1 | /**************************************************************************/ |
2 | /* rendering_device_vulkan.h */ |
3 | /**************************************************************************/ |
4 | /* This file is part of: */ |
5 | /* GODOT ENGINE */ |
6 | /* https://godotengine.org */ |
7 | /**************************************************************************/ |
8 | /* Copyright (c) 2014-present Godot Engine contributors (see AUTHORS.md). */ |
9 | /* Copyright (c) 2007-2014 Juan Linietsky, Ariel Manzur. */ |
10 | /* */ |
11 | /* Permission is hereby granted, free of charge, to any person obtaining */ |
12 | /* a copy of this software and associated documentation files (the */ |
13 | /* "Software"), to deal in the Software without restriction, including */ |
14 | /* without limitation the rights to use, copy, modify, merge, publish, */ |
15 | /* distribute, sublicense, and/or sell copies of the Software, and to */ |
16 | /* permit persons to whom the Software is furnished to do so, subject to */ |
17 | /* the following conditions: */ |
18 | /* */ |
19 | /* The above copyright notice and this permission notice shall be */ |
20 | /* included in all copies or substantial portions of the Software. */ |
21 | /* */ |
22 | /* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, */ |
23 | /* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF */ |
24 | /* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. */ |
25 | /* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY */ |
26 | /* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, */ |
27 | /* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE */ |
28 | /* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. */ |
29 | /**************************************************************************/ |
30 | |
31 | #ifndef RENDERING_DEVICE_VULKAN_H |
32 | #define RENDERING_DEVICE_VULKAN_H |
33 | |
34 | #include "core/object/worker_thread_pool.h" |
35 | #include "core/os/thread_safe.h" |
36 | #include "core/templates/local_vector.h" |
37 | #include "core/templates/oa_hash_map.h" |
38 | #include "core/templates/rid_owner.h" |
39 | #include "servers/rendering/rendering_device.h" |
40 | |
41 | #ifdef DEBUG_ENABLED |
42 | #ifndef _DEBUG |
43 | #define _DEBUG |
44 | #endif |
45 | #endif |
46 | #include "vk_mem_alloc.h" |
47 | |
48 | #ifdef USE_VOLK |
49 | #include <volk.h> |
50 | #else |
51 | #include <vulkan/vulkan.h> |
52 | #endif |
53 | |
54 | class VulkanContext; |
55 | |
56 | class RenderingDeviceVulkan : public RenderingDevice { |
57 | _THREAD_SAFE_CLASS_ |
58 | |
59 | // Miscellaneous tables that map |
60 | // our enums to enums used |
61 | // by vulkan. |
62 | |
63 | VkPhysicalDeviceLimits limits; |
64 | static const VkFormat vulkan_formats[DATA_FORMAT_MAX]; |
65 | static const char *named_formats[DATA_FORMAT_MAX]; |
66 | static const VkCompareOp compare_operators[COMPARE_OP_MAX]; |
67 | static const VkStencilOp stencil_operations[STENCIL_OP_MAX]; |
68 | static const VkSampleCountFlagBits rasterization_sample_count[TEXTURE_SAMPLES_MAX]; |
69 | static const VkLogicOp logic_operations[RenderingDevice::LOGIC_OP_MAX]; |
70 | static const VkBlendFactor blend_factors[RenderingDevice::BLEND_FACTOR_MAX]; |
71 | static const VkBlendOp blend_operations[RenderingDevice::BLEND_OP_MAX]; |
72 | static const VkSamplerAddressMode address_modes[SAMPLER_REPEAT_MODE_MAX]; |
73 | static const VkBorderColor sampler_border_colors[SAMPLER_BORDER_COLOR_MAX]; |
74 | static const VkImageType vulkan_image_type[TEXTURE_TYPE_MAX]; |
75 | |
76 | // Functions used for format |
77 | // validation, and ensures the |
78 | // user passes valid data. |
79 | |
80 | static int get_format_vertex_size(DataFormat p_format); |
81 | static uint32_t get_image_format_pixel_size(DataFormat p_format); |
82 | static void get_compressed_image_format_block_dimensions(DataFormat p_format, uint32_t &r_w, uint32_t &r_h); |
83 | uint32_t get_compressed_image_format_block_byte_size(DataFormat p_format); |
84 | static uint32_t get_compressed_image_format_pixel_rshift(DataFormat p_format); |
85 | static uint32_t get_image_format_required_size(DataFormat p_format, uint32_t p_width, uint32_t p_height, uint32_t p_depth, uint32_t p_mipmaps, uint32_t *r_blockw = nullptr, uint32_t *r_blockh = nullptr, uint32_t *r_depth = nullptr); |
86 | static uint32_t get_image_required_mipmaps(uint32_t p_width, uint32_t p_height, uint32_t p_depth); |
87 | static bool format_has_stencil(DataFormat p_format); |
88 | |
89 | /***************************/ |
90 | /**** ID INFRASTRUCTURE ****/ |
91 | /***************************/ |
92 | |
93 | enum IDType { |
94 | ID_TYPE_FRAMEBUFFER_FORMAT, |
95 | ID_TYPE_VERTEX_FORMAT, |
96 | ID_TYPE_DRAW_LIST, |
97 | ID_TYPE_SPLIT_DRAW_LIST, |
98 | ID_TYPE_COMPUTE_LIST, |
99 | ID_TYPE_MAX, |
100 | ID_BASE_SHIFT = 58 // 5 bits for ID types. |
101 | }; |
102 | |
103 | VkDevice device = VK_NULL_HANDLE; |
104 | |
105 | HashMap<RID, HashSet<RID>> dependency_map; // IDs to IDs that depend on it. |
106 | HashMap<RID, HashSet<RID>> reverse_dependency_map; // Same as above, but in reverse. |
107 | |
108 | void _add_dependency(RID p_id, RID p_depends_on); |
109 | void _free_dependencies(RID p_id); |
110 | |
111 | /*****************/ |
112 | /**** TEXTURE ****/ |
113 | /*****************/ |
114 | |
115 | // In Vulkan, the concept of textures does not exist, |
116 | // instead there is the image (the memory pretty much, |
117 | // the view (how the memory is interpreted) and the |
118 | // sampler (how it's sampled from the shader). |
119 | // |
120 | // Texture here includes the first two stages, but |
121 | // It's possible to create textures sharing the image |
122 | // but with different views. The main use case for this |
123 | // is textures that can be read as both SRGB/Linear, |
124 | // or slices of a texture (a mipmap, a layer, a 3D slice) |
125 | // for a framebuffer to render into it. |
126 | |
127 | struct Texture { |
128 | VkImage image = VK_NULL_HANDLE; |
129 | VmaAllocation allocation = nullptr; |
130 | VmaAllocationInfo allocation_info; |
131 | VkImageView view = VK_NULL_HANDLE; |
132 | |
133 | TextureType type; |
134 | DataFormat format; |
135 | TextureSamples samples; |
136 | uint32_t width = 0; |
137 | uint32_t height = 0; |
138 | uint32_t depth = 0; |
139 | uint32_t layers = 0; |
140 | uint32_t mipmaps = 0; |
141 | uint32_t usage_flags = 0; |
142 | uint32_t base_mipmap = 0; |
143 | uint32_t base_layer = 0; |
144 | |
145 | Vector<DataFormat> allowed_shared_formats; |
146 | |
147 | VkImageLayout layout; |
148 | |
149 | uint64_t used_in_frame = 0; |
150 | bool used_in_transfer = false; |
151 | bool used_in_raster = false; |
152 | bool used_in_compute = false; |
153 | |
154 | bool is_resolve_buffer = false; |
155 | |
156 | uint32_t read_aspect_mask = 0; |
157 | uint32_t barrier_aspect_mask = 0; |
158 | bool bound = false; // Bound to framebffer. |
159 | RID owner; |
160 | }; |
161 | |
162 | RID_Owner<Texture, true> texture_owner; |
163 | uint32_t texture_upload_region_size_px = 0; |
164 | |
165 | Vector<uint8_t> _texture_get_data_from_image(Texture *tex, VkImage p_image, VmaAllocation p_allocation, uint32_t p_layer, bool p_2d = false); |
