projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: add dump_shader to the NIR compiler options
[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 VkDeviceQueueCreateFlags flags;
577
578 uint32_t scratch_size;
579 uint32_t compute_scratch_size;
580 uint32_t esgs_ring_size;
581 uint32_t gsvs_ring_size;
582 bool has_tess_rings;
583 bool has_sample_positions;
584
585 struct radeon_winsys_bo *scratch_bo;
586 struct radeon_winsys_bo *descriptor_bo;
587 struct radeon_winsys_bo *compute_scratch_bo;
588 struct radeon_winsys_bo *esgs_ring_bo;
589 struct radeon_winsys_bo *gsvs_ring_bo;
590 struct radeon_winsys_bo *tess_rings_bo;
591 struct radeon_winsys_cs *initial_preamble_cs;
592 struct radeon_winsys_cs *initial_full_flush_preamble_cs;
593 struct radeon_winsys_cs *continue_preamble_cs;
594 };
595
596 struct radv_device {
597 VK_LOADER_DATA _loader_data;
598
599 VkAllocationCallbacks alloc;
600
601 struct radv_instance * instance;
602 struct radeon_winsys *ws;
603
604 struct radv_meta_state meta_state;
605
606 struct radv_queue *queues[RADV_MAX_QUEUE_FAMILIES];
607 int queue_count[RADV_MAX_QUEUE_FAMILIES];
608 struct radeon_winsys_cs *empty_cs[RADV_MAX_QUEUE_FAMILIES];
609
610 bool always_use_syncobj;
611 bool llvm_supports_spill;
612 bool has_distributed_tess;
613 bool pbb_allowed;
614 bool dfsm_allowed;
615 uint32_t tess_offchip_block_dw_size;
616 uint32_t scratch_waves;
617 uint32_t dispatch_initiator;
618
619 uint32_t gs_table_depth;
620
621 /* MSAA sample locations.
622 * The first index is the sample index.
623 * The second index is the coordinate: X, Y. */
624 float sample_locations_1x[1][2];
625 float sample_locations_2x[2][2];
626 float sample_locations_4x[4][2];
627 float sample_locations_8x[8][2];
628 float sample_locations_16x[16][2];
629
630 /* CIK and later */
631 uint32_t gfx_init_size_dw;
632 struct radeon_winsys_bo *gfx_init;
633
634 struct radeon_winsys_bo *trace_bo;
635 uint32_t *trace_id_ptr;
636
637 /* Whether to keep shader debug info, for tracing or VK_AMD_shader_info */
638 bool keep_shader_info;
639
640 struct radv_physical_device *physical_device;
641
642 /* Backup in-memory cache to be used if the app doesn't provide one */
643 struct radv_pipeline_cache * mem_cache;
644
645 /*
646 * use different counters so MSAA MRTs get consecutive surface indices,
647 * even if MASK is allocated in between.
648 */
649 uint32_t image_mrt_offset_counter;
650 uint32_t fmask_mrt_offset_counter;
651 struct list_head shader_slabs;
652 mtx_t shader_slab_mutex;
653
654 /* For detecting VM faults reported by dmesg. */
655 uint64_t dmesg_timestamp;
656
657 struct radv_device_extension_table enabled_extensions;
658 };
659
660 struct radv_device_memory {
661 struct radeon_winsys_bo *bo;
662 /* for dedicated allocations */
663 struct radv_image *image;
664 struct radv_buffer *buffer;
665 uint32_t type_index;
666 VkDeviceSize map_size;
667 void * map;
668 void * user_ptr;
669 };
670
671
672 struct radv_descriptor_range {
673 uint64_t va;
674 uint32_t size;
675 };
676
677 struct radv_descriptor_set {
678 const struct radv_descriptor_set_layout *layout;
679 uint32_t size;
680
681 struct radeon_winsys_bo *bo;
682 uint64_t va;
683 uint32_t *mapped_ptr;
684 struct radv_descriptor_range *dynamic_descriptors;
685
686 struct radeon_winsys_bo *descriptors[0];
687 };
688
689 struct radv_push_descriptor_set
690 {
691 struct radv_descriptor_set set;
692 uint32_t capacity;
693 };
694
695 struct radv_descriptor_pool_entry {
696 uint32_t offset;
697 uint32_t size;
698 struct radv_descriptor_set *set;
699 };
700
701 struct radv_descriptor_pool {
702 struct radeon_winsys_bo *bo;
703 uint8_t *mapped_ptr;
704 uint64_t current_offset;
705 uint64_t size;
706
707 uint8_t *host_memory_base;
708 uint8_t *host_memory_ptr;
709 uint8_t *host_memory_end;
710
711 uint32_t entry_count;
712 uint32_t max_entry_count;
713 struct radv_descriptor_pool_entry entries[0];
714 };
715
716 struct radv_descriptor_update_template_entry {
717 VkDescriptorType descriptor_type;
718
719 /* The number of descriptors to update */
720 uint32_t descriptor_count;
721
722 /* Into mapped_ptr or dynamic_descriptors, in units of the respective array */
723 uint32_t dst_offset;
724
725 /* In dwords. Not valid/used for dynamic descriptors */
726 uint32_t dst_stride;
727
728 uint32_t buffer_offset;
729
730 /* Only valid for combined image samplers and samplers */
731 uint16_t has_sampler;
732
733 /* In bytes */
734 size_t src_offset;
735 size_t src_stride;
736
737 /* For push descriptors */
738 const uint32_t *immutable_samplers;
739 };
740
741 struct radv_descriptor_update_template {
742 uint32_t entry_count;
