending_clear_aspects;
899 uint32_t cleared_views;
900 VkClearValue clear_value;
901 VkImageLayout current_layout;
902 };
903
904 struct radv_descriptor_state {
905 struct radv_descriptor_set *sets[MAX_SETS];
906 uint32_t dirty;
907 uint32_t valid;
908 struct radv_push_descriptor_set push_set;
909 bool push_dirty;
910 };
911
912 struct radv_cmd_state {
913 /* Vertex descriptors */
914 bool vb_prefetch_dirty;
915 uint64_t vb_va;
916 unsigned vb_size;
917
918 bool predicating;
919 uint32_t dirty;
920
921 struct radv_pipeline * pipeline;
922 struct radv_pipeline * emitted_pipeline;
923 struct radv_pipeline * compute_pipeline;
924 struct radv_pipeline * emitted_compute_pipeline;
925 struct radv_framebuffer * framebuffer;
926 struct radv_render_pass * pass;
927 const struct radv_subpass * subpass;
928 struct radv_dynamic_state dynamic;
929 struct radv_attachment_state * attachments;
930 VkRect2D render_area;
931
932 /* Index buffer */
933 struct radv_buffer *index_buffer;
934 uint64_t index_offset;
935 uint32_t index_type;
936 uint32_t max_index_count;
937 uint64_t index_va;
938 int32_t last_index_type;
939
940 int32_t last_primitive_reset_en;
941 uint32_t last_primitive_reset_index;
942 enum radv_cmd_flush_bits flush_bits;
943 unsigned active_occlusion_queries;
944 float offset_scale;
945 uint32_t trace_id;
946 uint32_t last_ia_multi_vgt_param;
947
948 uint32_t last_num_instances;
949 uint32_t last_first_instance;
950 uint32_t last_vertex_offset;
951 };
952
953 struct radv_cmd_pool {
954 VkAllocationCallbacks alloc;
955 struct list_head cmd_buffers;
956 struct list_head free_cmd_buffers;
957 uint32_t queue_family_index;
958 };
959
960 struct radv_cmd_buffer_upload {
961 uint8_t *map;
962 unsigned offset;
963 uint64_t size;
964 struct radeon_winsys_bo *upload_bo;
965 struct list_head list;
966 };
967
968 enum radv_cmd_buffer_status {
969 RADV_CMD_BUFFER_STATUS_INVALID,
970 RADV_CMD_BUFFER_STATUS_INITIAL,
971 RADV_CMD_BUFFER_STATUS_RECORDING,
972 RADV_CMD_BUFFER_STATUS_EXECUTABLE,
973 RADV_CMD_BUFFER_STATUS_PENDING,
974 };
975
976 struct radv_cmd_buffer {
977 VK_LOADER_DATA _loader_data;
978
979 struct radv_device * device;
980
981 struct radv_cmd_pool * pool;
982 struct list_head pool_link;
983
984 VkCommandBufferUsageFlags usage_flags;
985 VkCommandBufferLevel level;
986 enum radv_cmd_buffer_status status;
987 struct radeon_winsys_cs *cs;
988 struct radv_cmd_state state;
989 struct radv_vertex_binding vertex_bindings[MAX_VBS];
990 uint32_t queue_family_index;
991
992 uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
993 uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
994 VkShaderStageFlags push_constant_stages;
995 struct radv_descriptor_set meta_push_descriptors;
996
997 struct radv_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
998
999 struct radv_cmd_buffer_upload upload;
1000
1001 uint32_t scratch_size_needed;
1002 uint32_t compute_scratch_size_needed;
1003 uint32_t esgs_ring_size_needed;
1004 uint32_t gsvs_ring_size_needed;
1005 bool tess_rings_needed;
1006 bool sample_positions_needed;
1007
1008 VkResult record_result;
1009
1010 int ring_offsets_idx; /* just used for verification */
1011 uint32_t gfx9_fence_offset;
1012 struct radeon_winsys_bo *gfx9_fence_bo;
1013 uint32_t gfx9_fence_idx;
1014
1015 /**
1016 * Whether a query pool has been resetted and we have to flush caches.