166 | Error _texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, BitField<BarrierMask> p_post_barrier, bool p_use_setup_queue); |
167 | |
168 | /*****************/ |
169 | /**** SAMPLER ****/ |
170 | /*****************/ |
171 | |
172 | RID_Owner<VkSampler> sampler_owner; |
173 | |
174 | /***************************/ |
175 | /**** BUFFER MANAGEMENT ****/ |
176 | /***************************/ |
177 | |
178 | // These are temporary buffers on CPU memory that hold |
179 | // the information until the CPU fetches it and places it |
180 | // either on GPU buffers, or images (textures). It ensures |
181 | // updates are properly synchronized with whatever the |
182 | // GPU is doing. |
183 | // |
184 | // The logic here is as follows, only 3 of these |
185 | // blocks are created at the beginning (one per frame) |
186 | // they can each belong to a frame (assigned to current when |
187 | // used) and they can only be reused after the same frame is |
188 | // recycled. |
189 | // |
190 | // When CPU requires to allocate more than what is available, |
191 | // more of these buffers are created. If a limit is reached, |
192 | // then a fence will ensure will wait for blocks allocated |
193 | // in previous frames are processed. If that fails, then |
194 | // another fence will ensure everything pending for the current |
195 | // frame is processed (effectively stalling). |
196 | // |
197 | // See the comments in the code to understand better how it works. |
198 | |
199 | struct StagingBufferBlock { |
200 | VkBuffer buffer = VK_NULL_HANDLE; |
201 | VmaAllocation allocation = nullptr; |
202 | uint64_t frame_used = 0; |
203 | uint32_t fill_amount = 0; |
204 | }; |
205 | |
206 | Vector<StagingBufferBlock> staging_buffer_blocks; |
207 | int staging_buffer_current = 0; |
208 | uint32_t staging_buffer_block_size = 0; |
209 | uint64_t staging_buffer_max_size = 0; |
210 | bool staging_buffer_used = false; |
211 | |
212 | Error _staging_buffer_allocate(uint32_t p_amount, uint32_t p_required_align, uint32_t &r_alloc_offset, uint32_t &r_alloc_size, bool p_can_segment = true); |
213 | Error _insert_staging_block(); |
214 | |
215 | struct Buffer { |
216 | uint32_t size = 0; |
217 | uint32_t usage = 0; |
218 | VkBuffer buffer = VK_NULL_HANDLE; |
219 | VmaAllocation allocation = nullptr; |
220 | VkDescriptorBufferInfo buffer_info; // Used for binding. |
221 | Buffer() { |
222 | } |
223 | }; |
224 | |
225 | Error _buffer_allocate(Buffer *p_buffer, uint32_t p_size, uint32_t p_usage, VmaMemoryUsage p_mem_usage, VmaAllocationCreateFlags p_mem_flags); |
226 | Error _buffer_free(Buffer *p_buffer); |
227 | Error _buffer_update(Buffer *p_buffer, size_t p_offset, const uint8_t *p_data, size_t p_data_size, bool p_use_draw_command_buffer = false, uint32_t p_required_align = 32); |
228 | |
229 | void _full_barrier(bool p_sync_with_draw); |
230 | void _memory_barrier(VkPipelineStageFlags p_src_stage_mask, VkPipelineStageFlags p_dst_stage_mask, VkAccessFlags p_src_access, VkAccessFlags p_dst_access, bool p_sync_with_draw); |
231 | void _buffer_memory_barrier(VkBuffer buffer, uint64_t p_from, uint64_t p_size, VkPipelineStageFlags p_src_stage_mask, VkPipelineStageFlags p_dst_stage_mask, VkAccessFlags p_src_access, VkAccessFlags p_dst_access, bool p_sync_with_draw); |
232 | |
233 | /*********************/ |
234 | /**** FRAMEBUFFER ****/ |
235 | /*********************/ |
236 | |
237 | // In Vulkan, framebuffers work similar to how they |
238 | // do in OpenGL, with the exception that |
239 | // the "format" (vkRenderPass) is not dynamic |
240 | // and must be more or less the same as the one |
241 | // used for the render pipelines. |
242 | |
243 | struct FramebufferFormatKey { |
244 | Vector<AttachmentFormat> attachments; |
245 | Vector<FramebufferPass> passes; |
246 | uint32_t view_count = 1; |
247 | |
248 | bool operator<(const FramebufferFormatKey &p_key) const { |
249 | if (view_count != p_key.view_count) { |
250 | return view_count < p_key.view_count; |
251 | } |
252 | |
253 | uint32_t pass_size = passes.size(); |
254 | uint32_t key_pass_size = p_key.passes.size(); |
255 | if (pass_size != key_pass_size) { |
256 | return pass_size < key_pass_size; |
257 | } |
258 | const FramebufferPass *pass_ptr = passes.ptr(); |
259 | const FramebufferPass *key_pass_ptr = p_key.passes.ptr(); |
260 | |
261 | for (uint32_t i = 0; i < pass_size; i++) { |
262 | { // Compare color attachments. |
263 | uint32_t attachment_size = pass_ptr[i].color_attachments.size(); |
264 | uint32_t key_attachment_size = key_pass_ptr[i].color_attachments.size(); |
265 | if (attachment_size != key_attachment_size) { |
266 | return attachment_size < key_attachment_size; |
267 | } |
268 | const int32_t *pass_attachment_ptr = pass_ptr[i].color_attachments.ptr(); |
269 | const int32_t *key_pass_attachment_ptr = key_pass_ptr[i].color_attachments.ptr(); |
270 | |
271 | for (uint32_t j = 0; j < attachment_size; j++) { |
272 | if (pass_attachment_ptr[j] != key_pass_attachment_ptr[j]) { |
273 | return pass_attachment_ptr[j] < key_pass_attachment_ptr[j]; |
274 | } |
275 | } |
276 | } |
277 | { // Compare input attachments. |
278 | uint32_t attachment_size = pass_ptr[i].input_attachments.size(); |
279 | uint32_t key_attachment_size = key_pass_ptr[i].input_attachments.size(); |
280 | if (attachment_size != key_attachment_size) { |
281 | return attachment_size < key_attachment_size; |
282 | } |
283 | const int32_t *pass_attachment_ptr = pass_ptr[i].input_attachments.ptr(); |
284 | const int32_t *key_pass_attachment_ptr = key_pass_ptr[i].input_attachments.ptr(); |
285 | |
286 | for (uint32_t j = 0; j < attachment_size; j++) { |
287 | if (pass_attachment_ptr[j] != key_pass_attachment_ptr[j]) { |
288 | return pass_attachment_ptr[j] < key_pass_attachment_ptr[j]; |
289 | } |
290 | } |
291 | } |
292 | { // Compare resolve attachments. |
293 | uint32_t attachment_size = pass_ptr[i].resolve_attachments.size(); |
294 | uint32_t key_attachment_size = key_pass_ptr[i].resolve_attachments.size(); |
295 | if (attachment_size != key_attachment_size) { |
296 | return attachment_size < key_attachment_size; |
297 | } |
298 | const int32_t *pass_attachment_ptr = pass_ptr[i].resolve_attachments.ptr(); |
299 | const int32_t *key_pass_attachment_ptr = key_pass_ptr[i].resolve_attachments.ptr(); |
300 | |
301 | for (uint32_t j = 0; j < attachment_size; j++) { |
302 | if (pass_attachment_ptr[j] != key_pass_attachment_ptr[j]) { |
303 | return pass_attachment_ptr[j] < key_pass_attachment_ptr[j]; |
304 | } |
305 | } |
306 | } |
307 | { // Compare preserve attachments. |
308 | uint32_t attachment_size = pass_ptr[i].preserve_attachments.size(); |
309 | uint32_t key_attachment_size = key_pass_ptr[i].preserve_attachments.size(); |
310 | if (attachment_size != key_attachment_size) { |
311 | return attachment_size < key_attachment_size; |
312 | } |
313 | const int32_t *pass_attachment_ptr = pass_ptr[i].preserve_attachments.ptr(); |
314 | const int32_t *key_pass_attachment_ptr = key_pass_ptr[i].preserve_attachments.ptr(); |
315 | |
316 | for (uint32_t j = 0; j < attachment_size; j++) { |
317 | if (pass_attachment_ptr[j] != key_pass_attachment_ptr[j]) { |
318 | return pass_attachment_ptr[j] < key_pass_attachment_ptr[j]; |