743 VkPipelineBindPoint bind_point;
744 struct radv_descriptor_update_template_entry entry[0];
745 };
746
747 struct radv_buffer {
748 VkDeviceSize size;
749
750 VkBufferUsageFlags usage;
751 VkBufferCreateFlags flags;
752
753 /* Set when bound */
754 struct radeon_winsys_bo * bo;
755 VkDeviceSize offset;
756
757 bool shareable;
758 };
759
760 enum radv_dynamic_state_bits {
761 RADV_DYNAMIC_VIEWPORT = 1 << 0,
762 RADV_DYNAMIC_SCISSOR = 1 << 1,
763 RADV_DYNAMIC_LINE_WIDTH = 1 << 2,
764 RADV_DYNAMIC_DEPTH_BIAS = 1 << 3,
765 RADV_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
766 RADV_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
767 RADV_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
768 RADV_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
769 RADV_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
770 RADV_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
771 RADV_DYNAMIC_ALL = (1 << 10) - 1,
772 };
773
774 enum radv_cmd_dirty_bits {
775 /* Keep the dynamic state dirty bits in sync with
776 * enum radv_dynamic_state_bits */
777 RADV_CMD_DIRTY_DYNAMIC_VIEWPORT = 1 << 0,
778 RADV_CMD_DIRTY_DYNAMIC_SCISSOR = 1 << 1,
779 RADV_CMD_DIRTY_DYNAMIC_LINE_WIDTH = 1 << 2,
780 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BIAS = 1 << 3,
781 RADV_CMD_DIRTY_DYNAMIC_BLEND_CONSTANTS = 1 << 4,
782 RADV_CMD_DIRTY_DYNAMIC_DEPTH_BOUNDS = 1 << 5,
783 RADV_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK = 1 << 6,
784 RADV_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK = 1 << 7,
785 RADV_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE = 1 << 8,
786 RADV_CMD_DIRTY_DYNAMIC_DISCARD_RECTANGLE = 1 << 9,
787 RADV_CMD_DIRTY_DYNAMIC_ALL = (1 << 10) - 1,
788 RADV_CMD_DIRTY_PIPELINE = 1 << 10,
789 RADV_CMD_DIRTY_INDEX_BUFFER = 1 << 11,
790 RADV_CMD_DIRTY_FRAMEBUFFER = 1 << 12,
791 RADV_CMD_DIRTY_VERTEX_BUFFER = 1 << 13,
792 };
793
794 enum radv_cmd_flush_bits {
795 RADV_CMD_FLAG_INV_ICACHE = 1 << 0,
796 /* SMEM L1, other names: KCACHE, constant cache, DCACHE, data cache */
797 RADV_CMD_FLAG_INV_SMEM_L1 = 1 << 1,
798 /* VMEM L1 can optionally be bypassed (GLC=1). Other names: TC L1 */
799 RADV_CMD_FLAG_INV_VMEM_L1 = 1 << 2,
800 /* Used by everything except CB/DB, can be bypassed (SLC=1). Other names: TC L2 */
801 RADV_CMD_FLAG_INV_GLOBAL_L2 = 1 << 3,
802 /* Same as above, but only writes back and doesn't invalidate */
803 RADV_CMD_FLAG_WRITEBACK_GLOBAL_L2 = 1 << 4,
804 /* Framebuffer caches */
805 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META = 1 << 5,
806 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META = 1 << 6,
807 RADV_CMD_FLAG_FLUSH_AND_INV_DB = 1 << 7,
808 RADV_CMD_FLAG_FLUSH_AND_INV_CB = 1 << 8,
809 /* Engine synchronization. */
810 RADV_CMD_FLAG_VS_PARTIAL_FLUSH = 1 << 9,
811 RADV_CMD_FLAG_PS_PARTIAL_FLUSH = 1 << 10,
812 RADV_CMD_FLAG_CS_PARTIAL_FLUSH = 1 << 11,
813 RADV_CMD_FLAG_VGT_FLUSH = 1 << 12,
814
815 RADV_CMD_FLUSH_AND_INV_FRAMEBUFFER = (RADV_CMD_FLAG_FLUSH_AND_INV_CB |
816 RADV_CMD_FLAG_FLUSH_AND_INV_CB_META |
817 RADV_CMD_FLAG_FLUSH_AND_INV_DB |
818 RADV_CMD_FLAG_FLUSH_AND_INV_DB_META)
819 };
820
821 struct radv_vertex_binding {
822 struct radv_buffer * buffer;
823 VkDeviceSize offset;
824 };
825
826 struct radv_viewport_state {
827 uint32_t count;
828 VkViewport viewports[MAX_VIEWPORTS];
829 };
830
831 struct radv_scissor_state {
832 uint32_t count;
833 VkRect2D scissors[MAX_SCISSORS];
834 };
835
836 struct radv_discard_rectangle_state {
837 uint32_t count;
838 VkRect2D rectangles[MAX_DISCARD_RECTANGLES];
839 };
840
841 struct radv_dynamic_state {
842 /**
843 * Bitmask of (1 << VK_DYNAMIC_STATE_*).
844 * Defines the set of saved dynamic state.
845 */
846 uint32_t mask;
847
848 struct radv_viewport_state viewport;
849
850 struct radv_scissor_state scissor;
851
852 float line_width;
853
854 struct {
855 float bias;
856 float clamp;
857 float slope;
858 } depth_bias;
859
860 float blend_constants[4];
861
862 struct {
863 float min;
864 float max;
865 } depth_bounds;
866
867 struct {
868 uint32_t front;
869 uint32_t back;
870 } stencil_compare_mask;
871
872 struct {
873 uint32_t front;
874 uint32_t back;
875 } stencil_write_mask;
876
877 struct {
878 uint32_t front;
879 uint32_t back;
880 } stencil_reference;
881
882 struct radv_discard_rectangle_state discard_rectangle;
883 };
884
885 extern const struct radv_dynamic_state default_dynamic_state;
886
887 const char *
888 radv_get_debug_option_name(int id);
889
890 const char *
891 radv_get_perftest_option_name(int id);
892
893 /**
894 * Attachment state when recording a renderpass instance.
895 *
896 * The clear value is valid only if there exists a pending clear.