1017 */
1018 bool pending_reset_query;
1019 };
1020
1021 struct radv_image;
1022
1023 bool radv_cmd_buffer_uses_mec(struct radv_cmd_buffer *cmd_buffer);
1024
1025 void si_init_compute(struct radv_cmd_buffer *cmd_buffer);
1026 void si_init_config(struct radv_cmd_buffer *cmd_buffer);
1027
1028 void cik_create_gfx_config(struct radv_device *device);
1029
1030 void si_write_viewport(struct radeon_winsys_cs *cs, int first_vp,
1031 int count, const VkViewport *viewports);
1032 void si_write_scissors(struct radeon_winsys_cs *cs, int first,
1033 int count, const VkRect2D *scissors,
1034 const VkViewport *viewports, bool can_use_guardband);
1035 uint32_t si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer,
1036 bool instanced_draw, bool indirect_draw,
1037 uint32_t draw_vertex_count);
1038 void si_cs_emit_write_event_eop(struct radeon_winsys_cs *cs,
1039 bool predicated,
1040 enum chip_class chip_class,
1041 bool is_mec,
1042 unsigned event, unsigned event_flags,
1043 unsigned data_sel,
1044 uint64_t va,
1045 uint32_t old_fence,
1046 uint32_t new_fence);
1047
1048 void si_emit_wait_fence(struct radeon_winsys_cs *cs,
1049 bool predicated,
1050 uint64_t va, uint32_t ref,
1051 uint32_t mask);
1052 void si_cs_emit_cache_flush(struct radeon_winsys_cs *cs,
1053 enum chip_class chip_class,
1054 uint32_t *fence_ptr, uint64_t va,
1055 bool is_mec,
1056 enum radv_cmd_flush_bits flush_bits);
1057 void si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer);
1058 void si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer, uint64_t va);
1059 void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
1060 uint64_t src_va, uint64_t dest_va,
1061 uint64_t size);
1062 void si_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1063 unsigned size);
1064 void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1065 uint64_t size, unsigned value);
1066 void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer);
1067 bool
1068 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
1069 unsigned size,
1070 unsigned alignment,
1071 unsigned *out_offset,
1072 void **ptr);
1073 void
1074 radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
1075 const struct radv_subpass *subpass,
1076 bool transitions);
1077 bool
1078 radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
1079 unsigned size, unsigned alignmnet,
1080 const void *data, unsigned *out_offset);
1081
1082 void radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer);
1083 void radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer);
1084 void radv_cmd_buffer_resolve_subpass_cs(struct radv_cmd_buffer *cmd_buffer);
1085 void radv_cmd_buffer_resolve_subpass_fs(struct radv_cmd_buffer *cmd_buffer);
1086 void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs *cs, int nr_samples);
1087 unsigned radv_cayman_get_maxdist(int log_samples);
1088 void radv_device_init_msaa(struct radv_device *device);
1089 void radv_set_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1090 struct radv_image *image,
1091 VkClearDepthStencilValue ds_clear_value,
1092 VkImageAspectFlags aspects);
1093 void radv_set_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1094 struct radv_image *image,
1095 int idx,
1096 uint32_t color_values[2]);
1097 void radv_set_dcc_need_cmask_elim_pred(struct radv_cmd_buffer *cmd_buffer,
1098 struct radv_image *image,
1099 bool value);
1100 uint32_t radv_fill_buffer(struct radv_cmd_buffer *cmd_buffer,
1101 struct radeon_winsys_bo *bo,
1102 uint64_t offset, uint64_t size, uint32_t value);
1103 void radv_cmd_buffer_trace_emit(struct radv_cmd_buffer *cmd_buffer);
1104 bool radv_get_memory_fd(struct radv_device *device,
1105 struct radv_device_memory *memory,
1106 int *pFD);
1107
1108 static inline struct radv_descriptor_state *
1109 radv_get_descriptors_state(struct radv_cmd_buffer *cmd_buffer,
1110 VkPipelineBindPoint bind_point)
1111 {
1112 assert(bind_point == VK_PIPELINE_BIND_POINT_GRAPHICS ||
1113 bind_point == VK_PIPELINE_BIND_POINT_COMPUTE);
1114 return &cmd_buffer->descriptors[bind_point];
1115 }
1116
1117 /*
1118 * Takes x,y,z as exact numbers of invocations, instead of blocks.