319 | } |
320 | } |
321 | } |
322 | if (pass_ptr[i].depth_attachment != key_pass_ptr[i].depth_attachment) { |
323 | return pass_ptr[i].depth_attachment < key_pass_ptr[i].depth_attachment; |
324 | } |
325 | } |
326 | |
327 | int as = attachments.size(); |
328 | int bs = p_key.attachments.size(); |
329 | if (as != bs) { |
330 | return as < bs; |
331 | } |
332 | |
333 | const AttachmentFormat *af_a = attachments.ptr(); |
334 | const AttachmentFormat *af_b = p_key.attachments.ptr(); |
335 | for (int i = 0; i < as; i++) { |
336 | const AttachmentFormat &a = af_a[i]; |
337 | const AttachmentFormat &b = af_b[i]; |
338 | if (a.format != b.format) { |
339 | return a.format < b.format; |
340 | } |
341 | if (a.samples != b.samples) { |
342 | return a.samples < b.samples; |
343 | } |
344 | if (a.usage_flags != b.usage_flags) { |
345 | return a.usage_flags < b.usage_flags; |
346 | } |
347 | } |
348 | |
349 | return false; // Equal. |
350 | } |
351 | }; |
352 | |
353 | VkRenderPass _render_pass_create(const Vector<AttachmentFormat> &p_attachments, const Vector<FramebufferPass> &p_passes, InitialAction p_initial_action, FinalAction p_final_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, uint32_t p_view_count = 1, Vector<TextureSamples> *r_samples = nullptr); |
354 | // This is a cache and it's never freed, it ensures |
355 | // IDs for a given format are always unique. |
356 | RBMap<FramebufferFormatKey, FramebufferFormatID> framebuffer_format_cache; |
357 | struct FramebufferFormat { |
358 | const RBMap<FramebufferFormatKey, FramebufferFormatID>::Element *E; |
359 | VkRenderPass render_pass = VK_NULL_HANDLE; // Here for constructing shaders, never used, see section (7.2. Render Pass Compatibility from Vulkan spec). |
360 | Vector<TextureSamples> pass_samples; |
361 | uint32_t view_count = 1; // Number of views. |
362 | }; |
363 | |
364 | HashMap<FramebufferFormatID, FramebufferFormat> framebuffer_formats; |
365 | |
366 | struct Framebuffer { |
367 | FramebufferFormatID format_id = 0; |
368 | struct VersionKey { |
369 | InitialAction initial_color_action; |
370 | FinalAction final_color_action; |
371 | InitialAction initial_depth_action; |
372 | FinalAction final_depth_action; |
373 | uint32_t view_count; |
374 | |
375 | bool operator<(const VersionKey &p_key) const { |
376 | if (initial_color_action == p_key.initial_color_action) { |
377 | if (final_color_action == p_key.final_color_action) { |
378 | if (initial_depth_action == p_key.initial_depth_action) { |
379 | if (final_depth_action == p_key.final_depth_action) { |
380 | return view_count < p_key.view_count; |
381 | } else { |
382 | return final_depth_action < p_key.final_depth_action; |
383 | } |
384 | } else { |
385 | return initial_depth_action < p_key.initial_depth_action; |
386 | } |
387 | } else { |
388 | return final_color_action < p_key.final_color_action; |
389 | } |
390 | } else { |
391 | return initial_color_action < p_key.initial_color_action; |
392 | } |
393 | } |
394 | }; |
395 | |
396 | uint32_t storage_mask = 0; |
397 | Vector<RID> texture_ids; |
398 | InvalidationCallback invalidated_callback = nullptr; |
399 | void *invalidated_callback_userdata = nullptr; |
400 | |
401 | struct Version { |
402 | VkFramebuffer framebuffer = VK_NULL_HANDLE; |
403 | VkRenderPass render_pass = VK_NULL_HANDLE; // This one is owned. |
404 | uint32_t subpass_count = 1; |
405 | }; |
406 | |
407 | RBMap<VersionKey, Version> framebuffers; |
408 | Size2 size; |
409 | uint32_t view_count; |
410 | }; |
411 | |
412 | RID_Owner<Framebuffer, true> framebuffer_owner; |
413 | |
414 | /***********************/ |
415 | /**** VERTEX BUFFER ****/ |
416 | /***********************/ |
417 | |
418 | // Vertex buffers in Vulkan are similar to how |
419 | // they work in OpenGL, except that instead of |
420 | // an attribute index, there is a buffer binding |
421 | // index (for binding the buffers in real-time) |
422 | // and a location index (what is used in the shader). |
423 | // |
424 | // This mapping is done here internally, and it's not |
425 | // exposed. |
426 | |
427 | RID_Owner<Buffer, true> vertex_buffer_owner; |
428 | |
429 | struct VertexDescriptionKey { |
430 | Vector<VertexAttribute> vertex_formats; |
431 | bool operator==(const VertexDescriptionKey &p_key) const { |
432 | int vdc = vertex_formats.size(); |
433 | int vdck = p_key.vertex_formats.size(); |
434 | |
435 | if (vdc != vdck) { |
436 | return false; |
437 | } else { |
438 | const VertexAttribute *a_ptr = vertex_formats.ptr(); |
439 | const VertexAttribute *b_ptr = p_key.vertex_formats.ptr(); |
440 | for (int i = 0; i < vdc; i++) { |
441 | const VertexAttribute &a = a_ptr[i]; |
442 | const VertexAttribute &b = b_ptr[i]; |
443 | |
444 | if (a.location != b.location) { |
445 | return false; |
446 | } |
447 | if (a.offset != b.offset) { |
448 | return false; |
449 | } |
450 | if (a.format != b.format) { |
451 | return false; |
452 | } |
453 | if (a.stride != b.stride) { |
454 | return false; |
455 | } |
456 | if (a.frequency != b.frequency) { |
457 | return false; |
458 | } |
459 | } |
460 | return true; // They are equal. |
461 | } |
462 | } |
463 | |
464 | uint32_t hash() const { |
465 | int vdc = vertex_formats.size(); |
466 | uint32_t h = hash_murmur3_one_32(vdc); |
467 | const VertexAttribute *ptr = vertex_formats.ptr(); |
468 | for (int i = 0; i < vdc; i++) { |
469 | const VertexAttribute &vd = ptr[i]; |
470 | h = hash_murmur3_one_32(vd.location, h); |
471 | h = hash_murmur3_one_32(vd.offset, h); |
472 | h = hash_murmur3_one_32(vd.format, h); |
473 | h = hash_murmur3_one_32(vd.stride, h); |
474 | h = hash_murmur3_one_32(vd.frequency, h); |
475 | } |
476 | return hash_fmix32(h); |
477 | } |
478 | }; |
479 | |
480 | struct VertexDescriptionHash { |
481 | static _FORCE_INLINE_ uint32_t hash(const VertexDescriptionKey &p_key) { |
482 | return p_key.hash(); |
483 | } |
484 | }; |
485 | |
486 | // This is a cache and it's never freed, it ensures that |
487 | // ID used for a specific format always remain the same. |
488 | HashMap<VertexDescriptionKey, VertexFormatID, VertexDescriptionHash> vertex_format_cache; |
489 | |
490 | struct VertexDescriptionCache { |
491 | Vector<VertexAttribute> vertex_formats; |
492 | VkVertexInputBindingDescription *bindings = nullptr; |
493 | VkVertexInputAttributeDescription *attributes = nullptr; |
494 | VkPipelineVertexInputStateCreateInfo create_info; |
495 | }; |
496 | |
497 | HashMap<VertexFormatID, VertexDescriptionCache> vertex_formats; |
498 | |
499 | struct VertexArray { |
500 | RID buffer; |
501 | VertexFormatID description = 0; |
502 | int vertex_count = 0; |
503 | uint32_t max_instances_allowed = 0; |
504 | |
505 | Vector<VkBuffer> buffers; // Not owned, just referenced. |
506 | Vector<VkDeviceSize> offsets; |
507 | }; |
508 | |
509 | RID_Owner<VertexArray, true> vertex_array_owner; |
510 | |
511 | struct IndexBuffer : public Buffer { |
512 | uint32_t max_index = 0; // Used for validation. |
513 | uint32_t index_count = 0; |
514 | VkIndexType index_type = VK_INDEX_TYPE_NONE_NV; |
515 | bool supports_restart_indices = false; |
516 | }; |
517 | |
518 | RID_Owner<IndexBuffer, true> index_buffer_owner; |
519 | |
520 | struct IndexArray { |