897 */
898 struct radv_attachment_state {
899 VkImageAspectFlags pending_clear_aspects;
900 uint32_t cleared_views;
901 VkClearValue clear_value;
902 VkImageLayout current_layout;
903 };
904
905 struct radv_descriptor_state {
906 struct radv_descriptor_set *sets[MAX_SETS];
907 uint32_t dirty;
908 uint32_t valid;
909 struct radv_push_descriptor_set push_set;
910 bool push_dirty;
911 };
912
913 struct radv_cmd_state {
914 /* Vertex descriptors */
915 bool vb_prefetch_dirty;
916 uint64_t vb_va;
917 unsigned vb_size;
918
919 bool predicating;
920 uint32_t dirty;
921
922 struct radv_pipeline * pipeline;
923 struct radv_pipeline * emitted_pipeline;
924 struct radv_pipeline * compute_pipeline;
925 struct radv_pipeline * emitted_compute_pipeline;
926 struct radv_framebuffer * framebuffer;
927 struct radv_render_pass * pass;
928 const struct radv_subpass * subpass;
929 struct radv_dynamic_state dynamic;
930 struct radv_attachment_state * attachments;
931 VkRect2D render_area;
932
933 /* Index buffer */
934 struct radv_buffer *index_buffer;
935 uint64_t index_offset;
936 uint32_t index_type;
937 uint32_t max_index_count;
938 uint64_t index_va;
939 int32_t last_index_type;
940
941 int32_t last_primitive_reset_en;
942 uint32_t last_primitive_reset_index;
943 enum radv_cmd_flush_bits flush_bits;
944 unsigned active_occlusion_queries;
945 float offset_scale;
946 uint32_t trace_id;
947 uint32_t last_ia_multi_vgt_param;
948
949 uint32_t last_num_instances;
950 uint32_t last_first_instance;
951 uint32_t last_vertex_offset;
952 };
953
954 struct radv_cmd_pool {
955 VkAllocationCallbacks alloc;
956 struct list_head cmd_buffers;
957 struct list_head free_cmd_buffers;
958 uint32_t queue_family_index;
959 };
960
961 struct radv_cmd_buffer_upload {
962 uint8_t *map;
963 unsigned offset;
964 uint64_t size;
965 struct radeon_winsys_bo *upload_bo;
966 struct list_head list;
967 };
968
969 enum radv_cmd_buffer_status {
970 RADV_CMD_BUFFER_STATUS_INVALID,
971 RADV_CMD_BUFFER_STATUS_INITIAL,
972 RADV_CMD_BUFFER_STATUS_RECORDING,
973 RADV_CMD_BUFFER_STATUS_EXECUTABLE,
974 RADV_CMD_BUFFER_STATUS_PENDING,
975 };
976
977 struct radv_cmd_buffer {
978 VK_LOADER_DATA _loader_data;
979
980 struct radv_device * device;
981
982 struct radv_cmd_pool * pool;
983 struct list_head pool_link;
984
985 VkCommandBufferUsageFlags usage_flags;
986 VkCommandBufferLevel level;
987 enum radv_cmd_buffer_status status;
988 struct radeon_winsys_cs *cs;
989 struct radv_cmd_state state;
990 struct radv_vertex_binding vertex_bindings[MAX_VBS];
991 uint32_t queue_family_index;
992
993 uint8_t push_constants[MAX_PUSH_CONSTANTS_SIZE];
994 uint32_t dynamic_buffers[4 * MAX_DYNAMIC_BUFFERS];
995 VkShaderStageFlags push_constant_stages;
996 struct radv_descriptor_set meta_push_descriptors;
997
998 struct radv_descriptor_state descriptors[VK_PIPELINE_BIND_POINT_RANGE_SIZE];
999
1000 struct radv_cmd_buffer_upload upload;
1001
1002 uint32_t scratch_size_needed;
1003 uint32_t compute_scratch_size_needed;
1004 uint32_t esgs_ring_size_needed;
1005 uint32_t gsvs_ring_size_needed;
1006 bool tess_rings_needed;
1007 bool sample_positions_needed;
1008
1009 VkResult record_result;
1010
1011 int ring_offsets_idx; /* just used for verification */
1012 uint32_t gfx9_fence_offset;
1013 struct radeon_winsys_bo *gfx9_fence_bo;
1014 uint32_t gfx9_fence_idx;
1015
1016 /**
1017 * Whether a query pool has been resetted and we have to flush caches.