1119 *
1120 * Limitations: Can't call normal dispatch functions without binding or rebinding
1121 * the compute pipeline.
1122 */
1123 void radv_unaligned_dispatch(
1124 struct radv_cmd_buffer *cmd_buffer,
1125 uint32_t x,
1126 uint32_t y,
1127 uint32_t z);
1128
1129 struct radv_event {
1130 struct radeon_winsys_bo *bo;
1131 uint64_t *map;
1132 };
1133
1134 struct radv_shader_module;
1135
1136 #define RADV_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
1137 #define RADV_HASH_SHADER_SISCHED (1 << 1)
1138 #define RADV_HASH_SHADER_UNSAFE_MATH (1 << 2)
1139 void
1140 radv_hash_shaders(unsigned char *hash,
1141 const VkPipelineShaderStageCreateInfo **stages,
1142 const struct radv_pipeline_layout *layout,
1143 const struct radv_pipeline_key *key,
1144 uint32_t flags);
1145
1146 static inline gl_shader_stage
1147 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
1148 {
1149 assert(__builtin_popcount(vk_stage) == 1);
1150 return ffs(vk_stage) - 1;
1151 }
1152
1153 static inline VkShaderStageFlagBits
1154 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
1155 {
1156 return (1 << mesa_stage);
1157 }
1158
1159 #define RADV_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
1160
1161 #define radv_foreach_stage(stage, stage_bits) \
1162 for (gl_shader_stage stage, \
1163 __tmp = (gl_shader_stage)((stage_bits) & RADV_STAGE_MASK); \
1164 stage = __builtin_ffs(__tmp) - 1, __tmp; \
1165 __tmp &= ~(1 << (stage)))
1166
1167 unsigned radv_format_meta_fs_key(VkFormat format);
1168
1169 struct radv_multisample_state {
1170 uint32_t db_eqaa;
1171 uint32_t pa_sc_line_cntl;
1172 uint32_t pa_sc_mode_cntl_0;
1173 uint32_t pa_sc_mode_cntl_1;
1174 uint32_t pa_sc_aa_config;
1175 uint32_t pa_sc_aa_mask[2];
1176 unsigned num_samples;
1177 };
1178
1179 struct radv_prim_vertex_count {
1180 uint8_t min;
1181 uint8_t incr;
1182 };
1183
1184 struct radv_vertex_elements_info {
1185 uint32_t rsrc_word3[MAX_VERTEX_ATTRIBS];
1186 uint32_t format_size[MAX_VERTEX_ATTRIBS];
1187 uint32_t binding[MAX_VERTEX_ATTRIBS];
1188 uint32_t offset[MAX_VERTEX_ATTRIBS];
1189 uint32_t count;
1190 };
1191
1192 struct radv_ia_multi_vgt_param_helpers {
1193 uint32_t base;
1194 bool partial_es_wave;
1195 uint8_t primgroup_size;
1196 bool wd_switch_on_eop;
1197 bool ia_switch_on_eoi;
1198 bool partial_vs_wave;
1199 };
1200
1201 #define SI_GS_PER_ES 128
1202
1203 struct radv_pipeline {
1204 struct radv_device * device;
1205 struct radv_dynamic_state dynamic_state;
1206
1207 struct radv_pipeline_layout * layout;
1208
1209 bool needs_data_cache;
1210 bool need_indirect_descriptor_sets;
1211 struct radv_shader_variant * shaders[MESA_SHADER_STAGES];
1212 struct radv_shader_variant *gs_copy_shader;
1213 VkShaderStageFlags active_stages;
1214
1215 struct radeon_winsys_cs cs;
1216
1217 struct radv_vertex_elements_info vertex_elements;
1218
1219 uint32_t binding_stride[MAX_VBS];
1220
1221 uint32_t user_data_0[MESA_SHADER_STAGES];
1222 union {
1223 struct {
1224 struct radv_multisample_state ms;
1225 uint32_t spi_baryc_cntl;
1226 bool prim_restart_enable;
1227 unsigned esgs_ring_size;
1228 unsigned gsvs_ring_size;
1229 uint32_t vtx_base_sgpr;
1230 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param;