521 | uint32_t max_index = 0; // Remember the maximum index here too, for validation. |
522 | VkBuffer buffer; // Not owned, inherited from index buffer. |
523 | uint32_t offset = 0; |
524 | uint32_t indices = 0; |
525 | VkIndexType index_type = VK_INDEX_TYPE_NONE_NV; |
526 | bool supports_restart_indices = false; |
527 | }; |
528 | |
529 | RID_Owner<IndexArray, true> index_array_owner; |
530 | |
531 | /****************/ |
532 | /**** SHADER ****/ |
533 | /****************/ |
534 | |
535 | // Vulkan specifies a really complex behavior for the application |
536 | // in order to tell when descriptor sets need to be re-bound (or not). |
537 | // "When binding a descriptor set (see Descriptor Set Binding) to set |
538 | // number N, if the previously bound descriptor sets for sets zero |
539 | // through N-1 were all bound using compatible pipeline layouts, |
540 | // then performing this binding does not disturb any of the lower numbered sets. |
541 | // If, additionally, the previous bound descriptor set for set N was |
542 | // bound using a pipeline layout compatible for set N, then the bindings |
543 | // in sets numbered greater than N are also not disturbed." |
544 | // As a result, we need to figure out quickly when something is no longer "compatible". |
545 | // in order to avoid costly rebinds. |
546 | |
547 | struct UniformInfo { |
548 | UniformType type = UniformType::UNIFORM_TYPE_MAX; |
549 | bool writable = false; |
550 | int binding = 0; |
551 | uint32_t stages = 0; |
552 | int length = 0; // Size of arrays (in total elements), or ubos (in bytes * total elements). |
553 | |
554 | bool operator!=(const UniformInfo &p_info) const { |
555 | return (binding != p_info.binding || type != p_info.type || writable != p_info.writable || stages != p_info.stages || length != p_info.length); |
556 | } |
557 | |
558 | bool operator<(const UniformInfo &p_info) const { |
559 | if (binding != p_info.binding) { |
560 | return binding < p_info.binding; |
561 | } |
562 | if (type != p_info.type) { |
563 | return type < p_info.type; |
564 | } |
565 | if (writable != p_info.writable) { |
566 | return writable < p_info.writable; |
567 | } |
568 | if (stages != p_info.stages) { |
569 | return stages < p_info.stages; |
570 | } |
571 | return length < p_info.length; |
572 | } |
573 | }; |
574 | |
575 | struct UniformSetFormat { |
576 | Vector<UniformInfo> uniform_info; |
577 | bool operator<(const UniformSetFormat &p_format) const { |
578 | uint32_t size = uniform_info.size(); |
579 | uint32_t psize = p_format.uniform_info.size(); |
580 | |
581 | if (size != psize) { |
582 | return size < psize; |
583 | } |
584 | |
585 | const UniformInfo *infoptr = uniform_info.ptr(); |
586 | const UniformInfo *pinfoptr = p_format.uniform_info.ptr(); |
587 | |
588 | for (uint32_t i = 0; i < size; i++) { |
589 | if (infoptr[i] != pinfoptr[i]) { |
590 | return infoptr[i] < pinfoptr[i]; |
591 | } |
592 | } |
593 | |
594 | return false; |
595 | } |
596 | }; |
597 | |
598 | // Always grows, never shrinks, ensuring unique IDs, but we assume |
599 | // the amount of formats will never be a problem, as the amount of shaders |
600 | // in a game is limited. |
601 | RBMap<UniformSetFormat, uint32_t> uniform_set_format_cache; |
602 | |
603 | // Shaders in Vulkan are just pretty much |
604 | // precompiled blocks of SPIR-V bytecode. They |
605 | // are most likely not really compiled to host |
606 | // assembly until a pipeline is created. |
607 | // |
608 | // When supplying the shaders, this implementation |
609 | // will use the reflection abilities of glslang to |
610 | // understand and cache everything required to |
611 | // create and use the descriptor sets (Vulkan's |
612 | // biggest pain). |
613 | // |
614 | // Additionally, hashes are created for every set |
615 | // to do quick validation and ensuring the user |
616 | // does not submit something invalid. |
617 | |
618 | struct Shader { |
619 | struct Set { |
620 | Vector<UniformInfo> uniform_info; |
621 | VkDescriptorSetLayout descriptor_set_layout = VK_NULL_HANDLE; |
622 | }; |
623 | |
624 | uint32_t vertex_input_mask = 0; // Inputs used, this is mostly for validation. |
625 | uint32_t fragment_output_mask = 0; |
626 | |
627 | struct PushConstant { |
628 | uint32_t size = 0; |
629 | uint32_t vk_stages_mask = 0; |
630 | }; |
631 | |
632 | PushConstant push_constant; |
633 | |
634 | uint32_t compute_local_size[3] = { 0, 0, 0 }; |
635 | |
636 | struct SpecializationConstant { |
637 | PipelineSpecializationConstant constant; |
638 | uint32_t stage_flags = 0; |
639 | }; |
640 | |
641 | bool is_compute = false; |
642 | Vector<Set> sets; |
643 | Vector<uint32_t> set_formats; |
644 | Vector<VkPipelineShaderStageCreateInfo> pipeline_stages; |
645 | Vector<SpecializationConstant> specialization_constants; |
646 | VkPipelineLayout pipeline_layout = VK_NULL_HANDLE; |
647 | String name; // Used for debug. |
648 | }; |
649 | |
650 | String _shader_uniform_debug(RID p_shader, int p_set = -1); |
651 | |
652 | RID_Owner<Shader, true> shader_owner; |
653 | |
654 | /******************/ |
655 | /**** UNIFORMS ****/ |
656 | /******************/ |
657 | |
658 | // Descriptor sets require allocation from a pool. |
659 | // The documentation on how to use pools properly |
660 | // is scarce, and the documentation is strange. |
661 | // |
662 | // Basically, you can mix and match pools as you |
663 | // like, but you'll run into fragmentation issues. |
664 | // Because of this, the recommended approach is to |
665 | // create a pool for every descriptor set type, as |
666 | // this prevents fragmentation. |
667 | // |
668 | // This is implemented here as a having a list of |
669 | // pools (each can contain up to 64 sets) for each |
670 | // set layout. The amount of sets for each type |
671 | // is used as the key. |
672 | |
673 | enum { |
674 | MAX_DESCRIPTOR_POOL_ELEMENT = 65535 |
675 | }; |
676 | |
677 | struct DescriptorPoolKey { |
678 | union { |
679 | struct { |
680 | uint16_t uniform_type[UNIFORM_TYPE_MAX]; // Using 16 bits because, for sending arrays, each element is a pool set. |
681 | }; |
682 | struct { |
683 | uint64_t key1; |
684 | uint64_t key2; |
685 | uint64_t key3; |
686 | }; |
687 | }; |
688 | bool operator<(const DescriptorPoolKey &p_key) const { |
689 | if (key1 != p_key.key1) { |
690 | return key1 < p_key.key1; |
691 | } |
692 | if (key2 != p_key.key2) { |
693 | return key2 < p_key.key2; |
694 | } |
695 | |
696 | return key3 < p_key.key3; |
697 | } |
698 | DescriptorPoolKey() { |
699 | key1 = 0; |
700 | key2 = 0; |
701 | key3 = 0; |
702 | } |
703 | }; |
704 | |
705 | struct DescriptorPool { |
706 | VkDescriptorPool pool; |
707 | uint32_t usage; |
708 | }; |
709 | |
710 | RBMap<DescriptorPoolKey, HashSet<DescriptorPool *>> descriptor_pools; |
711 | uint32_t max_descriptors_per_pool = 0; |
712 | |
713 | DescriptorPool *_descriptor_pool_allocate(const DescriptorPoolKey &p_key); |
714 | void _descriptor_pool_free(const DescriptorPoolKey &p_key, DescriptorPool *p_pool); |
715 | |
716 | RID_Owner<Buffer, true> uniform_buffer_owner; |
717 | RID_Owner<Buffer, true> storage_buffer_owner; |
718 | |
719 | // Texture buffer needs a view. |
720 | struct TextureBuffer { |