1018 */
1019 bool pending_reset_query;
1020 };
1021
1022 struct radv_image;
1023
1024 bool radv_cmd_buffer_uses_mec(struct radv_cmd_buffer *cmd_buffer);
1025
1026 void si_init_compute(struct radv_cmd_buffer *cmd_buffer);
1027 void si_init_config(struct radv_cmd_buffer *cmd_buffer);
1028
1029 void cik_create_gfx_config(struct radv_device *device);
1030
1031 void si_write_viewport(struct radeon_winsys_cs *cs, int first_vp,
1032 int count, const VkViewport *viewports);
1033 void si_write_scissors(struct radeon_winsys_cs *cs, int first,
1034 int count, const VkRect2D *scissors,
1035 const VkViewport *viewports, bool can_use_guardband);
1036 uint32_t si_get_ia_multi_vgt_param(struct radv_cmd_buffer *cmd_buffer,
1037 bool instanced_draw, bool indirect_draw,
1038 uint32_t draw_vertex_count);
1039 void si_cs_emit_write_event_eop(struct radeon_winsys_cs *cs,
1040 bool predicated,
1041 enum chip_class chip_class,
1042 bool is_mec,
1043 unsigned event, unsigned event_flags,
1044 unsigned data_sel,
1045 uint64_t va,
1046 uint32_t old_fence,
1047 uint32_t new_fence);
1048
1049 void si_emit_wait_fence(struct radeon_winsys_cs *cs,
1050 bool predicated,
1051 uint64_t va, uint32_t ref,
1052 uint32_t mask);
1053 void si_cs_emit_cache_flush(struct radeon_winsys_cs *cs,
1054 enum chip_class chip_class,
1055 uint32_t *fence_ptr, uint64_t va,
1056 bool is_mec,
1057 enum radv_cmd_flush_bits flush_bits);
1058 void si_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer);
1059 void si_emit_set_predication_state(struct radv_cmd_buffer *cmd_buffer, uint64_t va);
1060 void si_cp_dma_buffer_copy(struct radv_cmd_buffer *cmd_buffer,
1061 uint64_t src_va, uint64_t dest_va,
1062 uint64_t size);
1063 void si_cp_dma_prefetch(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1064 unsigned size);
1065 void si_cp_dma_clear_buffer(struct radv_cmd_buffer *cmd_buffer, uint64_t va,
1066 uint64_t size, unsigned value);
1067 void radv_set_db_count_control(struct radv_cmd_buffer *cmd_buffer);
1068 bool
1069 radv_cmd_buffer_upload_alloc(struct radv_cmd_buffer *cmd_buffer,
1070 unsigned size,
1071 unsigned alignment,
1072 unsigned *out_offset,
1073 void **ptr);
1074 void
1075 radv_cmd_buffer_set_subpass(struct radv_cmd_buffer *cmd_buffer,
1076 const struct radv_subpass *subpass,
1077 bool transitions);
1078 bool
1079 radv_cmd_buffer_upload_data(struct radv_cmd_buffer *cmd_buffer,
1080 unsigned size, unsigned alignmnet,
1081 const void *data, unsigned *out_offset);
1082
1083 void radv_cmd_buffer_clear_subpass(struct radv_cmd_buffer *cmd_buffer);
1084 void radv_cmd_buffer_resolve_subpass(struct radv_cmd_buffer *cmd_buffer);
1085 void radv_cmd_buffer_resolve_subpass_cs(struct radv_cmd_buffer *cmd_buffer);
1086 void radv_cmd_buffer_resolve_subpass_fs(struct radv_cmd_buffer *cmd_buffer);
1087 void radv_cayman_emit_msaa_sample_locs(struct radeon_winsys_cs *cs, int nr_samples);
1088 unsigned radv_cayman_get_maxdist(int log_samples);
1089 void radv_device_init_msaa(struct radv_device *device);
1090 void radv_set_depth_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1091 struct radv_image *image,
1092 VkClearDepthStencilValue ds_clear_value,
1093 VkImageAspectFlags aspects);
1094 void radv_set_color_clear_regs(struct radv_cmd_buffer *cmd_buffer,
1095 struct radv_image *image,
1096 int idx,
1097 uint32_t color_values[2]);
1098 void radv_set_dcc_need_cmask_elim_pred(struct radv_cmd_buffer *cmd_buffer,
1099 struct radv_image *image,
1100 bool value);
1101 uint32_t radv_fill_buffer(struct radv_cmd_buffer *cmd_buffer,
1102 struct radeon_winsys_bo *bo,
1103 uint64_t offset, uint64_t size, uint32_t value);
1104 void radv_cmd_buffer_trace_emit(struct radv_cmd_buffer *cmd_buffer);
1105 bool radv_get_memory_fd(struct radv_device *device,
1106 struct radv_device_memory *memory,
1107 int *pFD);
1108
1109 static inline struct radv_descriptor_state *
1110 radv_get_descriptors_state(struct radv_cmd_buffer *cmd_buffer,
1111 VkPipelineBindPoint bind_point)
1112 {
1113 assert(bind_point == VK_PIPELINE_BIND_POINT_GRAPHICS ||
1114 bind_point == VK_PIPELINE_BIND_POINT_COMPUTE);
1115 return &cmd_buffer->descriptors[bind_point];
1116 }
1117
1118 /*
1119 * Takes x,y,z as exact numbers of invocations, instead of blocks.
1120 *
1121 * Limitations: Can't call normal dispatch functions without binding or rebinding
1122 * the compute pipeline.
1123 */
1124 void radv_unaligned_dispatch(
1125 struct radv_cmd_buffer *cmd_buffer,
1126 uint32_t x,
1127 uint32_t y,
1128 uint32_t z);
1129
1130 struct radv_event {
1131 struct radeon_winsys_bo *bo;
1132 uint64_t *map;
1133 };
1134
1135 struct radv_shader_module;
1136
1137 #define RADV_HASH_SHADER_IS_GEOM_COPY_SHADER (1 << 0)
1138 #define RADV_HASH_SHADER_SISCHED (1 << 1)
1139 #define RADV_HASH_SHADER_UNSAFE_MATH (1 << 2)
1140 void
1141 radv_hash_shaders(unsigned char *hash,
1142 const VkPipelineShaderStageCreateInfo **stages,
1143 const struct radv_pipeline_layout *layout,
1144 const struct radv_pipeline_key *key,
1145 uint32_t flags);
1146
1147 static inline gl_shader_stage
1148 vk_to_mesa_shader_stage(VkShaderStageFlagBits vk_stage)
1149 {
1150 assert(__builtin_popcount(vk_stage) == 1);
1151 return ffs(vk_stage) - 1;
1152 }
1153
1154 static inline VkShaderStageFlagBits
1155 mesa_to_vk_shader_stage(gl_shader_stage mesa_stage)
1156 {
1157 return (1 << mesa_stage);