1231 uint8_t vtx_emit_num;
1232 struct radv_prim_vertex_count prim_vertex_count;
1233 bool can_use_guardband;
1234 uint32_t needed_dynamic_state;
1235 } graphics;
1236 };
1237
1238 unsigned max_waves;
1239 unsigned scratch_bytes_per_wave;
1240 };
1241
1242 static inline bool radv_pipeline_has_gs(const struct radv_pipeline *pipeline)
1243 {
1244 return pipeline->shaders[MESA_SHADER_GEOMETRY] ? true : false;
1245 }
1246
1247 static inline bool radv_pipeline_has_tess(const struct radv_pipeline *pipeline)
1248 {
1249 return pipeline->shaders[MESA_SHADER_TESS_CTRL] ? true : false;
1250 }
1251
1252 struct radv_userdata_info *radv_lookup_user_sgpr(struct radv_pipeline *pipeline,
1253 gl_shader_stage stage,
1254 int idx);
1255
1256 struct radv_shader_variant *radv_get_vertex_shader(struct radv_pipeline *pipeline);
1257
1258 struct radv_graphics_pipeline_create_info {
1259 bool use_rectlist;
1260 bool db_depth_clear;
1261 bool db_stencil_clear;
1262 bool db_depth_disable_expclear;
1263 bool db_stencil_disable_expclear;
1264 bool db_flush_depth_inplace;
1265 bool db_flush_stencil_inplace;
1266 bool db_resummarize;
1267 uint32_t custom_blend_mode;
1268 };
1269
1270 VkResult
1271 radv_graphics_pipeline_create(VkDevice device,
1272 VkPipelineCache cache,
1273 const VkGraphicsPipelineCreateInfo *pCreateInfo,
1274 const struct radv_graphics_pipeline_create_info *extra,
1275 const VkAllocationCallbacks *alloc,
1276 VkPipeline *pPipeline);
1277
1278 struct vk_format_description;
1279 uint32_t radv_translate_buffer_dataformat(const struct vk_format_description *desc,
1280 int first_non_void);
1281 uint32_t radv_translate_buffer_numformat(const struct vk_format_description *desc,
1282 int first_non_void);
1283 uint32_t radv_translate_colorformat(VkFormat format);
1284 uint32_t radv_translate_color_numformat(VkFormat format,
1285 const struct vk_format_description *desc,
1286 int first_non_void);
1287 uint32_t radv_colorformat_endian_swap(uint32_t colorformat);
1288 unsigned radv_translate_colorswap(VkFormat format, bool do_endian_swap);
1289 uint32_t radv_translate_dbformat(VkFormat format);
1290 uint32_t radv_translate_tex_dataformat(VkFormat format,
1291 const struct vk_format_description *desc,
1292 int first_non_void);
1293 uint32_t radv_translate_tex_numformat(VkFormat format,
1294 const struct vk_format_description *desc,
1295 int first_non_void);
1296 bool radv_format_pack_clear_color(VkFormat format,
1297 uint32_t clear_vals[2],
1298 VkClearColorValue *value);
1299 bool radv_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
1300 bool radv_dcc_formats_compatible(VkFormat format1,
1301 VkFormat format2);
1302
1303 struct radv_fmask_info {
1304 uint64_t offset;
1305 uint64_t size;
1306 unsigned alignment;
1307 unsigned pitch_in_pixels;
1308 unsigned bank_height;
1309 unsigned slice_tile_max;
1310 unsigned tile_mode_index;
1311 unsigned tile_swizzle;
1312 };
1313
1314 struct radv_cmask_info {
1315 uint64_t offset;
1316 uint64_t size;
1317 unsigned alignment;
1318 unsigned slice_tile_max;
1319 };
1320
1321 struct radv_image {
1322 VkImageType type;
1323 /* The original VkFormat provided by the client. This may not match any
1324 * of the actual surface formats.