721 | Buffer buffer; |
722 | VkBufferView view = VK_NULL_HANDLE; |
723 | }; |
724 | |
725 | RID_Owner<TextureBuffer, true> texture_buffer_owner; |
726 | |
727 | // This structure contains the descriptor set. They _need_ to be allocated |
728 | // for a shader (and will be erased when this shader is erased), but should |
729 | // work for other shaders as long as the hash matches. This covers using |
730 | // them in shader variants. |
731 | // |
732 | // Keep also in mind that you can share buffers between descriptor sets, so |
733 | // the above restriction is not too serious. |
734 | |
735 | struct UniformSet { |
736 | uint32_t format = 0; |
737 | RID shader_id; |
738 | uint32_t shader_set = 0; |
739 | DescriptorPool *pool = nullptr; |
740 | DescriptorPoolKey pool_key; |
741 | VkDescriptorSet descriptor_set = VK_NULL_HANDLE; |
742 | //VkPipelineLayout pipeline_layout; // Not owned, inherited from shader. |
743 | struct AttachableTexture { |
744 | uint32_t bind; |
745 | RID texture; |
746 | }; |
747 | |
748 | LocalVector<AttachableTexture> attachable_textures; // Used for validation. |
749 | Vector<Texture *> mutable_sampled_textures; // Used for layout change. |
750 | Vector<Texture *> mutable_storage_textures; // Used for layout change. |
751 | InvalidationCallback invalidated_callback = nullptr; |
752 | void *invalidated_callback_userdata = nullptr; |
753 | }; |
754 | |
755 | RID_Owner<UniformSet, true> uniform_set_owner; |
756 | |
757 | /*******************/ |
758 | /**** PIPELINES ****/ |
759 | /*******************/ |
760 | |
761 | // Render pipeline contains ALL the |
762 | // information required for drawing. |
763 | // This includes all the rasterizer state |
764 | // as well as shader used, framebuffer format, |
765 | // etc. |
766 | // While the pipeline is just a single object |
767 | // (VkPipeline) a lot of values are also saved |
768 | // here to do validation (vulkan does none by |
769 | // default) and warn the user if something |
770 | // was not supplied as intended. |
771 | |
772 | struct RenderPipeline { |
773 | // Cached values for validation. |
774 | #ifdef DEBUG_ENABLED |
775 | struct Validation { |
776 | FramebufferFormatID framebuffer_format = 0; |
777 | uint32_t render_pass = 0; |
778 | uint32_t dynamic_state = 0; |
779 | VertexFormatID vertex_format = 0; |
780 | bool uses_restart_indices = false; |
781 | uint32_t primitive_minimum = 0; |
782 | uint32_t primitive_divisor = 0; |
783 | } validation; |
784 | #endif |
785 | // Actual pipeline. |
786 | RID shader; |
787 | Vector<uint32_t> set_formats; |
788 | VkPipelineLayout pipeline_layout = VK_NULL_HANDLE; // Not owned, needed for push constants. |
789 | VkPipeline pipeline = VK_NULL_HANDLE; |
790 | uint32_t push_constant_size = 0; |
791 | uint32_t push_constant_stages_mask = 0; |
792 | }; |
793 | |
794 | RID_Owner<RenderPipeline, true> render_pipeline_owner; |
795 | |
796 | struct { |
797 | uint32_t ; |
798 | uint32_t ; |
799 | uint64_t ; |
800 | uint32_t ; |
801 | uint32_t ; |
802 | uint32_t ; |
803 | uint8_t [VK_UUID_SIZE]; |
804 | uint8_t ; |
805 | }; |
806 | |
807 | struct PipelineCache { |
808 | String file_path; |
809 | PipelineCacheHeader = {}; |
810 | size_t current_size = 0; |
811 | LocalVector<uint8_t> buffer; |
812 | VkPipelineCache cache_object = VK_NULL_HANDLE; |
813 | }; |
814 | |
815 | PipelineCache pipelines_cache; |
816 | |
817 | WorkerThreadPool::TaskID pipelines_cache_save_task = WorkerThreadPool::INVALID_TASK_ID; |
818 | |
819 | void _load_pipeline_cache(); |
820 | void _update_pipeline_cache(bool p_closing = false); |
821 | static void _save_pipeline_cache(void *p_data); |
822 | |
823 | struct ComputePipeline { |
824 | RID shader; |
825 | Vector<uint32_t> set_formats; |
826 | VkPipelineLayout pipeline_layout = VK_NULL_HANDLE; // Not owned, needed for push constants. |
827 | VkPipeline pipeline = VK_NULL_HANDLE; |
828 | uint32_t push_constant_size = 0; |
829 | uint32_t push_constant_stages_mask = 0; |
830 | uint32_t local_group_size[3] = { 0, 0, 0 }; |
831 | }; |
832 | |
833 | RID_Owner<ComputePipeline, true> compute_pipeline_owner; |
834 | |
835 | /*******************/ |
836 | /**** DRAW LIST ****/ |
837 | /*******************/ |
838 | |
839 | // Draw list contains both the command buffer |
840 | // used for drawing as well as a LOT of |
841 | // information used for validation. This |
842 | // validation is cheap so most of it can |
843 | // also run in release builds. |
844 | |
845 | // When using split command lists, this is |
846 | // implemented internally using secondary command |
847 | // buffers. As they can be created in threads, |
848 | // each needs its own command pool. |
849 | |
850 | struct SplitDrawListAllocator { |
851 | VkCommandPool command_pool = VK_NULL_HANDLE; |
852 | Vector<VkCommandBuffer> command_buffers; // One for each frame. |
853 | }; |
854 | |
855 | Vector<SplitDrawListAllocator> split_draw_list_allocators; |
856 | |
857 | struct DrawList { |
858 | VkCommandBuffer command_buffer = VK_NULL_HANDLE; // If persistent, this is owned, otherwise it's shared with the ringbuffer. |
859 | Rect2i viewport; |
860 | bool viewport_set = false; |
861 | |
862 | struct SetState { |
863 | uint32_t pipeline_expected_format = 0; |
864 | uint32_t uniform_set_format = 0; |
865 | VkDescriptorSet descriptor_set = VK_NULL_HANDLE; |
866 | RID uniform_set; |
867 | bool bound = false; |
868 | }; |
869 | |
870 | struct State { |
871 | SetState sets[MAX_UNIFORM_SETS]; |
872 | uint32_t set_count = 0; |
873 | RID pipeline; |
874 | RID pipeline_shader; |
875 | VkPipelineLayout pipeline_layout = VK_NULL_HANDLE; |
876 | RID vertex_array; |
877 | RID index_array; |
878 | uint32_t pipeline_push_constant_stages = 0; |
879 | } state; |
880 | |
881 | #ifdef DEBUG_ENABLED |
882 | struct Validation { |
883 | bool active = true; // Means command buffer was not closed, so you can keep adding things. |
884 | // Actual render pass values. |
885 | uint32_t dynamic_state = 0; |
886 | VertexFormatID vertex_format = INVALID_ID; |
887 | uint32_t vertex_array_size = 0; |
888 | uint32_t vertex_max_instances_allowed = 0xFFFFFFFF; |
889 | bool index_buffer_uses_restart_indices = false; |
890 | uint32_t index_array_size = 0; |
891 | uint32_t index_array_max_index = 0; |
892 | uint32_t index_array_offset = 0; |
893 | Vector<uint32_t> set_formats; |
894 | Vector<bool> set_bound; |
895 | Vector<RID> set_rids; |
896 | // Last pipeline set values. |
897 | bool pipeline_active = false; |
898 | uint32_t pipeline_dynamic_state = 0; |
899 | VertexFormatID pipeline_vertex_format = INVALID_ID; |
900 | RID pipeline_shader; |
901 | bool pipeline_uses_restart_indices = false; |
902 | uint32_t pipeline_primitive_divisor = 0; |
903 | uint32_t pipeline_primitive_minimum = 0; |
904 | uint32_t pipeline_push_constant_size = 0; |
905 | bool pipeline_push_constant_supplied = false; |
906 | } validation; |
907 | #else |
908 | struct Validation { |
909 | uint32_t vertex_array_size = 0; |
910 | uint32_t index_array_size = 0; |
911 | uint32_t index_array_offset; |
912 | } validation; |
913 | #endif |
914 | }; |
915 | |
916 | DrawList *draw_list = nullptr; // One for regular draw lists, multiple for split. |
917 | uint32_t draw_list_subpass_count = 0; |