1158 }
1159
1160 #define RADV_STAGE_MASK ((1 << MESA_SHADER_STAGES) - 1)
1161
1162 #define radv_foreach_stage(stage, stage_bits) \
1163 for (gl_shader_stage stage, \
1164 __tmp = (gl_shader_stage)((stage_bits) & RADV_STAGE_MASK); \
1165 stage = __builtin_ffs(__tmp) - 1, __tmp; \
1166 __tmp &= ~(1 << (stage)))
1167
1168 unsigned radv_format_meta_fs_key(VkFormat format);
1169
1170 struct radv_multisample_state {
1171 uint32_t db_eqaa;
1172 uint32_t pa_sc_line_cntl;
1173 uint32_t pa_sc_mode_cntl_0;
1174 uint32_t pa_sc_mode_cntl_1;
1175 uint32_t pa_sc_aa_config;
1176 uint32_t pa_sc_aa_mask[2];
1177 unsigned num_samples;
1178 };
1179
1180 struct radv_prim_vertex_count {
1181 uint8_t min;
1182 uint8_t incr;
1183 };
1184
1185 struct radv_vertex_elements_info {
1186 uint32_t rsrc_word3[MAX_VERTEX_ATTRIBS];
1187 uint32_t format_size[MAX_VERTEX_ATTRIBS];
1188 uint32_t binding[MAX_VERTEX_ATTRIBS];
1189 uint32_t offset[MAX_VERTEX_ATTRIBS];
1190 uint32_t count;
1191 };
1192
1193 struct radv_ia_multi_vgt_param_helpers {
1194 uint32_t base;
1195 bool partial_es_wave;
1196 uint8_t primgroup_size;
1197 bool wd_switch_on_eop;
1198 bool ia_switch_on_eoi;
1199 bool partial_vs_wave;
1200 };
1201
1202 #define SI_GS_PER_ES 128
1203
1204 struct radv_pipeline {
1205 struct radv_device * device;
1206 struct radv_dynamic_state dynamic_state;
1207
1208 struct radv_pipeline_layout * layout;
1209
1210 bool needs_data_cache;
1211 bool need_indirect_descriptor_sets;
1212 struct radv_shader_variant * shaders[MESA_SHADER_STAGES];
1213 struct radv_shader_variant *gs_copy_shader;
1214 VkShaderStageFlags active_stages;
1215
1216 struct radeon_winsys_cs cs;
1217
1218 struct radv_vertex_elements_info vertex_elements;
1219
1220 uint32_t binding_stride[MAX_VBS];
1221
1222 uint32_t user_data_0[MESA_SHADER_STAGES];
1223 union {
1224 struct {
1225 struct radv_multisample_state ms;
1226 uint32_t spi_baryc_cntl;
1227 bool prim_restart_enable;
1228 unsigned esgs_ring_size;
1229 unsigned gsvs_ring_size;
1230 uint32_t vtx_base_sgpr;
1231 struct radv_ia_multi_vgt_param_helpers ia_multi_vgt_param;
1232 uint8_t vtx_emit_num;
1233 struct radv_prim_vertex_count prim_vertex_count;
1234 bool can_use_guardband;
1235 uint32_t needed_dynamic_state;
1236 } graphics;
1237 };
1238
1239 unsigned max_waves;
1240 unsigned scratch_bytes_per_wave;
1241 };
1242
1243 static inline bool radv_pipeline_has_gs(const struct radv_pipeline *pipeline)
1244 {
1245 return pipeline->shaders[MESA_SHADER_GEOMETRY] ? true : false;
1246 }
1247
1248 static inline bool radv_pipeline_has_tess(const struct radv_pipeline *pipeline)
1249 {
1250 return pipeline->shaders[MESA_SHADER_TESS_CTRL] ? true : false;
1251 }
1252
1253 struct radv_userdata_info *radv_lookup_user_sgpr(struct radv_pipeline *pipeline,
1254 gl_shader_stage stage,
1255 int idx);
1256
1257 struct radv_shader_variant *radv_get_vertex_shader(struct radv_pipeline *pipeline);
1258
1259 struct radv_graphics_pipeline_create_info {
1260 bool use_rectlist;
1261 bool db_depth_clear;
1262 bool db_stencil_clear;
1263 bool db_depth_disable_expclear;
1264 bool db_stencil_disable_expclear;
1265 bool db_flush_depth_inplace;
1266 bool db_flush_stencil_inplace;
1267 bool db_resummarize;
1268 uint32_t custom_blend_mode;
1269 };
1270
1271 VkResult
1272 radv_graphics_pipeline_create(VkDevice device,
1273 VkPipelineCache cache,
1274 const VkGraphicsPipelineCreateInfo *pCreateInfo,
1275 const struct radv_graphics_pipeline_create_info *extra,
1276 const VkAllocationCallbacks *alloc,
1277 VkPipeline *pPipeline);
1278
1279 struct vk_format_description;
1280 uint32_t radv_translate_buffer_dataformat(const struct vk_format_description *desc,
1281 int first_non_void);
1282 uint32_t radv_translate_buffer_numformat(const struct vk_format_description *desc,
1283 int first_non_void);
1284 uint32_t radv_translate_colorformat(VkFormat format);
1285 uint32_t radv_translate_color_numformat(VkFormat format,
1286 const struct vk_format_description *desc,
1287 int first_non_void);
1288 uint32_t radv_colorformat_endian_swap(uint32_t colorformat);
1289 unsigned radv_translate_colorswap(VkFormat format, bool do_endian_swap);
1290 uint32_t radv_translate_dbformat(VkFormat format);
1291 uint32_t radv_translate_tex_dataformat(VkFormat format,
1292 const struct vk_format_description *desc,
1293 int first_non_void);
1294 uint32_t radv_translate_tex_numformat(VkFormat format,
1295 const struct vk_format_description *desc,
1296 int first_non_void);
1297 bool radv_format_pack_clear_color(VkFormat format,
1298 uint32_t clear_vals[2],
1299 VkClearColorValue *value);
1300 bool radv_is_colorbuffer_format_supported(VkFormat format, bool *blendable);
1301 bool radv_dcc_formats_compatible(VkFormat format1,
1302 VkFormat format2);
1303
1304 struct radv_fmask_info {
1305 uint64_t offset;
1306 uint64_t size;
1307 unsigned alignment;
1308 unsigned pitch_in_pixels;
1309 unsigned bank_height;
1310 unsigned slice_tile_max;
1311 unsigned tile_mode_index;
1312 unsigned tile_swizzle;
1313 };
1314
1315 struct radv_cmask_info {
1316 uint64_t offset;
1317 uint64_t size;
1318 unsigned alignment;
1319 unsigned slice_tile_max;
1320 };
1321
1322 struct radv_image {
1323 VkImageType type;
1324 /* The original VkFormat provided by the client. This may not match any
1325 * of the actual surface formats.