1325 */
1326 VkFormat vk_format;
1327 VkImageAspectFlags aspects;
1328 VkImageUsageFlags usage; /**< Superset of VkImageCreateInfo::usage. */
1329 struct ac_surf_info info;
1330 VkImageTiling tiling; /** VkImageCreateInfo::tiling */
1331 VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
1332
1333 VkDeviceSize size;
1334 uint32_t alignment;
1335
1336 unsigned queue_family_mask;
1337 bool exclusive;
1338 bool shareable;
1339
1340 /* Set when bound */
1341 struct radeon_winsys_bo *bo;
1342 VkDeviceSize offset;
1343 uint64_t dcc_offset;
1344 uint64_t htile_offset;
1345 bool tc_compatible_htile;
1346 struct radeon_surf surface;
1347
1348 struct radv_fmask_info fmask;
1349 struct radv_cmask_info cmask;
1350 uint64_t clear_value_offset;
1351 uint64_t dcc_pred_offset;
1352
1353 /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
1354 VkDeviceMemory owned_memory;
1355 };
1356
1357 /* Whether the image has a htile that is known consistent with the contents of
1358 * the image. */
1359 bool radv_layout_has_htile(const struct radv_image *image,
1360 VkImageLayout layout,
1361 unsigned queue_mask);
1362
1363 /* Whether the image has a htile that is known consistent with the contents of
1364 * the image and is allowed to be in compressed form.
1365 *
1366 * If this is false reads that don't use the htile should be able to return
1367 * correct results.
1368 */
1369 bool radv_layout_is_htile_compressed(const struct radv_image *image,
1370 VkImageLayout layout,
1371 unsigned queue_mask);
1372
1373 bool radv_layout_can_fast_clear(const struct radv_image *image,
1374 VkImageLayout layout,
1375 unsigned queue_mask);
1376
1377 bool radv_layout_dcc_compressed(const struct radv_image *image,
1378 VkImageLayout layout,
1379 unsigned queue_mask);
1380
1381 static inline bool
1382 radv_vi_dcc_enabled(const struct radv_image *image, unsigned level)
1383 {
1384 return image->surface.dcc_size && level < image->surface.num_dcc_levels;
1385 }
1386
1387 static inline bool
1388 radv_htile_enabled(const struct radv_image *image, unsigned level)
1389 {
1390 return image->surface.htile_size && level == 0;
1391 }
1392
1393 unsigned radv_image_queue_family_mask(const struct radv_image *image, uint32_t family, uint32_t queue_family);
1394
1395 static inline uint32_t
1396 radv_get_layerCount(const struct radv_image *image,
1397 const VkImageSubresourceRange *range)
1398 {
1399 return range->layerCount == VK_REMAINING_ARRAY_LAYERS ?
1400 image->info.array_size - range->baseArrayLayer : range->layerCount;
1401 }
1402
1403 static inline uint32_t
1404 radv_get_levelCount(const struct radv_image *image,
1405 const VkImageSubresourceRange *range)
1406 {
1407 return range->levelCount == VK_REMAINING_MIP_LEVELS ?
1408 image->info.levels - range->baseMipLevel : range->levelCount;
1409 }
1410
1411 struct radeon_bo_metadata;
1412 void
1413 radv_init_metadata(struct radv_device *device,
1414 struct radv_image *image,
1415 struct radeon_bo_metadata *metadata);
1416
1417 struct radv_image_view {
1418 struct radv_image *image; /**< VkImageViewCreateInfo::image */
1419 struct radeon_winsys_bo *bo;
1420
1421 VkImageViewType type;
1422 VkImageAspectFlags aspect_mask;
1423 VkFormat vk_format;
1424 uint32_t base_layer;
1425 uint32_t layer_count;
1426 uint32_t base_mip;
1427 uint32_t level_count;
1428 VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
1429
1430 uint32_t descriptor[16];
1431
1432 /* Descriptor for use as a storage image as opposed to a sampled image.
1433 * This has a few differences for cube maps (e.g. type).