918 | uint32_t draw_list_count = 0; |
919 | VkRenderPass draw_list_render_pass = VK_NULL_HANDLE; |
920 | VkFramebuffer draw_list_vkframebuffer = VK_NULL_HANDLE; |
921 | #ifdef DEBUG_ENABLED |
922 | FramebufferFormatID draw_list_framebuffer_format = INVALID_ID; |
923 | #endif |
924 | uint32_t draw_list_current_subpass = 0; |
925 | |
926 | bool draw_list_split = false; |
927 | Vector<RID> draw_list_bound_textures; |
928 | Vector<RID> draw_list_storage_textures; |
929 | bool draw_list_unbind_color_textures = false; |
930 | bool draw_list_unbind_depth_textures = false; |
931 | |
932 | void _draw_list_insert_clear_region(DrawList *p_draw_list, Framebuffer *p_framebuffer, Point2i p_viewport_offset, Point2i p_viewport_size, bool p_clear_color, const Vector<Color> &p_clear_colors, bool p_clear_depth, float p_depth, uint32_t p_stencil); |
933 | Error _draw_list_setup_framebuffer(Framebuffer *p_framebuffer, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, VkFramebuffer *r_framebuffer, VkRenderPass *r_render_pass, uint32_t *r_subpass_count); |
934 | Error _draw_list_render_pass_begin(Framebuffer *framebuffer, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, const Vector<Color> &p_clear_colors, float p_clear_depth, uint32_t p_clear_stencil, Point2i viewport_offset, Point2i viewport_size, VkFramebuffer vkframebuffer, VkRenderPass render_pass, VkCommandBuffer command_buffer, VkSubpassContents subpass_contents, const Vector<RID> &p_storage_textures); |
935 | _FORCE_INLINE_ DrawList *_get_draw_list_ptr(DrawListID p_id); |
936 | Buffer *_get_buffer_from_owner(RID p_buffer, VkPipelineStageFlags &dst_stage_mask, VkAccessFlags &dst_access, BitField<BarrierMask> p_post_barrier); |
937 | Error _draw_list_allocate(const Rect2i &p_viewport, uint32_t p_splits, uint32_t p_subpass); |
938 | void _draw_list_free(Rect2i *r_last_viewport = nullptr); |
939 | |
940 | /**********************/ |
941 | /**** COMPUTE LIST ****/ |
942 | /**********************/ |
943 | |
944 | struct ComputeList { |
945 | VkCommandBuffer command_buffer = VK_NULL_HANDLE; // If persistent, this is owned, otherwise it's shared with the ringbuffer. |
946 | |
947 | struct SetState { |
948 | uint32_t pipeline_expected_format = 0; |
949 | uint32_t uniform_set_format = 0; |
950 | VkDescriptorSet descriptor_set = VK_NULL_HANDLE; |
951 | RID uniform_set; |
952 | bool bound = false; |
953 | }; |
954 | |
955 | struct State { |
956 | HashSet<Texture *> textures_to_sampled_layout; |
957 | SetState sets[MAX_UNIFORM_SETS]; |
958 | uint32_t set_count = 0; |
959 | RID pipeline; |
960 | RID pipeline_shader; |
961 | uint32_t local_group_size[3] = { 0, 0, 0 }; |
962 | VkPipelineLayout pipeline_layout = VK_NULL_HANDLE; |
963 | uint32_t pipeline_push_constant_stages = 0; |
964 | bool allow_draw_overlap; |
965 | } state; |
966 | |
967 | #ifdef DEBUG_ENABLED |
968 | struct Validation { |
969 | bool active = true; // Means command buffer was not closed, so you can keep adding things. |
970 | Vector<uint32_t> set_formats; |
971 | Vector<bool> set_bound; |
972 | Vector<RID> set_rids; |
973 | // Last pipeline set values. |
974 | bool pipeline_active = false; |
975 | RID pipeline_shader; |
976 | uint32_t invalid_set_from = 0; |
977 | uint32_t pipeline_push_constant_size = 0; |
978 | bool pipeline_push_constant_supplied = false; |
979 | } validation; |
980 | #endif |
981 | }; |
982 | |
983 | ComputeList *compute_list = nullptr; |
984 | |
985 | void _compute_list_add_barrier(BitField<BarrierMask> p_post_barrier, uint32_t p_barrier_flags, uint32_t p_access_flags); |
986 | |
987 | /**************************/ |
988 | /**** FRAME MANAGEMENT ****/ |
989 | /**************************/ |
990 | |
991 | // This is the frame structure. There are normally |
992 | // 3 of these (used for triple buffering), or 2 |
993 | // (double buffering). They are cycled constantly. |
994 | // |
995 | // It contains two command buffers, one that is |
996 | // used internally for setting up (creating stuff) |
997 | // and another used mostly for drawing. |
998 | // |
999 | // They also contains a list of things that need |
1000 | // to be disposed of when deleted, which can't |
1001 | // happen immediately due to the asynchronous |
1002 | // nature of the GPU. They will get deleted |
1003 | // when the frame is cycled. |
1004 | |
1005 | struct Frame { |
1006 | // List in usage order, from last to free to first to free. |
1007 | List<Buffer> buffers_to_dispose_of; |
1008 | List<Texture> textures_to_dispose_of; |
1009 | List<Framebuffer> framebuffers_to_dispose_of; |
1010 | List<VkSampler> samplers_to_dispose_of; |
1011 | List<Shader> shaders_to_dispose_of; |
1012 | List<VkBufferView> buffer_views_to_dispose_of; |
1013 | List<UniformSet> uniform_sets_to_dispose_of; |
1014 | List<RenderPipeline> render_pipelines_to_dispose_of; |
1015 | List<ComputePipeline> compute_pipelines_to_dispose_of; |
1016 | |
1017 | VkCommandPool command_pool = VK_NULL_HANDLE; |
1018 | VkCommandBuffer setup_command_buffer = VK_NULL_HANDLE; // Used at the beginning of every frame for set-up. |
1019 | VkCommandBuffer draw_command_buffer = VK_NULL_HANDLE; // Used at the beginning of every frame for set-up. |
1020 | |
1021 | struct Timestamp { |
1022 | String description; |
1023 | uint64_t value = 0; |
1024 | }; |
1025 | |
1026 | VkQueryPool timestamp_pool; |
1027 | |
1028 | TightLocalVector<String> timestamp_names; |
1029 | TightLocalVector<uint64_t> timestamp_cpu_values; |
1030 | uint32_t timestamp_count = 0; |
1031 | TightLocalVector<String> timestamp_result_names; |
1032 | TightLocalVector<uint64_t> timestamp_cpu_result_values; |
1033 | TightLocalVector<uint64_t> timestamp_result_values; |
1034 | uint32_t timestamp_result_count = 0; |
1035 | uint64_t index = 0; |
1036 | }; |
1037 | |
1038 | uint32_t max_timestamp_query_elements = 0; |
1039 | |
1040 | TightLocalVector<Frame> frames; // Frames available, for main device they are cycled (usually 3), for local devices only 1. |
1041 | int frame = 0; // Current frame. |
1042 | int frame_count = 0; // Total amount of frames. |
1043 | uint64_t frames_drawn = 0; |
1044 | RID local_device; |
1045 | bool local_device_processing = false; |
1046 | |
1047 | void _free_pending_resources(int p_frame); |
1048 | |
1049 | VmaAllocator allocator = nullptr; |
1050 | HashMap<uint32_t, VmaPool> small_allocs_pools; |
1051 | VmaPool _find_or_create_small_allocs_pool(uint32_t p_mem_type_index); |
1052 | |
1053 | VulkanContext *context = nullptr; |
1054 | |
1055 | uint64_t image_memory = 0; |
1056 | uint64_t buffer_memory = 0; |
1057 | |
1058 | void _free_internal(RID p_id); |
1059 | void _flush(bool p_current_frame); |
1060 | |
1061 | bool screen_prepared = false; |
1062 | |
1063 | template <class T> |
1064 | void _free_rids(T &p_owner, const char *p_type); |
1065 | |
1066 | void _finalize_command_bufers(); |
1067 | void _begin_frame(); |
1068 | |
1069 | #ifdef DEV_ENABLED |
1070 | HashMap<RID, String> resource_names; |
1071 | #endif |
1072 | |
1073 | VkSampleCountFlagBits _ensure_supported_sample_count(TextureSamples p_requested_sample_count) const; |
1074 | |
1075 | public: |
1076 | virtual RID texture_create(const TextureFormat &p_format, const TextureView &p_view, const Vector<Vector<uint8_t>> &p_data = Vector<Vector<uint8_t>>()); |