1326 */
1327 VkFormat vk_format;
1328 VkImageAspectFlags aspects;
1329 VkImageUsageFlags usage; /**< Superset of VkImageCreateInfo::usage. */
1330 struct ac_surf_info info;
1331 VkImageTiling tiling; /** VkImageCreateInfo::tiling */
1332 VkImageCreateFlags flags; /** VkImageCreateInfo::flags */
1333
1334 VkDeviceSize size;
1335 uint32_t alignment;
1336
1337 unsigned queue_family_mask;
1338 bool exclusive;
1339 bool shareable;
1340
1341 /* Set when bound */
1342 struct radeon_winsys_bo *bo;
1343 VkDeviceSize offset;
1344 uint64_t dcc_offset;
1345 uint64_t htile_offset;
1346 bool tc_compatible_htile;
1347 struct radeon_surf surface;
1348
1349 struct radv_fmask_info fmask;
1350 struct radv_cmask_info cmask;
1351 uint64_t clear_value_offset;
1352 uint64_t dcc_pred_offset;
1353
1354 /* For VK_ANDROID_native_buffer, the WSI image owns the memory, */
1355 VkDeviceMemory owned_memory;
1356 };
1357
1358 /* Whether the image has a htile that is known consistent with the contents of
1359 * the image. */
1360 bool radv_layout_has_htile(const struct radv_image *image,
1361 VkImageLayout layout,
1362 unsigned queue_mask);
1363
1364 /* Whether the image has a htile that is known consistent with the contents of
1365 * the image and is allowed to be in compressed form.
1366 *
1367 * If this is false reads that don't use the htile should be able to return
1368 * correct results.
1369 */
1370 bool radv_layout_is_htile_compressed(const struct radv_image *image,
1371 VkImageLayout layout,
1372 unsigned queue_mask);
1373
1374 bool radv_layout_can_fast_clear(const struct radv_image *image,
1375 VkImageLayout layout,
1376 unsigned queue_mask);
1377
1378 bool radv_layout_dcc_compressed(const struct radv_image *image,
1379 VkImageLayout layout,
1380 unsigned queue_mask);
1381
1382 static inline bool
1383 radv_vi_dcc_enabled(const struct radv_image *image, unsigned level)
1384 {
1385 return image->surface.dcc_size && level < image->surface.num_dcc_levels;
1386 }
1387
1388 static inline bool
1389 radv_htile_enabled(const struct radv_image *image, unsigned level)
1390 {
1391 return image->surface.htile_size && level == 0;
1392 }
1393
1394 unsigned radv_image_queue_family_mask(const struct radv_image *image, uint32_t family, uint32_t queue_family);
1395
1396 static inline uint32_t
1397 radv_get_layerCount(const struct radv_image *image,
1398 const VkImageSubresourceRange *range)
1399 {
1400 return range->layerCount == VK_REMAINING_ARRAY_LAYERS ?
1401 image->info.array_size - range->baseArrayLayer : range->layerCount;
1402 }
1403
1404 static inline uint32_t
1405 radv_get_levelCount(const struct radv_image *image,
1406 const VkImageSubresourceRange *range)
1407 {
1408 return range->levelCount == VK_REMAINING_MIP_LEVELS ?
1409 image->info.levels - range->baseMipLevel : range->levelCount;
1410 }
1411
1412 struct radeon_bo_metadata;
1413 void
1414 radv_init_metadata(struct radv_device *device,
1415 struct radv_image *image,
1416 struct radeon_bo_metadata *metadata);
1417
1418 struct radv_image_view {
1419 struct radv_image *image; /**< VkImageViewCreateInfo::image */
1420 struct radeon_winsys_bo *bo;
1421
1422 VkImageViewType type;
1423 VkImageAspectFlags aspect_mask;
1424 VkFormat vk_format;
1425 uint32_t base_layer;
1426 uint32_t layer_count;
1427 uint32_t base_mip;
1428 uint32_t level_count;
1429 VkExtent3D extent; /**< Extent of VkImageViewCreateInfo::baseMipLevel. */
1430
1431 uint32_t descriptor[16];
1432
1433 /* Descriptor for use as a storage image as opposed to a sampled image.
1434 * This has a few differences for cube maps (e.g. type).