1434 */
1435 uint32_t storage_descriptor[16];
1436 };
1437
1438 struct radv_image_create_info {
1439 const VkImageCreateInfo *vk_info;
1440 bool scanout;
1441 bool no_metadata_planes;
1442 };
1443
1444 VkResult radv_image_create(VkDevice _device,
1445 const struct radv_image_create_info *info,
1446 const VkAllocationCallbacks* alloc,
1447 VkImage *pImage);
1448
1449 VkResult
1450 radv_image_from_gralloc(VkDevice device_h,
1451 const VkImageCreateInfo *base_info,
1452 const VkNativeBufferANDROID *gralloc_info,
1453 const VkAllocationCallbacks *alloc,
1454 VkImage *out_image_h);
1455
1456 void radv_image_view_init(struct radv_image_view *view,
1457 struct radv_device *device,
1458 const VkImageViewCreateInfo* pCreateInfo);
1459
1460 struct radv_buffer_view {
1461 struct radeon_winsys_bo *bo;
1462 VkFormat vk_format;
1463 uint64_t range; /**< VkBufferViewCreateInfo::range */
1464 uint32_t state[4];
1465 };
1466 void radv_buffer_view_init(struct radv_buffer_view *view,
1467 struct radv_device *device,
1468 const VkBufferViewCreateInfo* pCreateInfo);
1469
1470 static inline struct VkExtent3D
1471 radv_sanitize_image_extent(const VkImageType imageType,
1472 const struct VkExtent3D imageExtent)
1473 {
1474 switch (imageType) {
1475 case VK_IMAGE_TYPE_1D:
1476 return (VkExtent3D) { imageExtent.width, 1, 1 };
1477 case VK_IMAGE_TYPE_2D:
1478 return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1479 case VK_IMAGE_TYPE_3D:
1480 return imageExtent;
1481 default:
1482 unreachable("invalid image type");
1483 }
1484 }
1485
1486 static inline struct VkOffset3D
1487 radv_sanitize_image_offset(const VkImageType imageType,
1488 const struct VkOffset3D imageOffset)
1489 {
1490 switch (imageType) {
1491 case VK_IMAGE_TYPE_1D:
1492 return (VkOffset3D) { imageOffset.x, 0, 0 };
1493 case VK_IMAGE_TYPE_2D:
1494 return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1495 case VK_IMAGE_TYPE_3D:
1496 return imageOffset;
1497 default:
1498 unreachable("invalid image type");
1499 }
1500 }
1501
1502 static inline bool
1503 radv_image_extent_compare(const struct radv_image *image,
1504 const VkExtent3D *extent)
1505 {
1506 if (extent->width != image->info.width ||
1507 extent->height != image->info.height ||
1508 extent->depth != image->info.depth)
1509 return false;
1510 return true;
1511 }
1512
1513 struct radv_sampler {
1514 uint32_t state[4];
1515 };
1516
1517 struct radv_color_buffer_info {
1518 uint64_t cb_color_base;
1519 uint64_t cb_color_cmask;
1520 uint64_t cb_color_fmask;
1521 uint64_t cb_dcc_base;
1522 uint32_t cb_color_pitch;
1523 uint32_t cb_color_slice;
1524 uint32_t cb_color_view;
1525 uint32_t cb_color_info;
1526 uint32_t cb_color_attrib;
1527 uint32_t cb_color_attrib2;
1528 uint32_t cb_dcc_control;
1529 uint32_t cb_color_cmask_slice;
1530 uint32_t cb_color_fmask_slice;
1531 };
1532
1533 struct radv_ds_buffer_info {
1534 uint64_t db_z_read_base;
1535 uint64_t db_stencil_read_base;
1536 uint64_t db_z_write_base;
1537 uint64_t db_stencil_write_base;
1538 uint64_t db_htile_data_base;
1539 uint32_t db_depth_info;
1540 uint32_t db_z_info;
1541 uint32_t db_stencil_info;
1542 uint32_t db_depth_view;
1543 uint32_t db_depth_size;
1544 uint32_t db_depth_slice;
1545 uint32_t db_htile_surface;
1546 uint32_t pa_su_poly_offset_db_fmt_cntl;
1547 uint32_t db_z_info2;
1548 uint32_t db_stencil_info2;
1549 float offset_scale;
1550 };
1551
1552 struct radv_attachment_info {
1553 union {
1554 struct radv_color_buffer_info cb;
1555 struct radv_ds_buffer_info ds;
1556 };
1557 struct radv_image_view *attachment;
1558 };
1559
1560 struct radv_framebuffer {
1561 uint32_t width;
1562 uint32_t height;
1563 uint32_t layers;