1077 | virtual RID texture_create_shared(const TextureView &p_view, RID p_with_texture); |
1078 | virtual RID texture_create_from_extension(TextureType p_type, DataFormat p_format, TextureSamples p_samples, uint64_t p_flags, uint64_t p_image, uint64_t p_width, uint64_t p_height, uint64_t p_depth, uint64_t p_layers); |
1079 | |
1080 | virtual RID texture_create_shared_from_slice(const TextureView &p_view, RID p_with_texture, uint32_t p_layer, uint32_t p_mipmap, uint32_t p_mipmaps = 1, TextureSliceType p_slice_type = TEXTURE_SLICE_2D, uint32_t p_layers = 0); |
1081 | virtual Error texture_update(RID p_texture, uint32_t p_layer, const Vector<uint8_t> &p_data, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); |
1082 | virtual Vector<uint8_t> texture_get_data(RID p_texture, uint32_t p_layer); |
1083 | |
1084 | virtual bool texture_is_format_supported_for_usage(DataFormat p_format, BitField<RenderingDevice::TextureUsageBits> p_usage) const; |
1085 | virtual bool texture_is_shared(RID p_texture); |
1086 | virtual bool texture_is_valid(RID p_texture); |
1087 | virtual TextureFormat texture_get_format(RID p_texture); |
1088 | virtual Size2i texture_size(RID p_texture); |
1089 | virtual uint64_t texture_get_native_handle(RID p_texture); |
1090 | |
1091 | virtual Error texture_copy(RID p_from_texture, RID p_to_texture, const Vector3 &p_from, const Vector3 &p_to, const Vector3 &p_size, uint32_t p_src_mipmap, uint32_t p_dst_mipmap, uint32_t p_src_layer, uint32_t p_dst_layer, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); |
1092 | virtual Error texture_clear(RID p_texture, const Color &p_color, uint32_t p_base_mipmap, uint32_t p_mipmaps, uint32_t p_base_layer, uint32_t p_layers, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); |
1093 | virtual Error texture_resolve_multisample(RID p_from_texture, RID p_to_texture, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); |
1094 | |
1095 | /*********************/ |
1096 | /**** FRAMEBUFFER ****/ |
1097 | /*********************/ |
1098 | |
1099 | virtual FramebufferFormatID framebuffer_format_create(const Vector<AttachmentFormat> &p_format, uint32_t p_view_count = 1); |
1100 | virtual FramebufferFormatID framebuffer_format_create_multipass(const Vector<AttachmentFormat> &p_attachments, const Vector<FramebufferPass> &p_passes, uint32_t p_view_count = 1); |
1101 | virtual FramebufferFormatID framebuffer_format_create_empty(TextureSamples p_samples = TEXTURE_SAMPLES_1); |
1102 | virtual TextureSamples framebuffer_format_get_texture_samples(FramebufferFormatID p_format, uint32_t p_pass = 0); |
1103 | |
1104 | virtual RID framebuffer_create(const Vector<RID> &p_texture_attachments, FramebufferFormatID p_format_check = INVALID_ID, uint32_t p_view_count = 1); |
1105 | virtual RID framebuffer_create_multipass(const Vector<RID> &p_texture_attachments, const Vector<FramebufferPass> &p_passes, FramebufferFormatID p_format_check = INVALID_ID, uint32_t p_view_count = 1); |
1106 | virtual RID framebuffer_create_empty(const Size2i &p_size, TextureSamples p_samples = TEXTURE_SAMPLES_1, FramebufferFormatID p_format_check = INVALID_ID); |
1107 | virtual bool framebuffer_is_valid(RID p_framebuffer) const; |
1108 | virtual void framebuffer_set_invalidation_callback(RID p_framebuffer, InvalidationCallback p_callback, void *p_userdata); |
1109 | |
1110 | virtual FramebufferFormatID framebuffer_get_format(RID p_framebuffer); |
1111 | |
1112 | /*****************/ |
1113 | /**** SAMPLER ****/ |
1114 | /*****************/ |
1115 | |
1116 | virtual RID sampler_create(const SamplerState &p_state); |
1117 | virtual bool sampler_is_format_supported_for_filter(DataFormat p_format, SamplerFilter p_sampler_filter) const; |
1118 | |
1119 | /**********************/ |
1120 | /**** VERTEX ARRAY ****/ |
1121 | /**********************/ |
1122 | |
1123 | virtual RID vertex_buffer_create(uint32_t p_size_bytes, const Vector<uint8_t> &p_data = Vector<uint8_t>(), bool p_use_as_storage = false); |
1124 | |
1125 | // Internally reference counted, this ID is warranted to be unique for the same description, but needs to be freed as many times as it was allocated. |
1126 | virtual VertexFormatID vertex_format_create(const Vector<VertexAttribute> &p_vertex_formats); |
1127 | virtual RID vertex_array_create(uint32_t p_vertex_count, VertexFormatID p_vertex_format, const Vector<RID> &p_src_buffers, const Vector<uint64_t> &p_offsets = Vector<uint64_t>()); |
1128 | |
1129 | virtual RID index_buffer_create(uint32_t p_size_indices, IndexBufferFormat p_format, const Vector<uint8_t> &p_data = Vector<uint8_t>(), bool p_use_restart_indices = false); |
1130 | |
1131 | virtual RID index_array_create(RID p_index_buffer, uint32_t p_index_offset, uint32_t p_index_count); |
1132 | |
1133 | /****************/ |
1134 | /**** SHADER ****/ |
1135 | /****************/ |
1136 | |
1137 | virtual String shader_get_binary_cache_key() const; |
1138 | virtual Vector<uint8_t> shader_compile_binary_from_spirv(const Vector<ShaderStageSPIRVData> &p_spirv, const String &p_shader_name = "" ); |
1139 | |
1140 | virtual RID shader_create_from_bytecode(const Vector<uint8_t> &p_shader_binary, RID p_placeholder = RID()); |
1141 | virtual RID shader_create_placeholder(); |
1142 | |
1143 | virtual uint32_t shader_get_vertex_input_attribute_mask(RID p_shader); |
1144 | |
1145 | /*****************/ |
1146 | /**** UNIFORM ****/ |
1147 | /*****************/ |
1148 | |
1149 | virtual RID uniform_buffer_create(uint32_t p_size_bytes, const Vector<uint8_t> &p_data = Vector<uint8_t>()); |
1150 | virtual RID storage_buffer_create(uint32_t p_size_bytes, const Vector<uint8_t> &p_data = Vector<uint8_t>(), BitField<StorageBufferUsage> p_usage = 0); |
1151 | virtual RID texture_buffer_create(uint32_t p_size_elements, DataFormat p_format, const Vector<uint8_t> &p_data = Vector<uint8_t>()); |
1152 | |
1153 | virtual RID uniform_set_create(const Vector<Uniform> &p_uniforms, RID p_shader, uint32_t p_shader_set); |
1154 | virtual bool uniform_set_is_valid(RID p_uniform_set); |
1155 | virtual void uniform_set_set_invalidation_callback(RID p_uniform_set, InvalidationCallback p_callback, void *p_userdata); |
1156 | |
1157 | virtual Error buffer_copy(RID p_src_buffer, RID p_dst_buffer, uint32_t p_src_offset, uint32_t p_dst_offset, uint32_t p_size, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); |
1158 | virtual Error buffer_update(RID p_buffer, uint32_t p_offset, uint32_t p_size, const void *p_data, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); // Works for any buffer. |
1159 | virtual Error buffer_clear(RID p_buffer, uint32_t p_offset, uint32_t p_size, BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); |
1160 | virtual Vector<uint8_t> buffer_get_data(RID p_buffer, uint32_t p_offset = 0, uint32_t p_size = 0); |
1161 | |
1162 | /*************************/ |
1163 | /**** RENDER PIPELINE ****/ |
1164 | /*************************/ |
1165 | |