1435 */
1436 uint32_t storage_descriptor[16];
1437 };
1438
1439 struct radv_image_create_info {
1440 const VkImageCreateInfo *vk_info;
1441 bool scanout;
1442 bool no_metadata_planes;
1443 };
1444
1445 VkResult radv_image_create(VkDevice _device,
1446 const struct radv_image_create_info *info,
1447 const VkAllocationCallbacks* alloc,
1448 VkImage *pImage);
1449
1450 VkResult
1451 radv_image_from_gralloc(VkDevice device_h,
1452 const VkImageCreateInfo *base_info,
1453 const VkNativeBufferANDROID *gralloc_info,
1454 const VkAllocationCallbacks *alloc,
1455 VkImage *out_image_h);
1456
1457 void radv_image_view_init(struct radv_image_view *view,
1458 struct radv_device *device,
1459 const VkImageViewCreateInfo* pCreateInfo);
1460
1461 struct radv_buffer_view {
1462 struct radeon_winsys_bo *bo;
1463 VkFormat vk_format;
1464 uint64_t range; /**< VkBufferViewCreateInfo::range */
1465 uint32_t state[4];
1466 };
1467 void radv_buffer_view_init(struct radv_buffer_view *view,
1468 struct radv_device *device,
1469 const VkBufferViewCreateInfo* pCreateInfo);
1470
1471 static inline struct VkExtent3D
1472 radv_sanitize_image_extent(const VkImageType imageType,
1473 const struct VkExtent3D imageExtent)
1474 {
1475 switch (imageType) {
1476 case VK_IMAGE_TYPE_1D:
1477 return (VkExtent3D) { imageExtent.width, 1, 1 };
1478 case VK_IMAGE_TYPE_2D:
1479 return (VkExtent3D) { imageExtent.width, imageExtent.height, 1 };
1480 case VK_IMAGE_TYPE_3D:
1481 return imageExtent;
1482 default:
1483 unreachable("invalid image type");
1484 }
1485 }
1486
1487 static inline struct VkOffset3D
1488 radv_sanitize_image_offset(const VkImageType imageType,
1489 const struct VkOffset3D imageOffset)
1490 {
1491 switch (imageType) {
1492 case VK_IMAGE_TYPE_1D:
1493 return (VkOffset3D) { imageOffset.x, 0, 0 };
1494 case VK_IMAGE_TYPE_2D:
1495 return (VkOffset3D) { imageOffset.x, imageOffset.y, 0 };
1496 case VK_IMAGE_TYPE_3D:
1497 return imageOffset;
1498 default:
1499 unreachable("invalid image type");
1500 }
1501 }
1502
1503 static inline bool
1504 radv_image_extent_compare(const struct radv_image *image,
1505 const VkExtent3D *extent)
1506 {
1507 if (extent->width != image->info.width ||
1508 extent->height != image->info.height ||
1509 extent->depth != image->info.depth)
1510 return false;
1511 return true;
1512 }
1513
1514 struct radv_sampler {
1515 uint32_t state[4];
1516 };
1517
1518 struct radv_color_buffer_info {
1519 uint64_t cb_color_base;
1520 uint64_t cb_color_cmask;
1521 uint64_t cb_color_fmask;
1522 uint64_t cb_dcc_base;
1523 uint32_t cb_color_pitch;
1524 uint32_t cb_color_slice;
1525 uint32_t cb_color_view;
1526 uint32_t cb_color_info;
1527 uint32_t cb_color_attrib;
1528 uint32_t cb_color_attrib2;
1529 uint32_t cb_dcc_control;
1530 uint32_t cb_color_cmask_slice;
1531 uint32_t cb_color_fmask_slice;
1532 };
1533
1534 struct radv_ds_buffer_info {
1535 uint64_t db_z_read_base;
1536 uint64_t db_stencil_read_base;
1537 uint64_t db_z_write_base;
1538 uint64_t db_stencil_write_base;
1539 uint64_t db_htile_data_base;
1540 uint32_t db_depth_info;
1541 uint32_t db_z_info;
1542 uint32_t db_stencil_info;
1543 uint32_t db_depth_view;
1544 uint32_t db_depth_size;
1545 uint32_t db_depth_slice;
1546 uint32_t db_htile_surface;
1547 uint32_t pa_su_poly_offset_db_fmt_cntl;
1548 uint32_t db_z_info2;
1549 uint32_t db_stencil_info2;
1550 float offset_scale;
1551 };
1552
1553 struct radv_attachment_info {
1554 union {
1555 struct radv_color_buffer_info cb;
1556 struct radv_ds_buffer_info ds;
1557 };
1558 struct radv_image_view *attachment;
1559 };
1560
1561 struct radv_framebuffer {
1562 uint32_t width;
1563 uint32_t height;
1564 uint32_t layers;
1565
1566 uint32_t attachment_count;
1567 struct radv_attachment_info attachments[0];
1568 };
1569
1570 struct radv_subpass_barrier {
1571 VkPipelineStageFlags src_stage_mask;
1572 VkAccessFlags src_access_mask;
1573 VkAccessFlags dst_access_mask;
1574 };
1575
1576 struct radv_subpass {
1577 uint32_t input_count;
1578 uint32_t color_count;
1579 VkAttachmentReference * input_attachments;
1580 VkAttachmentReference * color_attachments;
1581 VkAttachmentReference * resolve_attachments;
1582 VkAttachmentReference depth_stencil_attachment;
1583
1584 /** Subpass has at least one resolve attachment */
1585 bool has_resolve;
1586
1587 struct radv_subpass_barrier start_barrier;
1588
1589 uint32_t view_mask;
1590 };
1591
1592 struct radv_render_pass_attachment {
1593 VkFormat format;
1594 uint32_t samples;
1595 VkAttachmentLoadOp load_op;
1596 VkAttachmentLoadOp stencil_load_op;
1597 VkImageLayout initial_layout;
1598 VkImageLayout final_layout;
1599 uint32_t view_mask;
1600 };
1601
1602 struct radv_render_pass {
1603 uint32_t attachment_count;
1604 uint32_t subpass_count;
1605 VkAttachmentReference * subpass_attachments;
1606 struct radv_render_pass_attachment * attachments;
1607 struct radv_subpass_barrier end_barrier;
1608 struct radv_subpass subpasses[0];
1609 };
1610
1611 VkResult radv_device_init_meta(struct radv_device *device);
1612 void radv_device_finish_meta(struct radv_device *device);
1613
1614 struct radv_query_pool {
1615 struct radeon_winsys_bo *bo;
1616 uint32_t stride;
1617 uint32_t availability_offset;
1618 uint64_t size;
1619 char *ptr;
1620 VkQueryType type;
1621 uint32_t pipeline_stats_mask;
1622 };
1623
1624 struct radv_semaphore {
1625 /* use a winsys sem for non-exportable */
1626 struct radeon_winsys_sem *sem;
1627 uint32_t syncobj;
1628 uint32_t temp_syncobj;
1629 };
1630
1631 VkResult radv_alloc_sem_info(struct radv_winsys_sem_info *sem_info,
1632 int num_wait_sems,