1564
1565 uint32_t attachment_count;
1566 struct radv_attachment_info attachments[0];
1567 };
1568
1569 struct radv_subpass_barrier {
1570 VkPipelineStageFlags src_stage_mask;
1571 VkAccessFlags src_access_mask;
1572 VkAccessFlags dst_access_mask;
1573 };
1574
1575 struct radv_subpass {
1576 uint32_t input_count;
1577 uint32_t color_count;
1578 VkAttachmentReference * input_attachments;
1579 VkAttachmentReference * color_attachments;
1580 VkAttachmentReference * resolve_attachments;
1581 VkAttachmentReference depth_stencil_attachment;
1582
1583 /** Subpass has at least one resolve attachment */
1584 bool has_resolve;
1585
1586 struct radv_subpass_barrier start_barrier;
1587
1588 uint32_t view_mask;
1589 };
1590
1591 struct radv_render_pass_attachment {
1592 VkFormat format;
1593 uint32_t samples;
1594 VkAttachmentLoadOp load_op;
1595 VkAttachmentLoadOp stencil_load_op;
1596 VkImageLayout initial_layout;
1597 VkImageLayout final_layout;
1598 uint32_t view_mask;
1599 };
1600
1601 struct radv_render_pass {
1602 uint32_t attachment_count;
1603 uint32_t subpass_count;
1604 VkAttachmentReference * subpass_attachments;
1605 struct radv_render_pass_attachment * attachments;
1606 struct radv_subpass_barrier end_barrier;
1607 struct radv_subpass subpasses[0];
1608 };
1609
1610 VkResult radv_device_init_meta(struct radv_device *device);
1611 void radv_device_finish_meta(struct radv_device *device);
1612
1613 struct radv_query_pool {
1614 struct radeon_winsys_bo *bo;
1615 uint32_t stride;
1616 uint32_t availability_offset;
1617 uint64_t size;
1618 char *ptr;
1619 VkQueryType type;
1620 uint32_t pipeline_stats_mask;
1621 };
1622
1623 struct radv_semaphore {
1624 /* use a winsys sem for non-exportable */
1625 struct radeon_winsys_sem *sem;
1626 uint32_t syncobj;
1627 uint32_t temp_syncobj;
1628 };
1629
1630 VkResult radv_alloc_sem_info(struct radv_winsys_sem_info *sem_info,
1631 int num_wait_sems,
1632 const VkSemaphore *wait_sems,
1633 int num_signal_sems,
1634 const VkSemaphore *signal_sems,
1635 VkFence fence);
1636 void radv_free_sem_info(struct radv_winsys_sem_info *sem_info);
1637
1638 void radv_set_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
1639 VkPipelineBindPoint bind_point,
1640 struct radv_descriptor_set *set,
1641 unsigned idx);
1642
1643 void
1644 radv_update_descriptor_sets(struct radv_device *device,
1645 struct radv_cmd_buffer *cmd_buffer,
1646 VkDescriptorSet overrideSet,
1647 uint32_t descriptorWriteCount,
1648 const VkWriteDescriptorSet *pDescriptorWrites,
1649 uint32_t descriptorCopyCount,
1650 const VkCopyDescriptorSet *pDescriptorCopies);
1651
1652 void
1653 radv_update_descriptor_set_with_template(struct radv_device *device,
1654 struct radv_cmd_buffer *cmd_buffer,
1655 struct radv_descriptor_set *set,
1656 VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1657 const void *pData);
1658
1659 void radv_meta_push_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
1660 VkPipelineBindPoint pipelineBindPoint,
1661 VkPipelineLayout _layout,
1662 uint32_t set,
1663 uint32_t descriptorWriteCount,
1664 const VkWriteDescriptorSet *pDescriptorWrites);
1665
1666 void radv_initialise_cmask(struct radv_cmd_buffer *cmd_buffer,
1667 struct radv_image *image, uint32_t value);
1668 void radv_initialize_dcc(struct radv_cmd_buffer *cmd_buffer,
1669 struct radv_image *image, uint32_t value);
1670
1671 struct radv_fence {
1672 struct radeon_winsys_fence *fence;
1673 bool submitted;
1674 bool signalled;
1675
1676 uint32_t syncobj;
1677 uint32_t temp_syncobj;
1678 };
1679
1680 /* radv_nir_to_llvm.c */
1681 struct radv_shader_variant_info;
1682