1166 | virtual RID render_pipeline_create(RID p_shader, FramebufferFormatID p_framebuffer_format, VertexFormatID p_vertex_format, RenderPrimitive p_render_primitive, const PipelineRasterizationState &p_rasterization_state, const PipelineMultisampleState &p_multisample_state, const PipelineDepthStencilState &p_depth_stencil_state, const PipelineColorBlendState &p_blend_state, BitField<PipelineDynamicStateFlags> p_dynamic_state_flags = 0, uint32_t p_for_render_pass = 0, const Vector<PipelineSpecializationConstant> &p_specialization_constants = Vector<PipelineSpecializationConstant>()); |
1167 | virtual bool render_pipeline_is_valid(RID p_pipeline); |
1168 | |
1169 | /**************************/ |
1170 | /**** COMPUTE PIPELINE ****/ |
1171 | /**************************/ |
1172 | |
1173 | virtual RID compute_pipeline_create(RID p_shader, const Vector<PipelineSpecializationConstant> &p_specialization_constants = Vector<PipelineSpecializationConstant>()); |
1174 | virtual bool compute_pipeline_is_valid(RID p_pipeline); |
1175 | |
1176 | /****************/ |
1177 | /**** SCREEN ****/ |
1178 | /****************/ |
1179 | |
1180 | virtual int screen_get_width(DisplayServer::WindowID p_screen = 0) const; |
1181 | virtual int screen_get_height(DisplayServer::WindowID p_screen = 0) const; |
1182 | virtual FramebufferFormatID screen_get_framebuffer_format() const; |
1183 | |
1184 | /********************/ |
1185 | /**** DRAW LISTS ****/ |
1186 | /********************/ |
1187 | |
1188 | virtual DrawListID draw_list_begin_for_screen(DisplayServer::WindowID p_screen = 0, const Color &p_clear_color = Color()); |
1189 | |
1190 | virtual DrawListID draw_list_begin(RID p_framebuffer, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, const Vector<Color> &p_clear_color_values = Vector<Color>(), float p_clear_depth = 1.0, uint32_t p_clear_stencil = 0, const Rect2 &p_region = Rect2(), const Vector<RID> &p_storage_textures = Vector<RID>()); |
1191 | virtual Error draw_list_begin_split(RID p_framebuffer, uint32_t p_splits, DrawListID *r_split_ids, InitialAction p_initial_color_action, FinalAction p_final_color_action, InitialAction p_initial_depth_action, FinalAction p_final_depth_action, const Vector<Color> &p_clear_color_values = Vector<Color>(), float p_clear_depth = 1.0, uint32_t p_clear_stencil = 0, const Rect2 &p_region = Rect2(), const Vector<RID> &p_storage_textures = Vector<RID>()); |
1192 | |
1193 | virtual void draw_list_set_blend_constants(DrawListID p_list, const Color &p_color); |
1194 | virtual void draw_list_bind_render_pipeline(DrawListID p_list, RID p_render_pipeline); |
1195 | virtual void draw_list_bind_uniform_set(DrawListID p_list, RID p_uniform_set, uint32_t p_index); |
1196 | virtual void draw_list_bind_vertex_array(DrawListID p_list, RID p_vertex_array); |
1197 | virtual void draw_list_bind_index_array(DrawListID p_list, RID p_index_array); |
1198 | virtual void draw_list_set_line_width(DrawListID p_list, float p_width); |
1199 | virtual void draw_list_set_push_constant(DrawListID p_list, const void *p_data, uint32_t p_data_size); |
1200 | |
1201 | virtual void draw_list_draw(DrawListID p_list, bool p_use_indices, uint32_t p_instances = 1, uint32_t p_procedural_vertices = 0); |
1202 | |
1203 | virtual void draw_list_enable_scissor(DrawListID p_list, const Rect2 &p_rect); |
1204 | virtual void draw_list_disable_scissor(DrawListID p_list); |
1205 | |
1206 | virtual uint32_t draw_list_get_current_pass(); |
1207 | virtual DrawListID draw_list_switch_to_next_pass(); |
1208 | virtual Error draw_list_switch_to_next_pass_split(uint32_t p_splits, DrawListID *r_split_ids); |
1209 | |
1210 | virtual void draw_list_end(BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); |
1211 | |
1212 | /***********************/ |
1213 | /**** COMPUTE LISTS ****/ |
1214 | /***********************/ |
1215 | |
1216 | virtual ComputeListID compute_list_begin(bool p_allow_draw_overlap = false); |
1217 | virtual void compute_list_bind_compute_pipeline(ComputeListID p_list, RID p_compute_pipeline); |
1218 | virtual void compute_list_bind_uniform_set(ComputeListID p_list, RID p_uniform_set, uint32_t p_index); |
1219 | virtual void compute_list_set_push_constant(ComputeListID p_list, const void *p_data, uint32_t p_data_size); |
1220 | virtual void compute_list_add_barrier(ComputeListID p_list); |
1221 | |
1222 | virtual void compute_list_dispatch(ComputeListID p_list, uint32_t p_x_groups, uint32_t p_y_groups, uint32_t p_z_groups); |
1223 | virtual void compute_list_dispatch_threads(ComputeListID p_list, uint32_t p_x_threads, uint32_t p_y_threads, uint32_t p_z_threads); |
1224 | virtual void compute_list_dispatch_indirect(ComputeListID p_list, RID p_buffer, uint32_t p_offset); |
1225 | virtual void compute_list_end(BitField<BarrierMask> p_post_barrier = BARRIER_MASK_ALL_BARRIERS); |
1226 | |
1227 | virtual void barrier(BitField<BarrierMask> p_from = BARRIER_MASK_ALL_BARRIERS, BitField<BarrierMask> p_to = BARRIER_MASK_ALL_BARRIERS); |
1228 | virtual void full_barrier(); |
1229 | |
1230 | /**************/ |
1231 | /**** FREE ****/ |
1232 | /**************/ |
1233 | |
1234 | virtual void free(RID p_id); |
1235 | |
1236 | /****************/ |
1237 | /**** Timing ****/ |
1238 | /****************/ |
1239 | |
1240 | virtual void capture_timestamp(const String &p_name); |
1241 | virtual uint32_t get_captured_timestamps_count() const; |
1242 | virtual uint64_t get_captured_timestamps_frame() const; |
1243 | virtual uint64_t get_captured_timestamp_gpu_time(uint32_t p_index) const; |
1244 | virtual uint64_t get_captured_timestamp_cpu_time(uint32_t p_index) const; |
1245 | virtual String get_captured_timestamp_name(uint32_t p_index) const; |
1246 | |
1247 | /****************/ |
1248 | /**** Limits ****/ |
1249 | /****************/ |
1250 | |
1251 | virtual uint64_t limit_get(Limit p_limit) const; |
1252 | |
1253 | virtual void prepare_screen_for_drawing(); |
1254 | void initialize(VulkanContext *p_context, bool p_local_device = false); |
1255 | void finalize(); |
1256 | |
1257 | virtual void swap_buffers(); // For main device. |
1258 | |
1259 | virtual void submit(); // For local device. |
1260 | virtual void sync(); // For local device. |
1261 | |
1262 | virtual uint32_t get_frame_delay() const; |
1263 | |
1264 | virtual RenderingDevice *create_local_device(); |
1265 | |
1266 | virtual uint64_t get_memory_usage(MemoryType p_type) const; |
1267 | |
1268 | virtual void set_resource_name(RID p_id, const String p_name); |
1269 | |
1270 | virtual void draw_command_begin_label(String p_label_name, const Color p_color = Color(1, 1, 1, 1)); |
1271 | virtual void draw_command_insert_label(String p_label_name, const Color p_color = Color(1, 1, 1, 1)); |
1272 | virtual void draw_command_end_label(); |
1273 | |
1274 | virtual String get_device_vendor_name() const; |
1275 | virtual String get_device_name() const; |
1276 | virtual RenderingDevice::DeviceType get_device_type() const; |
1277 | virtual String get_device_api_version() const; |
1278 | virtual String get_device_pipeline_cache_uuid() const; |
1279 | |
1280 | virtual uint64_t get_driver_resource(DriverResource p_resource, RID p_rid = RID(), uint64_t p_index = 0); |
1281 | |
1282 | virtual bool has_feature(const Features p_feature) const; |
1283 | |
1284 | RenderingDeviceVulkan(); |
1285 | ~RenderingDeviceVulkan(); |
1286 | }; |
1287 | |
1288 | #endif // RENDERING_DEVICE_VULKAN_H |
1289 | |