1633 const VkSemaphore *wait_sems,
1634 int num_signal_sems,
1635 const VkSemaphore *signal_sems,
1636 VkFence fence);
1637 void radv_free_sem_info(struct radv_winsys_sem_info *sem_info);
1638
1639 void radv_set_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
1640 VkPipelineBindPoint bind_point,
1641 struct radv_descriptor_set *set,
1642 unsigned idx);
1643
1644 void
1645 radv_update_descriptor_sets(struct radv_device *device,
1646 struct radv_cmd_buffer *cmd_buffer,
1647 VkDescriptorSet overrideSet,
1648 uint32_t descriptorWriteCount,
1649 const VkWriteDescriptorSet *pDescriptorWrites,
1650 uint32_t descriptorCopyCount,
1651 const VkCopyDescriptorSet *pDescriptorCopies);
1652
1653 void
1654 radv_update_descriptor_set_with_template(struct radv_device *device,
1655 struct radv_cmd_buffer *cmd_buffer,
1656 struct radv_descriptor_set *set,
1657 VkDescriptorUpdateTemplateKHR descriptorUpdateTemplate,
1658 const void *pData);
1659
1660 void radv_meta_push_descriptor_set(struct radv_cmd_buffer *cmd_buffer,
1661 VkPipelineBindPoint pipelineBindPoint,
1662 VkPipelineLayout _layout,
1663 uint32_t set,
1664 uint32_t descriptorWriteCount,
1665 const VkWriteDescriptorSet *pDescriptorWrites);
1666
1667 void radv_initialise_cmask(struct radv_cmd_buffer *cmd_buffer,
1668 struct radv_image *image, uint32_t value);
1669 void radv_initialize_dcc(struct radv_cmd_buffer *cmd_buffer,
1670 struct radv_image *image, uint32_t value);
1671
1672 struct radv_fence {
1673 struct radeon_winsys_fence *fence;
1674 bool submitted;
1675 bool signalled;
1676
1677 uint32_t syncobj;
1678 uint32_t temp_syncobj;
1679 };
1680
1681 /* radv_nir_to_llvm.c */
1682 struct radv_shader_variant_info;
1683 struct radv_nir_compiler_options;
1684
1685 void radv_compile_gs_copy_shader(LLVMTargetMachineRef tm,
1686 struct nir_shader *geom_shader,
1687 struct ac_shader_binary *binary,
1688 struct ac_shader_config *config,
1689 struct radv_shader_variant_info *shader_info,
1690 const struct radv_nir_compiler_options *option);
1691
1692 void radv_compile_nir_shader(LLVMTargetMachineRef tm,
1693 struct ac_shader_binary *binary,
1694 struct ac_shader_config *config,
1695 struct radv_shader_variant_info *shader_info,
1696 struct nir_shader *const *nir,
1697 int nir_count,
1698 const struct radv_nir_compiler_options *options);
1699
1700 /* radv_shader_info.h */
1701 struct radv_shader_info;
1702
1703 void radv_nir_shader_info_pass(const struct nir_shader *nir,
1704 const struct radv_nir_compiler_options *options,
1705 struct radv_shader_info *info);
1706
1707 struct radeon_winsys_sem;
1708
1709 #define RADV_DEFINE_HANDLE_CASTS(__radv_type, __VkType) \
1710 \
1711 static inline struct __radv_type * \
1712 __radv_type ## _from_handle(__VkType _handle) \
1713 { \
1714 return (struct __radv_type *) _handle; \
1715 } \
1716 \
1717 static inline __VkType \
1718 __radv_type ## _to_handle(struct __radv_type *_obj) \
1719 { \
1720 return (__VkType) _obj; \
1721 }
1722
1723 #define RADV_DEFINE_NONDISP_HANDLE_CASTS(__radv_type, __VkType) \
1724 \
1725 static inline struct __radv_type * \
1726 __radv_type ## _from_handle(__VkType _handle) \
1727 { \
1728 return (struct __radv_type *)(uintptr_t) _handle; \
1729 } \
1730 \
1731 static inline __VkType \
1732 __radv_type ## _to_handle(struct __radv_type *_obj) \
1733 { \
1734 return (__VkType)(uintptr_t) _obj; \
1735 }
1736
1737 #define RADV_FROM_HANDLE(__radv_type, __name, __handle) \
1738 struct __radv_type *__name = __radv_type ## _from_handle(__handle)
1739
1740 RADV_DEFINE_HANDLE_CASTS(radv_cmd_buffer, VkCommandBuffer)
1741 RADV_DEFINE_HANDLE_CASTS(radv_device, VkDevice)
1742 RADV_DEFINE_HANDLE_CASTS(radv_instance, VkInstance)
1743 RADV_DEFINE_HANDLE_CASTS(radv_physical_device, VkPhysicalDevice)
1744 RADV_DEFINE_HANDLE_CASTS(radv_queue, VkQueue)
1745
1746 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_cmd_pool, VkCommandPool)
1747 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer, VkBuffer)
1748 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_buffer_view, VkBufferView)
1749 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_pool, VkDescriptorPool)
1750 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set, VkDescriptorSet)
1751 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_set_layout, VkDescriptorSetLayout)
1752 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_descriptor_update_template, VkDescriptorUpdateTemplateKHR)
1753 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_device_memory, VkDeviceMemory)
1754 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_fence, VkFence)
1755 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_event, VkEvent)
1756 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_framebuffer, VkFramebuffer)
1757 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image, VkImage)
1758 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_image_view, VkImageView);
1759 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_cache, VkPipelineCache)
1760 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline, VkPipeline)
1761 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_pipeline_layout, VkPipelineLayout)
1762 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_query_pool, VkQueryPool)
1763 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_render_pass, VkRenderPass)
1764 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_sampler, VkSampler)
1765 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_shader_module, VkShaderModule)
1766 RADV_DEFINE_NONDISP_HANDLE_CASTS(radv_semaphore, VkSemaphore)
1767
1768 #endif /* RADV_PRIVATE_H */