1683 void radv_compile_gs_copy_shader(LLVMTargetMachineRef tm,
1684 struct nir_shader *geom_shader,
1685 struct ac_shader_binary *binary,
1686 struct ac_shader_config *config,
1687 struct radv_shader_variant_info *shader_info,
1688 const struct ac_nir_compiler_options *options,
1689 bool dump_shader);
1690
1691 void radv_compile_nir_shader(LLVMTargetMachineRef tm,
1692 struct ac_shader_binary *binary,
1693 struct ac_shader_config *config,
1694 struct radv_shader_variant_info *shader_info,
1695 struct nir_shader *const *nir,
1696 int nir_count,
1697 const struct ac_nir_compiler_options *options,
1698 bool dump_shader);
1699
1700 struct radeon_winsys_sem;
1701
1702 #define RADV_DEFINE_HANDLE_CASTS(__radv_type, __VkType) \
1703 \
1704 static inline struct __radv_type * \
1705 __radv_type ## _from_handle(__VkType _handle) \
1706 { \
1707 return (struct __radv_type *) _handle; \
1708 } \
1709 \
1710 static inline __VkType \
1711 __radv_type ## _to_handle(struct __radv_type *_obj) \
1712 { \
1713 return (__VkType) _obj; \
1714 }
1715
1716 #define RADV_DEFINE_NONDISP_HANDLE_CASTS(__radv_type, __VkType) \
1717 \
1718 static inline struct __radv_type * \
1719 __radv_type ## _from_handle(__VkType _handle) \
1720 { \
1721 return (struct __radv_type *)(uintptr_t) _handle; \
1722 } \
1723 \
1724 static inline __VkType \
1725 __radv_type ## _to_handle(struct __radv_type *_obj) \
1726 { \
1727 return (__VkType)(uintptr_t) _obj; \
1728 }
1729
1730 #define RADV_FROM_HANDLE(__radv_type, __name, __handle) \
1731 struct __radv_type *__name = __radv_type ## _from_handle(__handle)
1732
1733 RADV_DEFINE_HANDLE_CASTS(radv_cmd_buffer, VkCommandBuffer)
1734 RADV_DEFINE_HANDLE_CASTS(radv_device, VkDevice)
1735 RADV_DEFINE_HANDLE_CASTS(radv_instance, VkInstance)
1736 RADV_DEFINE_HANDLE_CASTS(radv_physical_device, VkPhysicalDevice)
1737 RADV_DEFINE_HANDLE_CASTS(radv_queue, VkQueue)
1738
1739 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_cmd_pool, VkCommandPool)
1740 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer, VkBuffer)
1741 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer_view, VkBufferView)
1742 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_pool, VkDescriptorPool)
1743 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set, VkDescriptorSet)
1744 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set_layout, VkDescriptorSetLayout)
1745 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_update_template, VkDescriptorUpdateTemplateKHR)
1746 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_device_memory, VkDeviceMemory)
1747 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_fence, VkFence)
1748 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_event, VkEvent)
1749 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_framebuffer, VkFramebuffer)
1750 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image, VkImage)
1751 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image_view, VkImageView);
1752 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_cache, VkPipelineCache)
1753 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline, VkPipeline)
1754 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_layout, VkPipelineLayout)
1755 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_query_pool, VkQueryPool)
1756 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_render_pass, VkRenderPass)
1757 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_sampler, VkSampler)
1758 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_shader_module, VkShaderModule)
1759 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_semaphore, VkSemaphore)
1760
1761 #endif /* RADV_PRIVATE_H */